Noah D. Goodman

Joy He-Yueya, Gabriel Poesia, Rose E. Wang et al. · stanford

Automatically generating high-quality step-by-step solutions to math word problems has many applications in education. Recently, combining large language models (LLMs) with external tools to perform complex reasoning and calculation has emerged as a promising direction for solving math word problems, but prior approaches such as Program-Aided Language model (PAL) are biased towards simple procedural problems and less effective for problems that require declarative reasoning. We propose an approach that combines an LLM that can incrementally formalize word problems as a set of variables and equations with an external symbolic solver that can solve the equations. Our approach achieves comparable accuracy to the original PAL on the GSM8K benchmark of math word problems and outperforms PAL by an absolute 20% on ALGEBRA, a new dataset of more challenging word problems extracted from Algebra textbooks. Our work highlights the benefits of using declarative and incremental representations when interfacing with an external tool for solving complex math word problems. Our data and prompts are publicly available at https://github.com/joyheyueya/declarative-math-word-problem.

Alex Tamkin, Mohammad Taufeeque, Noah D. Goodman · stanford

Understanding neural networks is challenging in part because of the dense, continuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features. Codebook features are produced by finetuning neural networks with vector quantization bottlenecks at each layer, producing a network whose hidden features are the sum of a small number of discrete vector codes chosen from a larger codebook. Surprisingly, we find that neural networks can operate under this extreme bottleneck with only modest degradation in performance. This sparse, discrete bottleneck also provides an intuitive way of controlling neural network behavior: first, find codes that activate when the desired behavior is present, then activate those same codes during generation to elicit that behavior. We validate our approach by training codebook Transformers on several different datasets. First, we explore a finite state machine dataset with far more hidden states than neurons. In this setting, our approach overcomes the superposition problem by assigning states to distinct codes, and we find that we can make the neural network behave as if it is in a different state by activating the code for that state. Second, we train Transformer language models with up to 410M parameters on two natural language datasets. We identify codes in these models representing diverse, disentangled concepts (ranging from negative emotions to months of the year) and find that we can guide the model to generate different topics by activating the appropriate codes during inference. Overall, codebook features appear to be a promising unit of analysis and control for neural networks and interpretability. Our codebase and models are open-sourced at https://github.com/taufeeque9/codebook-features.

Lionel Wong, Gabriel Grand, Alexander K. Lew et al. · microsoft-research, mit

How does language inform our downstream thinking? In particular, how do humans make meaning from language--and how can we leverage a theory of linguistic meaning to build machines that think in more human-like ways? In this paper, we propose rational meaning construction, a computational framework for language-informed thinking that combines neural language models with probabilistic models for rational inference. We frame linguistic meaning as a context-sensitive mapping from natural language into a probabilistic language of thought (PLoT)--a general-purpose symbolic substrate for generative world modeling. Our architecture integrates two computational tools that have not previously come together: we model thinking with probabilistic programs, an expressive representation for commonsense reasoning; and we model meaning construction with large language models (LLMs), which support broad-coverage translation from natural language utterances to code expressions in a probabilistic programming language. We illustrate our framework through examples covering four core domains from cognitive science: probabilistic reasoning, logical and relational reasoning, visual and physical reasoning, and social reasoning. In each, we show that LLMs can generate context-sensitive translations that capture pragmatically-appropriate linguistic meanings, while Bayesian inference with the generated programs supports coherent and robust commonsense reasoning. We extend our framework to integrate cognitively-motivated symbolic modules (physics simulators, graphics engines, and planning algorithms) to provide a unified commonsense thinking interface from language. Finally, we explore how language can drive the construction of world models themselves. We hope this work will provide a roadmap towards cognitive models and AI systems that synthesize the insights of both modern and classical computational perspectives.

Ruocheng Wang, Eric Zelikman, Gabriel Poesia et al. · stanford

Inductive reasoning is a core problem-solving capacity: humans can identify underlying principles from a few examples, which robustly generalize to novel scenarios. Recent work evaluates large language models (LLMs) on inductive reasoning tasks by directly prompting them yielding "in context learning." This works well for straightforward inductive tasks but performs poorly on complex tasks such as the Abstraction and Reasoning Corpus (ARC). In this work, we propose to improve the inductive reasoning ability of LLMs by generating explicit hypotheses at multiple levels of abstraction: we prompt the LLM to propose multiple abstract hypotheses about the problem, in natural language, then implement the natural language hypotheses as concrete Python programs. These programs can be verified by running on observed examples and generalized to novel inputs. To reduce the hypothesis search space, we explore steps to filter the set of hypotheses to implement: we either ask the LLM to summarize them into a smaller set of hypotheses or ask human annotators to select a subset. We verify our pipeline's effectiveness on the ARC visual inductive reasoning benchmark, its variant 1D-ARC, string transformation dataset SyGuS, and list transformation dataset List Functions. On a random 100-problem subset of ARC, our automated pipeline using LLM summaries achieves 30% accuracy, outperforming the direct prompting baseline (accuracy of 17%). With the minimal human input of selecting from LLM-generated candidates, performance is boosted to 33%. Our ablations show that both abstract hypothesis generation and concrete program representations benefit LLMs on inductive reasoning tasks.

Eric Zelikman, Qian Huang, Gabriel Poesia et al.

Despite recent success in large language model (LLM) reasoning, LLMs struggle with hierarchical multi-step reasoning tasks like generating complex programs. For these tasks, humans often start with a high-level algorithmic design and implement each part gradually. We introduce Parsel, a framework enabling automatic implementation and validation of complex algorithms with code LLMs. With Parsel, we automatically decompose algorithmic tasks into hierarchical natural language function descriptions and then search over combinations of possible function implementations using tests. We show that Parsel can be used across domains requiring hierarchical reasoning, including program synthesis and robotic planning. We find that, using Parsel, LLMs solve more competition-level problems in the APPS dataset, resulting in pass rates over 75\% higher than prior results from directly sampling AlphaCode and Codex, while often using a smaller sample budget. Moreover, with automatically generated tests, we find that Parsel can improve the state-of-the-art pass@1 performance on HumanEval from 67\% to 85\%. We also find that LLM-generated robotic plans using Parsel are more than twice as likely to be considered accurate than directly generated plans. Lastly, we explore how Parsel addresses LLM limitations and discuss how Parsel may be useful for human programmers. We release our code at https://github.com/ezelikman/parsel

Jiayuan Mao, Xuelin Yang, Xikun Zhang et al. · stanford

Building machines that can reason about physical events and their causal relationships is crucial for flexible interaction with the physical world. However, most existing physical and causal reasoning benchmarks are exclusively based on synthetically generated events and synthetic natural language descriptions of causal relationships. This design brings up two issues. First, there is a lack of diversity in both event types and natural language descriptions; second, causal relationships based on manually-defined heuristics are different from human judgments. To address both shortcomings, we present the CLEVRER-Humans benchmark, a video reasoning dataset for causal judgment of physical events with human labels. We employ two techniques to improve data collection efficiency: first, a novel iterative event cloze task to elicit a new representation of events in videos, which we term Causal Event Graphs (CEGs); second, a data augmentation technique based on neural language generative models. We convert the collected CEGs into questions and answers to be consistent with prior work. Finally, we study a collection of baseline approaches for CLEVRER-Humans question-answering, highlighting the great challenges set forth by our benchmark.

Joy Hsu, Jiajun Wu, Noah D. Goodman · stanford

Euclidean geometry is among the earliest forms of mathematical thinking. While the geometric primitives underlying its constructions, such as perfect lines and circles, do not often occur in the natural world, humans rarely struggle to perceive and reason with them. Will computer vision models trained on natural images show the same sensitivity to Euclidean geometry? Here we explore these questions by studying few-shot generalization in the universe of Euclidean geometry constructions. We introduce Geoclidean, a domain-specific language for Euclidean geometry, and use it to generate two datasets of geometric concept learning tasks for benchmarking generalization judgements of humans and machines. We find that humans are indeed sensitive to Euclidean geometry and generalize strongly from a few visual examples of a geometric concept. In contrast, low-level and high-level visual features from standard computer vision models pretrained on natural images do not support correct generalization. Thus Geoclidean represents a novel few-shot generalization benchmark for geometric concept learning, where the performance of humans and of AI models diverge. The Geoclidean framework and dataset are publicly available for download.

Dilip Arumugam, Mark K. Ho, Noah D. Goodman et al. · stanford

Throughout the cognitive-science literature, there is widespread agreement that decision-making agents operating in the real world do so under limited information-processing capabilities and without access to unbounded cognitive or computational resources. Prior work has drawn inspiration from this fact and leveraged an information-theoretic model of such behaviors or policies as communication channels operating under a bounded rate constraint. Meanwhile, a parallel line of work also capitalizes on the same principles from rate-distortion theory to formalize capacity-limited decision making through the notion of a learning target, which facilitates Bayesian regret bounds for provably-efficient learning algorithms. In this paper, we aim to elucidate this latter perspective by presenting a brief survey of these information-theoretic models of capacity-limited decision making in biological and artificial agents.

Eric Zelikman, Yuhuai Wu, Jesse Mu et al.

Generating step-by-step "chain-of-thought" rationales improves language model performance on complex reasoning tasks like mathematics or commonsense question-answering. However, inducing language model rationale generation currently requires either constructing massive rationale datasets or sacrificing accuracy by using only few-shot inference. We propose a technique to iteratively leverage a small number of rationale examples and a large dataset without rationales, to bootstrap the ability to perform successively more complex reasoning. This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat. We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA. Thus, STaR lets a model improve itself by learning from its own generated reasoning.

Joy He-Yueya, Wanjing Anya Ma, Kanishk Gandhi et al. · stanford

Language models (LMs) are increasingly used to simulate human-like responses in scenarios where accurately mimicking a population's behavior can guide decision-making, such as in developing educational materials and designing public policies. The objective of these simulations is for LMs to capture the variations in human responses, rather than merely providing the expected correct answers. Prior work has shown that LMs often generate unrealistically accurate responses, but there are no established metrics to quantify how closely the knowledge distribution of LMs aligns with that of humans. To address this, we introduce "psychometric alignment," a metric that measures the extent to which LMs reflect human knowledge distribution. Assessing this alignment involves collecting responses from both LMs and humans to the same set of test items and using Item Response Theory to analyze the differences in item functioning between the groups. We demonstrate that our metric can capture important variations in populations that traditional metrics, like differences in accuracy, fail to capture. We apply this metric to assess existing LMs for their alignment with human knowledge distributions across three real-world domains. We find significant misalignment between LMs and human populations, though using persona-based prompts can improve alignment. Interestingly, smaller LMs tend to achieve greater psychometric alignment than larger LMs. Further, training LMs on human response data from the target distribution enhances their psychometric alignment on unseen test items, but the effectiveness of such training varies across domains.

Joy He-Yueya, Anikait Singh, Ge Gao et al.

Scientific breakthroughs often emerge from synthesizing prior ideas into novel contributions. While language models (LMs) show promise in scientific discovery, their ability to perform this targeted, literature-grounded synthesis remains underexplored. We introduce insight anticipation, a generation task in which a model predicts a downstream paper's core insight from its foundational parent papers. To evaluate this capability, we develop GiantsBench, a benchmark of 17k examples across eight scientific domains, where each example consists of a set of parent papers paired with the core insight of a downstream paper. We evaluate models using an LM judge that scores similarity between generated and ground-truth insights, and show that these similarity scores correlate with expert human ratings. Finally, we present GIANTS-4B, an LM trained via reinforcement learning (RL) to optimize insight anticipation using these similarity scores as a proxy reward. Despite its smaller open-source architecture, GIANTS-4B outperforms proprietary baselines and generalizes to unseen domains, achieving a 34% relative improvement in similarity score over gemini-3-pro. Human evaluations further show that GIANTS-4B produces insights that are more conceptually clear than those of the base model. In addition, SciJudge-30B, a third-party model trained to compare research abstracts by likely citation impact, predicts that insights generated by GIANTS-4B are more likely to lead to higher citations, preferring them over the base model in 68% of pairwise comparisons. We release our code, benchmark, and model to support future research in automated scientific discovery.

Kanishk Gandhi, Jan-Philipp Fränken, Tobias Gerstenberg et al.

As Large Language Models (LLMs) become increasingly integrated into our everyday lives, understanding their ability to comprehend human mental states becomes critical for ensuring effective interactions. However, despite the recent attempts to assess the Theory-of-Mind (ToM) reasoning capabilities of LLMs, the degree to which these models can align with human ToM remains a nuanced topic of exploration. This is primarily due to two distinct challenges: (1) the presence of inconsistent results from previous evaluations, and (2) concerns surrounding the validity of existing evaluation methodologies. To address these challenges, we present a novel framework for procedurally generating evaluations with LLMs by populating causal templates. Using our framework, we create a new social reasoning benchmark (BigToM) for LLMs which consists of 25 controls and 5,000 model-written evaluations. We find that human participants rate the quality of our benchmark higher than previous crowd-sourced evaluations and comparable to expert-written evaluations. Using BigToM, we evaluate the social reasoning capabilities of a variety of LLMs and compare model performances with human performance. Our results suggest that GPT4 has ToM capabilities that mirror human inference patterns, though less reliable, while other LLMs struggle.

Atticus Geiger, Zhengxuan Wu, Christopher Potts et al.

Causal abstraction is a promising theoretical framework for explainable artificial intelligence that defines when an interpretable high-level causal model is a faithful simplification of a low-level deep learning system. However, existing causal abstraction methods have two major limitations: they require a brute-force search over alignments between the high-level model and the low-level one, and they presuppose that variables in the high-level model will align with disjoint sets of neurons in the low-level one. In this paper, we present distributed alignment search (DAS), which overcomes these limitations. In DAS, we find the alignment between high-level and low-level models using gradient descent rather than conducting a brute-force search, and we allow individual neurons to play multiple distinct roles by analyzing representations in non-standard bases-distributed representations. Our experiments show that DAS can discover internal structure that prior approaches miss. Overall, DAS removes previous obstacles to conducting causal abstraction analyses and allows us to find conceptual structure in trained neural nets.

Ben Prystawski, Michael Y. Li, Noah D. Goodman

Humans have a powerful and mysterious capacity to reason. Working through a set of mental steps enables us to make inferences we would not be capable of making directly even though we get no additional data from the world. Similarly, when large language models generate intermediate steps (a chain of thought) before answering a question, they often produce better answers than they would directly. We investigate why and how chain-of-thought reasoning is useful in language models, testing the hypothesis that reasoning is effective when training data consists of overlapping local clusters of variables that influence each other strongly. These training conditions enable the chaining of accurate local inferences to estimate relationships between variables that were not seen together in training. We prove that there will exist a "reasoning gap", where reasoning through intermediate variables reduces bias, for the simple case of an autoregressive density estimator trained on local samples from a chain-structured probabilistic model. We then test our hypothesis experimentally in more complex models, training an autoregressive language model on samples from Bayes nets but only including a subset of variables in each sample. We test language models' ability to match conditional probabilities with and without intermediate reasoning steps, finding that intermediate steps are only helpful when the training data is locally structured with respect to dependencies between variables. The combination of locally structured observations and reasoning is much more data-efficient than training on all variables. Our results illustrate how the effectiveness of reasoning step by step is rooted in the local statistical structure of the training data.

Jan-Philipp Fränken, Sam Kwok, Peixuan Ye et al. · stanford

We explore the idea of aligning an AI assistant by inverting a model of users' (unknown) preferences from observed interactions. To validate our proposal, we run proof-of-concept simulations in the economic ultimatum game, formalizing user preferences as policies that guide the actions of simulated players. We find that the AI assistant accurately aligns its behavior to match standard policies from the economic literature (e.g., selfish, altruistic). However, the assistant's learned policies lack robustness and exhibit limited generalization in an out-of-distribution setting when confronted with a currency (e.g., grams of medicine) that was not included in the assistant's training distribution. Additionally, we find that when there is inconsistency in the relationship between language use and an unknown policy (e.g., an altruistic policy combined with rude language), the assistant's learning of the policy is slowed. Overall, our preliminary results suggest that developing simulation frameworks in which AI assistants need to infer preferences from diverse users can provide a valuable approach for studying practical alignment questions.

Eric Zelikman, Qian Huang, Percy Liang et al.

Language model training in distributed settings is limited by the communication cost of gradient exchanges. In this short note, we extend recent work from Malladi et al. (2023), using shared randomness to perform distributed fine-tuning with low bandwidth. The method is a natural decentralized extension of memory-efficient Simultaneous Perturbation Stochastic Approximation (SPSA). Each iteration, each machine seeds a Random Number Generator (RNG) to perform local reproducible perturbations on model weights and calculate and exchange scalar projected gradients, which are then used to update each model. By using a (machine, sample) identifier as the random seed, each model can regenerate one another's perturbations. As machines only exchange single-byte projected gradients, this is highly communication efficient. There are also potential privacy benefits, as projected gradients may be calculated on different training data, and models never access the other's data. Our approach not only drastically reduces communication bandwidth requirements but also accommodates dynamic addition or removal of machines during the training process and retains the memory-efficient and inference-only advantages of recent work. We perform proof-of-concept experiments to demonstrate the potential usefulness of this method, building off of rich literature on distributed optimization and memory-efficient training.

Gabriel Poesia, Noah D. Goodman

General mathematical reasoning is computationally undecidable, but humans routinely solve new problems. Moreover, discoveries developed over centuries are taught to subsequent generations quickly. What structure enables this, and how might that inform automated mathematical reasoning? We posit that central to both puzzles is the structure of procedural abstractions underlying mathematics. We explore this idea in a case study on 5 sections of beginning algebra on the Khan Academy platform. To define a computational foundation, we introduce Peano, a theorem-proving environment where the set of valid actions at any point is finite. We use Peano to formalize introductory algebra problems and axioms, obtaining well-defined search problems. We observe existing reinforcement learning methods for symbolic reasoning to be insufficient to solve harder problems. Adding the ability to induce reusable abstractions ("tactics") from its own solutions allows an agent to make steady progress, solving all problems. Furthermore, these abstractions induce an order to the problems, seen at random during training. The recovered order has significant agreement with the expert-designed Khan Academy curriculum, and second-generation agents trained on the recovered curriculum learn significantly faster. These results illustrate the synergistic role of abstractions and curricula in the cultural transmission of mathematics.

Ben Prystawski, Paul Thibodeau, Christopher Potts et al.

Probabilistic models of language understanding are valuable tools for investigating human language use. However, they need to be hand-designed for a particular domain. In contrast, large language models (LLMs) are trained on text that spans a wide array of domains, but they lack the structure and interpretability of probabilistic models. In this paper, we use chain-of-thought prompts to introduce structures from probabilistic models into LLMs. We explore this approach in the case of metaphor understanding. Our chain-of-thought prompts lead language models to infer latent variables and reason about their relationships in order to choose appropriate paraphrases for metaphors. The latent variables and relationships chosen are informed by theories of metaphor understanding from cognitive psychology. We apply these prompts to the two largest versions of GPT-3 and show that they can improve performance in a paraphrase selection task.

Zachary Kenton, Noah Y. Siegel, János Kramár et al.

Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a judge; consultancy, where a single AI tries to convince a judge that asks questions; and compare to a baseline of direct question-answering, where the judge just answers outright without the AI. We use large language models (LLMs) as both AI agents and as stand-ins for human judges, taking the judge models to be weaker than agent models. We benchmark on a diverse range of asymmetries between judges and agents, extending previous work on a single extractive QA task with information asymmetry, to also include mathematics, coding, logic and multimodal reasoning asymmetries. We find that debate outperforms consultancy across all tasks when the consultant is randomly assigned to argue for the correct/incorrect answer. Comparing debate to direct question answering, the results depend on the type of task: in extractive QA tasks with information asymmetry debate outperforms direct question answering, but in other tasks without information asymmetry the results are mixed. Previous work assigned debaters/consultants an answer to argue for. When we allow them to instead choose which answer to argue for, we find judges are less frequently convinced by the wrong answer in debate than in consultancy. Further, we find that stronger debater models increase judge accuracy, though more modestly than in previous studies.

Fei Fang, Kunal Sinha, Noah D. Goodman et al.

Speakers' referential expressions often depart from communicative ideals in ways that help illuminate the nature of pragmatic language use. Patterns of overmodification, in which a speaker uses a modifier that is redundant given their communicative goal, have proven especially informative in this regard. It seems likely that these patterns are shaped by the environment a speaker is exposed to in complex ways. Unfortunately, systematically manipulating these factors during human language acquisition is impossible. In this paper, we propose to address this limitation by adopting neural networks (NN) as learning agents. By systematically varying the environments in which these agents are trained, while keeping the NN architecture constant, we show that overmodification is more likely with environmental features that are infrequent or salient. We show that these findings emerge naturally in the context of a probabilistic model of pragmatic communication.

Ryan Burnell, Yumeya Yamamori, Orhan Firat et al.

Despite widespread discussion of AGI, there is no clear framework for measuring progress toward it. This ambiguity fuels subjective claims, makes it difficult to track progress, and risks hindering responsible governance. As a starting point to address this gap, we present a framework for understanding system capabilities in relation to human cognitive abilities. Drawing from decades of research in psychology, neuroscience, and cognitive science, we introduce a Cognitive Taxonomy that deconstructs general intelligence into 10 key cognitive faculties. We then propose a rigorous evaluation protocol in which a system's performance is measured across a suite of targeted, held-out cognitive tasks, generating a 'cognitive profile' that can be used to understand a system's strengths and weaknesses. We hope this framework will provide a practical roadmap and an initial step toward more rigorous, empirical evaluation of AGI.

Joy Hsu, Jiayuan Mao, Joshua B. Tenenbaum et al.

A unique aspect of human visual understanding is the ability to flexibly interpret abstract concepts: acquiring lifted rules explaining what they symbolize, grounding them across familiar and unfamiliar contexts, and making predictions or reasoning about them. While off-the-shelf vision-language models excel at making literal interpretations of images (e.g., recognizing object categories such as tree branches), they still struggle to make sense of such visual abstractions (e.g., how an arrangement of tree branches may form the walls of a maze). To address this challenge, we introduce Deep Schema Grounding (DSG), a framework that leverages explicit structured representations of visual abstractions for grounding and reasoning. At the core of DSG are schemas--dependency graph descriptions of abstract concepts that decompose them into more primitive-level symbols. DSG uses large language models to extract schemas, then hierarchically grounds concrete to abstract components of the schema onto images with vision-language models. The grounded schema is used to augment visual abstraction understanding. We systematically evaluate DSG and different methods in reasoning on our new Visual Abstractions Dataset, which consists of diverse, real-world images of abstract concepts and corresponding question-answer pairs labeled by humans. We show that DSG significantly improves the abstract visual reasoning performance of vision-language models, and is a step toward human-aligned understanding of visual abstractions.

Linas Nasvytis, Simon Jerome Han, Ben Prystawski et al.

Language models can use verifiable rewards to improve at a wide variety of reasoning tasks. However, both parametric (e.g. RLVR) and non-parametric (e.g. prompt optimization) approaches to doing so typically require hundreds of training samples and thousands of model rollouts, making them expensive in the best case and intractable in the worst. To address this challenge, we introduce Contrastive Reflection (CORE), a non-parametric learning algorithm that compares past reasoning traces to generate insights: short natural-language descriptions of reasoning strategies and constraints that capture differences between successful and unsuccessful problem attempts. Across four reasoning tasks, we demonstrate that CORE enables more rapid improvement than both parametric (GRPO) and non-parametric (GEPA, episodic RAG, and MemRL) methods, while using fewer rollouts. Under fixed rollout budgets with as few as five training samples, we then show that CORE also achieves comparable or greater performance gains than each baseline. Finally, we highlight how CORE is also substantially more context-efficient than non-parametric baselines, requiring fewer prompt tokens while storing learned knowledge as compact, interpretable natural-language insights. Our results therefore suggest that distilling contrasts between successful and unsuccessful reasoning traces into abstract and useful insights can provide a more efficient and interpretable route to model self-improvement than weight updates, prompt optimization, or direct reuse of stored reasoning traces.

Kanishk Gandhi, Zoe Lynch, Jan-Philipp Fränken et al.

Understanding emotions is fundamental to human interaction and experience. Humans easily infer emotions from situations or facial expressions, situations from emotions, and do a variety of other affective cognition. How adept is modern AI at these inferences? We introduce an evaluation framework for testing affective cognition in foundation models. Starting from psychological theory, we generate 1,280 diverse scenarios exploring relationships between appraisals, emotions, expressions, and outcomes. We evaluate the abilities of foundation models (GPT-4, Claude-3, Gemini-1.5-Pro) and humans (N = 567) across carefully selected conditions. Our results show foundation models tend to agree with human intuitions, matching or exceeding interparticipant agreement. In some conditions, models are ``superhuman'' -- they better predict modal human judgements than the average human. All models benefit from chain-of-thought reasoning. This suggests foundation models have acquired a human-like understanding of emotions and their influence on beliefs and behavior.

Zhengxuan Wu, Atticus Geiger, Aryaman Arora et al. · stanford

Interventions on model-internal states are fundamental operations in many areas of AI, including model editing, steering, robustness, and interpretability. To facilitate such research, we introduce $\textbf{pyvene}$, an open-source Python library that supports customizable interventions on a range of different PyTorch modules. $\textbf{pyvene}$ supports complex intervention schemes with an intuitive configuration format, and its interventions can be static or include trainable parameters. We show how $\textbf{pyvene}$ provides a unified and extensible framework for performing interventions on neural models and sharing the intervened upon models with others. We illustrate the power of the library via interpretability analyses using causal abstraction and knowledge localization. We publish our library through Python Package Index (PyPI) and provide code, documentation, and tutorials at https://github.com/stanfordnlp/pyvene.

Shubhra Mishra, Gabriel Poesia, Noah D. Goodman

Large Language Models (LLMs) solely trained on next-token prediction learn to solve a wide range of problems involving mathematical reasoning. But how does this ability evolve during training? We show the first analysis of how mathematical reasoning abilities of several open-weight LLMs develop during pre-training and post-training. To this end, we construct MathCAMPS, a synthetic dataset of novel mathematical reasoning problems grounded in 44 fine-grained skills taken from the Common Core curriculum from K to 8th grades. In one experiment, we show that mathematical skills are learned during pre-training in an order that measurably correlates with the human-designed curriculum, even though training data are randomly ordered. We also show a detailed analysis of which mathematical abilities benefit from instruction tuning, a widely used post-training method and, in contrast, which skills suffer. Our work paves the way for an empirical understanding of LLM training dynamics in relation to reasoning.

Steven Y. Feng, Noah D. Goodman, Michael C. Frank

While high-performing language models are typically trained on hundreds of billions of words, human children become fluent language users with a much smaller amount of data. What are the features of the data they receive, and how do these features support language modeling objectives? To investigate this question, we train GPT-2 and RoBERTa models on 29M words of English child-directed speech and a new matched, synthetic dataset (TinyDialogues), comparing to OpenSubtitles, Wikipedia, and a heterogeneous blend of datasets from the BabyLM challenge. We evaluate the syntactic and semantic knowledge of these models using developmentally-inspired evaluations. Through pretraining experiments, we test whether the global developmental ordering or the local discourse ordering of children's training data supports high performance relative to other datasets. The local properties of the data affect model results, but surprisingly, global properties do not. Further, child language input is not uniquely valuable for training language models. These findings support the hypothesis that, rather than proceeding from better data, the child's learning algorithm is substantially more data-efficient than current language modeling techniques.

Kanishk Gandhi, Shivam Garg, Noah D. Goodman et al.

Environments are the bottleneck for self-improving agents. Current terminal benchmarks were built for evaluation, not training; reinforcement learning requires a scalable pipeline, not just a dataset. We introduce Endless Terminals, a fully autonomous pipeline that procedurally generates terminal-use tasks without human annotation. The pipeline has four stages: generating diverse task descriptions, building and validating containerized environments, producing completion tests, and filtering for solvability. From this pipeline we obtain 3255 tasks spanning file operations, log management, data processing, scripting, and database operations. We train agents using vanilla PPO with binary episode level rewards and a minimal interaction loop: no retrieval, multi-agent coordination, or specialized tools. Despite this simplicity, models trained on Endless Terminals show substantial gains: on our held-out dev set, Llama-3.2-3B improves from 4.0% to 18.2%, Qwen2.5-7B from 10.7% to 53.3%, and Qwen3-8B-openthinker-sft from 42.6% to 59.0%. These improvements transfer to human-curated benchmarks: models trained on Endless Terminals show substantial gains on held out human curated benchmarks: on TerminalBench 2.0, Llama-3.2-3B improves from 0.0% to 2.2%, Qwen2.5-7B from 2.2% to 3.4%, and Qwen3-8B-openthinker-sft from 1.1% to 6.7%, in each case outperforming alternative approaches including models with more complex agentic scaffolds. These results demonstrate that simple RL succeeds when environments scale.

Gabriel Poesia, Kanishk Gandhi, Eric Zelikman et al.

Language models often achieve higher accuracy when reasoning step-by-step in complex tasks. However, even when arriving at a correct final answer, their rationales are often logically unsound or inconsistent. This is a major issue when reliable reasoning traces are needed, such when fine-tuning on model-generated reasoning for self-improvement. To tackle these issues, we introduce a class of tools for language models called \emph{guides}, that use state and incremental constraints to guide generation. A guide can be invoked by the model to constrain its own generation to a set of valid statements given by the tool. In turn, the model's choices can change the guide's state. We show how a general system for logical reasoning can be used as a guide, which we call \textsc{LogicGuide}. Given a reasoning problem in natural language, a model can formalize its assumptions for \textsc{LogicGuide} and guarantee that its step-by-step reasoning is sound. In experiments on PrOntoQA, ProofWriter and Syllogism Validity datasets, \textsc{LogicGuide} significantly improves the performance of GPT-3, GPT-3.5 Turbo and LLaMA (accuracy gains up to 35\%), while drastically reducing \emph{content effects} -- the interference between unwanted prior assumptions and reasoning, which humans and language models suffer from. We then explore bootstrapping GPT-3.5 Turbo and LLaMA using their own reasoning traces. We find that LogicGuide is critical: by training only on certified self-generated reasoning, models can self-improve, avoiding learning from their own hallucinations. Moreover, bootstrapped models enjoy significant boosts on ReClor, a challenging real-world reasoning dataset, even when not relying on formalization at inference time.

Michael Y. Li, Jubayer Ibn Hamid, Emily B. Fox et al.

Chain-of-thought reasoning has driven striking advances in language model capability, yet every reasoning step grows the KV cache, creating a bottleneck to scaling this paradigm further. Current approaches manage these constraints on the model's behalf using hand-designed criteria. A more scalable approach would let end-to-end learning subsume this design choice entirely, following a broader pattern in deep learning. After all, if a model can learn to reason, why can't it learn to forget? We introduce Neural Garbage Collection (NGC), in which a language model learns to forget while learning to reason, trained end-to-end from outcome-based task reward alone. As the model reasons, it periodically pauses, decides which KV cache entries to evict, and continues to reason conditioned on the remaining cache. By treating tokens in a chain-of-thought and cache-eviction decisions as discrete actions sampled from the language model, we can use reinforcement learning to jointly optimize how the model reasons and how it manages its own memory: what the model evicts shapes what it remembers, what it remembers shapes its reasoning, and the correctness of that reasoning determines its reward. Crucially, the model learns this behavior entirely from a single learning signal - the outcome-based task reward - without supervised fine-tuning or proxy objectives. On Countdown, AMC, and AIME tasks, NGC maintains strong accuracy relative to the full-cache upper bound at 2-3x peak KV cache size compression and substantially outperforms eviction baselines. Our results are a first step towards a broader vision where end-to-end optimization drives both capability and efficiency in language models.

Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh et al.

Test-time inference has emerged as a powerful paradigm for enabling language models to ``think'' longer and more carefully about complex challenges, much like skilled human experts. While reinforcement learning (RL) can drive self-improvement in language models on verifiable tasks, some models exhibit substantial gains while others quickly plateau. For instance, we find that Qwen-2.5-3B far exceeds Llama-3.2-3B under identical RL training for the game of Countdown. This discrepancy raises a critical question: what intrinsic properties enable effective self-improvement? We introduce a framework to investigate this question by analyzing four key cognitive behaviors -- verification, backtracking, subgoal setting, and backward chaining -- that both expert human problem solvers and successful language models employ. Our study reveals that Qwen naturally exhibits these reasoning behaviors, whereas Llama initially lacks them. In systematic experimentation with controlled behavioral datasets, we find that priming Llama with examples containing these reasoning behaviors enables substantial improvements during RL, matching or exceeding Qwen's performance. Importantly, the presence of reasoning behaviors, rather than correctness of answers, proves to be the critical factor -- models primed with incorrect solutions containing proper reasoning patterns achieve comparable performance to those trained on correct solutions. Finally, leveraging continued pretraining with OpenWebMath data, filtered to amplify reasoning behaviors, enables the Llama model to match Qwen's self-improvement trajectory. Our findings establish a fundamental relationship between initial reasoning behaviors and the capacity for improvement, explaining why some language models effectively utilize additional computation while others plateau.

Kanishk Gandhi, Denise Lee, Gabriel Grand et al.

Language models are rarely shown fruitful mistakes while training. They then struggle to look beyond the next token, suffering from a snowballing of errors and struggling to predict the consequence of their actions several steps ahead. In this paper, we show how language models can be taught to search by representing the process of search in language, as a flattened string -- a stream of search (SoS). We propose a unified language for search that captures an array of different symbolic search strategies. We demonstrate our approach using the simple yet difficult game of Countdown, where the goal is to combine input numbers with arithmetic operations to reach a target number. We pretrain a transformer-based language model from scratch on a dataset of streams of search generated by heuristic solvers. We find that SoS pretraining increases search accuracy by 25% over models trained to predict only the optimal search trajectory. We further finetune this model with two policy improvement methods: Advantage-Induced Policy Alignment (APA) and Self-Taught Reasoner (STaR). The finetuned SoS models solve 36% of previously unsolved problems, including problems that cannot be solved by any of the heuristic solvers. Our results indicate that language models can learn to solve problems via search, self-improve to flexibly use different search strategies, and potentially discover new ones.

Zhengxuan Wu, Atticus Geiger, Thomas Icard et al.

Obtaining human-interpretable explanations of large, general-purpose language models is an urgent goal for AI safety. However, it is just as important that our interpretability methods are faithful to the causal dynamics underlying model behavior and able to robustly generalize to unseen inputs. Distributed Alignment Search (DAS) is a powerful gradient descent method grounded in a theory of causal abstraction that has uncovered perfect alignments between interpretable symbolic algorithms and small deep learning models fine-tuned for specific tasks. In the present paper, we scale DAS significantly by replacing the remaining brute-force search steps with learned parameters -- an approach we call Boundless DAS. This enables us to efficiently search for interpretable causal structure in large language models while they follow instructions. We apply Boundless DAS to the Alpaca model (7B parameters), which, off the shelf, solves a simple numerical reasoning problem. With Boundless DAS, we discover that Alpaca does this by implementing a causal model with two interpretable boolean variables. Furthermore, we find that the alignment of neural representations with these variables is robust to changes in inputs and instructions. These findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models. Our tool is extensible to larger LLMs and is released publicly at `https://github.com/stanfordnlp/pyvene`.

Chinmaya Andukuri, Jan-Philipp Fränken, Tobias Gerstenberg et al.

When prompting language models to complete a task, users often leave important aspects unsaid. While asking questions could resolve this ambiguity (GATE; Li et al., 2023), models often struggle to ask good questions. We explore a language model's ability to self-improve (STaR; Zelikman et al., 2022) by rewarding the model for generating useful questions-a simple method we dub STaR-GATE. We generate a synthetic dataset of 25,500 unique persona-task prompts to simulate conversations between a pretrained language model-the Questioner-and a Roleplayer whose preferences are unknown to the Questioner. By asking questions, the Questioner elicits preferences from the Roleplayer. The Questioner is iteratively finetuned on questions that increase the probability of high-quality responses to the task, which are generated by an Oracle with access to the Roleplayer's latent preferences. After two iterations of self-improvement, the Questioner asks better questions, allowing it to generate responses that are preferred over responses from the initial model on 72% of tasks. Our results indicate that teaching a language model to ask better questions leads to better personalized responses.

Michael Y. Li, Emily B. Fox, Noah D. Goodman

Statistical model discovery is a challenging search over a vast space of models subject to domain-specific constraints. Efficiently searching over this space requires expertise in modeling and the problem domain. Motivated by the domain knowledge and programming capabilities of large language models (LMs), we introduce a method for language model driven automated statistical model discovery. We cast our automated procedure within the principled framework of Box's Loop: the LM iterates between proposing statistical models represented as probabilistic programs, acting as a modeler, and critiquing those models, acting as a domain expert. By leveraging LMs, we do not have to define a domain-specific language of models or design a handcrafted search procedure, which are key restrictions of previous systems. We evaluate our method in three settings in probabilistic modeling: searching within a restricted space of models, searching over an open-ended space, and improving expert models under natural language constraints (e.g., this model should be interpretable to an ecologist). Our method identifies models on par with human expert designed models and extends classic models in interpretable ways. Our results highlight the promise of LM-driven model discovery.

Jan-Philipp Fränken, Eric Zelikman, Rafael Rafailov et al.

When prompting a language model (LM), users often expect the model to adhere to a set of behavioral principles across diverse tasks, such as producing insightful content while avoiding harmful or biased language. Instilling such principles (i.e., a constitution) into a model is resource-intensive, technically challenging, and generally requires human preference labels or examples. We introduce SAMI, an iterative algorithm that finetunes a pretrained language model (without requiring preference labels or demonstrations) to increase the conditional mutual information between constitutions and self-generated responses given queries from a dataset. On single-turn dialogue and summarization, a SAMI-trained mistral-7b outperforms the initial pretrained model, with win rates between 66% and 77%. Strikingly, it also surpasses an instruction-finetuned baseline (mistral-7b-instruct) with win rates between 55% and 57% on single-turn dialogue. SAMI requires a model that writes the principles. To avoid dependence on strong models for writing principles, we align a strong pretrained model (mixtral-8x7b) using constitutions written by a weak instruction-finetuned model (mistral-7b-instruct), achieving a 65% win rate on summarization. Finally, we investigate whether SAMI generalizes to diverse summarization principles (e.g., "summaries should be scientific") and scales to stronger models (llama3-70b), finding that it achieves win rates of up to 68% for learned and 67% for held-out principles compared to the base model. Our results show that a pretrained LM can learn to follow constitutions without using preference labels, demonstrations, or human oversight.

Joy He-Yueya, Noah D. Goodman, Emma Brunskill · stanford

Creating effective educational materials generally requires expensive and time-consuming studies of student learning outcomes. To overcome this barrier, one idea is to build computational models of student learning and use them to optimize instructional materials. However, it is difficult to model the cognitive processes of learning dynamics. We propose an alternative approach that uses Language Models (LMs) as educational experts to assess the impact of various instructions on learning outcomes. Specifically, we use GPT-3.5 to evaluate the overall effect of instructional materials on different student groups and find that it can replicate well-established educational findings such as the Expertise Reversal Effect and the Variability Effect. This demonstrates the potential of LMs as reliable evaluators of educational content. Building on this insight, we introduce an instruction optimization approach in which one LM generates instructional materials using the judgments of another LM as a reward function. We apply this approach to create math word problem worksheets aimed at maximizing student learning gains. Human teachers' evaluations of these LM-generated worksheets show a significant alignment between the LM judgments and human teacher preferences. We conclude by discussing potential divergences between human and LM opinions and the resulting pitfalls of automating instructional design.

Aryaman Arora, Dan Jurafsky, Christopher Potts et al. · stanford

In-context learning (ICL) is a powerful technique for getting language models to perform complex tasks with no training updates. Prior work has established strong correlations between the number of in-context examples provided and the accuracy of the model's predictions. In this paper, we seek to explain this correlation by showing that ICL approximates a Bayesian learner. This perspective gives rise to a novel Bayesian scaling law for ICL. In experiments with \mbox{GPT-2} models of different sizes, our scaling law matches existing scaling laws in accuracy while also offering interpretable terms for task priors, learning efficiency, and per-example probabilities. To illustrate the analytic power that such interpretable scaling laws provide, we report on controlled synthetic dataset experiments designed to inform real-world studies of safety alignment. In our experimental protocol, we use SFT or DPO to suppress an unwanted existing model capability and then use ICL to try to bring that capability back (many-shot jailbreaking). We then study ICL on real-world instruction-tuned LLMs using capabilities benchmarks as well as a new many-shot jailbreaking dataset. In all cases, Bayesian scaling laws accurately predict the conditions under which ICL will cause suppressed behaviors to reemerge, which sheds light on the ineffectiveness of post-training at increasing LLM safety.

Zhengxuan Wu, Atticus Geiger, Jing Huang et al. · stanford

We respond to the recent paper by Makelov et al. (2023), which reviews subspace interchange intervention methods like distributed alignment search (DAS; Geiger et al. 2023) and claims that these methods potentially cause "interpretability illusions". We first review Makelov et al. (2023)'s technical notion of what an "interpretability illusion" is, and then we show that even intuitive and desirable explanations can qualify as illusions in this sense. As a result, their method of discovering "illusions" can reject explanations they consider "non-illusory". We then argue that the illusions Makelov et al. (2023) see in practice are artifacts of their training and evaluation paradigms. We close by emphasizing that, though we disagree with their core characterization, Makelov et al. (2023)'s examples and discussion have undoubtedly pushed the field of interpretability forward.

Polina Tsvilodub, Kanishk Gandhi, Haoran Zhao et al.

Humans naturally interpret numbers non-literally, effortlessly combining context, world knowledge, and speaker intent. We investigate whether large language models (LLMs) interpret numbers similarly, focusing on hyperbole and pragmatic halo effects. Through systematic comparison with human data and computational models of pragmatic reasoning, we find that LLMs diverge from human interpretation in striking ways. By decomposing pragmatic reasoning into testable components, grounded in the Rational Speech Act framework, we pinpoint where LLM processing diverges from human cognition -- not in prior knowledge, but in reasoning with it. This insight leads us to develop a targeted solution -- chain-of-thought prompting inspired by an RSA model makes LLMs' interpretations more human-like. Our work demonstrates how computational cognitive models can both diagnose AI-human differences and guide development of more human-like language understanding capabilities.

Satchel Grant, Noah D. Goodman, James L. McClelland

What types of numeric representations emerge in neural systems, and what would a satisfying answer to this question look like? In this work, we interpret Neural Network (NN) solutions to sequence based number tasks using a variety of methods to understand how well we can interpret them through the lens of interpretable Symbolic Algorithms (SAs) -- precise programs describable by rules and typed, mutable variables. We use autoregressive GRUs, LSTMs, and Transformers trained on tasks where the correct tokens depend on numeric information only latent in the task structure. We show through multiple causal and theoretical methods that we can interpret raw NN activity through the lens of simplified SAs when we frame the activity in terms of neural subspaces rather than individual neurons. Using Distributed Alignment Search (DAS), we find that, depending on network architecture, dimensionality, and task specifications, alignments with SA's can be very high, or they can be only approximate, or fail altogether. We extend our analytic toolkit to address the failure cases by expanding the DAS framework to a broader class of alignment functions that more flexibly capture NN activity in terms of interpretable variables from SAs, and we provide theoretic and empirical explorations of Linear Alignment Functions (LAFs) in contrast to the preexisting Orthogonal Alignment Functions (OAFs). Through analyses of specific cases we confirm the usefulness of causal interventions on neural subspaces for NN interpretability, and we show that recurrent models can develop graded, symbol-like number variables in their neural activity. We further show that shallow Transformers learn very different solutions than recurrent networks, and we prove that such models must use anti-Markovian solutions -- solutions that do not rely on cumulative, Markovian hidden states -- in the absence of sufficient attention layers.

Daniel Wurgaft, Ekdeep Singh Lubana, Core Francisco Park et al.

Recent work analyzing in-context learning (ICL) has identified a broad set of strategies that describe model behavior in different experimental conditions. We aim to unify these findings by asking why a model learns these disparate strategies in the first place. Specifically, we start with the observation that when trained to learn a mixture of tasks, as is popular in the literature, the strategies learned by a model for performing ICL can be captured by a family of Bayesian predictors: a memorizing predictor, which assumes a discrete prior on the set of seen tasks, and a generalizing predictor, where the prior matches the underlying task distribution. Adopting the normative lens of rational analysis, where a learner's behavior is explained as an optimal adaptation to data given computational constraints, we develop a hierarchical Bayesian framework that almost perfectly predicts Transformer next-token predictions throughout training -- without assuming access to its weights. Under this framework, pretraining is viewed as a process of updating the posterior probability of different strategies, and inference-time behavior as a posterior-weighted average over these strategies' predictions. Our framework draws on common assumptions about neural network learning dynamics, which make explicit a tradeoff between loss and complexity among candidate strategies: beyond how well it explains the data, a model's preference towards implementing a strategy is dictated by its complexity. This helps explain well-known ICL phenomena, while offering novel predictions: e.g., we show a superlinear trend in the timescale for transitioning from generalization to memorization as task diversity increases. Overall, our work advances an explanatory and predictive account of ICL grounded in tradeoffs between strategy loss and complexity.

Kanishk Gandhi, Michael Y. Li, Lyle Goodyear et al.

Understanding the world and explaining it with scientific theories is a central aspiration of artificial intelligence research. Proposing theories, designing experiments to test them, and then revising them based on data are fundamental to scientific discovery. Despite the significant promise of LLM-based scientific agents, no benchmarks systematically test LLM's ability to propose scientific models, collect experimental data, and revise them in light of new data. We introduce BoxingGym, a benchmark with 10 environments for systematically evaluating both experimental design (e.g. collecting data to test a scientific theory) and model discovery (e.g. proposing and revising scientific theories). To enable tractable and quantitative evaluation, we implement each environment as a generative probabilistic model with which a scientific agent can run interactive experiments. These probabilistic models are drawn from various real-world scientific domains ranging from psychology to ecology. To quantitatively evaluate a scientific agent's ability to collect informative experimental data, we compute the expected information gain (EIG), an information-theoretic quantity which measures how much an experiment reduces uncertainty about the parameters of a generative model. A good scientific theory is a concise and predictive explanation. Therefore, to quantitatively evaluate model discovery, we ask a scientific agent to explain their model and then assess whether this explanation enables another scientific agent to make reliable predictions about this environment. In addition to this explanation-based evaluation, we compute standard model evaluation metrics such as prediction errors. We find that current LLMs, such as GPT-4o, struggle with both experimental design and model discovery. We find that augmenting the LLM-based agent with an explicit statistical model does not reliably improve these results.

Daniel Wurgaft, Ben Prystawski, Kanishk Gandhi et al.

The think-aloud method, where participants voice their thoughts as they solve a task, is a valuable source of rich data about human reasoning processes. Yet, it has declined in popularity in contemporary cognitive science, largely because labor-intensive transcription and annotation preclude large sample sizes. Here, we develop methods to automate the transcription and annotation of verbal reports of reasoning using natural language processing tools, allowing for large-scale analysis of think-aloud data. In our study, 640 participants thought aloud while playing the Game of 24, a mathematical reasoning task. We automatically transcribed the recordings and coded the transcripts as search graphs, finding moderate inter-rater reliability with humans. We analyze these graphs and characterize consistency and variation in human reasoning traces. Our work demonstrates the value of think-aloud data at scale and serves as a proof of concept for the automated analysis of verbal reports.

Jan-Philipp Fränken, Kanishk Gandhi, Tori Qiu et al.

As AI systems like language models are increasingly integrated into decision-making processes affecting people's lives, it's critical to ensure that these systems have sound moral reasoning. To test whether they do, we need to develop systematic evaluations. We provide a framework that uses a language model to translate causal graphs that capture key aspects of moral dilemmas into prompt templates. With this framework, we procedurally generated a large and diverse set of moral dilemmas -- the OffTheRails benchmark -- consisting of 50 scenarios and 400 unique test items. We collected moral permissibility and intention judgments from human participants for a subset of our items and compared these judgments to those from two language models (GPT-4 and Claude-2) across eight conditions. We find that moral dilemmas in which the harm is a necessary means (as compared to a side effect) resulted in lower permissibility and higher intention ratings for both participants and language models. The same pattern was observed for evitable versus inevitable harmful outcomes. However, there was no clear effect of whether the harm resulted from an agent's action versus from having omitted to act. We discuss limitations of our prompt generation pipeline and opportunities for improving scenarios to increase the strength of experimental effects.

Michael Y. Li, Vivek Vajipey, Noah D. Goodman et al.

Understanding the world through models is a fundamental goal of scientific research. While large language model (LLM) based approaches show promise in automating scientific discovery, they often overlook the importance of criticizing scientific models. Criticizing models deepens scientific understanding and drives the development of more accurate models. Automating model criticism is difficult because it traditionally requires a human expert to define how to compare a model with data and evaluate if the discrepancies are significant--both rely heavily on understanding the modeling assumptions and domain. Although LLM-based critic approaches are appealing, they introduce new challenges: LLMs might hallucinate the critiques themselves. Motivated by this, we introduce CriticAL (Critic Automation with Language Models). CriticAL uses LLMs to generate summary statistics that capture discrepancies between model predictions and data, and applies hypothesis tests to evaluate their significance. We can view CriticAL as a verifier that validates models and their critiques by embedding them in a hypothesis testing framework. In experiments, we evaluate CriticAL across key quantitative and qualitative dimensions. In settings where we synthesize discrepancies between models and datasets, CriticAL reliably generates correct critiques without hallucinating incorrect ones. We show that both human and LLM judges consistently prefer CriticAL's critiques over alternative approaches in terms of transparency and actionability. Finally, we show that CriticAL's critiques enable an LLM scientist to improve upon human-designed models on real-world datasets.

Kanishk Gandhi, Agam Bhatia, Noah D. Goodman

To predict what someone will say is to model how they think. We study this through next-turn dialogue prediction: given a conversation, predict the next utterance produced by a person. We compare learning approaches along two dimensions: (1) whether the model is allowed to think before responding, and (2) how learning is rewarded either through an LLM-as-a-judge that scores semantic similarity and information completeness relative to the ground-truth response, or by directly maximizing the log-probability of the true human dialogue. We find that optimizing for judge-based rewards indeed increases judge scores throughout training, however it decreases the likelihood assigned to ground truth human responses and decreases the win rate when human judges choose the most human-like response among a real and synthetic option. This failure is amplified when the model is allowed to think before answering. In contrast, by directly maximizing the log-probability of observed human responses, the model learns to better predict what people actually say, improving on both log-probability and win rate evaluations. Treating chain-of-thought as a latent variable, we derive a lower bound on the log-probability. Optimizing this objective yields the best results on all our evaluations. These results suggest that thinking helps primarily when trained with a distribution-matching objective grounded in real human dialogue, and that scaling this approach to broader conversational data may produce models with a more nuanced understanding of human behavior.

Gabriel Poesia, David Broman, Nick Haber et al.

How did humanity coax mathematics from the aether? We explore the Platonic view that mathematics can be discovered from its axioms - a game of conjecture and proof. We describe Minimo (Mathematics from Intrinsic Motivation): an agent that jointly learns to pose challenging problems for itself (conjecturing) and solve them (theorem proving). Given a mathematical domain axiomatized in dependent type theory, we first combine methods for constrained decoding and type-directed synthesis to sample valid conjectures from a language model. Our method guarantees well-formed conjectures by construction, even as we start with a randomly initialized model. We use the same model to represent a policy and value function for guiding proof search. Our agent targets generating hard but provable conjectures - a moving target, since its own theorem proving ability also improves as it trains. We propose novel methods for hindsight relabeling on proof search trees to significantly improve the agent's sample efficiency in both tasks. Experiments on 3 axiomatic domains (propositional logic, arithmetic and group theory) demonstrate that our agent can bootstrap from only the axioms, self-improving in generating true and challenging conjectures and in finding proofs.

Eric Zelikman, Georges Harik, Yijia Shao et al.

When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%$\rightarrow$10.9%) and CommonsenseQA (36.3%$\rightarrow$47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way.

Kanishk Gandhi, Dorsa Sadigh, Noah D. Goodman

Strategic reasoning enables agents to cooperate, communicate, and compete with other agents in diverse situations. Existing approaches to solving strategic games rely on extensive training, yielding strategies that do not generalize to new scenarios or games without retraining. Large Language Models (LLMs), with their ability to comprehend and generate complex, context-rich language, could prove powerful as tools for strategic gameplay. This paper introduces an approach that uses pretrained LLMs with few-shot chain-of-thought examples to enable strategic reasoning for AI agents. Our approach uses systematically generated demonstrations of reasoning about states, values, and beliefs to prompt the model. Using extensive variations of simple matrix games, we show that strategies that are derived based on systematically generated prompts generalize almost perfectly to new game structures, alternate objectives, and hidden information. Additionally, we demonstrate our approach can lead to human-like negotiation strategies in realistic scenarios without any extra training or fine-tuning. Our results highlight the ability of LLMs, guided by systematic reasoning demonstrations, to adapt and excel in diverse strategic scenarios.