Noah Goodman

Megha Srivastava, Erdem Biyik, Suvir Mirchandani et al. · stanford

Recent works on shared autonomy and assistive-AI technologies, such as assistive robot teleoperation, seek to model and help human users with limited ability in a fixed task. However, these approaches often fail to account for humans' ability to adapt and eventually learn how to execute a control task themselves. Furthermore, in applications where it may be desirable for a human to intervene, these methods may inhibit their ability to learn how to succeed with full self-control. In this paper, we focus on the problem of assistive teaching of motor control tasks such as parking a car or landing an aircraft. Despite their ubiquitous role in humans' daily activities and occupations, motor tasks are rarely taught in a uniform way due to their high complexity and variance. We propose an AI-assisted teaching algorithm that leverages skill discovery methods from reinforcement learning (RL) to (i) break down any motor control task into teachable skills, (ii) construct novel drill sequences, and (iii) individualize curricula to students with different capabilities. Through an extensive mix of synthetic and user studies on two motor control tasks -- parking a car with a joystick and writing characters from the Balinese alphabet -- we show that assisted teaching with skills improves student performance by around 40% compared to practicing full trajectories without skills, and practicing with individualized drills can result in up to 25% further improvement. Our source code is available at https://github.com/Stanford-ILIAD/teaching

Jesse Mu, Xiang Lisa Li, Noah Goodman · stanford

Prompting is the primary way to utilize the multitask capabilities of language models (LMs), but prompts occupy valuable space in the input context window, and repeatedly encoding the same prompt is computationally inefficient. Finetuning and distillation methods allow for specialization of LMs without prompting, but require retraining the model for each task. To avoid this trade-off entirely, we present gisting, which trains an LM to compress prompts into smaller sets of "gist" tokens which can be cached and reused for compute efficiency. Gist models can be trained with no additional cost over standard instruction finetuning by simply modifying Transformer attention masks to encourage prompt compression. On decoder (LLaMA-7B) and encoder-decoder (FLAN-T5-XXL) LMs, gisting enables up to 26x compression of prompts, resulting in up to 40% FLOPs reductions, 4.2% wall time speedups, and storage savings, all with minimal loss in output quality.

Alex Tamkin, Dat Nguyen, Salil Deshpande et al. · stanford

Models can fail in unpredictable ways during deployment due to task ambiguity, when multiple behaviors are consistent with the provided training data. An example is an object classifier trained on red squares and blue circles: when encountering blue squares, the intended behavior is undefined. We investigate whether pretrained models are better active learners, capable of disambiguating between the possible tasks a user may be trying to specify. Intriguingly, we find that better active learning is an emergent property of the pretraining process: pretrained models require up to 5 times fewer labels when using uncertainty-based active learning, while non-pretrained models see no or even negative benefit. We find these gains come from an ability to select examples with attributes that disambiguate the intended behavior, such as rare product categories or atypical backgrounds. These attributes are far more linearly separable in pretrained model's representation spaces vs non-pretrained models, suggesting a possible mechanism for this behavior.

Rose E Wang, Esin Durmus, Noah Goodman et al. · stanford

Modern language models can generate high-quality short texts. However, they often meander or are incoherent when generating longer texts. These issues arise from the next-token-only language modeling objective. Recent work in self-supervised learning suggests that models can learn good latent representations via contrastive learning, which can be effective for discriminative tasks. Our work analyzes the application of contrastive representations for generative tasks, like long text generation. We propose one approach for leveraging constrastive representations, which we call Time Control (TC). TC first learns a contrastive representation of the target text domain, then generates text by decoding from these representations. Compared to domain-specific methods and fine-tuning GPT2 across a variety of text domains, TC performs competitively to methods specific for learning sentence representations on discourse coherence. On long text generation settings, TC preserves the text structure both in terms of ordering (up to $+15\%$ better) and text length consistency (up to $+90\%$ better).

Alex Tamkin, Kunal Handa, Avash Shrestha et al. · stanford

Language models have recently achieved strong performance across a wide range of NLP benchmarks. However, unlike benchmarks, real world tasks are often poorly specified, and agents must deduce the user's intended behavior from a combination of context, instructions, and examples. We investigate how both humans and models behave in the face of such task ambiguity by proposing AmbiBench, a new benchmark of six ambiguously-specified classification tasks. We evaluate humans and models on AmbiBench by seeing how well they identify the intended task using 1) instructions with varying degrees of ambiguity, and 2) different numbers of labeled examples. We find that the combination of model scaling (to 175B parameters) and training with human feedback data enables models to approach or exceed the accuracy of human participants across tasks, but that either one alone is not sufficient. In addition, we show how to dramatically improve the accuracy of language models trained without large-scale human feedback training by finetuning on a small number of ambiguous in-context examples, providing a promising direction for teaching models to generalize well in the face of ambiguity.

Jasmine Bayrooti, Noah Goodman, Alex Tamkin · stanford

Contrastive learning methods have been applied to a range of domains and modalities by training models to identify similar "views" of data points. However, specialized scientific modalities pose a challenge for this paradigm, as identifying good views for each scientific instrument is complex and time-intensive. In this paper, we focus on applying contrastive learning approaches to a variety of remote sensing datasets. We show that Viewmaker networks, a recently proposed method for generating views, are promising for producing views in this setting without requiring extensive domain knowledge and trial and error. We apply Viewmaker to four multispectral imaging problems, each with a different format, finding that Viewmaker can outperform cropping- and reflection-based methods for contrastive learning in every case when evaluated on downstream classification tasks. This provides additional evidence that domain-agnostic methods can empower contrastive learning to scale to real-world scientific domains. Open source code can be found at https://github.com/jbayrooti/divmaker.

Alex Tamkin, Margalit Glasgow, Xiluo He et al. · stanford

What role do augmentations play in contrastive learning? Recent work suggests that good augmentations are label-preserving with respect to a specific downstream task. We complicate this picture by showing that label-destroying augmentations can be useful in the foundation model setting, where the goal is to learn diverse, general-purpose representations for multiple downstream tasks. We perform contrastive learning experiments on a range of image and audio datasets with multiple downstream tasks (e.g. for digits superimposed on photographs, predicting the class of one vs. the other). We find that Viewmaker Networks, a recently proposed model for learning augmentations for contrastive learning, produce label-destroying augmentations that stochastically destroy features needed for different downstream tasks. These augmentations are interpretable (e.g. altering shapes, digits, or letters added to images) and surprisingly often result in better performance compared to expert-designed augmentations, despite not preserving label information. To support our empirical results, we theoretically analyze a simple contrastive learning setting with a linear model. In this setting, label-destroying augmentations are crucial for preventing one set of features from suppressing the learning of features useful for another downstream task. Our results highlight the need for analyzing the interaction between multiple downstream tasks when trying to explain the success of foundation models.

Zhening Li, Gabriel Poesia, Omar Costilla-Reyes et al. · mit

Humans tame the complexity of mathematical reasoning by developing hierarchies of abstractions. With proper abstractions, solutions to hard problems can be expressed concisely, thus making them more likely to be found. In this paper, we propose Learning Mathematical Abstractions (LEMMA): an algorithm that implements this idea for reinforcement learning agents in mathematical domains. LEMMA augments Expert Iteration with an abstraction step, where solutions found so far are revisited and rewritten in terms of new higher-level actions, which then become available to solve new problems. We evaluate LEMMA on two mathematical reasoning tasks--equation solving and fraction simplification--in a step-by-step fashion. In these two domains, LEMMA improves the ability of an existing agent, both solving more problems and generalizing more effectively to harder problems than those seen during training.

Atticus Geiger, Duligur Ibeling, Amir Zur et al. · stanford

Causal abstraction provides a theoretical foundation for mechanistic interpretability, the field concerned with providing intelligible algorithms that are faithful simplifications of the known, but opaque low-level details of black box AI models. Our contributions are (1) generalizing the theory of causal abstraction from mechanism replacement (i.e., hard and soft interventions) to arbitrary mechanism transformation (i.e., functionals from old mechanisms to new mechanisms), (2) providing a flexible, yet precise formalization for the core concepts of polysemantic neurons, the linear representation hypothesis, modular features, and graded faithfulness, and (3) unifying a variety of mechanistic interpretability methods in the common language of causal abstraction, namely, activation and path patching, causal mediation analysis, causal scrubbing, causal tracing, circuit analysis, concept erasure, sparse autoencoders, differential binary masking, distributed alignment search, and steering.

Belinda Z. Li, Alex Tamkin, Noah Goodman et al. · meta-ai, mit

Language models (LMs) can be directed to perform target tasks by using labeled examples or natural language prompts. But selecting examples or writing prompts for can be challenging--especially in tasks that involve unusual edge cases, demand precise articulation of nebulous preferences, or require an accurate mental model of LM behavior. We propose to use *LMs themselves* to guide the task specification process. In this paper, we introduce **Generative Active Task Elicitation (GATE)**: a learning framework in which models elicit and infer intended behavior through free-form, language-based interaction with users. We study GATE in three domains: email validation, content recommendation, and moral reasoning. In preregistered experiments, we show that LMs prompted to perform GATE (e.g., by generating open-ended questions or synthesizing informative edge cases) elicit responses that are often more informative than user-written prompts or labels. Users report that interactive task elicitation requires less effort than prompting or example labeling and surfaces novel considerations not initially anticipated by users. Our findings suggest that LM-driven elicitation can be a powerful tool for aligning models to complex human preferences and values.

Jared Moore, Noah Goodman, Nick Haber et al.

Large language models can shift human beliefs across high-stakes domains, but most persuasion studies rely on pre/post belief change. These endpoint measures identify whether persuasion occurred, yet miss where and how beliefs moved within a dialogue. We present PERSUASIONTRACE, a framework for studying persuasion in human-LLM interaction. Built on a web-based experimental platform, PERSUASIONTRACE contributes a tool for multi-turn persuasion studies and a process-level evaluation protocol: it records multi-turn belief reports from human or simulated targets of persuasion, annotates persuader turns with rhetorical dimensions (logos/pathos/ethos), and evaluates simulators by fidelity to real human belief dynamics. Using this framework, we find that human targets group into two clusters of multi-turn belief updates and exhibit susceptibility to rhetorical strategies, and that LLMs are persuasive across generic and personalized topics, text and audio modalities, and multi-turn interactions. Prior work has chiefly used vanilla-prompted LLMs to simulate human targets, but we show that these simulators fail to replicate human belief dynamics. We introduce a Bayesian-network simulated target that maintains an explicit latent belief state over time so each persuader message yields cognitively realistic belief updates. In human-likeness evaluation, our Bayesian target scores near a human reference (81 vs 80), while baseline LLM targets score substantially lower (64). PERSUASIONTRACE reframes persuasion evaluation from endpoint movement alone to process fidelity, providing a stronger basis for scientific analysis and safer optimization of persuasive systems.

Mike Wu, Noah Goodman

Probabilistic programs provide an expressive representation language for generative models. Given a probabilistic program, we are interested in the task of posterior inference: estimating a latent variable given a set of observed variables. Existing techniques for inference in probabilistic programs often require choosing many hyper-parameters, are computationally expensive, and/or only work for restricted classes of programs. Here we formulate inference as masked language modeling: given a program, we generate a supervised dataset of variables and assignments, and randomly mask a subset of the assignments. We then train a neural network to unmask the random values, defining an approximate posterior distribution. By optimizing a single neural network across a range of programs we amortize the cost of training, yielding a "foundation" posterior able to do zero-shot inference for new programs. The foundation posterior can also be fine-tuned for a particular program and dataset by optimizing a variational inference objective. We show the efficacy of the approach, zero-shot and fine-tuned, on a benchmark of STAN programs.

Megha Srivastava, Noah Goodman, Dorsa Sadigh

AI assistance continues to help advance applications in education, from language learning to intelligent tutoring systems, yet current methods for providing students feedback are still quite limited. Most automatic feedback systems either provide binary correctness feedback, which may not help a student understand how to improve, or require hand-coding feedback templates, which may not generalize to new domains. This can be particularly challenging for physical control tasks, where the rich diversity in student behavior and specialized domains make it challenging to leverage general-purpose assistive tools for providing feedback. We design and build CORGI, a model trained to generate language corrections for physical control tasks, such as learning to ride a bike. CORGI takes in as input a pair of student and expert trajectories, and then generates natural language corrections to help the student improve. We collect and train CORGI over data from three diverse physical control tasks (drawing, steering, and joint movement). Through both automatic and human evaluations, we show that CORGI can (i) generate valid feedback for novel student trajectories, (ii) outperform baselines on domains with novel control dynamics, and (iii) improve student learning in an interactive drawing task.

Julia White, Noah Goodman, Robert Hawkins

Language use differs dramatically from context to context. To some degree, modern language models like GPT-3 are able to account for such variance by conditioning on a string of previous input text, or prompt. Yet prompting is ineffective when contexts are sparse, out-of-sample, or extra-textual; for instance, accounting for when and where the text was produced or who produced it. In this paper, we introduce the mixed-effects transformer (MET), a novel approach for learning hierarchically-structured prefixes -- lightweight modules prepended to the input -- to account for structured variation. Specifically, we show how the popular class of mixed-effects models may be extended to transformer-based architectures using a regularized prefix-tuning procedure with dropout. We evaluate this approach on several domain-adaptation benchmarks, finding that it efficiently adapts to novel contexts with minimal data while still effectively generalizing to unseen contexts.

Ifdita Hasan Orney, Jubayer Ibn Hamid, Shreya S Ramanujam et al.

Exploration is a cornerstone of learning from experience: it enables agents to find solutions to complex problems, generalize to novel ones, and scale performance with test-time compute. In this paper, we present a framework for post-training language models (LMs) that explicitly encourages optimistic exploration and promotes a synergy between exploration and exploitation. The central idea is to train the LM to generate sets of responses that are collectively accurate under the reward function and exploratory in their reasoning strategies. We first develop a general recipe for optimizing LMs with set reinforcement learning (set RL) under arbitrary objective functions, showing how standard RL algorithms can be adapted to this setting through a modification to the advantage computation. We then propose Polychromic Exploratory Policy Optimization (Poly-EPO), which instantiates this framework with an objective that explicitly synergizes exploration and exploitation. Across a range of reasoning benchmarks, we show that Poly-EPO improves generalization, as evidenced by higher pass@$k$ coverage, preserves greater diversity in model generations, and effectively scales with test-time compute.

Peiyang Song, Pengrui Han, Noah Goodman

Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks. Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios. To systematically understand and address these shortcomings, we present the first comprehensive survey dedicated to reasoning failures in LLMs. We introduce a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further subdivided into informal (intuitive) and formal (logical) reasoning. In parallel, we classify reasoning failures along a complementary axis into three types: fundamental failures intrinsic to LLM architectures that broadly affect downstream tasks; application-specific limitations that manifest in particular domains; and robustness issues characterized by inconsistent performance across minor variations. For each reasoning failure, we provide a clear definition, analyze existing studies, explore root causes, and present mitigation strategies. By unifying fragmented research efforts, our survey provides a structured perspective on systemic weaknesses in LLM reasoning, offering valuable insights and guiding future research towards building stronger, more reliable, and robust reasoning capabilities. We additionally release a comprehensive collection of research works on LLM reasoning failures, as a GitHub repository at https://github.com/Peiyang-Song/Awesome-LLM-Reasoning-Failures, to provide an easy entry point to this area.

Rose E. Wang, Pawan Wirawarn, Noah Goodman et al.

Lectures are a learning experience for both students and teachers. Students learn from teachers about the subject material, while teachers learn from students about how to refine their instruction. However, online student feedback is unstructured and abundant, making it challenging for teachers to learn and improve. We take a step towards tackling this challenge. First, we contribute a dataset for studying this problem: SIGHT is a large dataset of 288 math lecture transcripts and 15,784 comments collected from the Massachusetts Institute of Technology OpenCourseWare (MIT OCW) YouTube channel. Second, we develop a rubric for categorizing feedback types using qualitative analysis. Qualitative analysis methods are powerful in uncovering domain-specific insights, however they are costly to apply to large data sources. To overcome this challenge, we propose a set of best practices for using large language models (LLMs) to cheaply classify the comments at scale. We observe a striking correlation between the model's and humans' annotation: Categories with consistent human annotations (>$0.9$ inter-rater reliability, IRR) also display higher human-model agreement (>$0.7$), while categories with less consistent human annotations ($0.7$-$0.8$ IRR) correspondingly demonstrate lower human-model agreement ($0.3$-$0.5$). These techniques uncover useful student feedback from thousands of comments, costing around $\$0.002$ per comment. We conclude by discussing exciting future directions on using online student feedback and improving automated annotation techniques for qualitative research.

Rose E. Wang, Mike Wu, Noah Goodman

Learning often involves interaction between multiple agents. Human teacher-student settings best illustrate how interactions result in efficient knowledge passing where the teacher constructs a curriculum based on their students' abilities. Prior work in machine teaching studies how the teacher should construct optimal teaching datasets assuming the teacher knows everything about the student. However, in the real world, the teacher doesn't have complete information about the student. The teacher must interact and diagnose the student, before teaching. Our work proposes a simple diagnosis algorithm which uses Gaussian processes for inferring student-related information, before constructing a teaching dataset. We apply this to two settings. One is where the student learns from scratch and the teacher must figure out the student's learning algorithm parameters, eg. the regularization parameters in ridge regression or support vector machines. Two is where the student has partially explored the environment and the teacher must figure out the important areas the student has not explored; we study this in the offline reinforcement learning setting where the teacher must provide demonstrations to the student and avoid sending redundant trajectories. Our experiments highlight the importance of diagosing before teaching and demonstrate how students can learn more efficiently with the help of an interactive teacher. We conclude by outlining where diagnosing combined with teaching would be more desirable than passive learning.

Eric Bigelow, Daniel Wurgaft, YingQiao Wang et al.

Large language models (LLMs) can be controlled at inference time through prompts (in-context learning) and internal activations (activation steering). Different accounts have been proposed to explain these methods, yet their common goal of controlling model behavior raises the question of whether these seemingly disparate methodologies can be seen as specific instances of a broader framework. Motivated by this, we develop a unifying, predictive account of LLM control from a Bayesian perspective. Specifically, we posit that both context- and activation-based interventions impact model behavior by altering its belief in latent concepts: steering operates by changing concept priors, while in-context learning leads to an accumulation of evidence. This results in a closed-form Bayesian model that is highly predictive of LLM behavior across context- and activation-based interventions in a set of domains inspired by prior work on many-shot in-context learning. This model helps us explain prior empirical phenomena - e.g., sigmoidal learning curves as in-context evidence accumulates - while predicting novel ones - e.g., additivity of both interventions in log-belief space, which results in distinct phases such that sudden and dramatic behavioral shifts can be induced by slightly changing intervention controls. Taken together, this work offers a unified account of prompt-based and activation-based control of LLM behavior, and a methodology for empirically predicting the effects of these interventions.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Rose E. Wang, Pawan Wirawarn, Omar Khattab et al.

Many online content portals allow users to ask questions to supplement their understanding (e.g., of lectures). While information retrieval (IR) systems may provide answers for such user queries, they do not directly assist content creators -- such as lecturers who want to improve their content -- identify segments that _caused_ a user to ask those questions. We introduce the task of backtracing, in which systems retrieve the text segment that most likely caused a user query. We formalize three real-world domains for which backtracing is important in improving content delivery and communication: understanding the cause of (a) student confusion in the Lecture domain, (b) reader curiosity in the News Article domain, and (c) user emotion in the Conversation domain. We evaluate the zero-shot performance of popular information retrieval methods and language modeling methods, including bi-encoder, re-ranking and likelihood-based methods and ChatGPT. While traditional IR systems retrieve semantically relevant information (e.g., details on "projection matrices" for a query "does projecting multiple times still lead to the same point?"), they often miss the causally relevant context (e.g., the lecturer states "projecting twice gets me the same answer as one projection"). Our results show that there is room for improvement on backtracing and it requires new retrieval approaches. We hope our benchmark serves to improve future retrieval systems for backtracing, spawning systems that refine content generation and identify linguistic triggers influencing user queries. Our code and data are open-sourced: https://github.com/rosewang2008/backtracing.

Daniel Wurgaft, Can Rager, Matthew Kowal et al.

Neural representations carry rich geometric structure; but does that structure causally shape behavior? To address this question, we intervene along paths through activation space defined by different geometries, and measure the behavioral trajectories they induce. In particular, we test whether interventions that respect the geometry of activation space will yield behaviors close to those the model exhibits naturally. Concretely, we first fit an activation manifold $M_h$ to representations and a behavior manifold $M_y$ to output probability distributions. We then test the link $M_h \leftrightarrow M_y$ via interventions: we find that steering along $M_h$, which we term manifold steering, yields behavioral trajectories that follow $M_y$, while linear steering -- which assumes a Euclidean geometry -- cuts through off-manifold regions and hence produces unnatural outputs. Moreover, optimizing interventions in activation space to produce paths along $M_y$ recovers activation trajectories that trace the curvature of $M_h$. We demonstrate this bidirectional relationship between the geometry of representation and behavior across tasks and modalities. In language models, we use reasoning tasks with cyclic and sequential geometries as well as in-context learning tasks with more complex graph geometries. In a video world model, we use a task with geometry corresponding to physical dynamics. Overall, our work shows that geometry in neural representation is not merely incidental, but is in fact the proper object for enabling principled control via intervention on internals. This recasts the core problem of steering from finding the right direction to finding the right geometry.

Alex Tamkin, Vincent Liu, Rongfei Lu et al.

Self-supervised learning algorithms, including BERT and SimCLR, have enabled significant strides in fields like natural language processing, computer vision, and speech processing. However, these algorithms are domain-specific, meaning that new self-supervised learning algorithms must be developed for each new setting, including myriad healthcare, scientific, and multimodal domains. To catalyze progress toward domain-agnostic methods, we introduce DABS: a Domain-Agnostic Benchmark for Self-supervised learning. To perform well on DABS, an algorithm is evaluated on seven diverse domains: natural images, multichannel sensor data, English text, speech recordings, multilingual text, chest x-rays, and images with text descriptions. Each domain contains an unlabeled dataset for pretraining; the model is then is scored based on its downstream performance on a set of labeled tasks in the domain. We also present e-Mix and ShED: two baseline domain-agnostic algorithms; their relatively modest performance demonstrates that significant progress is needed before self-supervised learning is an out-of-the-box solution for arbitrary domains. Code for benchmark datasets and baseline algorithms is available at https://github.com/alextamkin/dabs.

Oliver Zhang, Mike Wu, Jasmine Bayrooti et al.

Contrastive learning has demonstrated great capability to learn representations without annotations, even outperforming supervised baselines. However, it still lacks important properties useful for real-world application, one of which is uncertainty. In this paper, we propose a simple way to generate uncertainty scores for many contrastive methods by re-purposing temperature, a mysterious hyperparameter used for scaling. By observing that temperature controls how sensitive the objective is to specific embedding locations, we aim to learn temperature as an input-dependent variable, treating it as a measure of embedding confidence. We call this approach "Temperature as Uncertainty", or TaU. Through experiments, we demonstrate that TaU is useful for out-of-distribution detection, while remaining competitive with benchmarks on linear evaluation. Moreover, we show that TaU can be learned on top of pretrained models, enabling uncertainty scores to be generated post-hoc with popular off-the-shelf models. In summary, TaU is a simple yet versatile method for generating uncertainties for contrastive learning. Open source code can be found at: https://github.com/mhw32/temperature-as-uncertainty-public.

Mike Wu, Richard L. Davis, Benjamin W. Domingue et al.

Item Response Theory (IRT) is a ubiquitous model for understanding human behaviors and attitudes based on their responses to questions. Large modern datasets offer opportunities to capture more nuances in human behavior, potentially improving psychometric modeling leading to improved scientific understanding and public policy. However, while larger datasets allow for more flexible approaches, many contemporary algorithms for fitting IRT models may also have massive computational demands that forbid real-world application. To address this bottleneck, we introduce a variational Bayesian inference algorithm for IRT, and show that it is fast and scalable without sacrificing accuracy. Applying this method to five large-scale item response datasets from cognitive science and education yields higher log likelihoods and higher accuracy in imputing missing data than alternative inference algorithms. Using this new inference approach we then generalize IRT with expressive Bayesian models of responses, leveraging recent advances in deep learning to capture nonlinear item characteristic curves (ICC) with neural networks. Using an eigth-grade mathematics test from TIMSS, we show our nonlinear IRT models can capture interesting asymmetric ICCs. The algorithm implementation is open-source, and easily usable.

Alex Tamkin, Mike Wu, Noah Goodman

Many recent methods for unsupervised representation learning train models to be invariant to different "views," or distorted versions of an input. However, designing these views requires considerable trial and error by human experts, hindering widespread adoption of unsupervised representation learning methods across domains and modalities. To address this, we propose viewmaker networks: generative models that learn to produce useful views from a given input. Viewmakers are stochastic bounded adversaries: they produce views by generating and then adding an $\ell_p$-bounded perturbation to the input, and are trained adversarially with respect to the main encoder network. Remarkably, when pretraining on CIFAR-10, our learned views enable comparable transfer accuracy to the well-tuned SimCLR augmentations -- despite not including transformations like cropping or color jitter. Furthermore, our learned views significantly outperform baseline augmentations on speech recordings (+9% points, on average) and wearable sensor data (+17% points). Viewmakers can also be combined with handcrafted views: they improve robustness to common image corruptions and can increase transfer performance in cases where handcrafted views are less explored. These results suggest that viewmakers may provide a path towards more general representation learning algorithms -- reducing the domain expertise and effort needed to pretrain on a much wider set of domains. Code is available at https://github.com/alextamkin/viewmaker.

Mike Wu, Richard L. Davis, Benjamin W. Domingue et al.

Item Response Theory (IRT) is a ubiquitous model for understanding humans based on their responses to questions, used in fields as diverse as education, medicine and psychology. Large modern datasets offer opportunities to capture more nuances in human behavior, potentially improving test scoring and better informing public policy. Yet larger datasets pose a difficult speed / accuracy challenge to contemporary algorithms for fitting IRT models. We introduce a variational Bayesian inference algorithm for IRT, and show that it is fast and scaleable without sacrificing accuracy. Using this inference approach we then extend classic IRT with expressive Bayesian models of responses. Applying this method to five large-scale item response datasets from cognitive science and education yields higher log likelihoods and improvements in imputing missing data. The algorithm implementation is open-source, and easily usable.

Rohin Shah, Alex Irpan, Alexander Matt Turner et al. · deepmind

Artificial General Intelligence (AGI) promises transformative benefits but also presents significant risks. We develop an approach to address the risk of harms consequential enough to significantly harm humanity. We identify four areas of risk: misuse, misalignment, mistakes, and structural risks. Of these, we focus on technical approaches to misuse and misalignment. For misuse, our strategy aims to prevent threat actors from accessing dangerous capabilities, by proactively identifying dangerous capabilities, and implementing robust security, access restrictions, monitoring, and model safety mitigations. To address misalignment, we outline two lines of defense. First, model-level mitigations such as amplified oversight and robust training can help to build an aligned model. Second, system-level security measures such as monitoring and access control can mitigate harm even if the model is misaligned. Techniques from interpretability, uncertainty estimation, and safer design patterns can enhance the effectiveness of these mitigations. Finally, we briefly outline how these ingredients could be combined to produce safety cases for AGI systems.

Kunal Handa, Yarin Gal, Ellie Pavlick et al. · meta-ai, mit

Aligning AI systems to users' interests requires understanding and incorporating humans' complex values and preferences. Recently, language models (LMs) have been used to gather information about the preferences of human users. This preference data can be used to fine-tune or guide other LMs and/or AI systems. However, LMs have been shown to struggle with crucial aspects of preference learning: quantifying uncertainty, modeling human mental states, and asking informative questions. These challenges have been addressed in other areas of machine learning, such as Bayesian Optimal Experimental Design (BOED), which focus on designing informative queries within a well-defined feature space. But these methods, in turn, are difficult to scale and apply to real-world problems where simply identifying the relevant features can be difficult. We introduce OPEN (Optimal Preference Elicitation with Natural language) a framework that uses BOED to guide the choice of informative questions and an LM to extract features and translate abstract BOED queries into natural language questions. By combining the flexibility of LMs with the rigor of BOED, OPEN can optimize the informativity of queries while remaining adaptable to real-world domains. In user studies, we find that OPEN outperforms existing LM- and BOED-based methods for preference elicitation.

Taylor Sorensen, Pushkar Mishra, Roma Patel et al. · uw

Modelling human variation in rating tasks is crucial for personalization, pluralistic model alignment, and computational social science. We propose representing individuals using natural language value profiles -- descriptions of underlying values compressed from in-context demonstrations -- along with a steerable decoder model that estimates individual ratings from a rater representation. To measure the predictive information in a rater representation, we introduce an information-theoretic methodology and find that demonstrations contain the most information, followed by value profiles, then demographics. However, value profiles effectively compress the useful information from demonstrations (>70% information preservation) and offer advantages in terms of scrutability, interpretability, and steerability. Furthermore, clustering value profiles to identify similarly behaving individuals better explains rater variation than the most predictive demographic groupings. Going beyond test set performance, we show that the decoder predictions change in line with semantic profile differences, are well-calibrated, and can help explain instance-level disagreement by simulating an annotator population. These results demonstrate that value profiles offer novel, predictive ways to describe individual variation beyond demographics or group information.

Yoshua Bengio, Tegan Maharaj, Luke Ong et al. · cmu, mila

Rapidly improving AI capabilities and autonomy hold significant promise of transformation, but are also driving vigorous debate on how to ensure that AI is safe, i.e., trustworthy, reliable, and secure. Building a trusted ecosystem is therefore essential -- it helps people embrace AI with confidence and gives maximal space for innovation while avoiding backlash. The "2025 Singapore Conference on AI (SCAI): International Scientific Exchange on AI Safety" aimed to support research in this space by bringing together AI scientists across geographies to identify and synthesise research priorities in AI safety. This resulting report builds on the International AI Safety Report chaired by Yoshua Bengio and backed by 33 governments. By adopting a defence-in-depth model, this report organises AI safety research domains into three types: challenges with creating trustworthy AI systems (Development), challenges with evaluating their risks (Assessment), and challenges with monitoring and intervening after deployment (Control).

Cédric Colas, Tracey Mills, Ben Prystawski et al.

The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models -- revealing how structured, language-compatible representations might enable human-machine collaborative learning.

Sydney Levine, Matija Franklin, Tan Zhi-Xuan et al. · mit

AI systems will soon have to navigate human environments and make decisions that affect people and other AI agents whose goals and values diverge. Contractualist alignment proposes grounding those decisions in agreements that diverse stakeholders would endorse under the right conditions, yet securing such agreement at scale remains costly and slow -- even for advanced AI. We therefore propose Resource-Rational Contractualism (RRC): a framework where AI systems approximate the agreements rational parties would form by drawing on a toolbox of normatively-grounded, cognitively-inspired heuristics that trade effort for accuracy. An RRC-aligned agent would not only operate efficiently, but also be equipped to dynamically adapt to and interpret the ever-changing human social world.

Jesse Mu, Victor Zhong, Roberta Raileanu et al.

Reinforcement learning (RL) agents are particularly hard to train when rewards are sparse. One common solution is to use intrinsic rewards to encourage agents to explore their environment. However, recent intrinsic exploration methods often use state-based novelty measures which reward low-level exploration and may not scale to domains requiring more abstract skills. Instead, we explore natural language as a general medium for highlighting relevant abstractions in an environment. Unlike previous work, we evaluate whether language can improve over existing exploration methods by directly extending (and comparing to) competitive intrinsic exploration baselines: AMIGo (Campero et al., 2021) and NovelD (Zhang et al., 2021). These language-based variants outperform their non-linguistic forms by 47-85% across 13 challenging tasks from the MiniGrid and MiniHack environment suites.

Ananya Karthik, Mike Wu, Noah Goodman et al.

Contrastive learning has made considerable progress in computer vision, outperforming supervised pretraining on a range of downstream datasets. However, is contrastive learning the better choice in all situations? We demonstrate two cases where it is not. First, under sufficiently small pretraining budgets, supervised pretraining on ImageNet consistently outperforms a comparable contrastive model on eight diverse image classification datasets. This suggests that the common practice of comparing pretraining approaches at hundreds or thousands of epochs may not produce actionable insights for those with more limited compute budgets. Second, even with larger pretraining budgets we identify tasks where supervised learning prevails, perhaps because the object-centric bias of supervised pretraining makes the model more resilient to common corruptions and spurious foreground-background correlations. These results underscore the need to characterize tradeoffs of different pretraining objectives across a wider range of contexts and training regimes.

Julia White, Gabriel Poesia, Robert Hawkins et al.

An overarching goal of natural language processing is to enable machines to communicate seamlessly with humans. However, natural language can be ambiguous or unclear. In cases of uncertainty, humans engage in an interactive process known as repair: asking questions and seeking clarification until their uncertainty is resolved. We propose a framework for building a visually grounded question-asking model capable of producing polar (yes-no) clarification questions to resolve misunderstandings in dialogue. Our model uses an expected information gain objective to derive informative questions from an off-the-shelf image captioner without requiring any supervised question-answer data. We demonstrate our model's ability to pose questions that improve communicative success in a goal-oriented 20 questions game with synthetic and human answerers.

Rishi Bommasani, Drew A. Hudson, Ehsan Adeli et al.

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

Mike Wu, Noah Goodman, Chris Piech et al.

High-quality computer science education is limited by the difficulty of providing instructor feedback to students at scale. While this feedback could in principle be automated, supervised approaches to predicting the correct feedback are bottlenecked by the intractability of annotating large quantities of student code. In this paper, we instead frame the problem of providing feedback as few-shot classification, where a meta-learner adapts to give feedback to student code on a new programming question from just a few examples annotated by instructors. Because data for meta-training is limited, we propose a number of amendments to the typical few-shot learning framework, including task augmentation to create synthetic tasks, and additional side information to build stronger priors about each task. These additions are combined with a transformer architecture to embed discrete sequences (e.g. code) to a prototypical representation of a feedback class label. On a suite of few-shot natural language processing tasks, we match or outperform state-of-the-art performance. Then, on a collection of student solutions to exam questions from an introductory university course, we show that our approach reaches an average precision of 88% on unseen questions, surpassing the 82% precision of teaching assistants. Our approach was successfully deployed to deliver feedback to 16,000 student exam-solutions in a programming course offered by a tier 1 university. This is, to the best of our knowledge, the first successful deployment of a machine learning based feedback to open-ended student code.

Gabriel Poesia, WenXin Dong, Noah Goodman

Abstract symbolic reasoning, as required in domains such as mathematics and logic, is a key component of human intelligence. Solvers for these domains have important applications, especially to computer-assisted education. But learning to solve symbolic problems is challenging for machine learning algorithms. Existing models either learn from human solutions or use hand-engineered features, making them expensive to apply in new domains. In this paper, we instead consider symbolic domains as simple environments where states and actions are given as unstructured text, and binary rewards indicate whether a problem is solved. This flexible setup makes it easy to specify new domains, but search and planning become challenging. We introduce four environments inspired by the Mathematics Common Core Curriculum, and observe that existing Reinforcement Learning baselines perform poorly. We then present a novel learning algorithm, Contrastive Policy Learning (ConPoLe) that explicitly optimizes the InfoNCE loss, which lower bounds the mutual information between the current state and next states that continue on a path to the solution. ConPoLe successfully solves all four domains. Moreover, problem representations learned by ConPoLe enable accurate prediction of the categories of problems in a real mathematics curriculum. Our results suggest new directions for reinforcement learning in symbolic domains, as well as applications to mathematics education.

Megha Srivastava, Noah Goodman

Intelligent and adaptive online education systems aim to make high-quality education available for a diverse range of students. However, existing systems usually depend on a pool of hand-made questions, limiting how fine-grained and open-ended they can be in adapting to individual students. We explore targeted question generation as a controllable sequence generation task. We first show how to fine-tune pre-trained language models for deep knowledge tracing (LM-KT). This model accurately predicts the probability of a student answering a question correctly, and generalizes to questions not seen in training. We then use LM-KT to specify the objective and data for training a model to generate questions conditioned on the student and target difficulty. Our results show we succeed at generating novel, well-calibrated language translation questions for second language learners from a real online education platform.

Jesse Mu, Noah Goodman

To build agents that can collaborate effectively with others, recent research has trained artificial agents to communicate with each other in Lewis-style referential games. However, this often leads to successful but uninterpretable communication. We argue that this is due to the game objective: communicating about a single object in a shared visual context is prone to overfitting and does not encourage language useful beyond concrete reference. In contrast, human language conveys a rich variety of abstract ideas. To promote such skills, we propose games that require communicating generalizations over sets of objects representing abstract visual concepts, optionally with separate contexts for each agent. We find that these games greatly improve systematicity and interpretability of the learned languages, according to several metrics in the literature. Finally, we propose a method for identifying logical operations embedded in the emergent languages by learning an approximate compositional reconstruction of the language.

Mike Wu, Noah Goodman, Stefano Ermon

In traditional software programs, it is easy to trace program logic from variables back to input, apply assertion statements to block erroneous behavior, and compose programs together. Although deep learning programs have demonstrated strong performance on novel applications, they sacrifice many of the functionalities of traditional software programs. With this as motivation, we take a modest first step towards improving deep learning programs by jointly training a generative model to constrain neural network activations to "decode" back to inputs. We call this design a Decodable Neural Network, or DecNN. Doing so enables a form of compositionality in neural networks, where one can recursively compose DecNN with itself to create an ensemble-like model with uncertainty. In our experiments, we demonstrate applications of this uncertainty to out-of-distribution detection, adversarial example detection, and calibration -- while matching standard neural networks in accuracy. We further explore this compositionality by combining DecNN with pretrained models, where we show promising results that neural networks can be regularized from using protected features.

Alex Tamkin, Dan Jurafsky, Noah Goodman

Language exhibits structure at different scales, ranging from subwords to words, sentences, paragraphs, and documents. To what extent do deep models capture information at these scales, and can we force them to better capture structure across this hierarchy? We approach this question by focusing on individual neurons, analyzing the behavior of their activations at different timescales. We show that signal processing provides a natural framework for separating structure across scales, enabling us to 1) disentangle scale-specific information in existing embeddings and 2) train models to learn more about particular scales. Concretely, we apply spectral filters to the activations of a neuron across an input, producing filtered embeddings that perform well on part of speech tagging (word-level), dialog speech acts classification (utterance-level), or topic classification (document-level), while performing poorly on the other tasks. We also present a prism layer for training models, which uses spectral filters to constrain different neurons to model structure at different scales. Our proposed BERT + Prism model can better predict masked tokens using long-range context and produces multiscale representations that perform better at utterance- and document-level tasks. Our methods are general and readily applicable to other domains besides language, such as images, audio, and video.

Mike Wu, Noah Goodman

Contrastive approaches to representation learning have recently shown great promise. In contrast to generative approaches, these contrastive models learn a deterministic encoder with no notion of uncertainty or confidence. In this paper, we introduce a simple approach based on "contrasting distributions" that learns to assign uncertainty for pretrained contrastive representations. In particular, we train a deep network from a representation to a distribution in representation space, whose variance can be used as a measure of confidence. In our experiments, we show that this deep uncertainty model can be used (1) to visually interpret model behavior, (2) to detect new noise in the input to deployed models, (3) to detect anomalies, where we outperform 10 baseline methods across 11 tasks with improvements of up to 14% absolute, and (4) to classify out-of-distribution examples where our fully unsupervised model is competitive with supervised methods.

Mike Wu, Milan Mosse, Chengxu Zhuang et al.

Recent methods for learning unsupervised visual representations, dubbed contrastive learning, optimize the noise-contrastive estimation (NCE) bound on mutual information between two views of an image. NCE uses randomly sampled negative examples to normalize the objective. In this paper, we show that choosing difficult negatives, or those more similar to the current instance, can yield stronger representations. To do this, we introduce a family of mutual information estimators that sample negatives conditionally -- in a "ring" around each positive. We prove that these estimators lower-bound mutual information, with higher bias but lower variance than NCE. Experimentally, we find our approach, applied on top of existing models (IR, CMC, and MoCo) improves accuracy by 2-5% points in each case, measured by linear evaluation on four standard image datasets. Moreover, we find continued benefits when transferring features to a variety of new image distributions from the Meta-Dataset collection and to a variety of downstream tasks such as object detection, instance segmentation, and keypoint detection.

Mike Wu, Chengxu Zhuang, Milan Mosse et al.

In recent years, several unsupervised, "contrastive" learning algorithms in vision have been shown to learn representations that perform remarkably well on transfer tasks. We show that this family of algorithms maximizes a lower bound on the mutual information between two or more "views" of an image where typical views come from a composition of image augmentations. Our bound generalizes the InfoNCE objective to support negative sampling from a restricted region of "difficult" contrasts. We find that the choice of negative samples and views are critical to the success of these algorithms. Reformulating previous learning objectives in terms of mutual information also simplifies and stabilizes them. In practice, our new objectives yield representations that outperform those learned with previous approaches for transfer to classification, bounding box detection, instance segmentation, and keypoint detection. % experiments show that choosing more difficult negative samples results in a stronger representation, outperforming those learned with IR, LA, and CMC in classification, bounding box detection, instance segmentation, and keypoint detection. The mutual information framework provides a unifying comparison of approaches to contrastive learning and uncovers the choices that impact representation learning.

Alex Tamkin, Trisha Singh, Davide Giovanardi et al.

How does language model pretraining help transfer learning? We consider a simple ablation technique for determining the impact of each pretrained layer on transfer task performance. This method, partial reinitialization, involves replacing different layers of a pretrained model with random weights, then finetuning the entire model on the transfer task and observing the change in performance. This technique reveals that in BERT, layers with high probing performance on downstream GLUE tasks are neither necessary nor sufficient for high accuracy on those tasks. Furthermore, the benefit of using pretrained parameters for a layer varies dramatically with finetuning dataset size: parameters that provide tremendous performance improvement when data is plentiful may provide negligible benefits in data-scarce settings. These results reveal the complexity of the transfer learning process, highlighting the limitations of methods that operate on frozen models or single data samples.

Mike Wu, Noah Goodman

As deep neural networks become more adept at traditional tasks, many of the most exciting new challenges concern multimodality---observations that combine diverse types, such as image and text. In this paper, we introduce a family of multimodal deep generative models derived from variational bounds on the evidence (data marginal likelihood). As part of our derivation we find that many previous multimodal variational autoencoders used objectives that do not correctly bound the joint marginal likelihood across modalities. We further generalize our objective to work with several types of deep generative model (VAE, GAN, and flow-based), and allow use of different model types for different modalities. We benchmark our models across many image, label, and text datasets, and find that our multimodal VAEs excel with and without weak supervision. Additional improvements come from use of GAN image models with VAE language models. Finally, we investigate the effect of language on learned image representations through a variety of downstream tasks, such as compositionally, bounding box prediction, and visual relation prediction. We find evidence that these image representations are more abstract and compositional than equivalent representations learned from only visual data.

Jesse Mu, Percy Liang, Noah Goodman

By describing the features and abstractions of our world, language is a crucial tool for human learning and a promising source of supervision for machine learning models. We use language to improve few-shot visual classification in the underexplored scenario where natural language task descriptions are available during training, but unavailable for novel tasks at test time. Existing models for this setting sample new descriptions at test time and use those to classify images. Instead, we propose language-shaped learning (LSL), an end-to-end model that regularizes visual representations to predict language. LSL is conceptually simpler, more data efficient, and outperforms baselines in two challenging few-shot domains.

Ali Malik, Mike Wu, Vrinda Vasavada et al.

Access to high-quality education at scale is limited by the difficulty of providing student feedback on open-ended assignments in structured domains like computer programming, graphics, and short response questions. This problem has proven to be exceptionally difficult: for humans, it requires large amounts of manual work, and for computers, until recently, achieving anything near human-level accuracy has been unattainable. In this paper, we present generative grading: a novel computational approach for providing feedback at scale that is capable of accurately grading student work and providing nuanced, interpretable feedback. Our approach uses generative descriptions of student cognition, written as probabilistic programs, to synthesise millions of labelled example solutions to a problem; we then learn to infer feedback for real student solutions based on this cognitive model. We apply our methods to three settings. In block-based coding, we achieve a 50% improvement upon the previous best results for feedback, achieving super-human accuracy. In two other widely different domains -- graphical tasks and short text answers -- we achieve major improvement over the previous state of the art by about 4x and 1.5x respectively, approaching human accuracy. In a real classroom, we ran an experiment where we used our system to augment human graders, yielding doubled grading accuracy while halving grading time.