Dan Klein

Collin Burns, Haotian Ye, Dan Klein et al. · anthropic

Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output errors that human evaluators can't detect. We propose circumventing this issue by directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. Specifically, we introduce a method for accurately answering yes-no questions given only unlabeled model activations. It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question-answering datasets, it outperforms zero-shot accuracy by 4\% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers. Our results provide an initial step toward discovering what language models know, distinct from what they say, even when we don't have access to explicit ground truth labels.

Jessy Lin, Yuqing Du, Olivia Watkins et al. · deepmind, utoronto

To interact with humans and act in the world, agents need to understand the range of language that people use and relate it to the visual world. While current agents can learn to execute simple language instructions, we aim to build agents that leverage diverse language -- language like "this button turns on the TV" or "I put the bowls away" -- that conveys general knowledge, describes the state of the world, provides interactive feedback, and more. Our key idea is that agents should interpret such diverse language as a signal that helps them predict the future: what they will observe, how the world will behave, and which situations will be rewarded. This perspective unifies language understanding with future prediction as a powerful self-supervised learning objective. We instantiate this in Dynalang, an agent that learns a multimodal world model to predict future text and image representations, and learns to act from imagined model rollouts. While current methods that learn language-conditioned policies degrade in performance with more diverse types of language, we show that Dynalang learns to leverage environment descriptions, game rules, and instructions to excel on tasks ranging from game-playing to navigating photorealistic home scans. Finally, we show that our method enables additional capabilities due to learning a generative model: Dynalang can be pretrained on text-only data, enabling learning from offline datasets, and generate language grounded in an environment.

Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar et al. · berkeley

We present a framework that formulates visual question answering as modular code generation. In contrast to prior work on modular approaches to VQA, our approach requires no additional training and relies on pre-trained language models (LMs), visual models pre-trained on image-caption pairs, and fifty VQA examples used for in-context learning. The generated Python programs invoke and compose the outputs of the visual models using arithmetic and conditional logic. Our approach improves accuracy on the COVR dataset by at least 3% and on the GQA dataset by roughly 2% compared to the few-shot baseline that does not employ code generation.

Jessy Lin, Daniel Fried, Dan Klein et al. · cmu

In classic instruction following, language like "I'd like the JetBlue flight" maps to actions (e.g., selecting that flight). However, language also conveys information about a user's underlying reward function (e.g., a general preference for JetBlue), which can allow a model to carry out desirable actions in new contexts. We present a model that infers rewards from language pragmatically: reasoning about how speakers choose utterances not only to elicit desired actions, but also to reveal information about their preferences. On a new interactive flight-booking task with natural language, our model more accurately infers rewards and predicts optimal actions in unseen environments, in comparison to past work that first maps language to actions (instruction following) and then maps actions to rewards (inverse reinforcement learning).

Yichen Wang, Kevin Yang, Xiaoming Liu et al. · berkeley

Existing LLM-based systems for writing long-form stories or story outlines frequently suffer from unnatural pacing, whether glossing over important events or over-elaborating on insignificant details, resulting in a jarring experience for the reader. We propose a CONCrete Outline ConTrol (CONCOCT) system to improve pacing when automatically generating story outlines. We first train a concreteness evaluator to judge which of two events is more concrete (low-level-detailed). This evaluator can then be used to control pacing in hierarchical outline generation; in this work, we explore a vaguest-first expansion procedure that aims for uniform pacing. We further use the evaluator to filter new outline items based on predicted concreteness. Compared to a baseline hierarchical outline generator, humans judge CONCOCT's pacing to be more consistent over 57% of the time across multiple outline lengths; the gains also translate to downstream stories. All code, data, and models are open-sourced.

Kevin Yang, Dan Klein, Asli Celikyilmaz et al. · berkeley, meta-ai

We propose Reinforcement Learning from Contrastive Distillation (RLCD), a method for aligning language models to follow principles expressed in natural language (e.g., to be more harmless) without using human feedback. RLCD creates preference pairs from two contrasting model outputs, one using a positive prompt designed to encourage following the given principles, and one using a negative prompt designed to encourage violating them. Using two different prompts causes model outputs to be more differentiated on average, resulting in cleaner preference labels in the absence of human annotations. We then use the preference pairs to train a preference model, which is in turn used to improve a base unaligned language model via reinforcement learning. Empirically, RLCD outperforms RLAIF (Bai et al., 2022b) and context distillation (Huang et al., 2022) baselines across three diverse alignment tasks--harmlessness, helpfulness, and story outline generation--and when using both 7B and 30B model scales for simulating preference data.

Evonne Ng, Sanjay Subramanian, Dan Klein et al. · berkeley

We present a framework for generating appropriate facial responses from a listener in dyadic social interactions based on the speaker's words. Given an input transcription of the speaker's words with their timestamps, our approach autoregressively predicts a response of a listener: a sequence of listener facial gestures, quantized using a VQ-VAE. Since gesture is a language component, we propose treating the quantized atomic motion elements as additional language token inputs to a transformer-based large language model. Initializing our transformer with the weights of a language model pre-trained only on text results in significantly higher quality listener responses than training a transformer from scratch. We show that our generated listener motion is fluent and reflective of language semantics through quantitative metrics and a qualitative user study. In our evaluation, we analyze the model's ability to utilize temporal and semantic aspects of spoken text. Project page: https://people.eecs.berkeley.edu/~evonne_ng/projects/text2listen/

Eric Wallace, Nicholas Tomlin, Albert Xu et al. · berkeley

We present the Berkeley Crossword Solver, a state-of-the-art approach for automatically solving crossword puzzles. Our system works by generating answer candidates for each crossword clue using neural question answering models and then combines loopy belief propagation with local search to find full puzzle solutions. Compared to existing approaches, our system improves exact puzzle accuracy from 71% to 82% on crosswords from The New York Times and obtains 99.9% letter accuracy on themeless puzzles. Additionally, in 2021, a hybrid of our system and the existing Dr.Fill system outperformed all human competitors for the first time at the American Crossword Puzzle Tournament. To facilitate research on question answering and crossword solving, we analyze our system's remaining errors and release a dataset of over six million question-answer pairs.

Ruiqi Zhong, Charlie Snell, Dan Klein et al. · microsoft-research

Can non-programmers annotate natural language utterances with complex programs that represent their meaning? We introduce APEL, a framework in which non-programmers select among candidate programs generated by a seed semantic parser (e.g., Codex). Since they cannot understand the candidate programs, we ask them to select indirectly by examining the programs' input-ouput examples. For each utterance, APEL actively searches for a simple input on which the candidate programs tend to produce different outputs. It then asks the non-programmers only to choose the appropriate output, thus allowing us to infer which program is correct and could be used to fine-tune the parser. As a first case study, we recruited human non-programmers to use APEL to re-annotate SPIDER, a text-to-SQL dataset. Our approach achieved the same annotation accuracy as the original expert annotators (75%) and exposed many subtle errors in the original annotations.

Nicholas Tomlin, Andre He, Dan Klein · berkeley

We present a new dataset containing 10K human-annotated games of Go and show how these natural language annotations can be used as a tool for model interpretability. Given a board state and its associated comment, our approach uses linear probing to predict mentions of domain-specific terms (e.g., ko, atari) from the intermediate state representations of game-playing agents like AlphaGo Zero. We find these game concepts are nontrivially encoded in two distinct policy networks, one trained via imitation learning and another trained via reinforcement learning. Furthermore, mentions of domain-specific terms are most easily predicted from the later layers of both models, suggesting that these policy networks encode high-level abstractions similar to those used in the natural language annotations.

Catherine Chen, Zejiang Shen, Dan Klein et al. · berkeley, mit

Recent work has shown that infusing layout features into language models (LMs) improves processing of visually-rich documents such as scientific papers. Layout-infused LMs are often evaluated on documents with familiar layout features (e.g., papers from the same publisher), but in practice models encounter documents with unfamiliar distributions of layout features, such as new combinations of text sizes and styles, or new spatial configurations of textual elements. In this work we test whether layout-infused LMs are robust to layout distribution shifts. As a case study we use the task of scientific document structure recovery, segmenting a scientific paper into its structural categories (e.g., "title", "caption", "reference"). To emulate distribution shifts that occur in practice we re-partition the GROTOAP2 dataset. We find that under layout distribution shifts model performance degrades by up to 20 F1. Simple training strategies, such as increasing training diversity, can reduce this degradation by over 35% relative F1; however, models fail to reach in-distribution performance in any tested out-of-distribution conditions. This work highlights the need to consider layout distribution shifts during model evaluation, and presents a methodology for conducting such evaluations.

Andre He, Nicholas Tomlin, Dan Klein · berkeley

We present a state-of-the-art neural approach to the unsupervised reconstruction of ancient word forms. Previous work in this domain used expectation-maximization to predict simple phonological changes between ancient word forms and their cognates in modern languages. We extend this work with neural models that can capture more complicated phonological and morphological changes. At the same time, we preserve the inductive biases from classical methods by building monotonic alignment constraints into the model and deliberately underfitting during the maximization step. We evaluate our performance on the task of reconstructing Latin from a dataset of cognates across five Romance languages, achieving a notable reduction in edit distance from the target word forms compared to previous methods.

Hao Fang, Anusha Balakrishnan, Harsh Jhamtani et al. · microsoft-research, mit

In a real-world dialogue system, generated text must be truthful and informative while remaining fluent and adhering to a prescribed style. Satisfying these constraints simultaneously is difficult for the two predominant paradigms in language generation: neural language modeling and rule-based generation. We describe a hybrid architecture for dialogue response generation that combines the strengths of both paradigms. The first component of this architecture is a rule-based content selection model defined using a new formal framework called dataflow transduction, which uses declarative rules to transduce a dialogue agent's actions and their results (represented as dataflow graphs) into context-free grammars representing the space of contextually acceptable responses. The second component is a constrained decoding procedure that uses these grammars to constrain the output of a neural language model, which selects fluent utterances. Our experiments show that this system outperforms both rule-based and learned approaches in human evaluations of fluency, relevance, and truthfulness.

Boyi Li, Rodolfo Corona, Karttikeya Mangalam et al. · berkeley

Are multimodal inputs necessary for grammar induction? Recent work has shown that multimodal training inputs can improve grammar induction. However, these improvements are based on comparisons to weak text-only baselines that were trained on relatively little textual data. To determine whether multimodal inputs are needed in regimes with large amounts of textual training data, we design a stronger text-only baseline, which we refer to as LC-PCFG. LC-PCFG is a C-PFCG that incorporates em-beddings from text-only large language models (LLMs). We use a fixed grammar family to directly compare LC-PCFG to various multi-modal grammar induction methods. We compare performance on four benchmark datasets. LC-PCFG provides an up to 17% relative improvement in Corpus-F1 compared to state-of-the-art multimodal grammar induction methods. LC-PCFG is also more computationally efficient, providing an up to 85% reduction in parameter count and 8.8x reduction in training time compared to multimodal approaches. These results suggest that multimodal inputs may not be necessary for grammar induction, and emphasize the importance of strong vision-free baselines for evaluating the benefit of multimodal approaches.

Kevin Yang, Yuandong Tian, Nanyun Peng et al.

We consider the problem of automatically generating longer stories of over two thousand words. Compared to prior work on shorter stories, long-range plot coherence and relevance are more central challenges here. We propose the Recursive Reprompting and Revision framework (Re3) to address these challenges by (a) prompting a general-purpose language model to construct a structured overarching plan, and (b) generating story passages by repeatedly injecting contextual information from both the plan and current story state into a language model prompt. We then revise by (c) reranking different continuations for plot coherence and premise relevance, and finally (d) editing the best continuation for factual consistency. Compared to similar-length stories generated directly from the same base model, human evaluators judged substantially more of Re3's stories as having a coherent overarching plot (by 14% absolute increase), and relevant to the given initial premise (by 20%).

Charlie Snell, Dan Klein, Ruiqi Zhong

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

Ruiqi Zhong, Peter Zhang, Steve Li et al.

Mining large corpora can generate useful discoveries but is time-consuming for humans. We formulate a new task, D5, that automatically discovers differences between two large corpora in a goal-driven way. The task input is a problem comprising a research goal "$\textit{comparing the side effects of drug A and drug B}$" and a corpus pair (two large collections of patients' self-reported reactions after taking each drug). The output is a language description (discovery) of how these corpora differ (patients taking drug A "$\textit{mention feelings of paranoia}$" more often). We build a D5 system, and to quantitatively measure its performance, we 1) contribute a meta-dataset, OpenD5, aggregating 675 open-ended problems ranging across business, social sciences, humanities, machine learning, and health, and 2) propose a set of unified evaluation metrics: validity, relevance, novelty, and significance. With the dataset and the unified metrics, we confirm that language models can use the goals to propose more relevant, novel, and significant candidate discoveries. Finally, our system produces discoveries previously unknown to the authors on a wide range of applications in OpenD5, including temporal and demographic differences in discussion topics, political stances and stereotypes in speech, insights in commercial reviews, and error patterns in NLP models.

Jonathan Pei, Kevin Yang, Dan Klein

We propose Prefix-Adaptive Decoding (PREADD), a flexible method for controlled text generation. Unlike existing methods that use auxiliary expert models to control for attributes, PREADD does not require an external model, instead relying on linearly combining output logits from multiple prompts. Specifically, PREADD contrasts the output logits generated using a raw prompt against those generated using a prefix-prepended prompt, enabling both positive and negative control with respect to any attribute encapsulated by the prefix. We evaluate PREADD on three tasks -- toxic output mitigation, gender bias reduction, and sentiment control -- and find that PREADD outperforms not only prompting baselines, but also an auxiliary-expert control method, by 12% or more in relative gain on our main metrics for each task.

Rodolfo Corona, Shizhan Zhu, Dan Klein et al.

Natural language applied to natural 2D images describes a fundamentally 3D world. We present the Voxel-informed Language Grounder (VLG), a language grounding model that leverages 3D geometric information in the form of voxel maps derived from the visual input using a volumetric reconstruction model. We show that VLG significantly improves grounding accuracy on SNARE, an object reference game task. At the time of writing, VLG holds the top place on the SNARE leaderboard, achieving SOTA results with a 2.0% absolute improvement.

Lakshya A Agrawal, Donghyun Lee, Shangyin Tan et al.

Can a single LLM-based optimization system match specialized tools across fundamentally different domains? We show that when optimization problems are formulated as improving a text artifact evaluated by a scoring function, a single AI-based optimization system-supporting single-task search, multi-task search with cross-problem transfer, and generalization to unseen inputs-achieves state-of-the-art results across six diverse tasks. Our system discovers agent architectures that nearly triple Gemini Flash's ARC-AGI accuracy (32.5% to 89.5%), finds scheduling algorithms that cut cloud costs by 40%, generates CUDA kernels where 87% match or beat PyTorch, and outperforms AlphaEvolve's reported circle packing solution (n=26). Ablations across three domains reveal that actionable side information yields faster convergence and substantially higher final scores than score-only feedback, and that multi-task search outperforms independent optimization given equivalent per-problem budget through cross-task transfer, with benefits scaling with the number of related tasks. Together, we show for the first time that text optimization with LLM-based search is a general-purpose problem-solving paradigm, unifying tasks traditionally requiring domain-specific algorithms under a single framework. We open-source optimize\_anything with support for multiple backends as part of the GEPA project at https://github.com/gepa-ai/gepa .

Chancharik Mitra, Abrar Anwar, Rodolfo Corona et al.

When connecting objects and their language referents in an embodied 3D environment, it is important to note that: (1) an object can be better characterized by leveraging comparative information between itself and other objects, and (2) an object's appearance can vary with camera position. As such, we present the Multi-view Approach to Grounding in Context (MAGiC), which selects an object referent based on language that distinguishes between two similar objects. By pragmatically reasoning over both objects and across multiple views of those objects, MAGiC improves over the state-of-the-art model on the SNARE object reference task with a relative error reduction of 12.9\% (representing an absolute improvement of 2.7\%). Ablation studies show that reasoning jointly over object referent candidates and multiple views of each object both contribute to improved accuracy. Code: https://github.com/rcorona/magic_snare/

Kevin Yang, Dan Klein, Nanyun Peng et al.

We propose the Detailed Outline Control (DOC) framework for improving long-range plot coherence when automatically generating several-thousand-word-long stories. DOC consists of two complementary components: a detailed outliner and a detailed controller. The detailed outliner creates a more detailed, hierarchically structured outline, shifting creative burden from the main drafting procedure to the planning stage. The detailed controller ensures the more detailed outline is still respected during generation by controlling story passages to align with outline details. In human evaluations of automatically generated stories, DOC substantially outperforms a strong Re3 baseline (Yang et al., 2022) on plot coherence (22.5% absolute gain), outline relevance (28.2%), and interestingness (20.7%). Humans also judged DOC to be much more controllable in an interactive generation setting.

Ruiqi Zhong, Heng Wang, Dan Klein et al.

To make sense of massive data, we often fit simplified models and then interpret the parameters; for example, we cluster the text embeddings and then interpret the mean parameters of each cluster. However, these parameters are often high-dimensional and hard to interpret. To make model parameters directly interpretable, we introduce a family of statistical models -- including clustering, time series, and classification models -- parameterized by natural language predicates. For example, a cluster of text about COVID could be parameterized by the predicate "discusses COVID". To learn these statistical models effectively, we develop a model-agnostic algorithm that optimizes continuous relaxations of predicate parameters with gradient descent and discretizes them by prompting language models (LMs). Finally, we apply our framework to a wide range of problems: taxonomizing user chat dialogues, characterizing how they evolve across time, finding categories where one language model is better than the other, clustering math problems based on subareas, and explaining visual features in memorable images. Our framework is highly versatile, applicable to both textual and visual domains, can be easily steered to focus on specific properties (e.g. subareas), and explains sophisticated concepts that classical methods (e.g. n-gram analysis) struggle to produce.

Olivia Huang, Eve Fleisig, Dan Klein

Current practices regarding data collection for natural language processing on Amazon Mechanical Turk (MTurk) often rely on a combination of studies on data quality and heuristics shared among NLP researchers. However, without considering the perspectives of MTurk workers, these approaches are susceptible to issues regarding workers' rights and poor response quality. We conducted a critical literature review and a survey of MTurk workers aimed at addressing open questions regarding best practices for fair payment, worker privacy, data quality, and considering worker incentives. We found that worker preferences are often at odds with received wisdom among NLP researchers. Surveyed workers preferred reliable, reasonable payments over uncertain, very high payments; reported frequently lying on demographic questions; and expressed frustration at having work rejected with no explanation. We also found that workers view some quality control methods, such as requiring minimum response times or Master's qualifications, as biased and largely ineffective. Based on the survey results, we provide recommendations on how future NLP studies may better account for MTurk workers' experiences in order to respect workers' rights and improve data quality.

Charlie Snell, Eric Wallace, Dan Klein et al.

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

Shangyin Tan, Lakshya A Agrawal, Arnav Singhvi et al.

Composing language models (LMs) into multi-step language programs and automatically optimizing their modular prompts is now a mainstream paradigm for building AI systems, but the tradeoffs in this space have only scarcely been studied before. We introduce LangProBe, the first large-scale benchmark for evaluating the architectures and optimization strategies for language programs, with over 2000 combinations of tasks, architectures, optimizers, and choices of LMs. Using LangProBe, we are the first to study the impact of program architectures and optimizers (and their compositions together and with different models) on tradeoffs of quality and cost. We find that optimized language programs offer strong cost--quality Pareto improvement over raw calls to models, but simultaneously demonstrate that human judgment (or empirical decisions) about which compositions to pursue is still necessary for best performance. We will open source the code and evaluation data for LangProBe.

Mert Cemri, Melissa Z. Pan, Shuyi Yang et al. · berkeley

Despite enthusiasm for Multi-Agent LLM Systems (MAS), their performance gains on popular benchmarks are often minimal. This gap highlights a critical need for a principled understanding of why MAS fail. Addressing this question requires systematic identification and analysis of failure patterns. We introduce MAST-Data, a comprehensive dataset of 1600+ annotated traces collected across 7 popular MAS frameworks. MAST-Data is the first multi-agent system dataset to outline the failure dynamics in MAS for guiding the development of better future systems. To enable systematic classification of failures for MAST-Data, we build the first Multi-Agent System Failure Taxonomy (MAST). We develop MAST through rigorous analysis of 150 traces, guided closely by expert human annotators and validated by high inter-annotator agreement (kappa = 0.88). This process identifies 14 unique modes, clustered into 3 categories: (i) system design issues, (ii) inter-agent misalignment, and (iii) task verification. To enable scalable annotation, we develop an LLM-as-a-Judge pipeline with high agreement with human annotations. We leverage MAST and MAST-Data to analyze failure patterns across models (GPT4, Claude 3, Qwen2.5, CodeLlama) and tasks (coding, math, general agent), demonstrating improvement headrooms from better MAS design. Our analysis provides insights revealing that identified failures require more sophisticated solutions, highlighting a clear roadmap for future research. We publicly release our comprehensive dataset (MAST-Data), the MAST, and our LLM annotator to facilitate widespread research and development in MAS.

Noah Ziems, Dilara Soylu, Lakshya A Agrawal et al.

Group Relative Policy Optimization (GRPO) has proven to be an effective tool for post-training language models (LMs). However, AI systems are increasingly expressed as modular programs that mix together multiple LM calls with distinct prompt templates and other tools, and it is not clear how best to leverage GRPO to improve these systems. We begin to address this challenge by defining mmGRPO, a simple multi-module generalization of GRPO that groups LM calls by module across rollouts and handles variable-length and interrupted trajectories. We find that mmGRPO, composed with automatic prompt optimization, improves accuracy by 11% on average across classification, many-hop search, and privacy-preserving delegation tasks against the post-trained LM, and by 5% against prompt optimization on its own. We open-source mmGRPO in DSPy as the dspy.GRPO optimizer.

Sanjay Subramanian, Evonne Ng, Lea Müller et al.

Language is often used to describe physical interaction, yet most 3D human pose estimation methods overlook this rich source of information. We bridge this gap by leveraging large multimodal models (LMMs) as priors for reconstructing contact poses, offering a scalable alternative to traditional methods that rely on human annotations or motion capture data. Our approach extracts contact-relevant descriptors from an LMM and translates them into tractable losses to constrain 3D human pose optimization. Despite its simplicity, our method produces compelling reconstructions for both two-person interactions and self-contact scenarios, accurately capturing the semantics of physical and social interactions. Our results demonstrate that LMMs can serve as powerful tools for contact prediction and pose estimation, offering an alternative to costly manual human annotations or motion capture data. Our code is publicly available at https://prosepose.github.io.

Alexander Wan, Eric Wallace, Sheng Shen et al.

Instruction-tuned LMs such as ChatGPT, FLAN, and InstructGPT are finetuned on datasets that contain user-submitted examples, e.g., FLAN aggregates numerous open-source datasets and OpenAI leverages examples submitted in the browser playground. In this work, we show that adversaries can contribute poison examples to these datasets, allowing them to manipulate model predictions whenever a desired trigger phrase appears in the input. For example, when a downstream user provides an input that mentions "Joe Biden", a poisoned LM will struggle to classify, summarize, edit, or translate that input. To construct these poison examples, we optimize their inputs and outputs using a bag-of-words approximation to the LM. We evaluate our method on open-source instruction-tuned LMs. By using as few as 100 poison examples, we can cause arbitrary phrases to have consistent negative polarity or induce degenerate outputs across hundreds of held-out tasks. Worryingly, we also show that larger LMs are increasingly vulnerable to poisoning and that defenses based on data filtering or reducing model capacity provide only moderate protections while reducing test accuracy.

Kevin Yang, Tianjun Zhang, Chris Cummins et al.

Path planning, the problem of efficiently discovering high-reward trajectories, often requires optimizing a high-dimensional and multimodal reward function. Popular approaches like CEM and CMA-ES greedily focus on promising regions of the search space and may get trapped in local maxima. DOO and VOOT balance exploration and exploitation, but use space partitioning strategies independent of the reward function to be optimized. Recently, LaMCTS empirically learns to partition the search space in a reward-sensitive manner for black-box optimization. In this paper, we develop a novel formal regret analysis for when and why such an adaptive region partitioning scheme works. We also propose a new path planning method LaP3 which improves the function value estimation within each sub-region, and uses a latent representation of the search space. Empirically, LaP3 outperforms existing path planning methods in 2D navigation tasks, especially in the presence of difficult-to-escape local optima, and shows benefits when plugged into the planning components of model-based RL such as PETS. These gains transfer to highly multimodal real-world tasks, where we outperform strong baselines in compiler phase ordering by up to 39% on average across 9 tasks, and in molecular design by up to 0.4 on properties on a 0-1 scale. Code is available at https://github.com/yangkevin2/neurips2021-lap3.

Alexander Wan, Eric Wallace, Dan Klein · berkeley

Retrieval-augmented language models are being increasingly tasked with subjective, contentious, and conflicting queries such as "is aspartame linked to cancer". To resolve these ambiguous queries, one must search through a large range of websites and consider "which, if any, of this evidence do I find convincing?". In this work, we study how LLMs answer this question. In particular, we construct ConflictingQA, a dataset that pairs controversial queries with a series of real-world evidence documents that contain different facts (e.g., quantitative results), argument styles (e.g., appeals to authority), and answers (Yes or No). We use this dataset to perform sensitivity and counterfactual analyses to explore which text features most affect LLM predictions. Overall, we find that current models rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone. Taken together, these results highlight the importance of RAG corpus quality (e.g., the need to filter misinformation), and possibly even a shift in how LLMs are trained to better align with human judgements.

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu et al.

Large language models (LLMs) are increasingly adapted to downstream tasks via reinforcement learning (RL) methods like Group Relative Policy Optimization (GRPO), which often require thousands of rollouts to learn new tasks. We argue that the interpretable nature of language can often provide a much richer learning medium for LLMs, compared with policy gradients derived from sparse, scalar rewards. To test this, we introduce GEPA (Genetic-Pareto), a prompt optimizer that thoroughly incorporates natural language reflection to learn high-level rules from trial and error. Given any AI system containing one or more LLM prompts, GEPA samples system-level trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts. As a result of GEPA's design, it can often turn even just a few rollouts into a large quality gain. Across four tasks, GEPA outperforms GRPO by 10% on average and by up to 20%, while using up to 35x fewer rollouts. GEPA also outperforms the leading prompt optimizer, MIPROv2, by over 10% across two LLMs, and demonstrates promising results as an inference-time search strategy for code optimization.

Daniel Nahmias, Gal Engelberg, Dan Klein et al.

Spear-phishing attacks present a significant security challenge, with large language models (LLMs) escalating the threat by generating convincing emails and facilitating target reconnaissance. To address this, we propose a detection approach based on a novel document vectorization method that utilizes an ensemble of LLMs to create representation vectors. By prompting LLMs to reason and respond to human-crafted questions, we quantify the presence of common persuasion principles in the email's content, producing prompted contextual document vectors for a downstream supervised machine learning model. We evaluate our method using a unique dataset generated by a proprietary system that automates target reconnaissance and spear-phishing email creation. Our method achieves a 91\% F1 score in identifying LLM-generated spear-phishing emails, with the training set comprising only traditional phishing and benign emails. Key contributions include a novel document vectorization method utilizing LLM reasoning, a publicly available dataset of high-quality spear-phishing emails, and the demonstrated effectiveness of our method in detecting such emails. This methodology can be utilized for various document classification tasks, particularly in adversarial problem domains.

Ritwik Gupta, Rodolfo Corona, Jiaxin Ge et al.

Large language models (LLMs) exhibit the ability to generalize given few-shot examples in their input prompt, an emergent capability known as in-context learning (ICL). We investigate whether LLMs use ICL to perform structured reasoning in ways that are consistent with a Bayesian framework or rely on pattern matching. Using a controlled setting of biased coin flips, we find that: (1) LLMs often possess biased priors, causing initial divergence in zero-shot settings, (2) in-context evidence outweighs explicit bias instructions, (3) LLMs broadly follow Bayesian posterior updates, with deviations primarily due to miscalibrated priors rather than flawed updates, and (4) attention magnitude has negligible effect on Bayesian inference. With sufficient demonstrations of biased coin flips via ICL, LLMs update their priors in a Bayesian manner.

Yizhou Chi, Kevin Yang, Dan Klein

We present THOUGHTSCULPT, a general reasoning and search method for tasks with outputs that can be decomposed into components. THOUGHTSCULPT explores a search tree of potential solutions using Monte Carlo Tree Search (MCTS), building solutions one action at a time and evaluating according to any domain-specific heuristic, which in practice is often simply an LLM evaluator. Critically, our action space includes revision actions: THOUGHTSCULPT may choose to revise part of its previous output rather than continuing to build the rest of its output. Empirically, THOUGHTSCULPT outperforms state-of-the-art reasoning methods across three challenging tasks: Story Outline Improvement (up to +30% interestingness), Mini-Crosswords Solving (up to +16% word success rate), and Constrained Generation (up to +10% concept coverage).

Yizhou Chi, Jessy Lin, Kevin Lin et al.

Users often make ambiguous requests that require clarification. We study the problem of asking clarification questions in an information retrieval setting, where systems often face ambiguous search queries and it is challenging to turn the uncertainty in the retrieval model into a natural language question. We present CLARINET, a system that asks informative clarification questions by choosing questions whose answers would maximize certainty in the correct candidate. Our approach works by augmenting a large language model (LLM) to condition on a retrieval distribution, finetuning end-to-end to generate the question that would have maximized the rank of the true candidate at each turn. When evaluated on a real-world retrieval dataset of users searching for books, our system outperforms traditional heuristics such as information gain on retrieval success by 17% and vanilla-prompted LLMs by 39% relative.

Vivek Verma, David Huang, William Chen et al.

We present gg-bench, a collection of game environments designed to evaluate general reasoning capabilities in language models. Unlike most static benchmarks, gg-bench is a data generating process where new evaluation instances can be generated at will. In particular, gg-bench is synthetically generated by (1) using a large language model (LLM) to generate natural language descriptions of novel games, (2) using the LLM to implement each game in code as a Gym environment, and (3) training reinforcement learning (RL) agents via self-play on the generated games. We evaluate language models by their winrate against these RL agents by prompting models with the game description, current board state, and a list of valid moves, after which models output the moves they wish to take. gg-bench is challenging: state-of-the-art LLMs such as GPT-4o and Claude 3.7 Sonnet achieve winrates of 7-9% on gg-bench using in-context learning, while reasoning models such as o1, o3-mini and DeepSeek-R1 achieve average winrates of 31-36%. We release the generated games, data generation process, and evaluation code in order to support future modeling work and expansion of our benchmark.

Eve Fleisig, Matthias Orlikowski, Philipp Cimiano et al.

For machine learning datasets to accurately represent diverse opinions in a population, they must preserve variation in data labels while filtering out spam or low-quality responses. How can we balance annotator reliability and representation? We empirically evaluate how a range of heuristics for annotator filtering affect the preservation of variation on subjective tasks. We find that these methods, designed for contexts in which variation from a single ground-truth label is considered noise, often remove annotators who disagree instead of spam annotators, introducing suboptimal tradeoffs between accuracy and label diversity. We find that conservative settings for annotator removal (<5%) are best, after which all tested methods increase the mean absolute error from the true average label. We analyze performance on synthetic spam to observe that these methods often assume spam annotators are more random than real spammers tend to be: most spammers are distributionally indistinguishable from real annotators, and the minority that are distinguishable tend to give relatively fixed answers, not random ones. Thus, tasks requiring the preservation of variation reverse the intuition of existing spam filtering methods: spammers tend to be less random than non-spammers, so metrics that assume variation is spam fare worse. These results highlight the need for spam removal methods that account for label diversity.

Austen Liao, Nicholas Tomlin, Dan Klein

Game-playing agents like AlphaGo have achieved superhuman performance through self-play, which is theoretically guaranteed to yield optimal policies in competitive games. However, most language tasks are partially or fully cooperative, so it is an open question whether techniques like self-play can effectively be used to improve language models. We empirically investigate this question in a negotiation game setting known as Deal or No Deal (DoND). Crucially, the objective in DoND can be modified to produce a fully cooperative game, a strictly competitive one, or anything in between. We finetune language models in self-play over multiple rounds of filtered behavior cloning in DoND for each of these objectives and evaluate them in self-play and in collaboration with humans. We find that language models improve substantially in self-play, achieving 14-17x higher scores in task reward after finetuning. Further, the trained models generalize to both cooperation and competition with humans, scoring 2.5-6x higher than base models. We view these results as an early promising sign for language model self-play in cooperative settings, despite a lack of theoretical guarantees.

Eve Fleisig, Genevieve Smith, Madeline Bossi et al.

We present a large-scale study of linguistic bias exhibited by ChatGPT covering ten dialects of English (Standard American English, Standard British English, and eight widely spoken non-"standard" varieties from around the world). We prompted GPT-3.5 Turbo and GPT-4 with text by native speakers of each variety and analyzed the responses via detailed linguistic feature annotation and native speaker evaluation. We find that the models default to "standard" varieties of English; based on evaluation by native speakers, we also find that model responses to non-"standard" varieties consistently exhibit a range of issues: stereotyping (19% worse than for "standard" varieties), demeaning content (25% worse), lack of comprehension (9% worse), and condescending responses (15% worse). We also find that if these models are asked to imitate the writing style of prompts in non-"standard" varieties, they produce text that exhibits lower comprehension of the input and is especially prone to stereotyping. GPT-4 improves on GPT-3.5 in terms of comprehension, warmth, and friendliness, but also exhibits a marked increase in stereotyping (+18%). The results indicate that GPT-3.5 Turbo and GPT-4 can perpetuate linguistic discrimination toward speakers of non-"standard" varieties.

Kayo Yin, Terry Regier, Dan Klein

Communicative efficiency is a key topic in linguistics and cognitive psychology, with many studies demonstrating how the pressure to communicate with minimal effort guides the form of natural language. However, this phenomenon is rarely explored in signed languages. This paper shows how handshapes in American Sign Language (ASL) reflect these efficiency pressures and provides new evidence of communicative efficiency in the visual-gestural modality. We focus on hand configurations in native ASL signs and signs borrowed from English to compare efficiency pressures from both ASL and English usage. First, we develop new methodologies to quantify the articulatory effort needed to produce handshapes and the perceptual effort required to recognize them. Then, we analyze correlations between communicative effort and usage statistics in ASL or English. Our findings reveal that frequent ASL handshapes are easier to produce and that pressures for communicative efficiency mostly come from ASL usage, rather than from English lexical borrowing.

Eve Fleisig, Su Lin Blodgett, Dan Klein et al.

Longstanding data labeling practices in machine learning involve collecting and aggregating labels from multiple annotators. But what should we do when annotators disagree? Though annotator disagreement has long been seen as a problem to minimize, new perspectivist approaches challenge this assumption by treating disagreement as a valuable source of information. In this position paper, we examine practices and assumptions surrounding the causes of disagreement--some challenged by perspectivist approaches, and some that remain to be addressed--as well as practical and normative challenges for work operating under these assumptions. We conclude with recommendations for the data labeling pipeline and avenues for future research engaging with subjectivity and disagreement.

Kevin Lin, Kyle Lo, Joseph E. Gonzalez et al.

When re-finding items, users who forget or are uncertain about identifying details often rely on creative strategies for expressing their information needs -- complex queries that describe content elements (e.g., book characters or events), information beyond the document text (e.g., descriptions of book covers), or personal context (e.g., when they read a book). This retrieval setting, called tip of the tongue (TOT), is especially challenging for models heavily reliant on lexical and semantic overlap between query and document text. In this work, we introduce a simple yet effective framework for handling such complex queries by decomposing the query into individual clues, routing those as sub-queries to specialized retrievers, and ensembling the results. This approach allows us to take advantage of off-the-shelf retrievers (e.g., CLIP for retrieving images of book covers) or incorporate retriever-specific logic (e.g., date constraints). We show that our framework incorportating query decompositions into retrievers can improve gold book recall up to 7% relative again for Recall@5 on a new collection of 14,441 real-world query-book pairs from an online community for resolving TOT inquiries.

Vivek Verma, Eve Fleisig, Nicholas Tomlin et al.

We introduce Ghostbuster, a state-of-the-art system for detecting AI-generated text. Our method works by passing documents through a series of weaker language models, running a structured search over possible combinations of their features, and then training a classifier on the selected features to predict whether documents are AI-generated. Crucially, Ghostbuster does not require access to token probabilities from the target model, making it useful for detecting text generated by black-box models or unknown model versions. In conjunction with our model, we release three new datasets of human- and AI-generated text as detection benchmarks in the domains of student essays, creative writing, and news articles. We compare Ghostbuster to a variety of existing detectors, including DetectGPT and GPTZero, as well as a new RoBERTa baseline. Ghostbuster achieves 99.0 F1 when evaluated across domains, which is 5.9 F1 higher than the best preexisting model. It also outperforms all previous approaches in generalization across writing domains (+7.5 F1), prompting strategies (+2.1 F1), and language models (+4.4 F1). We also analyze the robustness of our system to a variety of perturbations and paraphrasing attacks and evaluate its performance on documents written by non-native English speakers.

Vyoma Raman, Eve Fleisig, Dan Klein

The impact of AI models on marginalized communities has traditionally been measured by identifying performance differences between specified demographic subgroups. Though this approach aims to center vulnerable groups, it risks obscuring patterns of harm faced by intersectional subgroups or shared across multiple groups. To address this, we draw on theories of marginalization from disability studies and related disciplines, which state that people farther from the norm face greater adversity, to consider the "margins" in the domain of toxicity detection. We operationalize the "margins" of a dataset by employing outlier detection to identify text about people with demographic attributes distant from the "norm". We find that model performance is consistently worse for demographic outliers, with mean squared error (MSE) between outliers and non-outliers up to 70.4% worse across toxicity types. It is also worse for text outliers, with a MSE up to 68.4% higher for outliers than non-outliers. We also find text and demographic outliers to be particularly susceptible to errors in the classification of severe toxicity and identity attacks. Compared to analysis of disparities using traditional demographic breakdowns, we find that our outlier analysis frequently surfaces greater harms faced by a larger, more intersectional group, which suggests that outlier analysis is particularly beneficial for identifying harms against those groups.

Vivek Verma, Nicholas Tomlin, Dan Klein

The uniform information density (UID) hypothesis states that humans tend to distribute information roughly evenly across an utterance or discourse. Early evidence in support of the UID hypothesis came from Genzel & Charniak (2002), which proposed an entropy rate constancy principle based on the probability of English text under n-gram language models. We re-evaluate the claims of Genzel & Charniak (2002) with neural language models, failing to find clear evidence in support of entropy rate constancy. We conduct a range of experiments across datasets, model sizes, and languages and discuss implications for the uniform information density hypothesis and linguistic theories of efficient communication more broadly.

Eve Fleisig, Rediet Abebe, Dan Klein

Though majority vote among annotators is typically used for ground truth labels in natural language processing, annotator disagreement in tasks such as hate speech detection may reflect differences in opinion across groups, not noise. Thus, a crucial problem in hate speech detection is determining whether a statement is offensive to the demographic group that it targets, when that group may constitute a small fraction of the annotator pool. We construct a model that predicts individual annotator ratings on potentially offensive text and combines this information with the predicted target group of the text to model the opinions of target group members. We show gains across a range of metrics, including raising performance over the baseline by 22% at predicting individual annotators' ratings and by 33% at predicting variance among annotators, which provides a metric for model uncertainty downstream. We find that annotator ratings can be predicted using their demographic information and opinions on online content, without the need to track identifying annotator IDs that link each annotator to their ratings. We also find that use of non-invasive survey questions on annotators' online experiences helps to maximize privacy and minimize unnecessary collection of demographic information when predicting annotators' opinions.

Ruiqi Zhong, Charlie Snell, Dan Klein et al.

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

Daniel Fried, Justin T. Chiu, Dan Klein

We present a grounded neural dialogue model that successfully collaborates with people in a partially-observable reference game. We focus on a setting where two agents each observe an overlapping part of a world context and need to identify and agree on some object they share. Therefore, the agents should pool their information and communicate pragmatically to solve the task. Our dialogue agent accurately grounds referents from the partner's utterances using a structured reference resolver, conditions on these referents using a recurrent memory, and uses a pragmatic generation procedure to ensure the partner can resolve the references the agent produces. We evaluate on the OneCommon spatial grounding dialogue task (Udagawa and Aizawa 2019), involving a number of dots arranged on a board with continuously varying positions, sizes, and shades. Our agent substantially outperforms the previous state of the art for the task, obtaining a 20% relative improvement in successful task completion in self-play evaluations and a 50% relative improvement in success in human evaluations.