# ul2_unifying_language_learning_paradigms__7d7aeb81.pdf

Published as a conference paper at ICLR 2023

UL2: UNIFYING LANGUAGE LEARNING PARADIGMS

Yi Tay , Mostafa Dehghani , Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng Denny Zhou, Neil Houlsby, Donald Metzler Google Research, Brain Team {yitay,dehghani}@google.com

Existing pre-trained models are generally geared towards a particular class of problems. To date, there seems to be still no consensus on what the right architecture and pre-training setup should be. This paper presents a unified framework for pre-training models that are universally effective across datasets and setups. We begin by disentangling architectural archetypes with pre-training objectives two concepts that are commonly conflated. Next, we present a generalized and unified perspective for self-supervision in NLP and show how different pre-training objectives can be cast as one another and how interpolating between different objectives can be effective. We then propose Mixture-of-Denoisers (Mo D), a pretraining objective that combines diverse pre-training paradigms together. We furthermore introduce a notion of mode switching, wherein downstream fine-tuning is associated with specific pre-training schemes. We conduct extensive ablative experiments to compare multiple pre-training objectives and find that our method pushes the Pareto-frontier by outperforming T5 and/or GPT-like models across multiple diverse setups. Finally, by scaling our model up to 20B parameters, we achieve SOTA performance on 50 well-established supervised NLP tasks ranging from language generation (with automated and human evaluation), language understanding, text classification, question answering, commonsense reasoning, long text reasoning, structured knowledge grounding and information retrieval. Our model also achieve strong results at in-context learning, outperforming 175B GPT-3 on zero-shot Super GLUE and tripling the performance of T5-XXL on oneshot summarization. Finally, we show that UL2 20B works well with chain-ofthought prompting and reasoning tasks, making it an appealing choice for research into reasoning at a small to medium scale of 20B parameters. We publicly release Flax-based T5X model checkpoints for the 20B model.

1 INTRODUCTION

Note: This is a static copy of this paper as of the ICLR submission. Please use the arxiv version for future updates : https://arxiv.org/abs/2205.05131. TYVM.

There is a wide spectrum of pre-trained model options for NLP researchers and practitioners these days (Devlin et al., 2018; Brown et al., 2020; Raffel et al., 2019; Radford et al., 2019; Liu et al., 2019; Yang et al., 2019; Thoppilan et al., 2022; Fedus et al., 2021; Du et al., 2021; Chowdhery et al., 2022). When faced with the question of what model should one use, the answer is often it depends, followed by on what task?

Answering this can be overwhelming, comprising of a number of fine-grained follow-up questions like, encoder-only or encoder-decoder? , span corruption or language model? . Pressing further, the answer always seems to depend on the target downstream task. This paper questions and rethinks this thought process, specifically answering the questions of why should the choice of the pre-trained LM depend on the downstream task? and how can we pre-train models that work universally well across many tasks?.

This paper proposes a step towards making a universally applicable language model possible. We present a framework for Unifying Language Learning Paradigms or UL2 in short, that is consistently

Published as a conference paper at ICLR 2023

0 2 4 6 8 10 1-Shot GEM XSum, SGD, TOT Avg. Rouge-L

Finetuned Super GLUE Score

GPT-like (Dec) Prefix LM (Dec)

Span Corrupt (Dec)

Prefix LM (Enc Dec)

T5 (Enc Dec)

Uni LM (Enc Dec)

UL2 (Enc Dec)

Figure 1: In both decoder-only and encoder-decoder setups, UL2 strikes a significantly improved balance in performance between fine-tuned discriminative tasks and prompt-based 1-shot openended text generation than previous methods. Note: Dec and Enc Dec are compute matched but Enc Dec models have double the parameters.

effective across a very diverse set of tasks and setups. Figure 1 shows an example of how UL2 can perform universally well, unlike other models that often have to make a trade-off.

The appeal of a universal model is clear, i.e., as this not only allows concentrated effort in improving and scaling a single model, instead of diversifying resources across N models. Moreover, under resource constrained settings where only a few models can be served (e.g., on device), it would be preferable to have a single pretrained model that can perform well on many types of tasks.

At the core of UL2 is a the newly proposed Mixture-of-Denoisers (Mo D), a pre-training objective that enables strong performance across tasks. Mo D is a mixture of several well-established denoising objectives along with new ones; namely X-denoising (extreme denoising) which considers extreme span lengths and corruption rates, S-denoising (sequential denoising) that strictly follows sequence order, and R-denoising (regular denoising) that is a standard span corruption objective introduced in (Raffel et al., 2019). We show that Mo D is conceptually simple but highly effective for a diverse set of tasks.

Our approach exploits the realization that most (if not all) well-studied pre-training objectives differ in the type of context a model is conditioned on. For example, the span corruption objective is akin to invoking multiple regions of prefix language modeling (PLM) (Liu et al., 2018; Raffel et al., 2019) whereby prefixes are contiguous segments of non-corrupted tokens and targets have full access to prefixes of all PLM segments. The setting where the span approaches the full sequence length is approximately a language modeling objective conditioned on long-range context. Thus, we are able to design a pre-training objective that smoothly interpolates these different paradigms (span corruption vs language modeling vs prefix language modeling).

It is also easy to see that each denoiser is difficult in different ways. They also differ in the nature of extrapolation (or interpolation). For example, bounding a model by bidirectional context (or the future) (ie.., span corruption) makes the task easier and becomes more akin to fact completion. Meanwhile, Prefix LM/LM objectives are generally more open ended . These behaviours can be easily observed by monitoring the cross entropy losses of these different denoising objectives.

Given the Mo D formulation, we conjecture that it is beneficial for our model to not only distinguish between different denoisers during pre-training but also to adaptively switch modes when learning

Published as a conference paper at ICLR 2023

downstream tasks. We introduce mode switching, a new concept that associates pre-training tasks with dedicated sentinel tokens and allows dynamic mode switching via discrete prompting. Our model is able to switch modes between R,S and X denoisers on-demand after being pre-trained.

We then disentangle the architecture from the self-supervision scheme. While it might be a common misconception, as previously noted in Raffel et al. (2019), that a pre-trained model is strongly characterized by its backbone architecture (e.g., decoder-only vs. encoder-decoder), we find that the choice of the denoiser has significantly more impact. Mo D supports either backbone, similar to how T5 s span corruption may be trained with a decoder-only model. As such, UL2 is agnostic to architecture. We argue that the choice of backbone architecture is mainly a trade-off across different efficiency metrics.

We conduct systematic and ablative experiments on a suite of 9 diverse tasks aimed to capture different problem formulations (supervised and prompt-based in-context few-shot learning). We experiment with the Super GLUE suite (Wang et al., 2019), and three tasks from the GEM benchmark (Gehrmann et al., 2021). In addition, we evaluate on open text generation, as well as prompt-based one-shot settings on all tasks. In this ablative setup, our experimental results show that UL2 outperforms T5 and GPT-like baselines on all 9 setups. On average, UL2 outperforms a T5 baseline by +43.6% and a language model by +76.1%. Among all the other competitive baselines considered, UL2 is the only method that outperforms T5 and GPT-like models on all tasks.

We scale UL2 up to a moderate scale setting of approximately 20B (19.5 to be exact) parameters. We publicly release T5X-based Flax checkpoints of the trained 20B UL2 model. We run experiments across a very diverse suite of 50+ NLP tasks ranging from language generation (with automated and human evaluation), language understanding, text classification, question answering, commonsense reasoning, long text reasoning, structured knowledge grounding and information retrieval. Our results show that UL2 achieves SOTA on a vast majority of tasks and setups.

Additionally, we conduct zero/few-shot experiments and show that UL2 outperforms GPT-3 175B on zero shot Super GLUE. When compared with newer state-of-the-art models like GLa M (Du et al., 2021), Pa LM (Chowdhery et al., 2022) and ST-Mo E (Zoph et al., 2022), UL2 remains competitive at a compute-matched setup despite only training on C4 corpus which is known to be less effective than specially curated datasets used in (Du et al., 2021; Chowdhery et al., 2022). We delve into understanding trade-offs between zero-shot and finetuning performance and show that UL2 is Pareto-efficient with respect to both learning paradigms. On one-shot summarization, UL2 triples the performance of an LM adapted T5 XXL model and is competitive with (or outperforms) Pa LM and La MDA at the same compute cost.

Finally, we conduct experiments on reasoning tasks using the recent chain-of-thought (Co T) prompting (Wei et al., 2022b) and shows that UL2 20B benefits from Co T, making it an appealing and accessible open source option to conduct research on the reasoning capabilities language models.

2 UNIFYING LANGUAGE LEARNING PARADIGMS (UL2)

This section describes the UL2 framework and the proposed pre-training objectives which we study for the remainder of the paper.

He dealt in archetypes before anyone knew such

things existed, and his ability to take an emotion or a

situation and push it to the limit helped create a cadre of

plays that have been endlessly staged and copied.

Apart from this, Romeo and Juliet inspired Malorie

Blackman's Noughts & Crosses, there are references to

Hamlet in Lunar Park by Bret Easton Ellis and The

Tempest was the cue for The Magus by John Fowles.

He dealt in archetypes before anyone knew such

things existed, and his ability to take an emotion or a

situation and push it to the limit helped create a cadre of

plays that have been endlessly staged and copied.

Apart from this, Romeo and Juliet inspired Malorie

Blackman's Noughts & Crosses, there are references to

Hamlet in Lunar Park by Bret Easton Ellis and The

Tempest was the cue for The Magus by John Fowles.

He dealt in archetypes before anyone knew such

things existed, and his ability to take an emotion or a

situation and push it to the limit helped create a cadre of

plays that have been endlessly staged and copied.

Apart from this, Romeo and Juliet inspired Malorie

Blackman's Noughts & Crosses, there are references to

Hamlet in Lunar Park by Bret Easton Ellis and The

Tempest was the cue for The Magus by John Fowles.

He dealt in archetypes before anyone knew such

things existed, and his ability to take an emotion or a

situation and push it to the limit helped create a cadre of

plays that have been endlessly staged and copied.

Apart from this, Romeo and Juliet inspired Malorie

Blackman's Noughts & Crosses, there are references to

Hamlet in Lunar Park by Bret Easton Ellis and The

Tempest was the cue for The Magus by John Fowles.

<S> <S> <S>

<S> 24 <S> 24 <S>

3 3 4 4 2 4

3 3 4 4 2 4

<S> <S> <S>

S-Denoising X-Denoising R-Denoising

Target: Target:

Figure 2: Mixture of denoisers for training UL2. Greyed out rectangles are masked tokens that are shifted to targets for prediction.

Published as a conference paper at ICLR 2023

2.1 UNIFIED PERSPECTIVE FOR PRE-TRAINING TASKS

Many pre-training tasks can be simply formulated as an input-to-target task, wherein the input refers to any form of memory or context that the model conditions on, and the target is the model s expected output. Language models use all previous time-steps as inputs to the model to predict the next token, which is the target. In span corruption, the model leverages all uncorrupted tokens from the past and future as inputs for predicting the corrupted span (targets). Prefix-LMs are LMs that use past tokens as inputs, but consume the inputs bidirectionally: this offer more modelling power than unidirectional encoding of inputs in vanilla LM.

Given this perspective, we can approximately reduce one pre-training objective to another. For instance, in the span corruption objective, when the corrupted span, i.e., target, is equal to the entire sequence, the problem becomes effectively1 a language modeling problem. With this in mind, using span corruption, by setting the span length to be large, we can effectively mimic the LM objective in local regions.

We define a notation that covers all of the different denoising tasks that we use in this paper. The inputs and targets of the denoising tasks are generated by a SPANCORRUPT function that is parameterized by three values (µ, r, n), where µ is the mean span length, r is the corruption rate, and n which is number of corrupted spans. Note that n may be a function of the input length, L, and the span length µ, e.g. Lµ, but in some cases, we use a fixed value of n. Given an input text, SPANCORRUPT introduces corruptions to the spans of lengths that are drawn from a (normal or uniform) distribution with mean of µ. After corruption, the input text is then fed to the denoising task and the corrupted spans are used as targets to be recovered.

As an example, to construct an objective analogous to causal language modeling using this formulation, one would simply set (µ = L, r = 1.0, n = 1), i.e. a single span with its span length equal to the length of the sequence. To express one similar to Prefix LM objective, one would set (µ = L P, r = 1.0 P/L, n = 1) where P is the length of the prefix, with the additional constraint that the single corrupted span always reaches the end of the sequence.

We note that this inputs-to-targets formulation can be applied to both encoder-decoder models and single-stack transformer models (e.g., decoder models). We opt to select models that predict the next target token instead of those that do so in-place (e.g., predict the current masked token in BERT) because the next-target formulation is more general and can subsume more tasks instead of using a special CLS tokens and task-specific projection heads.

2.2 MIXTURE OF DENOISERS

We conjecture that a strong universal model has to be exposed to solving diverse set of problems during pre-training. Given that pre-training is done using self-supervision, we argue that such diversity should be injected to the objective of the model, otherwise the model might suffer from lack a certain ability, like long-coherent text generation.

Motivated by this, as well as current class of objective functions, we define three main paradigms that are used during pre-training:

R-Denoiser - The regular denoising is the standard span corruption introduced in Raffel et al. (2019) that uses a range of 2 to 5 tokens as the span length, which masks about 15% of input tokens. These spans are short and potentially useful to acquire knowledge instead of learning to generate fluent text.

S-Denoiser - A specific case of denoising where we observe a strict sequential order when framing the inputs-to-targets task, i.e., prefix language modeling. To do so, we simply partition the input sequence into two sub-sequences of tokens as context and target such that the targets do not rely on future information. This is unlike standard span corruption where there could be a target token with earlier position than a context token. Note that similar to the Prefix-LM setup, the context (prefix) retains a bidirectional receptive field. We note that S-Denoising with very short memory or no memory is in similar spirit to standard causal language modeling.

1This is roughly approximate since the model still conditions on a sentinel token.

Published as a conference paper at ICLR 2023

X-Denoiser - An extreme version of denoising where the model must recover a large part of the input, given a small to moderate part of it. This simulates a situation where a model needs to generate long target from a memory with relatively limited information. To do so, we opt to include examples with aggressive denoising where approximately 50% of the input sequence is masked. This is by increasing the span length and/or corruption rate. We consider a pre-training task to be extreme if it has a long span (e.g., 12 tokens) or have a large corruption rate (e.g., 30%). X-denoising is motivated by being an interpolation between regular span corruption and language model like objectives.

This set of denoisers has strong connections with previously used objective functions: R-Denoising is the T5 span corruption objective, S-Denoising is connected to causal language models that are GPT-like, and X-Denoising can expose the model to a combination of objectives from T5 and Causal LMs. Notably, X-denoisers are also connected to improve sample efficiency since more tokens are learned to be predicted in each sample, in similar spirit to LMs. We propose blending all these tasks in a uniform fashion and have a hybrid self-supervised objective. The final objective is a mixture of 7 denoisers that are configured as follows:

Denoiser Setting

R (µ = 3, r = 0.15, n) (µ = 8, r = 0.15, n) S (µ = L4, r = 0.25, 1) X (µ = 3, r = 0.5, n) (µ = 8, r = 0.5, n) (µ = 64, r = 0.15, n) (µ = 64, r = 0.5, n)

Table 1: Configuration of UL2 s mixture-of-denoisers used in the paper.

For Xand R-Denoisers, the span length is sampled from a normal distribution with mean of µ. For S-Denoisers, we use a uniform distribution, fix the number of corrupted spans to 1, and have an additional constraint that the corrupted span should end at the end of the original input text, i.e. no un-cropped token should appear after the corrupted part. This is roughly equivalent to seq2seq denoising or the Prefix LM pre-training objective.

Since LM is a special case of Prefix-LM, we did not find it necessary to include a casual LM task into the mixture. All tasks have an approximate equal participation in the mixture. We also explore an alternative where we increase number of S-denoisers up to 50% of denoisers in the Mixture and all other denoisers take up the remainder. We present detailed ablation studies of various design choices in the later sections.

Finally, the mixing in Mixture-of-Denoisers is what makes it universally powerful. Alone, some of the denoiser types do not perform well. For instance, the original T5 paper explored an option with 50% corruption rate (X-denoising) and found that to not work well.

The implementation of UL2 s mixture of denoiser is simple and easy to implement using a library like seqio2 (Roberts et al., 2022). See appendix for more details on implementation.

2.3 MODE SWITCHING

We introduce the notion of paradigm-shifting via mode switching. During pre-training, we feed the model an extra paradigm token, i.e., {[R], [S], [X]} that helps the model switch gears and operate on a mode that is more suitable for the given task. For fine-tuning and downstream few-shot learning, to trigger the model to learn better solutions, we also add a paradigm token with respect to the setups and requirements of the downstream task. Mode switching in fact binds downstream behavior to one of the modes we used during upstream training.

2.4 MODEL ARCHITECTURE

UL2 adopts an architecture-agnostic philosophy. We argue that the choice between both architectures (encoder-decoder vs decoder-only) is a more of an efficiency trade-off and that architecture

2https://github.com/google/seqio

Published as a conference paper at ICLR 2023

Table 2: Experimental results on a suite of language understanding and generation tasks on both supervised and one-shot setup. Models are pretrained on 32B tokens.

Supervised Finetuning In-context One-shot Obj Arch Params SG XS SGD TOT SG XS SGD TOT LM

CLM Dec 167M 62.24 28.18 55.44 59.40 39.22 1.16 1.40 0.20 -2.35 PLM Dec 167M 62.44 28.21 55.55 59.52 42.54 1.08 3.70 6.40 -2.54 SC Dec 167M 67.67 29.14 55.48 60.47 38.53 1.16 2.20 1.60 -3.62 SCLM Dec 167M 63.36 29.02 55.71 60.00 40.78 3.03 1.27 0.10 -2.38 UL2 Dec 167M 65.50 28.90 55.80 60.39 42.30 8.01 6.30 5.80 -2.34

PLM ED 335M 69.30 31.95 55.70 60.91 38.18 6.50 7.11 3.90 -2.42 SC ED 335M 72.00 31.05 55.80 61.25 38.51 7.49 1.43 2.10 -7.23 SCLM ED 335M 72.50 31.69 55.70 60.94 39.74 5.13 8.70 7.30 -2.40 Uni LM ED 335M 71.10 31.00 55.83 61.03 39.86 6.70 6.50 4.10 -2.65 UL2 ED 335M 73.10 31.86 56.10 61.50 41.30 11.51 6.63 6.50 -2.55

choice should not be conflated with the pretraining objective. Hence, we have both a UL2 decoder and UL2 encoder-decoder in similar spirit to how there are multiple sizes per model. We discuss this efficiency trade-off in detail in our experiment section. UL2 adopts a pretty standard vanilla T5 Transformer that have been enhanced with modifications that have withstood the test of time, i.e., GLU layers (Shazeer, 2020) and T5-style relative attention. To not further conflate architectural modifications with pretraining contributions, the backbone of the model remains similar to a T5-like model. This is also in light of results such as (Narang et al., 2021).

3 ABLATIVE EXPERIMENTS

This section describes our ablative experimental setup (e.g., baselines, datasets, implementation details) and results. Our overall findings show that UL2 outperforms T5-like and GPT-like models on 9 out of 9 tasks.

Baselines For pre-training objectives, we compare with the following pre-training baselines: (1) Causal Language Model, (2) Prefix Language model, (3) Span Corruption (SC), (4) Span Corruption + LM (SCLM), and (5) Uni LM (Dong et al., 2019). For all objectives, we explore both singlestack and encoder-decoder architectures. All architectures are inputs-to-targets either implemented in encoder-decoder or decoder-only model structures since we consider BERT-style masked language modeling pretraining to have already been effectively subsumed by this style of pretraining, as empirically made evident in (Raffel et al., 2019). Task-specific classification heads are also not recommended, since they clearly go against the principle of having a universal model (and are also very cumbersome).

Experimental Setup We conduct our experiments on a diverse set of supervised and promptbased few-shot learning tasks. To enable the comparison between models from this perspective, we need an aggregate performance score. However, metrics on different tasks we include are widely different in nature take, for example, F1 and perplexity. To address this, we opt to report and use the normalized relative gain with respect to baselines as an overall metric. For this purpose, we use the standard language model (decoder-only) (GPT-like) and standard span denoising encoderdecoder (T5) as prime baselines and report all methods against their relative performance against these well-established candidates. We believe this is the most suitable method for comparing these models since it is easy to reason about how much a new model is generally better than a popular setting (e.g., GPT or T5-like). The full details of our experimental setup, including metrics, datasets and implementtion details can be found in the Appendix.

3.1 OVERVIEW OF ABLATIVE EXPERIMENTAL RESULTS

Table 2 reports the raw results on all the benchmark tasks and datasets. To facilitate easier comparison across setups, we also report relative comparisons against well-established baselines such as T5 and GPT models. This is reported in Tables 3 and 4 respectively.

Published as a conference paper at ICLR 2023

Table 3: Relative performance compared to standard encoder-decoder span corruption model (T5). Results in this table are expressed in terms of relative percentage improvements over a baseline. Model with denotes the main compared baseline. Overall score column is normalized to be weighted equally across tasks.

Supervised One-shot Obj Arch SG XS SGD TOT SGL XS SGD TOT LM All Win

CLM Dec -13.6 -9.2 -0.7 -3.0 +1.8 -91.7 -2.2 -90.5 +208 -31.7 2/9 PLM Dec -13.3 -9.2 -0.5 -2.8 +10.5 -85.6 +158 +205 +185 -11.0 4/9 SC Dec -5.6 -6.2 -0.6 -1.3 +0.05 -84.5 +54 -23.8 +99 -20.6 3/9 SCLM Dec -6.0 -6.5 -0.2 -2.0 +5.9 -59.6 -11.3 -95 +204 -16.1 2/9 Uni LM Dec -10.1 -8.2 -0.2 -2.3 -5.3 -69.1 +382 +110 +200 -16.1 3/9 UL2 Dec -9.0 -6.9 0.0 -1.4 +9.8 +6.9 +340 +176 +209 +14.1 5/9

PLM ED -3.7 +2.9 -0.2 -0.6 -0.86 -13.3 +397 +86 +199 +16.7 5/9 SC ED 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - SCLM ED +0.7 +2.1 -0.2 -0.5 +3.2 -31.6 +508 +248 +201 +28.3 7/9 Uni LM ED -1.2 -0.2 +0.1 -0.4 +3.5 -11.0 +355 +95 +173 +19.8 5/9 UL2 ED +1.5 +2.6 +0.5 +0.4 +7.2 +53.6 +363 +210 +184 +43.6 9/9

Table 4: Relative performance compared to standard decoder causal language model (GPT-like). Results in this table are expressed in terms of relative percentage improvements over a baseline. Model with denotes the main compared baseline. Overall score column is normalized to be weighted equally across tasks.

Supervised One-shot Obj Arch SG XS SGD TOT SG XS SGD TOT LM All Win

CLM Dec 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - PLM Dec +0.3 +0.1 +0.2 +0.2 +8.5 +74.3 +164 +3100 -8.0 +21.4 8/9 Uni LM Dec +4.0 +1.1 +0.5 +0.7 -7.0 +274 +393 +2100 -2.5 +21.0 7/9 SC Dec +8.7 +3.4 +0.1 +1.8 -1.8 +87.0 +57.1 +700 -54.2 +13.9 7/9 SCLM Dec +1.8 +3.0 +0.5 +1.0 +4.0 +387 -9.3 -50 -1.3 +15.8 6/9 UL2 Dec +5.2 +2.6 +0.6 +1.7 +7.9 +1190 +350 +2800 +0.3 +45.7 9/9

PLM ED +11.3 +13.4 +0.5 +2.5 -2.6 +946 +408 +1850 -2.9 +48.6 7/9 SC ED +16.5 +10.2 +0.6 +3.1 -1.8 +1107 +2.3 +950 -208 +31.7 7/9 SCLM ED +15.7 +12.5 +0.5 +2.6 +1.3 +726 +522 +3550 -2.2 +60.3 8/9 Uni LM ED +14.2 +10.0 +0.7 +2.7 +1.6 +974 +365 +1950 -12.9 +52.6 8/9 UL2 ED +17.4 +13.1 +1.2 +3.5 +5.3 +1754 +373 +3150 -8.3 +76.1 8/9

Decoder Vs Encoder-Decoder Before we dive into the results of this segment, we would like to remind readers that there is no easy way to compare decoder-only models with encoder-decoder models. In short, we can either compare them in a compute-matched setup or a parameter-matched way. Therefore, the encoder-decoder models in these set of results have approximately twice the number of parameters as the decoder models but have similar speeds.

We note that this may slightly favor encoder-decoders since this can be interpreted form of model sparsity. Moving back to the results, when using T5 as the reference baseline, we note that, with the exception of UL2 Decoder, none of the pre-trained decoders models outperform T5. Additionally, there is a 10% to 30% degradation in overall relative performance. The best decoder baseline model here is the Prefix-LM decoder model, which is about 10% worse than the T5 baseline. It is clear from these results that encoder-decoder models should be preferred over decoder-only models if and only if there is no concern about storage, i.e., parameter counts are generally less important than actual throughput (see (Dehghani et al., 2021a) for a detailed discussion).

When there is a parameter constraint, the Prefix-LM decoder makes for a suitable alternative. Finally, an interesting data point is how we were able to push the UL2 decoder to outperform the T5 encoder-decoder setup by +14.6%. That said, this UL2 decoder does not outperform our UL2 encoder-decoder. However, this reinforces our point that the self-supervision objective may be intrinsically more important than the backbone architecture and negotiating architectural choices is mainly about efficiency trade-offs that can be studied independently.

Published as a conference paper at ICLR 2023

Is GPT and/or T5 the optimal setup? Based on the relative comparisons against a GPT-like (causal LM + decoder) and T5-like (span corruption + encoder decoder) setup, we are able to easily identify if the well-established setups are indeed optimal or already close to optimal. Firstly, the causal LM (GPT-like) setup appears to be the worse configuration as it is outperformed by all our baselines. We thus make the straightforward recommendation of always at least training with Prefix LM or Uni LM whenever possible. The best decoder-only model (with the exception of UL2) is the Prefix-LM pre-training that keeps a memory prefix for a language model to condition on. Regarding Prefix-LM pre-training, it is interesting that Prefix-LM actually outperforms the T5 span corrupt setup by +16.7%. The Prefix-LM encoder-decoder model is indeed less effective than the default T5 model on Super GLUE but is on a whole, stronger especially when it comes to one-shot or open text-generation. Overall, between the Prefix-LM and the span corruption encoder-decoder model (T5), it is unclear to which is the universally superior model as there are gives and takes across the different sub-tasks although it is worthy noting the Prefix-LM Enc Dec model only sacrifices a minor degradation in certain tasks for a huge multifold increase in other tasks.

On the Performance of Uni LM and SCLM On the encoder-decoder setup, both the Uni LM and SCLM objective performs better than the standard span corruption objective in terms of aggregated and normalized overall gain. This shows that, in general, mixing pre-training objectives is helpful. On the decoder setup, there is an overall gain of +9.4% gain for Uni LM and +16.1% for SCLM compared to the baseline causal LM. In terms of individual tasks, Uni LM and SCLM both outperforms T5 on 6 out of 9 tasks. It is also noteworthy that SCLM performs the best out of all models on 1shot generation (SGD and TOTTO).

On the Performance of the Proposed UL2 Finally, we note that UL2 performs the best when compared against both the GPT-like model and the T5-like model. Overall, UL2 outperforms by T5 +43.4% and +76.2% when compared to the GPT-like CLM decoder model. This is the highest relative (overall) gain compared to all other alternatives. We also note that on all individual tasks, UL2 outperforms T5 on all 9 out of 9 considered tasks. Hence, UL2 is a universally better option compared to the span corruption T5 model. While UL2 doesn t always outperform all baselines on all individual tasks, UL2 is very consistent. Even when it loses to another method on a task, the loss is relatively marginal (e.g., 6.5 vs 7.3 on one-shot TOTTO). Conversely, when UL2 outperforms a baseline like T5, the gain can be as large as +363%. UL2 remains the most consistently strong method. The consistent improvement also suggests that it can be used as a more consistent replacement to T5 and GPT-like models.

4 SCALING TO 20B PARAMETERS

We are also interested to evaluate UL2 in a scaled up setting. We trained a 20B UL2 model that we eventually released publicly. This section describes the highlights of our experimental results for UL2 20B.

4.0.1 EXPERIMENTS ON SUPERVISED FINETUNING

We conduct experiments over 50 NLP tasks under the supervised finetuning paradigm. These tasks range from language generation (with automated and human evaluation), language understanding, text classification, question answering, commonsense reasoning, long text reasoning, structured knowledge grounding and information retrieval. The full list of 50 tasks and their corresponding experimental results can be found in the Appendix in 11. Overall, our experimental results show that UL2 achieves state-of-the-art performance on around 50+ NLP tasks and setups. Finally, we note that UL2 20B does pretty well on human evaluation on GENIE tasks, outperforming sota on several metrics. This ascertains that the generation quality of UL2 is reasonably solid.

4.0.2 RESULTS ON ZERO-SHOT AND ONE-SHOT LEARNING

Table 5 reports results on zero-shot Super GLUE. We note that UL2 20B outperforms GPT-3 and other compute-matched models on zero-shot NLU.

Published as a conference paper at ICLR 2023

Table 5: Results on zero-shot learning on Super GLUE dataset. We compare with GPT-3, GLa M and Pa LM (Chowdhery et al., 2022). We also include models that are relatively compute-matched with UL20B such as T5-XXL with LM adaptation (Lester et al., 2021), GPT-3 13B and GLa M-8B dense. Notably, UL20B outperforms GPT-3 175B and all other models in a similar compute class on average score.

Model Bool Q CB RTE Re CORD WSC Wi C COPA Multi RC Avg

ST-Mo E-32B (269B) 40.8 41.1 52.7 50.0 57.5 50.0 56.0 30.3 47.6 GPT-3 175B 60.5 46.4 63.5 90.2 65.4 0.0 91.0 72.9 61.2 GLa M-Mo E 1.2T 83.0 33.9 68.8 90.3 84.9 50.5 90.0 45.1 68.3 Pa LM 540B 88.0 51.8 72.9 92.9 89.1 59.1 93.0 83.5 78.8

T5-XXL 44.3 37.5 48.8 85.8 59.3 50.9 70.0 23.0 52.5 GPT-3 13B 66.2 19.6 62.8 89.0 64.4 0.0 84.0 71.4 57.2 Pa LM-Dense 8B 68.3 41.1 54.2 87.8 78.9 47.0 86.0 47.5 63.9

UL2 20B (single ckpt) 63.1 41.1 60.7 88.1 79.9 49.8 85.0 36.2 63.0 UL2 20B (best) 63.1 50.0 60.7 88.1 80.6 55.2 88.0 36.2 65.2

Table 6: Results on One-Shot Summarization on XSUM.

Model Rouge-1 Rouge-2 Rouge-L

La MDA 137B - 5.4 - Pa LM 62B - 11.2 - Pa LM 540B - 12.2 -

Pa LM 8B - 7.9 - T5 XXL 11B 0.6 0.1 0.6 T5 XXL 11B + LM 13.3 2.3 10.7 UL2 20B 25.5 8.6 19.8

Table 14 reports results on 1-shot summarization. Our results show that the performance of UL2 20B is about 3x the performance of LM adapted T5 XXL model. Moreover, UL2 20B outperform La MDA 137B and has better performance compared to Pa LM 8B which is approximately compute-matched with UL2. The best result, however, is still the larger 540B and 62B Pa LM models.

4.0.3 RESULTS ON CHAIN-OF-THOUGHT REASONING

We conduct experiments (full details in Appendix) to show that UL2 20B is capable of performing chain-of-thought reasoning (Wei et al., 2022b). Whereas most prior success on chain-of-thought has been shown on large and non-public models, UL2 is a comparatively smaller and publicly available.

Table 7: Chain-of-thought prompting and self-consistency (SC) results on five arithmetic reasoning benchmarks. Full details in Appendix.

Model GSM8K SVAMP ASDiv AQu A MAWPS Average

UL2 20B: standard prompting 4.1 10.1 16.0 20.5 16.6 13.5 UL2 20B: Co T prompting 4.4 12.5 16.9 23.6 19.1 15.3 UL2 20B: Co T prompting + calc. 6.9 28.3 34.3 23.6 42.7 27.2 UL2 20B: Co T prompting + calc. + SC 10.2 41.4 43.5 26.9 57.9 36.0

5 CONCLUSION

We proposed a new paradigm for training universally effective models. UL2 is characterized by two key ideas. Firstly, we propose a new Mixture of Denoisers (Mo D) pretraining that frames multiple pretraining tasks as span corruption, diversifies and then mixes them. Secondly, we introduce mode switching, a way of associating downstream task behaviour to upstream pretraining. Extensive ablative experiments show that UL2 consistently outperforms GPT-like and T5 models on a wide range of supervised and few-shot tasks, outperforming T5 on 9 out of 9 tasks and by a normalized overall gain of +76.1%. Finally, we scale UL2 up to 20B parameters and conduct experiments on a diverse suite of 50 to 60 NLP tasks and setups. UL2 achieves sota performance on 50 of them. Pretrained checkpoints will be released at https://github.com/anonymous.

Published as a conference paper at ICLR 2023

Armen Aghajanyan, Akshat Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better fine-tuning by reducing representational collapse. ar Xiv preprint ar Xiv:2008.03156, 2020.

Armen Aghajanyan, Dmytro Okhonko, Mike Lewis, Mandar Joshi, Hu Xu, Gargi Ghosh, and Luke Zettlemoyer. Htlm: Hyper-text pre-training and prompting of language models. ar Xiv preprint ar Xiv:2107.06955, 2021.

Rami Aly, Zhijiang Guo, Michael Sejr Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. The fact extraction and VERification over unstructured and structured information (FEVEROUS) shared task. In Proceedings of the Fourth Workshop on Fact Extraction and VERification (FEVER), pp. 1 13, Dominican Republic, November 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.fever-1.1. URL https://aclanthology.org/2021.fever-1.1.

Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q Tran, Dara Bahri, Jianmo Ni, et al. Ext5: Towards extreme multi-task scaling for transfer learning. ar Xiv preprint ar Xiv:2111.10952, 2021.

Shreyan Bakshi, Soumya Batra, Peyman Heidari, Ankit Arun, Shashank Jain, and Michael White. Structure-to-text generation with self-training, acceptability classifiers and context-conditioning for the gem shared task. In Proceedings of the 1st Workshop on Natural Language Generation, Evaluation, and Metrics (GEM 2021), pp. 136 147, 2021.

Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen-tau Yih, and Yejin Choi. Abductive commonsense reasoning. ar Xiv preprint ar Xiv:1908.05739, 2019.

Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pp. 7432 7439, 2020.

Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Nuanced metrics for measuring unintended bias with real data for text classification. Co RR, abs/1903.04561, 2019. URL http://arxiv.org/abs/1903.04561.

James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake Vander Plas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+Num Py programs, 2018. URL http: //github.com/google/jax.

Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. ar Xiv preprint ar Xiv:2005.14165, 2020.

Thiago Castro Ferreira, Claire Gardent, Nikolai Ilinykh, Chris van der Lee, Simon Mille, Diego Moussallem, and Anastasia Shimorina. The 2020 bilingual, bi-directional webnlg+ shared task overview and evaluation results (webnlg+ 2020). In Proceedings of the 3rd Web NLG Workshop on Natural Language Generation from the Semantic Web (Web NLG+ 2020), pp. 55 76, Dublin, Ireland (Virtual), 2020. Association for Computational Linguistics.

Mingda Chen, Zewei Chu, Sam Wiseman, and Kevin Gimpel. Summscreen: A dataset for abstractive screenplay summarization. ar Xiv preprint ar Xiv:2104.07091, 2021.

Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. ar Xiv preprint ar Xiv:1909.02164, 2019.

Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Wang. Hybridqa: A dataset of multi-hop question answering over tabular and textual data. ar Xiv preprint ar Xiv:2004.07347, 2020.

Published as a conference paper at ICLR 2023

Aakanksha Chowdhery, Sharan Narang, and Jacob Devlin. Palm: Scaling language modeling with pathways. ar Xiv preprint, 2022.

Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the AI2 reasoning challenge. Co RR, abs/1803.05457, 2018. URL http://arxiv.org/abs/1803.05457.

Jordan Clive, Kris Cao, and Marek Rei. Control prefixes for text generation. Co RR, abs/2110.08329, 2021. URL https://arxiv.org/abs/2110.08329.

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. ar Xiv preprint ar Xiv:2110.14168, 2021. URL https://arxiv.org/abs/2110.14168.

Andrew M Dai and Quoc V Le. Semi-supervised sequence learning. Advances in neural information processing systems, 28:3079 3087, 2015.

Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A Smith, and Matt Gardner. A dataset of information-seeking questions and answers anchored in research papers. ar Xiv preprint ar Xiv:2105.03011, 2021.

Mostafa Dehghani, Anurag Arnab, Lucas Beyer, Ashish Vaswani, and Yi Tay. The efficiency misnomer. ar Xiv preprint ar Xiv:2110.12894, 2021a.

Mostafa Dehghani, Yi Tay, Alexey A Gritsenko, Zhe Zhao, Neil Houlsby, Fernando Diaz, Donald Metzler, and Oriol Vinyals. The benchmark lottery. 2021b.

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ar Xiv preprint ar Xiv:1810.04805, 2018.

Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language understanding and generation. ar Xiv preprint ar Xiv:1905.03197, 2019.

Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. ar Xiv preprint ar Xiv:2112.06905, 2021.

Ondˇrej Duˇsek, David M Howcroft, and Verena Rieser. Semantic Noise Matters for Neural Natural Language Generation. In Proceedings of the 12th International Conference on Natural Language Generation (INLG 2019), pp. 421 426, Tokyo, Japan, 2019. URL https://www.aclweb. org/anthology/W19-8652/.

Julian Martin Eisenschlos, Maharshi Gor, Thomas M uller, and William W Cohen. Mate: Multi-view attention for table transformer efficiency. ar Xiv preprint ar Xiv:2109.04312, 2021.

Alexander R Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir R Radev. Multi-news: A largescale multi-document summarization dataset and abstractive hierarchical model. ar Xiv preprint ar Xiv:1906.01749, 2019.

William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. ar Xiv preprint ar Xiv:2101.03961, 2021.

Wikimedia Foundation. Acl 2019 fourth conference on machine translation (wmt19), shared task: Machine translation of news. URL http://www.statmt.org/wmt19/ translation-task.html.

Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Ammanamanchi, Aremu Anuoluwapo, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna Clinciu, Dipanjan Das, Kaustubh D Dhole, et al. The gem benchmark: Natural language generation, its evaluation and metrics. ar Xiv preprint ar Xiv:2102.01672, 2021.

Sebastian Gehrmann, Elizabeth Clark, and Thibault Sellam. Repairing the cracked foundation: A survey of obstacles in evaluation practices for generated text. ar Xiv preprint ar Xiv:2202.06935, 2022.

Published as a conference paper at ICLR 2023

Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. Did aristotle use a laptop? A question answering benchmark with implicit reasoning strategies. TACL, 2021. doi: 10.1162/tacl a 00370. URL https://aclanthology.org/2021.tacl-1.21.

Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. Samsum corpus: A humanannotated dialogue dataset for abstractive summarization. ar Xiv preprint ar Xiv:1911.12237, 2019.

Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. ar Xiv preprint ar Xiv:2006.03654, 2020.

Yun He, Huaixiu Steven Zheng, Yi Tay, Jai Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, Ya Guang Li, Zhao Chen, Donald Metzler, et al. Hyperprompt: Prompt-based task-conditioning of transformers. ar Xiv preprint ar Xiv:2203.00759, 2022.

Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In Advances in neural information processing systems, pp. 1693 1701, 2015.

Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. ar Xiv preprint ar Xiv:1801.06146, 2018.

Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Cosmos qa: Machine reading comprehension with contextual commonsense reasoning. ar Xiv preprint ar Xiv:1909.00277, 2019.

Luyang Huang, Shuyang Cao, Nikolaus Parulian, Heng Ji, and Lu Wang. Efficient attentions for long document summarization. ar Xiv preprint ar Xiv:2104.02112, 2021.

Mohit Iyyer, Wen-tau Yih, and Ming-Wei Chang. Search-based neural structured learning for sequential question answering. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1821 1831, 2017.

Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64 77, 2020.

Norman P Jouppi, Doe Hyun Yoon, George Kurian, Sheng Li, Nishant Patil, James Laudon, Cliff Young, and David Patterson. A domain-specific supercomputer for training deep neural networks. Communications of the ACM, 63(7):67 78, 2020.

Jared Kaplan, Sam Mc Candlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. ar Xiv preprint ar Xiv:2001.08361, 2020.

Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. ar Xiv preprint ar Xiv:2005.00700, 2020.

Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, and Daniel S. Weld. GENIE: A leaderboard for human-in-the-loop evaluation of text generation. Co RR, abs/2101.06561, 2021. URL https://arxiv.org/abs/2101. 06561.

Daniel Khashabi, Yeganeh Kordi, and Hannaneh Hajishirzi. Unifiedqa-v2: Stronger generalization via broader cross-format training. ar Xiv preprint ar Xiv:2202.12359, 2022.

Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. Qasc: A dataset for question answering via sentence composition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp. 8082 8090, 2020.

Tom aˇs Koˇcisk y, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, G abor Melis, and Edward Grefenstette. The Narrative QA Reading Comprehension Challenge. Transactions of the Association for Computational Linguistics, 2018.

Published as a conference paper at ICLR 2023

Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. MAWPS: A math word problem repository. NAACL, 2016. doi: 10.18653/v1/N16-1136. URL https://aclanthology.org/N16-1136.

Yuta Koreeda and Christopher D Manning. Contractnli: A dataset for document-level natural language inference for contracts. ar Xiv preprint ar Xiv:2110.01799, 2021.

Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin Kenton Lee, Kristina Toutanova, Llion Jones Matthew Kelcey, Ming-Wei Chang, Andrew M Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural Questions: a Benchmark for Question Answering Research. In Transactions of the ACL, 2019.

Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. RACE: Large-scale Re Ading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 785 794, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. doi: 10.18653/v1/D17-1082. URL https://aclanthology.org/D17-1082.

Alyssa Lees, Vinh Q Tran, Yi Tay, Jeffrey Sorensen, Jai Gupta, Donald Metzler, and Lucy Vasserman. A new generation of perspective api: Efficient multilingual character-level transformers. ar Xiv preprint ar Xiv:2202.11176, 2022.

Dmitry Lepikhin, Hyouk Joong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. ar Xiv preprint ar Xiv:2006.16668, 2020.

Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. ar Xiv preprint ar Xiv:2104.08691, 2021.

Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. Mtop: A comprehensive multilingual task-oriented semantic parsing benchmark. ar Xiv preprint ar Xiv:2008.09335, 2020.

Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. Common Gen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pp. 1823 1840, Online, November 2020. Association for Computational Linguistics. URL https: //www.aclweb.org/anthology/2020.findings-emnlp.165.

Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems. ACL, 2017. doi: 10.18653/v1/ P17-1015. URL https://aclanthology.org/P17-1015.

Peter J Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. Generating wikipedia by summarizing long sequences. ar Xiv preprint ar Xiv:1801.10198, 2018.

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ar Xiv preprint ar Xiv:1907.11692, 2019.

Nicholas Lourie, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Unicorn on rainbow: A universal commonsense reasoning model on a new multitask benchmark. ar Xiv preprint ar Xiv:2103.13009, 2021.

Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pp. 142 150, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. URL http: //www.aclweb.org/anthology/P11-1015.

Published as a conference paper at ICLR 2023

Shen Yun Miao, Chao Chun Liang, and Keh Yih Su. A diverse corpus for evaluating and developing English math word problem solvers. ACL, 2020. doi: 10.18653/v1/2020.acl-main.92. URL https://aclanthology.org/2020.acl-main.92.

Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. ar Xiv preprint ar Xiv:1809.02789, 2018.

Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pp. 3111 3119, 2013.

Linyong Nan, Dragomir Radev, Rui Zhang, Amrit Rau, Abhinand Sivaprasad, Chiachun Hsieh, Xiangru Tang, Aadit Vyas, Neha Verma, Pranav Krishna, et al. Dart: Open-domain structured data record to text generation. ar Xiv preprint ar Xiv:2007.02871, 2020.

Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kry sci nski, Nick Schoelkopf, Riley Kong, Xiangru Tang, et al. Fetaqa: Free-form table question answering. ar Xiv preprint ar Xiv:2104.00369, 2021.

Sharan Narang, Hyung Won Chung, Yi Tay, William Fedus, Thibault Fevry, Michael Matena, Karishma Malkan, Noah Fiedel, Noam Shazeer, Zhenzhong Lan, et al. Do transformer modifications transfer across implementations and applications? ar Xiv preprint ar Xiv:2102.11972, 2021.

Shashi Narayan, Shay B. Cohen, and Mirella Lapata. Don t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. Ar Xiv, abs/1808.08745, 2018.

Shashi Narayan, Yao Zhao, Joshua Maynez, Gonc alo Sim oes, Vitaly Nikolaev, and Ryan Mc Donald. Planning with learned entity prompts for abstractive summarization. Transactions of the Association for Computational Linguistics, 9:1475 1492, 2021. doi: 10.1162/tacl a 00438. URL https://aclanthology.org/2021.tacl-1.88.

ME Peters M Neumann, M Iyyer, M Gardner, C Clark, K Lee, and L Zettlemoyer. Deep contextualized word representations. ar Xiv preprint ar Xiv:1802.05365, 2018.

Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. Adversarial nli: A new benchmark for natural language understanding. ar Xiv preprint ar Xiv:1910.14599, 2019.

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. ar Xiv preprint ar Xiv:2203.02155, 2022. URL http://go/ arxiv/2203.02155.

Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, et al. Quality: Question answering with long input texts, yes! ar Xiv preprint ar Xiv:2112.08608, 2021.

Ankur P Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. Totto: A controlled table-to-text generation dataset. ar Xiv preprint ar Xiv:2004.14373, 2020.

Panupong Pasupat and Percy Liang. Compositional semantic parsing on semi-structured tables. ar Xiv preprint ar Xiv:1508.00305, 2015.

Arkil Patel, Satwik Bhattamishra, and Navin Goyal. Are NLP models really able to solve simple math word problems? NAACL, 2021. URL https://aclanthology.org/2021. naacl-main.168.pdf.

Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532 1543, 2014.

Published as a conference paper at ICLR 2023

Guanghui Qin, Yukun Feng, and Benjamin Van Durme. The nlp task effectiveness of long-range transformers. ar Xiv preprint ar Xiv:2202.07856, 2022.

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. 2019.

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. ar Xiv preprint ar Xiv:1910.10683, 2019.

Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta, and Pranav Khaitan. Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset. ar Xiv preprint ar Xiv:1909.05855, 2019.

Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, Curtis Hawthorne, Aitor Lewkowycz, Alex Salcianu, Marc van Zee, Jacob Austin, Sebastian Goodman, Livio Baldini Soares, Haitang Hu, Sasha Tsvyashchenko, Aakanksha Chowdhery, Jasmijn Bastings, Jannis Bulian, Xavier Garcia, Jianmo Ni, Andrew Chen, Kathleen Kenealy, Jonathan H. Clark, Stephan Lee, Dan Garrette, James Lee-Thorp, Colin Raffel, Noam Shazeer, Marvin Ritter, Maarten Bosma, Alexandre Passos, Jeremy Maitin-Shepard, Noah Fiedel, Mark Omernick, Brennan Saeta, Ryan Sepassi, Alexander Spiridonov, Joshua Newlan, and Andrea Gesmundo. Scaling up models and data with t5x and seqio, 2022. URL https://arxiv.org/abs/2203.17189.

Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. Getting closer to ai complete question answering: A set of prerequisite real tasks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pp. 8722 8731, 2020.

Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot task generalization. ar Xiv preprint ar Xiv:2110.08207, 2021.

Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. Socialiqa: Commonsense reasoning about social interactions. ar Xiv preprint ar Xiv:1904.09728, 2019.

Tal Schuster, Adam Fisch, and Regina Barzilay. Get your vitamin c! robust fact verification with contrastive evidence. ar Xiv preprint ar Xiv:2103.08541, 2021a.

Tal Schuster, Ashwin Kalyan, Alex Polozov, and Adam Tauman Kalai. Programming puzzles. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021b. URL https://openreview.net/forum?id=fe_h Cc4RBrg.

Uri Shaham, Elad Segal, Maor Ivgi, Avia Efrat, Ori Yoran, Adi Haviv, Ankit Gupta, Wenhan Xiong, Mor Geva, Jonathan Berant, and Omer Levy. Scrolls: Standardized comparison over long language sequences, 2022.

Noam Shazeer. Glu variants improve transformer. ar Xiv preprint ar Xiv:2002.05202, 2020.

Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International Conference on Machine Learning, pp. 4596 4604. PMLR, 2018.

Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick Le Gresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism. ar Xiv preprint ar Xiv:1909.08053, 2019.

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adri a Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. ar Xiv preprint ar Xiv:2206.04615, 2022.

Alon Talmor and Jonathan Berant. The web as a knowledge-base for answering complex questions. ar Xiv preprint ar Xiv:1803.06643, 2018.

Published as a conference paper at ICLR 2023

Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsenseqa: A question answering challenge targeting commonsense knowledge. ar Xiv preprint ar Xiv:1811.00937, 2018.

Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsense QA: A question answering challenge targeting commonsense knowledge. NAACL, 2019. doi: 10.18653/v1/ N19-1421. URL https://aclanthology.org/N19-1421.

Alon Talmor, Ori Yoran, Ronan Le Bras, Chandra Bhagavatula, Yoav Goldberg, Yejin Choi, and Jonathan Berant. Commonsense QA 2.0: Exposing the limits of AI through gamification. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. URL https://openreview.net/forum?id=q F7Fl UT5dxa.

Yi Tay, Zhe Zhao, Dara Bahri, Donald Metzler, and Da-Cheng Juan. Hypergrid: Efficient multi-task transformers with grid-wise decomposable hyper projections. ar Xiv preprint ar Xiv:2007.05891, 2020.

Yi Tay, Mostafa Dehghani, Jai Gupta, Dara Bahri, Vamsi Aribandi, Zhen Qin, and Donald Metzler. Are pre-trained convolutions better than pre-trained transformers? ar Xiv preprint ar Xiv:2105.03322, 2021a.

Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. Scale efficiently: Insights from pre-training and fine-tuning transformers. ar Xiv preprint ar Xiv:2109.10686, 2021b.

Yi Tay, Vinh Q Tran, Sebastian Ruder, Jai Gupta, Hyung Won Chung, Dara Bahri, Zhen Qin, Simon Baumgartner, Cong Yu, and Donald Metzler. Charformer: Fast character transformers via gradient-based subword tokenization. ar Xiv preprint ar Xiv:2106.12672, 2021c.

Yi Tay, Vinh Q Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. Transformer memory as a differentiable search index. ar Xiv preprint ar Xiv:2202.06991, 2022.

Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, Ya Guang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. Lamda: Language models for dialog applications, 2022.

Hui Wan. Multi-task learning with multi-head attention for multi-choice reading comprehension. ar Xiv preprint ar Xiv:2003.04992, 2020.

Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. ar Xiv preprint ar Xiv:1804.07461, 2018.

Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Superglue: A stickier benchmark for general-purpose language understanding systems. ar Xiv preprint ar Xiv:1905.00537, 2019.

Boxin Wang, Shuohang Wang, Yu Cheng, Zhe Gan, Ruoxi Jia, Bo Li, and Jingjing Liu. Infobert: Improving robustness of language models from an information theoretic perspective. ar Xiv preprint ar Xiv:2010.02329, 2020.

Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy, Julien Launay, and Colin Raffel. What language model architecture and pretraining objective work best for zero-shot generalization? ar Xiv preprint ar Xiv:2204.05832, 2022a.

Published as a conference paper at ICLR 2023

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. ar Xiv preprint ar Xiv:2203.11171, 2022b. URL https://arxiv.org/abs/2203.11171.

Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. Transactions on Machine Learning Research (TMLR), 2022a. URL http: //go/arxiv/2206.07682.

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. Conference on Neural Information Processing Systems (Neur IPS), 2022b. URL http://go/ arxiv/2201.11903.

Alexander Wettig, Tianyu Gao, Zexuan Zhong, and Danqi Chen. Should you mask 15% in masked language modeling? ar Xiv preprint ar Xiv:2202.08005, 2022.

Wen Xiao, Iz Beltagy, Giuseppe Carenini, and Arman Cohan. Primer: Pyramid-based masked sentence pre-training for multi-document summarization. ar Xiv preprint ar Xiv:2110.08499, 2021.

Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. Unifiedskg: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. ar Xiv preprint ar Xiv:2201.05966, 2022.

Wenhan Xiong, Jiawei Wu, Hong Wang, Vivek Kulkarni, Mo Yu, Shiyu Chang, Xiaoxiao Guo, and William Yang Wang. Tweetqa: A social media focused question answering dataset. ar Xiv preprint ar Xiv:1907.06292, 2019.

Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mt5: A massively multilingual pre-trained text-to-text transformer. ar Xiv preprint ar Xiv:2010.11934, 2020.

Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. ar Xiv preprint ar Xiv:2105.13626, 2021.

Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019.

Wenpeng Yin, Dragomir Radev, and Caiming Xiong. Docnli: A large-scale dataset for documentlevel natural language inference. ar Xiv preprint ar Xiv:2106.09449, 2021.

Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? ar Xiv preprint ar Xiv:1905.07830, 2019.

Xiang Zhang, Junbo Zhao, and Yann Le Cun. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015.

Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia Mutuma, Rahul Jha, Ahmed Hassan Awadallah, Asli Celikyilmaz, Yang Liu, Xipeng Qiu, et al. Qmsum: A new benchmark for query-based multi-domain meeting summarization. ar Xiv preprint ar Xiv:2104.05938, 2021.

Victor Zhong, Caiming Xiong, and Richard Socher. Seq2sql: Generating structured queries from natural language using reinforcement learning. Co RR, abs/1709.00103, 2017.

Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. Designing effective sparse expert models, 2022.

Published as a conference paper at ICLR 2023

A EXPERIMENTAL SETUP FOR ABLATIVE EXPERIMENTS

A.0.1 DATASETS AND TASKS

The datasets we use are Super GLUE (Wang et al., 2019), comprising of 8 NLU sub-tasks. We also conduct experiments on 3 datasets from the GEM benchmark (Gehrmann et al., 2021) that focuses on language generation problems. We arbitrarily select XSUM (summarization), To TTo (table-to-text generation) (Parikh et al., 2020) and Schema Guided Dialog (SGD) (Rastogi et al., 2019) from the GEM benchmark. For all these tasks, we evaluate on both supervised fine-tuning and prompt-based one-shot learning. Finally we also compare our models on their general ability for text generation using perplexity scores on the C4 validation set. We believe our suite of tasks gives us good coverage across many setups in the literature including supervised and conditional few-shot learning.

A.0.2 METRICS AND HOLISTIC EVALUATION

For Super GLUE, we report well-established metrics such as accuracy, F1 or Exact Match, whenever appropriate. For GEM benchmark, we use the Rouge-L metric. For language modeling we report negative log perplexity. The universality of the models, i.e., their collective performance across all range of tasks, is a main evaluation criteria here. To enable the comparison between models from this perspective, we need an aggregate performance score. However, metrics on different tasks we include are widely different in nature take, for example, F1 and perplexity. To address this, we opt to report and use the normalized relative gain with respect to baselines as an overall metric. For this purpose, we use the standard language model (decoder-only) (GPT-like) and standard span denoising encoder-decoder (T5) as prime baselines and report all methods against their relative performance against these well-established candidates. We believe this is the most suitable method for comparing these models since it is easy to reason about how much a new model is generally better than a popular setting (e.g., GPT or T5-like). We also highlight the fact that the overall gain is normalized, so this becomes harder to exploit or be susceptible to benchmark lottery effects (Dehghani et al., 2021b).

A.0.3 IMPLEMENTATION DETAILS

Our experiments are all conducted in JAX/Flax (Bradbury et al., 2018) using the open source T5X3 framework (Roberts et al., 2022) and Flaxformer4. We pre-train all models for 500K steps with a batch size of 128 and a sequence length of 512 inputs and 512 targets using the C4 corpus. The total approximate tokens seen during pre-training is approximately 32 billion tokens. Each pretraining run is typically trained using 64 to 128 TPUv4 chips (Jouppi et al., 2020). We optimize our model with the Adafactor (Shazeer & Stern, 2018) optimizer with an inverse square root learning rate. To understand the trade-off of different backbone architectures, we run all baseline pre-training objectives with both the decoder-only architecture and encoder-decoder architecture. We report key experiment results using a base architecture of approximately 167M parameters for the decoder model and 335M parameters for the encoder-decoder model. All models use a standard Transformer that uses Swi GLU layers as described in (Shazeer, 2020). We utilize the default T5 English 32K sentencepiece for all models. Within the context of decoder-only models, except for the case of the decoder model trained on causal LM, our experiments always use a bidirectional receptive field only in it s input segment and autoregressive decoding at the targets segment. This is essentially the a Prefix LM-type architecture5 (Raffel et al., 2019) which we find to be consistently better than a full causal decoder model.

A.1 MODE SWITCHING ABLATIONS

In order to ascertain that our mode switching capabilities have an effective on performance, we conduct ablation experiments.

3https://github.com/google-research/t5x. 4https://github.com/google/flaxformer 5Not to be confused with the Prefix LM pretraining objective.

Published as a conference paper at ICLR 2023

Table 9: Ablation study for Mixture-of-Denoisers. Span, Rate and SD are in percentages (%). We report Super GLUE score (SG) and XSUM Rouge-L (XS).

Ablation Method Supervised One-shot Name Span (µ) Rate (r) SD% SG XS SG XS

A - - 100 69.3 31.1 38.2 6.5 B 3 50 0 72.0 32.0 38.5 7.5

C 3,8,12 15,50 14 71.9 32.1 38.6 4.1 D 3,8,12,32 15,50 11 71.0 32.2 42.7 10.6 E 3,8,32,64 15,50 11 73.1 32.2 40.7 10.4 F 3,8,64 15,50 17 70.6 31.6 41.3 11.5 G 3,8,32,64 15 25 69.2 31.6 42.4 10.1 H 8, 64 15 25 72.5 31.2 39.2 10.9 I 3,8,12, 32 15,50 50 71.2 32.0 38.1 11.7 J 3,8,64 15,50 50 71.3 31.6 38.1 11.8 K 3,8,12 15,50 0 73.7 32.0 39.3 2.6 L 3,8,64 15,50 0 70.1 32.1 38.0 7.3

Table 8: Effect of different paradigm prompts on 1-shot evaluation, using a Encoder-Decoder architecture pre-trained using UL2 on 7B tokens.

Model/Prompt 1Shot XSum 1Shot Super GLUE

Baseline T5 6.9/0.6/6.1 33.9 UL2 / None 13.2/1.4/10.8 38.3 UL2 / [R] 13.5/1.5/11.1 38.5 UL2 / [S] 11.6/1.2/10.0 38.5 UL2 / [X] 8.9/0.9/7.6 38.7

We conduct experiments on one-shot XSum and one-shot Super GLUE. Table 8 reports the result of varying the paradigm prompt to the model. Firstly, we observe that the prompt has quite substantial effect on model performance i.e., using the right or wrong prompt can lead to a 48% gap in performance (on XSum, Rouge-1). Super GLUE, on the other hand, is less sensitive to prompting. On Super GLUE, using prompts are almost always better than not using prompts during one-shot eval. However, for XSum, getting the prompt right seems to be crucial for good performance.

A.2 MIXTURE-OF-DENOISERS ABLATIONS

We conduct extensive experiments to verify the effectiveness of individual objectives within the Mo D objective. Table 9 reports results for these ablations. We report results for varying the mean span, and corruption rate, along with the percentage of S-denoising used (denoted by % SD)). Note that the total number of denoisers in a mixture is Span Corrupt Rate + 1. We label these configurations from Var-A through Var-J to refer to them easily.

X-Denoising is Complementarily Effective but Does Not Suffice as a Standalone We observe that mixing Extreme Denoising is effective. Most of the best results across the board come from mixtures with long spans (e.g., 32 or 64). When compared with variants without long spans (Var-D vs. Var-C), we see that Var-D is strictly better. We also draw the readers attention to Var-H, which is a variant that only employs long spans. In general, Var-H performs poorly, suggesting that extreme denoising complements regular denoising but does not suffice in isolation. This also corroborates the result from Raffel et al. (2019) that shows that a 50% corruption rate does not perform well. This slightly conflicts with the finding of (Wettig et al., 2022) although our architectures use a inputs-totargets form of pretraining instead of BERT-style masked language modeling.

Small Amounts of S-Denoisers is Preferred We explore a setting where we scale S-denoisers to 50% of the entire Mo D mixture. We find that this generally hurts performance. Hence, we make a conclusion that S-denoisers are necessary but only small amounts of S-denoisers ( 20%)

Published as a conference paper at ICLR 2023

are preferred. Var-K and Var-L also explore the case where there is no S-denoising at all. While performance on one task substantially improves (Super GLUE), another substantially degrades (oneshot XSUM). Meanwhile for Var-L which is identical to var-F (but without S-denoising), performs on a whole, substantially worse. Hence, we showed that S-denoising is crucial.

A.3 MODESTLY SCALING MODEL SIZE AND PRETRAINING DATA

We conduct additional experiments by scaling up both 1) the model size and 2) pre-training dataset size. Concretely, we scale the UL2 Encoder-Decoder model up to approximately 1B parameters and increase the number of pre-training tokens to 0.5 trillion tokens. Our motivation is to conduct a sanity check that the proposed formulation also works at different model scales and to observe if there are differences and implications at operating at a larger scale. Moreover, it has also become a staple for language model research to derive scaling laws (Kaplan et al., 2020; Tay et al., 2021b). Table 10 reports results in this scaled setting. At large scale, we find that the proposed of the UL2 encoder-decoder model is still competitive. A key difference now is that UL2 drops the Super GLUE suite against T5 (1B). However, this is compensated by not only out-performing on 7 out of 8 tasks but also improving performance by 2-4 times on one-shot evaluation. The gains on supervised finetuning is smaller, but still noticeable across the board on XSUM, SGD and TOT.

Table 10: Experiments with moderately scaled up models in terms of model compute (e.g., 1B for Enc Dec and 0.5B for decoder-only) and dataset size (0.5T tokens).

Finetuning In-context Learning Model SG XS SGD TOT SG XS SGD TOT

GPT-like 62.3 37.1/15.7/30.2 56.0 60.3 36.4 1.2/0.1/1.1 3.5 0.0 T5 84.7 43.0/20.8/35.6 56.0 62.1 29.4 8.9/0.8/7.8 2.1 1.4 UL2 83.3 43.3/21.0/35.9 56.5 62.6 45.4 15.4/2.5/11.1 9.6 7.8

B 20B UL2 EXPERIMENTAL SETUP

This section describes the setup for our 20B experiments.

B.1 PRETRAINING AND MODEL CONFIGURATION

We follow the same training protocol in earlier experiments by pretraining on the C4 corpus but by also scaling the number of tokens the model sees during pretraining. We use a batch size of 1024 and 512 TPUv4 chips for pretraining this model. The model is trained on a total of 1 trillion tokens on C4 (2 million steps). The sequence length is set to 512/512 for inputs and targets. Dropout is set to 0 during pretraining. Pre-training took approximately slight more than one month for about 1 trillion tokens. We use the same mixture of denoisers as earlier sections. The model has 32 encoder layers and 32 decoder layers, dmodel of 4096 and dff of 16384. The dimension of each head is 256 for a total of 16 heads. Our model uses a model parallelism of 8. We retain the same sentencepiece tokenizer as T5 of 32k vocab size. Hence, UL20B can be interpreted as a model that is quite similar to T5 but trained with a different objective and slightly different scaling knobs. Similar to earlier experiments, UL20B is trained with Jax and T5X infrastructure. We release and open source T5Xbased model checkpoints of this 20B model.

B.1.1 SETUP AND IMPLEMENTATION DETAILS

We conduct experiments on both finetuning and in-context learning. For supervised finetuning, our models are continuously finetuned after N pretraining steps where N is typically from 50k to 100k. In other words, after each Nk steps of pretraining, we finetune on each downstream task and record its results. This is generally done in a manual fashion. While some tasks were finetuned on earlier pretrained checkpoints as the model was still pretraining, many were finetuned on checkpoints nearer to convergence that we release. As we continiously finetune, we stop finetuning on a task once it has reached sota to save compute. Finetuning is generally done in a per-task basis and not co-trained.

Published as a conference paper at ICLR 2023

Details of tasks where co-training is performed is found in the appendix. We leave the combination of massive multi-task training (Aribandi et al., 2021) and UL2 to future work.

For supervised finetuning, we generally adopt a learning rate in the range of {5 10 5, 1 10 5 1 10 4} using the Adafactor optimizer . The general recipe is that we reset Adafactor optimizer states and/or adopt a loss normalization based on the number of real target tokens. This is reminiscent of the Pa LM finetuning setup (Chowdhery et al., 2022). Batch size is generally a range of 32 to 128 although we did not find much impact of batch size on finetuning performance. Many of the evaluated tasks were not tuned much and we only ran once or twice before performing leaderboard submissions.

B.1.2 DATASETS FOR SUPERVISED FINETUNING

To demonstrate the universality of the approach, we consider a total of nearly 50+ NLP tasks. We list our categorization of tasks below. Note that the categorization of tasks are generally soft in nature and some tasks may cross into different categorization boundaries.

Language Generation - We consider summarization and data-to-text generation tasks. We use CNN/Dailymail (Hermann et al., 2015), XSUM (Narayan et al., 2018), Multi News (Fabbri et al., 2019), SAMSum (Gliwa et al., 2019), Web NLG (Castro Ferreira et al., 2020) (English), E2E (Duˇsek et al., 2019) and Common Gen (Lin et al., 2020) to evaluate our models. For Web NLG, E2E and Common Gen, we use the versions from the GEM benchmark (Gehrmann et al., 2021).

Language Generation with Human Evaluation - We evaluate on a variety of text generation tasks using human evaluation, via the GENIE leaderboard (Khashabi et al., 2021). These tasks include a NLG (Bhagavatula et al., 2019), ARC-DA (Clark et al., 2018), WMT19 (Foundation), and XSUM (Narayan et al., 2018).

Language Understanding, Classification and Question Answering - We use Reading Comprehension, Question Answering, Text Classification and natural language inference datasets. Concretely, we use RACE (Reading comprehension) (Lai et al., 2017), QASC (Khot et al., 2020), Open Book QA (Mihaylov et al., 2018), Tweet QA (Xiong et al., 2019), Qu AIL (Rogers et al., 2020), IMDB (Maas et al., 2011), Agnews (Zhang et al., 2015), Doc NLI (Yin et al., 2021), Adversarial NLI (Nie et al., 2019), Vitamin C (Schuster et al., 2021a), Civil Comments and Wikipedia Toxicity detection datasets (Borkan et al., 2019). We also use standard Super GLUE (Wang et al., 2019) and GLUE (Wang et al., 2018) datasets.

Commonsense Reasoning - We use Hella Swag (Zellers et al., 2019), Social IQA/SIQA (Sap et al., 2019), Physical IQA/PIQA (Bisk et al., 2020), Cosmos QA (Huang et al., 2019), Abductive NLI (Bhagavatula et al., 2019), Commonsense QA (Talmor et al., 2018), Commonsense QA2 (Talmor et al., 2021).

Long Range Reasoning - We use the Scrolls benchmark (Shaham et al., 2022) which comprises of seven component tasks including Gov Report (Huang et al., 2021), Sum Scr (Chen et al., 2021), QMSUm (Zhong et al., 2021), QASPER (Dasigi et al., 2021), Narrative QA (Koˇcisk y et al., 2018), Qua LITY (Pang et al., 2021), and Contract NLI (Koreeda & Manning, 2021).

Structured Knowledge Grounding - We use several component tasks from Unified SKG (Xie et al., 2022), namely Wiki TQ (Pasupat & Liang, 2015), Comp WQ (Talmor & Berant, 2018), Feta QA (Nan et al., 2021), Hybrid QA (Chen et al., 2020), Wiki SQL (Zhong et al., 2017), Tab Fat (Chen et al., 2019), Feverous (Aly et al., 2021), SQA (Iyyer et al., 2017), MTOP (Li et al., 2020) and DART (Nan et al., 2020). We select datasets that are relatively convenient to perform evaluation and uses mainstream metrics such as accuracy or exact match instead of obscure ones or those that require significant domain specific post-processing.

Information Retrieval - IR is the task of retrieving relevant documents given queries. We use the setup of the latest next generation IR paradigm, i.e., differentiable search index (Tay et al., 2022) for our experiments. We use the same NQ (Kwiatkowski et al., 2019) splits in the DSI paper.

Published as a conference paper at ICLR 2023

For each dataset, we report the best previous sota result. For generation tasks, we generally report ROUGE-2 following the advice of (Gehrmann et al., 2022). For the rest of the datasets, we report the dominant metric that is reported in prior work. For BLEU scores, we use sacrebleu. For commonsense reasoning tasks, we do not compare against approaches that use external knowledge bases as they are orthogonal and out of scope for this paper. For most part, GLUE is generally considered to be saturated and there are many unpublished results on the GLUE leaderboard. For this reason, we make a very reasonable decision of considering (Raffel et al., 2019) to be the state-of-the-art since we believe that there has not been any real advance on the GLUE benchmark since the T5 model (Raffel et al., 2019). GLUE results, given how saturated it already is, are provided as a reference and should be taken with a pinch of salt.

B.1.3 SUMMARY OF SUPERVISED FINETUNING RESULTS

This section describes the overview results of our experiments.

Table 11: Summary of UL20B results compared to state-of-the-art. (l) denotes leaderboard submission. ( ) denotes the best published we could find on the leaderboard. (e) denotes SOTA used an ensembled approach. Because we evaluate finetuning and in-context trade-offs for Super GLUE, Super GLUE scores have their own dedicated section below.

Dataset Metric Eval Sota Reference SOTA Ours

CNN/DM Rouge-2 Test Zoph et al. 21.7 21.9 XSUM Rouge-2 Test Zoph et al. 27.1 26.6 Multi News Rouge-2 Test Xiao et al. 21.1 21.7 SAMSum Rouge-2 Test Narayan et al. 28.3 29.6 Gigaword Rouge 2 Test Aghajanyan et al. 20.7 20.7 Web NLG (en) Rouge-2 Test Bakshi et al. 53.5 55.4 E2E-NLG Rouge-2 Test Xue et al. 45.8 46.5 Common Gen Rouge-2 Dev Gehrmann et al. 32.5 37.4 Schema-Guided Dialog Rouge-2 Test Gehrmann et al. 36.8 44.1

GENIE - a NLG Human (H) Test Khashabi et al. 76.0 77.0(l)

GENIE - ARC-DA (w/o IR) Human Test Khashabi et al. 72.0 72.0(l)

GENIE - WMT19 Human Test Khashabi et al. 71.0 67.0(l)6

GENIE - XSUM H-Overall Test Clive et al. 51.0 50.0(l)

GENIE - XSUM H-Concise Test Clive et al. 53.0 53.0(l)

GENIE - XSUM H-Fluency Test Clive et al. 51.0 52.0(l)

GENIE - XSUM H-No-Hallucination Test Clive et al. 53.0 54.0(l)

GENIE - XSUM H-Informativeness Test Clive et al. 49.0 49.0(l)

SIQA Accuracy Test Lourie et al. 83.2 83.3(l)

PIQA Accuracy Test Lourie et al. 90.1 90.7(l)

CSQA Accuracy Dev Lourie et al. 79.1 84.9 CSQA2 Accuracy Test Lourie et al. 69.6( ) 70.1(l)

QASC (w/o IR) Accuracy Dev Khashabi et al. 81.8 83.8 QASC (w IR) Accuracy Test Khashabi et al. 89.6 90.7(l)

Tweet QA BLEU-1 Dev Khashabi et al. 77.5 78.4 Qu AIL Accuracy Test Khashabi et al. 74.2 87.2 Adversarial QA (Bert) F1 Dev Khashabi et al. 53.6 70.1 Adversarial QA (Roberta) F1 Dev Khashabi et al. 45.5 57.5 Adversarial QA (Bidaf) F1 Dev Khashabi et al. 71.5 77.5 MCScript Accuracy Test Khashabi et al. 95.1 97.3 MCScript 2.0 Accuracy Test Khashabi et al. 94.6 97.9 RACE Accuracy Test Shoeybi et al. 90.9(e) 90.9 DREAM Accuracy Test Wan 91.8 91.8 OBQA Accuracy Test Khashabi et al. 87.2 87.2(l)

Continued on next page

6This task is German-to-English translation. Our submission is pretrained on only English C4 then finetuned on only the provided WMT19 data (no German pretraining, parallel data or backtranslation.)

Published as a conference paper at ICLR 2023

Table 11 continued from previous page

Cosmos QA Accuracy Test Lourie et al. 91.8 91.6(l)

Winogrande XL Accuracy Test Lourie et al. 91.3 90.1(l)

Doc NLI Accuracy Test Qin et al. 76.9 88.2 Adversarial NLI (r3) Accuracy Test Wang et al. 47.7 53.5 Vitamin C Accuracy Test Schuster et al. 90.8 91.1 Hellaswag Accuracy Test Lourie et al. 93.9 94.1(l)

QQP F1 Dev Raffel et al. 90.1 90.6 QNLI Accuracy Dev Raffel et al. 96.1 96.5 Co LA Matthews Dev Raffel et al. 68.6 71.5 STSB Spearman Dev Raffel et al. 92.1 92.3 Abductive NLI Accuracy Test He et al. 89.8( ) 87.5(l)

Multi NLI Accuracy Dev Raffel et al. 92.1 91.9

IMDB Accuracy Test Yang et al. 96.2 97.3 Ag News Error Test Yang et al. 4.45 4.42 Civil Comments F1 Dev Tay et al. 87.8 87.9 Wikipedia Toxicity F1 Dev Tay et al. 96.5 97.0 SST-2 Acc Dev Raffel et al. 97.3 97.0

Scrolls Challenge Aggregate Test Shaham et al. 29.2 37.9(l)

Sum Scr Rouge (Avg) Test Shaham et al. 16.3 20.0(l)

QMSum Rouge (Avg) Test Shaham et al. 19.9 20.0(l)

QASPER F1 Test Shaham et al. 26.6 37.6(l)

Narrative QA F1 Test Shaham et al. 18.5 24.2(l)

QUALITY EM Test Shaham et al. 26.0 45.8(l)

Contract NLI EM Test Shaham et al. 77.4 88.7(l)

Gov Rep Rouge (Avg) Test Shaham et al. 37.2 36.2(l)

Wiki TQ Accuracy Test Xie et al. 49.3 54.6 Comp Web Q Accuracy Test Xie et al. 73.3 75.9 Feta QA BLEU-4 Test Xie et al. 33.4 35.8 Hybrid QA Accuracy Dev Eisenschlos et al. 60.8 61.0 Wiki SQL Accuracy Test Xie et al. 86.0 87.3 Tab Fat Accuracy Test Xie et al. 83.4 87.1 Feverous Accuracy Dev Xie et al. 82.4 85.6 SQA Sent.Acc Test Xie et al. 62.4 70.5 MTOP Match Test Xie et al. 86.8 87.5 DART BLEU-4 Test Aghajanyan et al. 47.2 50.4

DSI-NQ HITS@10 Dev Tay et al. 70.3 73.8

B.1.4 TRADEOFFS BETWEEN FINETUNING AND PROMPT-BASED ZERO-SHOT LEARNING (SUPERGLUE)

In this section, we explore finetuning and in-context learning trade-offs on the Super GLUE benchmark. We conduct experiments on Super GLUE with UL20B. While UL20B does not achieve SOTA on this benchmark, we note that UL20B at least remains competitive and outperforms T5-11B. This section reassures that UL2 indeed scales and matches/slightly outperforms T5-11B on Super GLUE (while strongly outperforming T5-XXL on many other in-context tasks). UL20B still lacks behind the SOTA model ST-Mo E-32B given two main reasons. Firstly, ST-Mo E-32B has 200B+ parameters and is costs equivalent to a 32B dense model. Secondly, ST-Mo E-32B is trained solely on span corruption using an encoder-decoder architecture which is known to be very advantageous on NLU finetuning.

Published as a conference paper at ICLR 2023

Table 12: Results on Super GLUE dev set. We compare with T5-11B (Raffel et al., 2019), STMo E-32B (Zoph et al., 2022) and Pa LM-8B, Pa LM-62B and Pa LM-540B (Chowdhery et al., 2022). Scores reported are the peak validation scores per task.

Model Bool Q CB Co PA Multi RC Record RTE Wi C WSC Avg

Pa LM 62B 90.6 96.4/95.7 98.0 87.7/61.9 93.0/92.4 89.5 75.9 96.2 89.2 Pa LM 540B 92.2 100/100 100 90.1/69.2 94.0/94.6 95.7 78.8 100 92.6 ST-Mo E 32B269B 93.1 100/100 100 90.4/69.9 95.0/95.6 95.7 81.0 100 93.2

Pa LM 8B 87.6 96.4/92.1 86.0 81.6/64.0 89.7/89.3 84.5 73.4 88.5 83.4 T5 11B 90.8 94.9/96.4 98.0 87.4/66.1 93.8/93.2 93.9 77.3 96.2 89.9 UL2 20B 90.8 98.7/98.2 99.0 88.4/64.8 93.7/93.2 92.1 77.3 98.1 90.7

Table 13: Results on zero-shot learning on Super GLUE dataset. We compare with GPT-3, GLa M and Pa LM (Chowdhery et al., 2022). We also include models that are relatively compute-matched with UL20B such as T5-XXL with LM adaptation (Lester et al., 2021), GPT-3 13B and GLa M-8B dense. Notably, UL20B outperforms GPT-3 175B and all other models in a similar compute class on average score.

Model Bool Q CB RTE Re CORD WSC Wi C COPA Multi RC Avg

ST-Mo E-32B (269B) 40.8 41.1 52.7 50.0 57.5 50.0 56.0 30.3 47.6 GPT-3 175B 60.5 46.4 63.5 90.2 65.4 0.0 91.0 72.9 61.2 GLa M-Mo E 1.2T 83.0 33.9 68.8 90.3 84.9 50.5 90.0 45.1 68.3 Pa LM 540B 88.0 51.8 72.9 92.9 89.1 59.1 93.0 83.5 78.8

T5-XXL 44.3 37.5 48.8 85.8 59.3 50.9 70.0 23.0 52.5 GPT-3 13B 66.2 19.6 62.8 89.0 64.4 0.0 84.0 71.4 57.2 GLa M-Dense 8B 73.6 33.9 44.0 89.2 80.7 44.0 86.0 39.0 61.3 GLa M-Mo E 64E 72.2 40.7 60.3 88.9 81.8 49.5 86.0 52.4 66.5 Pa LM-Dense 8B 68.3 41.1 54.2 87.8 78.9 47.0 86.0 47.5 63.9

UL2 20B (single ckpt) 63.1 41.1 60.7 88.1 79.9 49.8 85.0 36.2 63.0 UL2 20B (best) 63.1 50.0 60.7 88.1 80.6 55.2 88.0 36.2 65.2

B.1.5 GENERATIVE FEW-SHOT: XSUM SUMMARIZATION

Finally, we conduct additional few-shot in-context one-shot learning using the XSum dataset We compare our model with the baseline T5-XXL, T5-XXL with LM Adaptation (Lester et al., 2021), La MDA 137B (Thoppilan et al., 2022), and Pa LM (8B, 62B, 540B) (Chowdhery et al., 2022). We run T5-XXL ourselves in the same experimental setup but report results from (Chowdhery et al., 2022) for the other models.

Table 14: Results on One-Shot Summarization on XSUM.

Model Rouge-1 Rouge-2 Rouge-L

La MDA 137B - 5.4 - Pa LM 62B - 11.2 - Pa LM 540B - 12.2 -

Pa LM 8B - 7.9 - T5 XXL 11B 0.6 0.1 0.6 T5 XXL 11B + LM 13.3 2.3 10.7 UL2 20B 25.5 8.6 19.8

Table 14 reports results on 1-shot summarization. We note that T5-XXL performs poorly on this task. Even with LM adaptation, the Rouge-2 score is only 2.3, which substantially lacks behind decoder-only causal language models (e.g., Pa LM 8B models). Notably, the off-the-shelf T5-XXL is not able to generate meaningful summaries even with prompting just because it is only trained with span corruption so it is intuitive that some form of adaptation is required for generative fewshot settings. Here is is worth that the performance of UL2 20B is about 3x the performance of

Published as a conference paper at ICLR 2023

LM adapted T5 XXL model. Moreover, UL2 20B outperform La MDA 137B and has improved performance compared to Pa LM 8B. The best result, however, is still the larger 540B and 62B Pa LM models.

B.2 UL2 FOR CHAIN-OF-THOUGHT PROMPTING

It has recently been shown that language models at scale can perform multi-step reasoning tasks such as math word problems or commonsense reasoning via chain-of-thought prompting, which prompts the model to generate a step-by-step reasoning path before giving the final answer (Wei et al., 2022b). Notably, chain-of-thought (Co T) prompting does not require any additional fine-tuning of the model.

A crucial consideration of Co T prompting is that it is an emergent ability of scale (Wei et al., 2022a) it requires a sufficiently large language model to improve performance, and actually hurts performance for small language models. Hence, the successful use cases of chain-of-thought prompting use either La MDA 137B (Thoppilan et al., 2022), Pa LM 540B (Chowdhery et al., 2022), or Open AI models (Brown et al., 2020; Ouyang et al., 2022). These models, however, are compute intensive and not available as public checkpoints.

Here we demonstrate that UL2 20B can successfully leverage Co T prompting to solve multi-step arithmetic and commonsense tasks. We use the same benchmark tasks and prompts from Wei et al. (2022b). In Table 15 below, we see that on five arithmetic reasoning datasets, Co T prompting outperforms standard prompting (directly outputting the answer without a chain of thought) for UL2 20B. Similar to Wei et al. (2022b), we also show that Co T prompting can be augmented by using an external calculator ( calc. ) to perform arithmetic computational only (+, , , /) to further improve performance by a large margin. In addition, we add self-consistency (Wang et al., 2022b) (denoted as SC ) on top of Co T prompting and observed significant gains consistently across all benchmarks, with an average improvement of 22.5% compared to standard prompting.

Table 15: Chain-of-thought prompting and self-consistency (SC) results on five arithmetic reasoning benchmarks. GSM8K: (Cobbe et al., 2021). SVAMP: (Patel et al., 2021). ASDiv: (Miao et al., 2020). AQu A: (Ling et al., 2017). MAWPS: (Koncel-Kedziorski et al., 2016).

Model GSM8K SVAMP ASDiv AQu A MAWPS Average

UL2 20B: standard prompting 4.1 10.1 16.0 20.5 16.6 13.5 UL2 20B: Co T prompting 4.4 12.5 16.9 23.6 19.1 15.3 UL2 20B: Co T prompting + calc. 6.9 28.3 34.3 23.6 42.7 27.2 UL2 20B: Co T prompting + calc. + SC 10.2 41.4 43.5 26.9 57.9 36.0

In addition to arithmetic reasoning, Table 16 shows the performance of Co T prompting using UL2 20B compared to standard prompting on five commonsense reasoning benchmarks. Co T prompting plus self-consistency outperforms standard prompting in four of the five benchmarks, with an average improvement of 14.4%.

Table 16: Chain-of-thought prompting and self-consistency (SC) results on five commonsense reasoning benchmarks. CSQA: (Talmor et al., 2019). Strategy QA: (Geva et al., 2021). Date Understanding and Sports Understanding: (Srivastava et al., 2022). ARC-easy/challenge: (Clark et al., 2018).

Model CSQA Strategy QA Date Sports ARC-e ARC-c Average

UL2 20B: standard prompting 34.2 59.0 13.5 57.9 32.2 29.8 37.8 UL2 20B: Co T prompting 51.4 53.3 14.0 65.3 61.6 42.9 48.1 UL2 20B: Co T prompting + SC 55.7 54.9 16.3 66.8 69.8 49.5 52.2

Overall, we have shown that whereas prior Co T work have required large pre-trained models such as Pa LM 540B, UL2 20B is a relatively smaller model that can also perform multi-step reasoning. We hypothesize that the mixture of denoisers may contribute to the ability of UL2 to leverage Co T

Published as a conference paper at ICLR 2023

prompting at 20B parameters, although we leave further investigation of what unlocks emergent chain-of-thought reasoning to future work.

C REFERENCE BACKGROUND: PRE-TRAINED LANGUAGE MODELS

In this section, we discuss background surrounding pretrained language models, pretraining objectives and other unified pretraining proposals.

C.1 PRE-TRAINED LANGUAGE MODELS

Learning pre-trained representations for language is a far-reaching pillar of modern NLP research, dating back to (Mikolov et al., 2013; Pennington et al., 2014; Neumann et al., 2018; Dai & Le, 2015; Howard & Ruder, 2018). The first pre-trained Transformer, GPT, was proposed by (Radford et al., 2019) and was trained as a causal language model. Subsequently, BERT (Devlin et al., 2018) demonstrated the importance of bidirectional modeling for many downstream tasks. BERT introduced masked language modeling (MLM), a denoising objective that reconstructs the input in-place using bidirectional receptive fields. XLNet Yang et al. (2019) introduced the Permutation Language Modeling to account for dependencies between masked tokens during training. A number of additional papers (e.g., Ro BERTA (Liu et al., 2019), Span BERT (Joshi et al., 2020)) suggested further improvements to the pre-training process.

At the same time, two-stack encoder-decoder architectures such as T5 (Raffel et al., 2019) gained popularity due to their improved performance on classification and sequence-to-sequence ( seq2seq ) tasks. However, so far, these models have shown limited performance on open-text generation and prompt-based inference (i.e., in-context learning), which motivates the use of decoderonly models that are trained with different objectives (e.g., GPT-3 (Brown et al., 2020), GLa M (Du et al., 2021), La MDa (Thoppilan et al., 2022) and Pa LM (Chowdhery et al., 2022)). In this work, we aim to bridge the performance gap between the two by means of a general training paradigm that suits both architectures.

Decoder-only vs Encoder-only The key similarities of decoder-only and encoder-only architectures is that decoder-only architectures operate with an input-to-target paradigm or targets-only paradigm if Causal LM is used over Prefix LM used. For both architectures, the objective is always to predict the next token (LM) and are therefore autoregressive models. Notably this is different from position-wise masked LM denoising (sometimes known as autoencoding), which have been popularized by encoder-only BERT-style models. These class of models are very restricted in their generative capabilities. On top of that, task specific classification heads are also typically employed for downstream tasks. Because of the cumbersomeness of task specific classification heads, we strongly do not recommend using this class of autoencoding models moving forward and consider them somewhat deprecated. Caveats do apply. For instance, regression is the probably only reason why one would add a task specific head (Lees et al., 2022), or to squeeze out some efficiency gains from eliminating a full vocabulary. Either way, one can always start from a encoder-decoder and chop off the decoder later so there is no good reason to use an encoder-only model. Hence the only real objective consideration here is between decoder-only and encoder-decoder architectures.

Decoder-only vs Encoder-Decoder The line between decoder-only and encoder-decoder models is less clear. Prefix LM models are almost encoder-decoder models with shared parameters (but not quite). From an inductive bias point of view, there are multiple differences. Encoder-Decoder models process input and targets independently with a different set of parameters. This is a form of sparsity where different set of parameters are used for different tokens. Encoder-Decoder models also have a cross attention component that connects input tokens to target tokens. Meanwhile, decoder-only models process inputs and targets by concatenating them. Hence, the representations of inputs and targets are concurrently build layer by layer as the input/targets propagate up the network. Conversely, the decoder in Encoder-decoder models generally only looks at the fully processed encoder input. Overall, the inductive bias of Prefix LM decoder-only models and Encoder Decoder models could be pretty similar modulo the subtle differences stated above. The distinct property is that Encoder-Decoder models are generally approximately 2x parameters of a decoderonly model when compute-matched.

Published as a conference paper at ICLR 2023

Sparse Models On a side note, there have also been also an emerging trend of sparse pretrained models that achieve state-of-the-art performance. Sparse mixture-of-expert models such as the Switch Transformer (Fedus et al., 2021), GLa M (Du et al., 2021) and/or GShard (Lepikhin et al., 2020) have also demonstrated a lot of promise. While orthogonal to the topic of pretraining objectives, sparse models achieve a very different flop-per-parameter ratio compared to dense models - a core recurring motif in the debate surrounding encoder-decoder models vs decoder-only models.

C.2 PRE-TRAINING OBJECTIVES FOR LARGE LANGUAGE MODELS

While recent research demonstrates the potential of large supervised multi-task pre-training (Aribandi et al., 2021; Sanh et al., 2021; Wang et al., 2022a), most pre-training objectives rely on the vast availability of unsupervised data and use self-training techniques. As mentioned above, different architectures typically leverage different objectives. Decoder-only models are typically trained with causal language model objectives to mimic auto-regressive generation (Radford et al., 2019). Raffel et al. (2019) explored many objectives for encoder-decoder models and found span corruption to be effective. (Wang et al., 2022a) conducts a systematic study of different architectures combined with three different pretraining objectives (causal LM, prefix LM and span corruption) and analyzed their impact on zero-shot generalization. Related to our proposed X-denoisers, (Wettig et al., 2022) studies the effect of corruption rate in BERT-style masked language modeling and hypothesizes that this improves sample efficiency along with benefitting larger models. Notably, the benefits of heightened corruption rates as a standalone denoiser is still unclear, as noted by (Raffel et al., 2019) and also apparent in our own ablations. Pre-training (or denoising) is generally applied on the subword level (Raffel et al., 2019; Devlin et al., 2018) but it is worth to note that it has also been applied on the character or byte-level (Xue et al., 2021; Tay et al., 2021c). In these setups, the corrupted spans are generally much larger than subword-based denoising.

C.3 UNIFIED PRE-TRAINING PROPOSALS

Uni LM (Dong et al., 2019) proposed to train on multiple language modeling objectives using a single Transformer model. Specifically, Uni LM trains on unidirectional LM, bidirectional LM and seq2seq LM. This is quite similar to combining auto-regressive LMs with BERT and prefix-LM models. Notably, Uni LM trains using a cloze-type formulation which adds explicit mask tokens to the inputs. Losses are then computed by the difference of the predicted token and target token in a position-wise fashion. Aside from pretraining unification, there have been a recent trend of thematic unification, i.e., unifying common tasks into one model. Examples of these include UNICORN (Lourie et al., 2021) for commonsense reasoning, Unified QA (Khashabi et al., 2020; 2022) for question answering, Programming Puzzles (Schuster et al., 2021b) for problem solving, and Unified SKG (Xie et al., 2022) for Structured Knowledge Grounding.

C.4 IMPLEMENTATION DETAILS AND UL2 CODE

This section aims to give more insight to how UL2 pretraining is implemented. Our implementation is actually pretty simple. It is simply a mixture of different pretraining objectives that is implemented in seqio7. Most of our experiments were run with simply mixing different seqio tasks with seqio s Mixture Registry. However, one could also implement a generalized UL2 objective with the following function which could be cleaner.

def ul2_objective(dataset: tf.data.Dataset, sequence_length: seqio.preprocessors.Sequence Length Type, output_features: seqio.preprocessors.Output Features Type, use_prefix_lm_task: bool = False, rates: Optional[Sequence[float]] = None, mean_noise_span_lengths: Sequence[float] = (3.0,), noise_densities: Sequence[float] = (0.15,), shard_ds: bool = True, optional_task_prefixes: Optional[Sequence[str]] = None, input_feature_key: str = "inputs", merge_examples_to_reduce_padding: bool = True,

7https://github.com/google/seqio

Published as a conference paper at ICLR 2023

reserved_for_packing: bool = None, seed: int = 7) -> tf.data.Dataset: """UL2-like pre-training objectives.

This preprocessor amounts to calling the span_corruption function

several times with different values of noise_density and

mean_noise_span_length . We either shard or copy the dataset, then apply each function to each

shard. Add S-denoising (prefix LM) using use_prefix_lm_task.

dataset: A tf.data.Dataset with dictionaries containing the key

input_feature_key . sequence_length: dict mapping of feature key to int length for that

feature. output_features: mapping of keys to features. use_prefix_lm_task: <bool> If True, include Prefix LM in the task mix. rates: <Optional<List<float>> List of rates per task. If None, tasks

sampled uniformly. mean_noise_span_lengths: List of mean number of tokens per masked

span per example. noise_densities: List of what fraction of the tokens to mask. shard_ds: <bool> If True, shard dataset per objective. optional_task_prefixes: <Optional<list<str>> Strings to prepend for

corruption scheme. NOTE: If including prefix LM task, it must be the last prefix. input_feature_key: which feature to use from the dataset as the input

text tokens. merge_examples_to_reduce_padding: if True, combines multiple input

examples to reduce padding. reserved_for_packing: if specified, reduces the desired inputs length

by the specified amount to enable multiple examples to be packed together downstream. seed: tf.int64 for controlling the random choice of spans.

a dataset """

if optional_task_prefixes: # Ensure each task has a prefix.

num_tasks = len(noise_densities) + int(use_prefix_lm_task) valid_number_of_prefixes = num_tasks == len(optional_task_prefixes) if not valid_number_of_prefixes:

raise Value Error("Number of task prefixes must match number of

tasks.") inputs_length = sequence_length[input_feature_key] input_lengths, targets_lengths = [], [] sequence_lengths = {x: y for x, y in sequence_length.items()} if reserved_for_packing: inputs_length -= reserved_for_packing for x, y in sequence_length.items(): sequence_lengths[x] = y - reserved_for_packing hyperparams = list(zip(mean_noise_span_lengths, noise_densities)) for mean_noise_span_length, noise_density in hyperparams: input_length, targets_length = t5.data.preprocessors.random_spans_helper( extra_tokens_per_span_inputs=1,

Published as a conference paper at ICLR 2023

extra_tokens_per_span_targets=1, inputs_length=inputs_length, mean_noise_span_length=mean_noise_span_length, noise_density=noise_density) input_lengths.append(input_length) targets_lengths.append(targets_length)

if sequence_length["targets"] < targets_length: upper_bound = max(targets_lengths) raise Value Error( f Expected max targets length for span corruption ({upper_bound})

is f greater than configured targets length f"({sequence_length[ targets ]})")

ds = dataset ds = t5.data.preprocessors.select_random_chunk( ds, output_features=output_features, feature_key="targets", max_length=65536) if merge_examples_to_reduce_padding: ds = t5.data.preprocessors.reduce_concat_tokens( ds, feature_key="targets", batch_size=128) num_shards = len(input_lengths) + int(use_prefix_lm_task) if shard_ds: ds_shards = [ds.shard(num_shards, i) for i in range(num_shards)] else: ds_shards = [ds for _ in range(num_shards)] processed_ds = [] hyperparams = zip(input_lengths, hyperparams, range(num_shards)) for input_length, (noise_span_length, noise_density), i in hyperparams: ds = ds_shards[i] ds = t5.data.preprocessors.split_tokens( ds, feature_key="targets", min_tokens_per_segment=None, max_tokens_per_segment=input_length) ds = t5.data.preprocessors.denoise( ds, output_features, inputs_fn=t5.data.preprocessors.noise_span_to_unique_sentinel, targets_fn=t5.data.preprocessors.nonnoise_span_to_unique_sentinel, noise_density=noise_density, noise_mask_fn=functools.partial( t5.data.preprocessors.random_spans_noise_mask, mean_noise_span_length=noise_span_length), input_feature_key=input_feature_key) if optional_task_prefixes: ds = prepend_prompt( ds, output_features, prompt_mode=optional_task_prefixes[i], mode=optional_task_prefixes[i]) processed_ds.append(ds) if use_prefix_lm_task: ds = ds_shards[-1] ds = t5.data.preprocessors.prefix_lm(ds, sequence_lengths, output_features) if optional_task_prefixes: ds = prepend_prompt( ds, output_features, prompt_mode=optional_task_prefixes[-1], mode=optional_task_prefixes[-1])

Published as a conference paper at ICLR 2023

processed_ds.append(ds)

ds = tf.data.experimental.sample_from_datasets(processed_ds, rates, seed) return ds

C.5 DETAILS OF SUPERVISED FINETUNING SOTA RUNS

Most of our supervised finetuning runs were finetuned as single tasks. The only exception was that:

We finetuned GLUE as a single mixture with proportionate sampling. This has become standard and defacto setup (Raffel et al., 2019; He et al., 2022; Tay et al., 2020; 2021b).

We finetuned Super GLUE as a single mixture which is also a standard setup these days (Fedus et al., 2021; Raffel et al., 2019; Chowdhery et al., 2022).

SIQA, PIQA, Abductive NLI, Winogrande XL and Cosmos QA were co-trained in a proportionate mixture similar to (Lourie et al., 2021) under the Rainbow benchmark.

For CSQA, CSQA2. OBQA, and ARC-DA we co-trained with the rainbow mixture to obtain results on these three datasets.

All other tasks were single-task finetuned.

Published as a conference paper at ICLR 2023

C.6 QUALITATIVE SAMPLES

This section lists some qualitative samples that are randomly sampled (and not cherrypicked) on 1shot XSUM summarization.

Summarize: John Martin, 48, denies murdering Russian pianist Natalia Strelchenko, 38, at their Manchester home on their second wedding anniversary last August. ; Manchester Crown Court heard he took the diazepam believing it to be his prescribed anti-depressant medication. ; He said everything is very blurry . ; It is alleged the double bass player strangled and beat Ms Strelchenko, who was also known by the surname Strelle, in a loss of temper at their home in Newton Heath on 30 August. ; Giving evidence, Mr Martin, also known as Jon Skogsbakken, said he had been suffering with depression and that he had unwittingly taken unmarked diazepam tablets for around six weeks. ; He said he guessed they had been prescribed to his wife. ; Mr Martin told the jury he had suffered with depression since 2005 and his GP would prescribe him Escitalopram. ; He told the court that taking the diazepam was the biggest mistake of my life . ; Mr Martin said: I would never have taken the two together, absolutely not , if he was aware the tablets could have a paradoxical and aggressive response and when taken with alcohol it could increase the effects. ; The court heard following an argument with Ms Strelchenko, he drank four cans of cider at their home before going to the working men s club where he had another pint of cider and then returned home. ; He said: I just sat down on the floor and poured myself a glass of wine. I started to drink it. Natalia was angry because I was drinking wine and she tried to take it away from me. ; Everything is very blurry for me. I can t recall anything after that point. ; I still love her very much. I would never wish to do such terrible things to her. I recall virtually nothing. ; Mr Denney QC said: Do you accept that on evidence you killed Natalia? ; Mr Martin replied: Well according to what the witnesses described...I perfectly understand that I must be the man behind it. ; Mr Martin, who had worked for the computer firm IBM in his home country of Norway, met Ms Strelchenko in 2007. ; She had performed piano recitals at New York s Carnegie Hall and London s Wigmore Hall. ; The trial continues. Output: A man has denied murdering his Russian pianist wife after taking a cocktail of drugs and alcohol.

Summarize: The pair, dressed in Team GB tracksuits and wearing plastic medals, admitted to Sky News they blagged their way on to the float and joined the celebrations. ; The float was carrying Britain s gold medal-winning women s hockey team and athletes from the modern pentathlon. ; The BOA said: It didn t spoil anyone s enjoyment of a wonderful event. ; We are aware of the matter and are disappointed anyone would want to detract from the athletes celebration, it added. ; Four hundred athletes took part in the Manchester celebrations with more 150,000 people lining the streets. ; A second parade is to take place in Trafalgar Square in London from 13:00 BST after which the athletes will visit Buckingham Palace. ; Greater Manchester Police said it did not provide security for the floats and had received no reports regarding the incident. ; Manchester City Council is yet to comment. ; Team GB set an Olympic record in Rio 2016 by winning 67 medals - two more than they managed at London 2012. ; The Paralympians also beat their previous performance with 147 medals - 27 more than four years ago. Output: Two men have been arrested after attempting to join the Olympic and Paralympic parade in Manchester.

Summarize: Laura Stewart, 20, was walking on North Hanover Street with her friend Mhairi Convy, 18, when they were hit by a Range Rover on 17 December. ; The funeral of Ms Stewart, an accounts student from Cumbernauld, was held at Our Lady and St Helen s Parish Church in her home town. ; About 200 people attended the ceremony to pay their respects. ; Outside the church floral tributes were laid, including the name Laura written out in red and white flowers opposite the word sister in purple flowers. ; Tributes left with the flowers included a card saying with all our love for always, mum and dad and another to a beautiful girl who was a loving friend and sister . ; The order of service showed a photo of the student with the words Courage grows strong at a wound - the motto of the Stewart clan. ; Mourners wept as the white coffin with purple handles was carried out of the church to the hymn Walk With Me Oh My Lord. ; Father Gerald Sharkey led the procession out of the church through the snow. The body was then driven to Eastfield Cemetery in Cumbernauld for a private burial. ; Ms Stewart and Ms Convy, of Lennoxtown,

Published as a conference paper at ICLR 2023

studied at Glasgow College of Commerce. ; Following the crash, the two women were taken to Glasgow Royal Infirmary, where they died. ; A 36-year-old male pedestrian and the 50-year-old male driver of the Range Rover were also treated in hospital after the crash. ; Strathclyde Police are investigating the deaths. Output: The funeral of a student who died in a car crash in Glasgow has taken place.

Summarize: Eleven of Scotland s 14 territorial health boards were hit by the ransomware attack linked to other IT attacks around the world. ; The hack has encrypted information on NHS computers, denying access unless a payment is made. ; First Minister Nicola Sturgeon said work to fix the systems affected had entered a recovery phase ; She called the attack hugely concerning but said there was no evidence as yet to suggest outdated computer systems were responsible for the breach. ; Scotland s Health Secretary Shona Robison said there was a level of confidence that systems would be back up and running by Monday and that no breach of patient confidentiality had been detected. ; The health boards which have been affected are: ; The Scottish Ambulance Service has also been affected, along with NHS National Services Scotland. ; The incidents are thought to be part of a wider attack affecting organisations in about 100 countries around the world. ; IT problems have also caused disruption in about 30 health authorities in England, while the NHS in Wales and Northern Ireland are so far unaffected. ; The Scottish government said most incidents had been confined to desktop computers in GP surgeries, dental practices and other primary care centres. ; A spokesman said the only acute hospital sites so far affected had been in NHS Lanarkshire. ; BBC Scotland understands that computer systems at Hairmyres Hospital in East Kilbride were compromised. ; Ms Robison told the BBC s Good Morning Scotland programme that IT specialists had been working non-stop to get GP systems back up and running. ; She said: People are working very, very hard and have worked through the night. The update I ve got this morning is that we re very much into recovery phase now, with a lot of work going on to get systems back up running. ; The GP systems, which of course were the main problem across our health boards - work is going on, and there is a level of confidence that many will be back up and running before GP surgeries open on Monday morning. ; Ms Sturgeon told the BBC: Obviously cyber-attacks of this nature are hugely concerning and I think they underline the vulnerability not just of the public sector, but also of society generally to cyber-attacks, but they also underline the importance of all organisations making sure that they have all appropriate measures in place to protect against those kinds of attack. ; She added: Obviously there is a lot of investigation into exactly why the health service has been affected in the way it has. I think it s important to stress that this has been a global international attack. ; Shona Robison said NHS Lanarkshire had been more affected in terms of its acute hospitals, but said manual systems had worked safely and insisted that patients had not been negatively impacted by the breach. ; The intention today is to begin testing those IT systems and to gradually and safely try to bring those back on over the course of the weekend. ; Ms Robison added: Throughout all of this there s been no breach to patient confidentiality that has been detected to date so patients should be reassured by that. ; Patient services, including emergency services, are continuing to operate across Scotland, the Scottish government confirmed. ; Ms Robison emphasised there had been no impact on the majority of the out-of-hours systems across Scotland, with NHS 24 working as normal along with the Scottish Ambulance Service, where the only issue had been with desktop PCs that were non-patient facing . ; All the other parts of the system that people would use over the weekend are working as normal, Ms Robison said. ; In NHS Lanarkshire, non-emergency patients have been urged to stay away from its hospitals as it deals with the ransomware attack. ; Dr Helen Mackie, chief of medical services at Hairmyres Hospital, said staff had reverted to paper and manual records for patients. ; She has urged any patients turning up at the authorities emergency departments to bring their medication with them because medics may have problems accessing their records. ; She said: Help us by bringing as much information with you, So bring your medicines, bring any information you have about health care. Relatives can really help us as well because they re a wealth of information for us. ; Dr Mackie said the IT issues were leading to the cancellation of planned out-patient appoints for tests such as CT scans. However, she said that any emergency diagnostic tests would continue to take place. ; She added: It s business as usual in terms of emergency care in our Lanarkshire hospitals. We are asking the public to help us by only coming to emergency departments if it is an emergency. But please be reassured that all our emergency access to treatment and care is up and fully running. Output: Eleven of Scotland s 14 territorial health boards were hit by the ransomware attack

Published as a conference paper at ICLR 2023

linked to other IT attacks around the world.

Summarize: Two-time major winner Mc Ilroy, 25, and Danish former world number one Wozniacki, 23, announced their engagement on New Year s Eve. ; Media playback is not supported on this device ; The problem is mine, Mc Ilroy said. The wedding invitations issued at the weekend made me realise that I wasn t ready for all that marriage entails. ; The couple had been together for more than two years. ; Mc Ilroy is currently at Wentworth for the BMW PGA Championship, the European Tour s flagship event which starts on Thursday. ; Wozniacki is set to compete in the French Open in Paris starting on 25 May. ; In a statement issued through his management company, Mc Ilroy added: There is no right way to end a relationship that has been so important to two people. ; I wish Caroline all the happiness she deserves and thank her for the great times we have had. ; The news has shocked the worlds of both golf and tennis, particularly with the timing of the split. Mc Ilroy is due to start a crucial run of tournaments, while Wozniacki is playing the upcoming French Open. ; Speaking at Wentworth on Wednesday, Mc Ilroy said the decision to end the engagement had been mutual and amicable. ; He said it was a difficult time, but that the move was best for both of them. ; Mc Ilroy won the US Open in 2011 and US PGA Championship the following year. ; Wozniacki, who has reached the final of one grand slam tournament - the US Open in 2009 - is currently ranked 13th in the world. ; When Mc Ilroy announced their engagement, he tweeted: I have a feeling it is going to be a great year. Output: Rory Mc Ilroy and Caroline Wozniacki have ended their engagement.