Towards System 2 Reasoning in LLMs: Learning How to Think With Meta Chain-of-Thought

We propose a novel framework, Meta Chain-of-Thought (Meta-CoT), which extends traditional Chain-of-Thought (CoT) by explicitly modeling the underlying reasoning required to arrive at a particular CoT. We present empirical evidence from state-of-the-art models exhibiting behaviors consistent with in-context search, and explore methods for producing Meta-CoT via process supervision, synthetic data generation, and search algorithms. Finally, we outline a concrete pipeline for training a model to produce Meta-CoTs, incorporating instruction tuning with linearized search traces and reinforcement learning post-training. Finally, we discuss open research questions, including scaling laws, verifier roles, and the potential for discovering novel reasoning algorithms. This work provides a theoretical and practical roadmap to enable Meta-CoT in LLMs, paving the way for more powerful and human-like reasoning in artificial intelligence.

Predicting Emergent Capabilities by Finetuning

A fundamental open challenge in modern LLM scaling is the lack of understanding around emergent capabilities. In particular, language model pretraining loss is known to be highly predictable as a function of compute. However, downstream capabilities are far less predictable -- sometimes even exhibiting emergent jumps -- which makes it challenging to anticipate the capabilities of future models. In this work, we first pose the task of emergence prediction: given access to current LLMs that have random few-shot accuracy on a task, can we predict whether future models (GPT-N+1) will have non-trivial accuracy on that task? We then discover a simple insight for this problem: finetuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models. To operationalize this insight, we can finetune LLMs with varying amounts of data and fit a parametric function that predicts when emergence will occur (i.e., "emergence laws"). We validate this approach using four standard NLP benchmarks where large-scale open-source LLMs already demonstrate emergence (MMLU, GSM8K, CommonsenseQA, and CoLA). Using only small-scale LLMs, we find that, in some cases, we can accurately predict whether models trained with up to 4x more compute have emerged. Finally, we present a case study of two realistic uses for emergence prediction.

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

Enabling LLMs to improve their outputs by using more test-time computation is a critical step towards building generally self-improving agents that can operate on open-ended natural language. In this paper, we study the scaling of inference-time computation in LLMs, with a focus on answering the question: if an LLM is allowed to use a fixed but non-trivial amount of inference-time compute, how much can it improve its performance on a challenging prompt? Answering this question has implications not only on the achievable performance of LLMs, but also on the future of LLM pretraining and how one should tradeoff inference-time and pre-training compute. Despite its importance, little research attempted to understand the scaling behaviors of various test-time inference methods. Moreover, current work largely provides negative results for a number of these strategies. In this work, we analyze two primary mechanisms to scale test-time computation: (1) searching against dense, process-based verifier reward models; and (2) updating the model's distribution over a response adaptively, given the prompt at test time. We find that in both cases, the effectiveness of different approaches to scaling test-time compute critically varies depending on the difficulty of the prompt. This observation motivates applying a "compute-optimal" scaling strategy, which acts to most effectively allocate test-time compute adaptively per prompt. Using this compute-optimal strategy, we can improve the efficiency of test-time compute scaling by more than 4x compared to a best-of-N baseline. Additionally, in a FLOPs-matched evaluation, we find that on problems where a smaller base model attains somewhat non-trivial success rates, test-time compute can be used to outperform a 14x larger model.

LMRL Gym: Benchmarks for Multi-Turn Reinforcement Learning with Language Models

Large language models (LLMs) provide excellent text-generation capabilities, but standard prompting and generation methods generally do not lead to intentional or goal-directed agents and might necessitate considerable prompt tuning. This becomes particularly apparent in multi-turn conversations: even the best current LLMs rarely ask clarifying questions, engage in explicit information gathering, or take actions now that lead to better decisions after multiple turns. Reinforcement learning has the potential to leverage the powerful modeling capabilities of LLMs, as well as their internal representation of textual interactions, to create capable goal-directed language agents. This can enable intentional and temporally extended interactions, such as with humans, through coordinated persuasion and carefully crafted questions, or in goal-directed play through text games to bring about desired final outcomes. However, enabling this requires the community to develop stable and reliable reinforcement learning algorithms that can effectively train LLMs. Developing such algorithms requires tasks that can gauge progress on algorithm design, provide accessible and reproducible evaluations for multi-turn interactions, and cover a range of task properties and challenges in improving reinforcement learning algorithms. Our paper introduces the LMRL-Gym benchmark for evaluating multi-turn RL for LLMs, together with an open-source research framework containing a basic toolkit for getting started on multi-turn RL with offline value-based and policy-based RL methods. Our benchmark consists of 8 different language tasks, which require multiple rounds of language interaction and cover a range of tasks in open-ended dialogue and text games.

The False Promise of Imitating Proprietary LLMs

An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). This approach looks to cheaply imitate the proprietary model's capabilities using a weaker open-source model. In this work, we critically analyze this approach. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B--13B), data sources, and imitation data amounts (0.3M--150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models -- they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT's style but not its factuality. Overall, we conclude that model imitation is a false promise: there exists a substantial capabilities gap between open and closed LMs that, with current methods, can only be bridged using an unwieldy amount of imitation data or by using more capable base LMs. In turn, we argue that the highest leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems.

Learning by Distilling Context

Language models significantly benefit from context tokens, such as prompts or scratchpads. They perform better when prompted with informative instructions, and they acquire new reasoning capabilities by generating a scratch-pad before predicting the final answers. However, they do not \textit{internalize} these performance gains, which disappear when the context tokens are gone. Our work proposes to apply context distillation so that a language model can improve itself by internalizing these gains. Concretely, given a synthetic unlabeled input for the target task, we condition the model on ``[instructions] + [task-input]'' to predict ``[scratch-pad] + [final answer]''; then we fine-tune the same model to predict its own ``[final answer]'' conditioned on the ``[task-input]'', without seeing the ``[instructions]'' or using the ``[scratch-pad]''. We show that context distillation is a general method to train language models, and it can effectively internalize 3 types of training signals. First, it can internalize abstract task instructions and explanations, so we can iteratively update the model parameters with new instructions and overwrite old ones. Second, it can internalize step-by-step reasoning for complex tasks (e.g., 8-digit addition), and such a newly acquired capability proves to be useful for other downstream tasks. Finally, it can internalize concrete training examples, and it outperforms directly learning with gradient descent by 9\% on the SPIDER Text-to-SQL dataset; furthermore, combining context distillation operations can internalize more training examples than the context window size allows.

Active Programming by Example with a Natural Language Prior

We introduce APEL, a new framework that enables non-programmers to indirectly annotate natural language utterances with executable meaning representations, such as SQL programs. Based on a natural language utterance, we first run a seed semantic parser to generate a prior over a list of candidate programs. To obtain information about which candidate is correct, we synthesize an input on which the more likely programs tend to produce different outputs, and ask an annotator which output is appropriate for the utterance. Hence, the annotator does not have to directly inspect the programs. To further reduce effort required from annotators, we aim to synthesize simple input databases that nonetheless have high information gain. With human annotators and Bayesian inference to handle annotation errors, we outperform Codex's top-1 performance (59%) and achieve the same accuracy as the original expert annotators (75%), by soliciting answers for each utterance on only 2 databases with an average of 9 records each. In contrast, it would be impractical to solicit outputs on the original 30K-record databases provided by SPIDER

Context-Aware Language Modeling for Goal-Oriented Dialogue Systems

Goal-oriented dialogue systems face a trade-off between fluent language generation and task-specific control. While supervised learning with large language models is capable of producing realistic text, how to steer such responses towards completing a specific task without sacrificing language quality remains an open question. In this work, we formulate goal-oriented dialogue as a partially observed Markov decision process, interpreting the language model as a representation of both the dynamics and the policy. This view allows us to extend techniques from learning-based control, such as task relabeling, to derive a simple and effective method to finetune language models in a goal-aware way, leading to significantly improved task performance. We additionally introduce a number of training strategies that serve to better focus the model on the task at hand. We evaluate our method, Context-Aware Language Models (CALM), on a practical flight-booking task using AirDialogue. Empirically, CALM outperforms the state-of-the-art method by 7% in terms of task success, matching human-level task performance.

Summarizing Differences between Text Distributions with Natural Language

How do two distributions of texts differ? Humans are slow at answering this, since discovering patterns might require tediously reading through hundreds of samples. We propose to automatically summarize the differences by "learning a natural language hypothesis": given two distributions $D_{0}$ and $D_{1}$, we search for a description that is more often true for $D_{1}$, e.g., "is military-related." To tackle this problem, we fine-tune GPT-3 to propose descriptions with the prompt: "[samples of $D_{0}$] + [samples of $D_{1}$] + the difference between them is _____". We then re-rank the descriptions by checking how often they hold on a larger set of samples with a learned verifier. On a benchmark of 54 real-world binary classification tasks, while GPT-3 Curie (13B) only generates a description similar to human annotation 7% of the time, the performance reaches 61% with fine-tuning and re-ranking, and our best system using GPT-3 Davinci (175B) reaches 76%. We apply our system to describe distribution shifts, debug dataset shortcuts, summarize unknown tasks, and label text clusters, and present analyses based on automatically generated descriptions.

Approximating How Single Head Attention Learns

Why do models often attend to salient words, and how does this evolve throughout training? We approximate model training as a two stage process: early on in training when the attention weights are uniform, the model learns to translate individual input word `i` to `o` if they co-occur frequently. Later, the model learns to attend to `i` while the correct output is $o$ because it knows `i` translates to `o`. To formalize, we define a model property, Knowledge to Translate Individual Words (KTIW) (e.g. knowing that `i` translates to `o`), and claim that it drives the learning of the attention. This claim is supported by the fact that before the attention mechanism is learned, KTIW can be learned from word co-occurrence statistics, but not the other way around. Particularly, we can construct a training distribution that makes KTIW hard to learn, the learning of the attention fails, and the model cannot even learn the simple task of copying the input words to the output. Our approximation explains why models sometimes attend to salient words, and inspires a toy example where a multi-head attention model can overcome the above hard training distribution by improving learning dynamics rather than expressiveness.