Alane Suhr

Melanie Sclar, Yejin Choi, Yulia Tsvetkov et al. · berkeley, uw

As large language models (LLMs) are adopted as a fundamental component of language technologies, it is crucial to accurately characterize their performance. Because choices in prompt design can strongly influence model behavior, this design process is critical in effectively using any modern pre-trained generative language model. In this work, we focus on LLM sensitivity to a quintessential class of meaning-preserving design choices: prompt formatting. We find that several widely used open-source LLMs are extremely sensitive to subtle changes in prompt formatting in few-shot settings, with performance differences of up to 76 accuracy points when evaluated using LLaMA-2-13B. Sensitivity remains even when increasing model size, the number of few-shot examples, or performing instruction tuning. Our analysis suggests that work evaluating LLMs with prompting-based methods would benefit from reporting a range of performance across plausible prompt formats, instead of the currently-standard practice of reporting performance on a single format. We also show that format performance only weakly correlates between models, which puts into question the methodological validity of comparing models with an arbitrarily chosen, fixed prompt format. To facilitate systematic analysis we propose FormatSpread, an algorithm that rapidly evaluates a sampled set of plausible prompt formats for a given task, and reports the interval of expected performance without accessing model weights. Furthermore, we present a suite of analyses that characterize the nature of this sensitivity, including exploring the influence of particular atomic perturbations and the internal representation of particular formats.

Yanai Elazar, Akshita Bhagia, Ian Magnusson et al. · allen-ai, berkeley

Large text corpora are the backbone of language models. However, we have a limited understanding of the content of these corpora, including general statistics, quality, social factors, and inclusion of evaluation data (contamination). In this work, we propose What's In My Big Data? (WIMBD), a platform and a set of sixteen analyses that allow us to reveal and compare the contents of large text corpora. WIMBD builds on two basic capabilities -- count and search -- at scale, which allows us to analyze more than 35 terabytes on a standard compute node. We apply WIMBD to ten different corpora used to train popular language models, including C4, The Pile, and RedPajama. Our analysis uncovers several surprising and previously undocumented findings about these corpora, including the high prevalence of duplicate, synthetic, and low-quality content, personally identifiable information, toxic language, and benchmark contamination. For instance, we find that about 50% of the documents in RedPajama and LAION-2B-en are duplicates. In addition, several datasets used for benchmarking models trained on such corpora are contaminated with respect to important benchmarks, including the Winograd Schema Challenge and parts of GLUE and SuperGLUE. We open-source WIMBD's code and artifacts to provide a standard set of evaluations for new text-based corpora and to encourage more analyses and transparency around them.

Zeqiu Wu, Yushi Hu, Weijia Shi et al. · allen-ai, berkeley

Language models (LMs) often exhibit undesirable text generation behaviors, including generating false, toxic, or irrelevant outputs. Reinforcement learning from human feedback (RLHF) - where human preference judgments on LM outputs are transformed into a learning signal - has recently shown promise in addressing these issues. However, such holistic feedback conveys limited information on long text outputs; it does not indicate which aspects of the outputs influenced user preference; e.g., which parts contain what type(s) of errors. In this paper, we use fine-grained human feedback (e.g., which sentence is false, which sub-sentence is irrelevant) as an explicit training signal. We introduce Fine-Grained RLHF, a framework that enables training and learning from reward functions that are fine-grained in two respects: (1) density, providing a reward after every segment (e.g., a sentence) is generated; and (2) incorporating multiple reward models associated with different feedback types (e.g., factual incorrectness, irrelevance, and information incompleteness). We conduct experiments on detoxification and long-form question answering to illustrate how learning with such reward functions leads to improved performance, supported by both automatic and human evaluation. Additionally, we show that LM behaviors can be customized using different combinations of fine-grained reward models. We release all data, collected human feedback, and codes at https://FineGrainedRLHF.github.io.

Alisa Liu, Zhaofeng Wu, Julian Michael et al. · allen-ai, berkeley

Ambiguity is an intrinsic feature of natural language. Managing ambiguity is a key part of human language understanding, allowing us to anticipate misunderstanding as communicators and revise our interpretations as listeners. As language models (LMs) are increasingly employed as dialogue interfaces and writing aids, handling ambiguous language is critical to their success. We characterize ambiguity in a sentence by its effect on entailment relations with another sentence, and collect AmbiEnt, a linguist-annotated benchmark of 1,645 examples with diverse kinds of ambiguity. We design a suite of tests based on AmbiEnt, presenting the first evaluation of pretrained LMs to recognize ambiguity and disentangle possible meanings. We find that the task remains extremely challenging, including for GPT-4, whose generated disambiguations are considered correct only 32% of the time in human evaluation, compared to 90% for disambiguations in our dataset. Finally, to illustrate the value of ambiguity-sensitive tools, we show that a multilabel NLI model can flag political claims in the wild that are misleading due to ambiguity. We encourage the field to rediscover the importance of ambiguity for NLP.

Kolby Nottingham, Prithviraj Ammanabrolu, Alane Suhr et al. · berkeley

Reinforcement learning (RL) agents typically learn tabula rasa, without prior knowledge of the world. However, if initialized with knowledge of high-level subgoals and transitions between subgoals, RL agents could utilize this Abstract World Model (AWM) for planning and exploration. We propose using few-shot large language models (LLMs) to hypothesize an AWM, that will be verified through world experience, to improve sample efficiency of RL agents. Our DECKARD agent applies LLM-guided exploration to item crafting in Minecraft in two phases: (1) the Dream phase where the agent uses an LLM to decompose a task into a sequence of subgoals, the hypothesized AWM; and (2) the Wake phase where the agent learns a modular policy for each subgoal and verifies or corrects the hypothesized AWM. Our method of hypothesizing an AWM with LLMs and then verifying the AWM based on agent experience not only increases sample efficiency over contemporary methods by an order of magnitude but is also robust to and corrects errors in the LLM, successfully blending noisy internet-scale information from LLMs with knowledge grounded in environment dynamics.

Melanie Sclar, Sachin Kumar, Peter West et al. · allen-ai, berkeley

Theory of Mind (ToM)$\unicode{x2014}$the ability to reason about the mental states of other people$\unicode{x2014}$is a key element of our social intelligence. Yet, despite their ever more impressive performance, large-scale neural language models still lack basic theory of mind capabilities out-of-the-box. We posit that simply scaling up models will not imbue them with theory of mind due to the inherently symbolic and implicit nature of the phenomenon, and instead investigate an alternative: can we design a decoding-time algorithm that enhances theory of mind of off-the-shelf neural language models without explicit supervision? We present SymbolicToM, a plug-and-play approach to reason about the belief states of multiple characters in reading comprehension tasks via explicit symbolic representation. More concretely, our approach tracks each entity's beliefs, their estimation of other entities' beliefs, and higher-order levels of reasoning, all through graphical representations, allowing for more precise and interpretable reasoning than previous approaches. Empirical results on the well-known ToMi benchmark (Le et al., 2019) demonstrate that SymbolicToM dramatically enhances off-the-shelf neural networks' theory of mind in a zero-shot setting while showing robust out-of-distribution performance compared to supervised baselines. Our work also reveals spurious patterns in existing theory of mind benchmarks, emphasizing the importance of out-of-distribution evaluation and methods that do not overfit a particular dataset.

Anya Ji, Noriyuki Kojima, Noah Rush et al. · berkeley

We introduce KiloGram, a resource for studying abstract visual reasoning in humans and machines. Drawing on the history of tangram puzzles as stimuli in cognitive science, we build a richly annotated dataset that, with >1k distinct stimuli, is orders of magnitude larger and more diverse than prior resources. It is both visually and linguistically richer, moving beyond whole shape descriptions to include segmentation maps and part labels. We use this resource to evaluate the abstract visual reasoning capacities of recent multi-modal models. We observe that pre-trained weights demonstrate limited abstract reasoning, which dramatically improves with fine-tuning. We also observe that explicitly describing parts aids abstract reasoning for both humans and models, especially when jointly encoding the linguistic and visual inputs. KiloGram is available at https://lil.nlp.cornell.edu/kilogram .

Alane Suhr, Yoav Artzi · berkeley

We propose and deploy an approach to continually train an instruction-following agent from feedback provided by users during collaborative interactions. During interaction, human users instruct an agent using natural language, and provide realtime binary feedback as they observe the agent following their instructions. We design a contextual bandit learning approach, converting user feedback to immediate reward. We evaluate through thousands of human-agent interactions, demonstrating 15.4% absolute improvement in instruction execution accuracy over time. We also show our approach is robust to several design variations, and that the feedback signal is roughly equivalent to the learning signal of supervised demonstration data.

Wenting Zhao, Justin T Chiu, Jena D. Hwang et al. · allen-ai, berkeley

Language technologies that accurately model the dynamics of events must perform commonsense reasoning. Existing work evaluating commonsense reasoning focuses on making inferences about common, everyday situations. To instead investigate the ability to model unusual, unexpected, and unlikely situations, we explore the task of uncommonsense abductive reasoning. Given a piece of context with an unexpected outcome, this task requires reasoning abductively to generate an explanation that makes the unexpected outcome more likely in the context. To this end, we curate and release a new English language corpus called UNcommonsense. We characterize the performance differences between human explainers and the best-performing large language models, finding that model-enhanced human-written explanations achieve the highest quality by trading off between specificity and diversity. Finally, we experiment with several imitation learning algorithms to train open and accessible language models on this task. When compared with the vanilla supervised fine-tuning approach, these methods consistently reduce lose rates on both common and uncommonsense abductive reasoning judged by human evaluators.

Harman Singh, Xiuyu Li, Kusha Sareen et al. · berkeley

Test-time scaling for complex reasoning tasks shows that leveraging inference-time compute, by methods such as independently sampling and aggregating multiple solutions, results in significantly better task outcomes. However, a critical bottleneck is verification: sampling is only effective if correct solutions can be reliably identified among candidates. While existing approaches typically evaluate candidates independently via scalar scoring, we demonstrate that models are substantially stronger at pairwise self-verification. Leveraging this insight, we introduce $V_1$, a framework that unifies generation and verification through efficient pairwise ranking. $V_1$ comprises two components: $V_1$-Infer, an uncertainty-guided algorithm using a tournament-based ranking that dynamically allocates self-verification compute to candidate pairs whose relative correctness is most uncertain; and $V_1$-PairRL, an RL framework that jointly trains a single model as both generator and pairwise self-verifier, ensuring the verifier adapts to the generator's evolving distribution. On code generation (LiveCodeBench, CodeContests, SWE-Bench) and math reasoning (AIME, HMMT) benchmarks, $V_1$-Infer improves Pass@1 by up to $10%$ over pointwise verification and outperforms recent test-time scaling methods while being significantly more efficient. Furthermore, $V_1$-PairRL achieves $7$--$9%$ test-time scaling gains over standard RL and pointwise joint training, and improves base Pass@1 by up to 8.7% over standard RL in a code-generation setting.

Haiwen Feng, Long Lian, Lisa Dunlap et al.

A key challenge in evaluating VLMs is testing models' ability to analyze visual content independently from their textual priors. Recent benchmarks such as BLINK probe visual perception through visual prompting, where questions about visual content are paired with coordinates to which the question refers, with the coordinates explicitly marked in the image itself. While these benchmarks are an important part of VLM evaluation, we find that existing models are surprisingly fragile to seemingly irrelevant details of visual prompting: simply changing a visual marker from red to blue can completely change rankings among models on a leaderboard. By evaluating nine commonly-used open- and closed-source VLMs on two visually prompted tasks, we demonstrate how details in benchmark setup, including visual marker design and dataset size, have a significant influence on model performance and leaderboard rankings. These effects can even be exploited to lift weaker models above stronger ones; for instance, slightly increasing the size of the visual marker results in open-source InternVL3-8B ranking alongside or better than much larger proprietary models like Gemini 2.5 Pro. We further show that low-level inference choices that are often ignored in benchmarking, such as JPEG compression levels in API calls, can also cause model lineup changes. These details have substantially larger impacts on visually prompted benchmarks than on conventional semantic VLM evaluations. To mitigate this instability, we curate existing datasets to create VPBench, a larger visually prompted benchmark with 16 visual marker variants. We open-source VPBench and our analysis framework at: https://lisadunlap.github.io/vpbench/.

Anya Ji, George Ma, Téa Wright et al.

Recent progress in generative models has significantly advanced image editing capabilities, yet precise and intuitive user control remains difficult. Specifically, users often struggle to communicate both exact spatial layouts and specific semantic details simultaneously. While natural language instructions effectively convey high-level semantics like texture and color, they lack spatial specificity. Conversely, freehand scribbles provide rough spatial boundaries but cannot express detailed visual attributes. Consequently, achieving precise control requires combining both modalities. However, existing models struggle to jointly interpret abstract scribbles alongside text due to a lack of specialized training data. In this work, we introduce ScribbleEdit, a large-scale synthetic dataset designed to bridge this gap by combining natural language instructions with freehand scribble inputs for more accurate, controllable edits. We construct this dataset through a synthetic pipeline that automatically generates source-target image pairs via inpainting, which are then paired with human-drawn scribbles and VLM-generated text instructions. Using ScribbleEdit, we evaluate and finetune both diffusion-based and autoregressive unified multimodal image editing models. Our experiments reveal that while off-the-shelf models struggle with abstract scribble inputs, finetuning on our synthetic dataset significantly improves their ability to generate spatially aligned and semantically consistent edits.

Zineng Tang, Long Lian, Seun Eisape et al.

Despite the recent success of image-text contrastive models like CLIP and SigLIP, these models often struggle with vision-centric tasks that demand high-fidelity image understanding, such as counting, depth estimation, and fine-grained object recognition. These models, by performing language alignment, tend to prioritize high-level semantics over visual understanding, weakening their image understanding. On the other hand, vision-focused models are great at processing visual information but struggle to understand language, limiting their flexibility for language-driven tasks. In this work, we introduce TULIP, an open-source, drop-in replacement for existing CLIP-like models. Our method leverages generative data augmentation, enhanced image-image and text-text contrastive learning, and image/text reconstruction regularization to learn fine-grained visual features while preserving global semantic alignment. Our approach, scaling to over 1B parameters, outperforms existing state-of-the-art (SOTA) models across multiple benchmarks, establishing a new SOTA zero-shot performance on ImageNet-1K, delivering up to a $2\times$ enhancement over SigLIP on RxRx1 in linear probing for few-shot classification, and improving vision-language models, achieving over $3\times$ higher scores than SigLIP on MMVP. Our code/checkpoints are available at https://tulip-berkeley.github.io

Nicholas Tomlin, Naitian Zhou, Eve Fleisig et al.

Cooperative video games, where multiple participants must coordinate by communicating and reasoning under uncertainty in complex environments, yield a rich source of language data. We collect the Portal Dialogue Corpus: a corpus of 11.5 hours of spoken human dialogue in the co-op mode of the popular Portal 2 virtual puzzle game, comprising 24.5K total utterances. We analyze player language and behavior, identifying a number of linguistic phenomena that rarely appear in most existing chitchat or task-oriented dialogue corpora, including complex spatial reference, clarification and repair, and ad-hoc convention formation. To support future analyses of language use in complex, situated, collaborative problem-solving scenarios, we publicly release the corpus, which comprises player videos, audio, transcripts, game state data, and both manual and automatic annotations of language data.

Yuexiang Zhai, Hao Bai, Zipeng Lin et al. · berkeley

Large vision-language models (VLMs) fine-tuned on specialized visual instruction-following data have exhibited impressive language reasoning capabilities across various scenarios. However, this fine-tuning paradigm may not be able to efficiently learn optimal decision-making agents in multi-step goal-directed tasks from interactive environments. To address this challenge, we propose an algorithmic framework that fine-tunes VLMs with reinforcement learning (RL). Specifically, our framework provides a task description and then prompts the VLM to generate chain-of-thought (CoT) reasoning, enabling the VLM to efficiently explore intermediate reasoning steps that lead to the final text-based action. Next, the open-ended text output is parsed into an executable action to interact with the environment to obtain goal-directed task rewards. Finally, our framework uses these task rewards to fine-tune the entire VLM with RL. Empirically, we demonstrate that our proposed framework enhances the decision-making capabilities of VLM agents across various tasks, enabling 7b models to outperform commercial models such as GPT4-V or Gemini. Furthermore, we find that CoT reasoning is a crucial component for performance improvement, as removing the CoT reasoning results in a significant decrease in the overall performance of our method.

Jiayi Pan, Xingyao Wang, Graham Neubig et al. · berkeley

We present SWE-Gym, the first environment for training real-world software engineering (SWE) agents. SWE-Gym contains 2,438 real-world Python task instances, each comprising a codebase with an executable runtime environment, unit tests, and a task specified in natural language. We use SWE-Gym to train language model based SWE agents, achieving up to 19% absolute gains in resolve rate on the popular SWE-Bench Verified and Lite test sets. We also experiment with inference-time scaling through verifiers trained on agent trajectories sampled from SWE-Gym. When combined with our fine-tuned SWE agents, we achieve 32.0% and 26.0% on SWE-Bench Verified and Lite, respectively, reflecting a new state-of-the-art for open-weight SWE agents. To facilitate further research, we publicly release SWE-Gym, models, and agent trajectories.

Jiayi Pan, Yichi Zhang, Nicholas Tomlin et al. · berkeley

We show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve around 75% relative improvement in device control settings.

Aly Lidayan, Jakob Bjorner, Satvik Golechha et al.

As the length of sequential decision-making tasks increases, it becomes computationally impractical to keep full interaction histories in context. We introduce a general framework for LLM agents to maintain concise contexts through multi-step interaction: Acting through Belief Bottlenecks Expressed in Language (ABBEL), and methods to further improve ABBEL agents with RL post-training. ABBEL replaces long multi-step interaction history by a belief state, i.e., a natural language summary of what has been discovered about task-relevant unknowns. Under ABBEL, at each step the agent first updates a prior belief with the most recent observation from the environment to form a posterior belief, then uses only the posterior to select an action. We systematically evaluate frontier models under ABBEL across six diverse multi-step environments, finding that ABBEL supports generating interpretable beliefs while maintaining near-constant memory use over interaction steps. However, bottleneck approaches are generally prone to error propagation, which we observe causing inferior performance when compared to the full context setting due to errors in belief updating. Therefore, we train LLMs to generate and act on beliefs within the ABBEL framework via reinforcement learning (RL). We experiment with belief grading, to reward higher quality beliefs, as well as belief length penalties to reward more compressed beliefs. Our experiments demonstrate the ability of RL to improve ABBEL's performance beyond the full context setting, while using less memory than contemporaneous approaches.

Jiayi Pan, Xiuyu Li, Long Lian et al. · berkeley

Scaling inference-time computation has substantially improved the reasoning capabilities of language models. However, existing methods have significant limitations: serialized chain-of-thought approaches generate overly long outputs, leading to increased latency and exhausted context windows, while parallel methods such as self-consistency suffer from insufficient coordination, resulting in redundant computations and limited performance gains. To address these shortcomings, we propose Adaptive Parallel Reasoning (APR), a novel reasoning framework that enables language models to orchestrate both serialized and parallel computations end-to-end. APR generalizes existing reasoning methods by enabling adaptive multi-threaded inference using spawn() and join() operations. A key innovation is our end-to-end reinforcement learning strategy, optimizing both parent and child inference threads to enhance task success rate without requiring predefined reasoning structures. Experiments on the Countdown reasoning task demonstrate significant benefits of APR: (1) higher performance within the same context window (83.4% vs. 60.0% at 4k context); (2) superior scalability with increased computation (80.1% vs. 66.6% at 20k total tokens); (3) improved accuracy at equivalent latency (75.2% vs. 57.3% at approximately 5,000ms). APR represents a step towards enabling language models to autonomously optimize their reasoning processes through adaptive allocation of computation.

Jiaxin Ge, Zora Zhiruo Wang, Xuhui Zhou et al. · allen-ai, cmu

Designing structured visuals such as presentation slides is essential for communicative needs, necessitating both content creation and visual planning skills. In this work, we tackle the challenge of automated slide generation, where models produce slide presentations from natural language (NL) instructions. We first introduce the SlidesBench benchmark, the first benchmark for slide generation with 7k training and 585 testing examples derived from 310 slide decks across 10 domains. SlidesBench supports evaluations that are (i)reference-based to measure similarity to a target slide, and (ii)reference-free to measure the design quality of generated slides alone. We benchmark end-to-end image generation and program generation methods with a variety of models, and find that programmatic methods produce higher-quality slides in user-interactable formats. Built on the success of program generation, we create AutoPresent, an 8B Llama-based model trained on 7k pairs of instructions paired with code for slide generation, and achieve results comparable to the closed-source model GPT-4o. We further explore iterative design refinement where the model is tasked to self-refine its own output, and we found that this process improves the slide's quality. We hope that our work will provide a basis for future work on generating structured visuals.

Josh Barua, Sanjay Subramanian, Kayo Yin et al.

In translation, a concept represented by a single word in a source language can have multiple variations in a target language. The task of lexical selection requires using context to identify which variation is most appropriate for a source text. We work with native speakers of nine languages to create DTAiLS, a dataset of 1,377 sentence pairs that exhibit cross-lingual concept variation when translating from English. We evaluate recent LLMs and neural machine translation systems on DTAiLS, with the best-performing model, GPT-4, achieving from 67 to 85% accuracy across languages. Finally, we use language models to generate English rules describing target-language concept variations. Providing weaker models with high-quality lexical rules improves accuracy substantially, in some cases reaching or outperforming GPT-4.

Lingjun Mao, Zineng Tang, Alane Suhr

We study the perception of color illusions by vision-language models. Color illusion, where a person's visual system perceives color differently from actual color, is well-studied in human vision. However, it remains underexplored whether vision-language models (VLMs), trained on large-scale human data, exhibit similar perceptual biases when confronted with such color illusions. We propose an automated framework for generating color illusion images, resulting in RCID (Realistic Color Illusion Dataset), a dataset of 19,000 realistic illusion images. Our experiments show that all studied VLMs exhibit perceptual biases similar human vision. Finally, we train a model to distinguish both human perception and actual pixel differences.

Long Lian, Sida Wang, Felix Juefei-Xu et al.

Scaling inference-time computation has enabled Large Language Models (LLMs) to achieve strong reasoning performance, but inherently sequential decoding leads to substantial latency, especially on complex tasks. Recent work on adaptive parallel reasoning aims to improve inference efficiency by decomposing the problem-solving process into concurrent reasoning threads when beneficial. However, existing methods on realistic tasks are either limited to supervised behavior cloning or exhibit significant accuracy drops compared to widely-used sequential long chain-of-thought (CoT) baselines. Moreover, many require customized inference engines, complicating deployment. We introduce ThreadWeaver, a framework for adaptive parallel reasoning that achieves accuracy on par with popular sequential reasoning models of comparable size while significantly reducing inference latency. ThreadWeaver's performance stems from three key innovations: 1) a two-stage parallel trajectory generator that produces large-scale, high-quality CoT data with parallel annotations for supervised fine-tuning; 2) a trie-based training-inference co-design that enables parallel reasoning on any off-the-shelf autoregressive inference engine without modifying position embeddings or KV caches; and 3) a parallelization-aware reinforcement learning framework that teaches the model to balance accuracy with effective parallelization. Across six challenging mathematical reasoning benchmarks, ThreadWeaver trained atop Qwen3-8B achieves accuracy comparable to cutting-edge sequential reasoning models (71.9% on average and 79.9% on AIME24) while delivering up to 1.53x average speedup in token latency, establishing a new Pareto frontier between accuracy and efficiency.

Josh Barua, Seun Eisape, Kayo Yin et al.

While large reasoning models have shown remarkable ability to generate long chains-of-thought (CoTs) in English, we still lack understanding of how these long-form reasoning abilities transfer to the vast majority of the world's languages. In this work, we systematically investigate four key stages of model development--scaling, pretraining, post-training, and inference--to understand how long CoT capabilities extend beyond English. We compare two reasoning settings across nine non-English target languages: En-CoT, where models process target-language inputs, but reason in English; and Target-CoT, where models both process inputs and generate long CoTs in the target language. We find that scaling reasoning model size improves multilingual task performance in En-CoT, but Target-CoT performance lags behind. This gap widens for tasks requiring long, multi-step CoTs such as mathematical reasoning. Shifting to pretraining, we find that adding a specialized reasoning stage enhances En-CoT performance but degrades Target-CoT, whereas broad multilingual pretraining improves both modes simultaneously. Given the scarcity of high-quality reasoning traces in languages other than English, we explore synthetic data curation approaches for post-training. We demonstrate that fine-tuning on reasoning traces automatically translated from gold English traces outperforms fine-tuning on target-language traces distilled from large reasoning models. Finally, we report disparities in inference efficiency between languages and uncover language-specific failure modes in CoTs. We release models, datasets, and code to foster further research.

Hao Bai, Yifei Zhou, Mert Cemri et al.

Training corpuses for vision language models (VLMs) typically lack sufficient amounts of decision-centric data. This renders off-the-shelf VLMs sub-optimal for decision-making tasks such as in-the-wild device control through graphical user interfaces (GUIs). While training with static demonstrations has shown some promise, we show that such methods fall short for controlling real GUIs due to their failure to deal with real-world stochasticity and non-stationarity not captured in static observational data. This paper introduces a novel autonomous RL approach, called DigiRL, for training in-the-wild device control agents through fine-tuning a pre-trained VLM in two stages: offline RL to initialize the model, followed by offline-to-online RL. To do this, we build a scalable and parallelizable Android learning environment equipped with a VLM-based evaluator and develop a simple yet effective RL approach for learning in this domain. Our approach runs advantage-weighted RL with advantage estimators enhanced to account for stochasticity along with an automatic curriculum for deriving maximal learning signal. We demonstrate the effectiveness of DigiRL using the Android-in-the-Wild (AitW) dataset, where our 1.3B VLM trained with RL achieves a 49.5% absolute improvement -- from 17.7 to 67.2% success rate -- over supervised fine-tuning with static human demonstration data. These results significantly surpass not only the prior best agents, including AppAgent with GPT-4V (8.3% success rate) and the 17B CogAgent trained with AitW data (38.5%), but also the prior best autonomous RL approach based on filtered behavior cloning (57.8%), thereby establishing a new state-of-the-art for digital agents for in-the-wild device control.

Anna Effenberger, Eva Yan, Rhia Singh et al.

We analyze language change over time in a collaborative, goal-oriented instructional task, where utility-maximizing participants form conventions and increase their expertise. Prior work studied such scenarios mostly in the context of reference games, and consistently found that language complexity is reduced along multiple dimensions, such as utterance length, as conventions are formed. In contrast, we find that, given the ability to increase instruction utility, instructors increase language complexity along these previously studied dimensions to better collaborate with increasingly skilled instruction followers.

Noriyuki Kojima, Alane Suhr, Yoav Artzi

We study continual learning for natural language instruction generation, by observing human users' instruction execution. We focus on a collaborative scenario, where the system both acts and delegates tasks to human users using natural language. We compare user execution of generated instructions to the original system intent as an indication to the system's success communicating its intent. We show how to use this signal to improve the system's ability to generate instructions via contextual bandit learning. In interaction with real users, our system demonstrates dramatic improvements in its ability to generate language over time.

Alane Suhr, Claudia Yan, Charlotte Schluger et al.

We study a collaborative scenario where a user not only instructs a system to complete tasks, but also acts alongside it. This allows the user to adapt to the system abilities by changing their language or deciding to simply accomplish some tasks themselves, and requires the system to effectively recover from errors as the user strategically assigns it new goals. We build a game environment to study this scenario, and learn to map user instructions to system actions. We introduce a learning approach focused on recovery from cascading errors between instructions, and modeling methods to explicitly reason about instructions with multiple goals. We evaluate with a new evaluation protocol using recorded interactions and online games with human users, and observe how users adapt to the system abilities.

Alane Suhr, Yoav Artzi

NLVR2 (Suhr et al., 2019) was designed to be robust for language bias through a data collection process that resulted in each natural language sentence appearing with both true and false labels. The process did not provide a similar measure of control for visual bias. This technical report analyzes the potential for visual bias in NLVR2. We show that some amount of visual bias likely exists. Finally, we identify a subset of the test data that allows to test for model performance in a way that is robust to such potential biases. We show that the performance of existing models (Li et al., 2019; Tan and Bansal 2019) is relatively robust to this potential bias. We propose to add the evaluation on this subset of the data to the NLVR2 evaluation protocol, and update the official release to include it. A notebook including an implementation of the code used to replicate this analysis is available at http://nlvr.ai/NLVR2BiasAnalysis.html.

Howard Chen, Alane Suhr, Dipendra Misra et al.

We study the problem of jointly reasoning about language and vision through a navigation and spatial reasoning task. We introduce the Touchdown task and dataset, where an agent must first follow navigation instructions in a real-life visual urban environment, and then identify a location described in natural language to find a hidden object at the goal position. The data contains 9,326 examples of English instructions and spatial descriptions paired with demonstrations. Empirical analysis shows the data presents an open challenge to existing methods, and qualitative linguistic analysis shows that the data displays richer use of spatial reasoning compared to related resources.

Alane Suhr, Stephanie Zhou, Ally Zhang et al.

We introduce a new dataset for joint reasoning about natural language and images, with a focus on semantic diversity, compositionality, and visual reasoning challenges. The data contains 107,292 examples of English sentences paired with web photographs. The task is to determine whether a natural language caption is true about a pair of photographs. We crowdsource the data using sets of visually rich images and a compare-and-contrast task to elicit linguistically diverse language. Qualitative analysis shows the data requires compositional joint reasoning, including about quantities, comparisons, and relations. Evaluation using state-of-the-art visual reasoning methods shows the data presents a strong challenge.

Alane Suhr, Yoav Artzi

We propose a learning approach for mapping context-dependent sequential instructions to actions. We address the problem of discourse and state dependencies with an attention-based model that considers both the history of the interaction and the state of the world. To train from start and goal states without access to demonstrations, we propose SESTRA, a learning algorithm that takes advantage of single-step reward observations and immediate expected reward maximization. We evaluate on the SCONE domains, and show absolute accuracy improvements of 9.8%-25.3% across the domains over approaches that use high-level logical representations.

Alane Suhr, Srinivasan Iyer, Yoav Artzi

We propose a context-dependent model to map utterances within an interaction to executable formal queries. To incorporate interaction history, the model maintains an interaction-level encoder that updates after each turn, and can copy sub-sequences of previously predicted queries during generation. Our approach combines implicit and explicit modeling of references between utterances. We evaluate our model on the ATIS flight planning interactions, and demonstrate the benefits of modeling context and explicit references.

Stephanie Zhou, Alane Suhr, Yoav Artzi

Natural language provides a widely accessible and expressive interface for robotic agents. To understand language in complex environments, agents must reason about the full range of language inputs and their correspondence to the world. Such reasoning over language and vision is an open problem that is receiving increasing attention. While existing data sets focus on visual diversity, they do not display the full range of natural language expressions, such as counting, set reasoning, and comparisons. We propose a simple task for natural language visual reasoning, where images are paired with descriptive statements. The task is to predict if a statement is true for the given scene. This abstract describes our existing synthetic images corpus and our current work on collecting real vision data.