Jeff Da

Jack Hessel, Ana Marasović, Jena D. Hwang et al. · allen-ai, uw

Large neural networks can now generate jokes, but do they really "understand" humor? We challenge AI models with three tasks derived from the New Yorker Cartoon Caption Contest: matching a joke to a cartoon, identifying a winning caption, and explaining why a winning caption is funny. These tasks encapsulate progressively more sophisticated aspects of "understanding" a cartoon; key elements are the complex, often surprising relationships between images and captions and the frequent inclusion of indirect and playful allusions to human experience and culture. We investigate both multimodal and language-only models: the former are challenged with the cartoon images directly, while the latter are given multifaceted descriptions of the visual scene to simulate human-level visual understanding. We find that both types of models struggle at all three tasks. For example, our best multimodal models fall 30 accuracy points behind human performance on the matching task, and, even when provided ground-truth visual scene descriptors, human-authored explanations are preferred head-to-head over the best machine-authored ones (few-shot GPT-4) in more than 2/3 of cases. We release models, code, leaderboard, and corpus, which includes newly-gathered annotations describing the image's locations/entities, what's unusual in the scene, and an explanation of the joke.

Chaithanya Bandi, Ben Hertzberg, Geobio Boo et al.

The Model Context Protocol (MCP) is rapidly becoming the standard interface for Large Language Models (LLMs) to discover and invoke external tools. However, existing evaluations often fail to capture the complexity of real-world scenarios, relying on restricted toolsets, simplistic workflows, or subjective LLM-as-a-judge metrics. We introduce MCP-Atlas, a large-scale benchmark for evaluating tool-use competency, comprising 36 real MCP servers and 220 tools. It includes 1,000 tasks designed to assess tool-use competency in realistic, multi-step workflows. Tasks use natural language prompts that avoid naming specific tools or servers, requiring agents to identify and orchestrate 3-6 tool calls across multiple servers. We score tasks using a claims-based rubric that awards partial credit based on the factual claims satisfied in the model's final answer, complemented by internal diagnostics on tool discovery, parameterization, syntax, error recovery, and efficiency. Evaluation results on frontier models reveal that top models achieve pass rates exceeding 50%, with primary failures arising from inadequate tool usage and task understanding. We release the task schema, containerized harness, and a 500-task public subset of the benchmark dataset to facilitate reproducible comparisons and advance the development of robust, tool-augmented agents.

Vaskar Nath, Dylan Slack, Jeff Da et al.

Techniques that learn improved representations via offline data or self-supervised objectives have shown impressive results in traditional reinforcement learning (RL). Nevertheless, it is unclear how improved representation learning can benefit reinforcement learning from human feedback (RLHF) on language models (LMs). In this work, we propose training reward models (RMs) in a contrastive, $\textit{goal-conditioned}$ fashion by increasing the representation similarity of future states along sampled preferred trajectories and decreasing the similarity along randomly sampled dispreferred trajectories. This objective significantly improves RM performance by up to 0.09 AUROC across challenging benchmarks, such as MATH and GSM8k. These findings extend to general alignment as well -- on the Helpful-Harmless dataset, we observe $2.3\%$ increase in accuracy. Beyond improving reward model performance, we show this way of training RM representations enables improved $\textit{steerability}$ because it allows us to evaluate the likelihood of an action achieving a particular goal-state (e.g., whether a solution is correct or helpful). Leveraging this insight, we find that we can filter up to $55\%$ of generated tokens during majority voting by discarding trajectories likely to end up in an "incorrect" state, which leads to significant cost savings. We additionally find that these representations can perform fine-grained control by conditioning on desired future goal-states. For example, we show that steering a Llama 3 model towards helpful generations with our approach improves helpfulness by $9.6\%$ over a supervised-fine-tuning trained baseline. Similarly, steering the model towards complex generations improves complexity by $21.6\%$ over the baseline. Overall, we find that training RMs in this contrastive, goal-conditioned fashion significantly improves performance and enables model steerability.

Niklas Lauffer, Xiang Deng, Srivatsa Kundurthy et al.

A popular paradigm for training LM agents relies on imitation learning, fine-tuning on expert trajectories. However, we show that the off-policy nature of imitation learning for multi-turn LM agents suffers from the fundamental limitation known as covariate shift: as the student policy's behavior diverges from the expert's, it encounters states not present in the training data, reducing the effectiveness of fine-tuning. Taking inspiration from the classic DAgger algorithm, we propose a novel data generation methodology for addressing covariate shift for multi-turn LLM training. We introduce on-policy expert corrections (OECs), partially on-policy data generated by starting rollouts with a student model and then switching to an expert model part way through the trajectory. We explore the effectiveness of our data generation technique in the domain of software engineering (SWE) tasks, a multi-turn setting where LLM agents must interact with a development environment to fix software bugs. Our experiments compare OEC data against various other on-policy and imitation learning approaches on SWE agent problems and train models using a common rejection sampling (i.e., using environment reward) combined with supervised fine-tuning technique. Experiments find that OEC trajectories show a relative 14% and 13% improvement over traditional imitation learning in the 7b and 32b setting, respectively, on SWE-bench verified. Our results demonstrate the need for combining expert demonstrations with on-policy data for effective multi-turn LM agent training.

Hugh Zhang, Jeff Da, Dean Lee et al.

Large language models (LLMs) have achieved impressive success on many benchmarks for mathematical reasoning. However, there is growing concern that some of this performance actually reflects dataset contamination, where data closely resembling benchmark questions leaks into the training data, instead of true reasoning ability. To investigate this claim rigorously, we commission Grade School Math 1000 (GSM1k). GSM1k is designed to mirror the style and complexity of the established GSM8k benchmark, the gold standard for measuring elementary mathematical reasoning. We ensure that the two benchmarks are comparable across important metrics such as human solve rates, number of steps in solution, answer magnitude, and more. When evaluating leading open- and closed-source LLMs on GSM1k, we observe accuracy drops of up to 8%, with several families of models showing evidence of systematic overfitting across almost all model sizes. Further analysis suggests a positive relationship (Spearman's r^2 = 0.36) between a model's probability of generating an example from GSM8k and its performance gap between GSM8k and GSM1k, suggesting that some models may have partially memorized GSM8k. Nevertheless, many models, especially those on the frontier, show minimal signs of overfitting, and all models broadly demonstrate generalization to novel math problems guaranteed to not be in their training data.

Mohit Raghavendra, Soham Dan, Miguel Romero Calvo et al.

We introduce SWE Atlas, a benchmark suite for coding agents spanning three professional software engineering workflows: Codebase Q&A (124 tasks), Test Writing (90 tasks), and Refactoring (70 tasks). SWE Atlas differs from prior SWE benchmarks in three key ways: it targets underrepresented but practically important task categories, uses comprehensive category-specific evaluation protocols, and adopts under-specified, agentic task formulations that better reflect real-world usage. Its evaluation framework combines programmatic checks with rubric-based assessment. This goes beyond functional correctness, evaluating software engineering quality, including test and refactor completeness, maintainability, reusable abstractions, and codebase hygiene. We evaluate a range of frontier and open-weight models on SWE Atlas and find that GPT-5.4 and Opus 4.7 achieve the strongest overall performance, while even the best open-weight models score poorly. Our analysis suggests that top models rely on extensive codebase exploration and runtime-driven reasoning. However, even top models consistently struggle with subtle edge cases, complex runtime analysis, and adherence to software engineering best practices. Overall, SWE Atlas provides a complementary evaluation suite for measuring both correctness and engineering quality in coding agents.

Jeff Da, Clinton Wang, Xiang Deng et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has been widely adopted as the de facto method for enhancing the reasoning capabilities of large language models and has demonstrated notable success in verifiable domains like math and competitive programming tasks. However, the efficacy of RLVR diminishes significantly when applied to agentic environments. These settings, characterized by multi-step, complex problem solving, lead to high failure rates even for frontier LLMs, as the reward landscape is too sparse for effective model training via conventional RLVR. In this work, we introduce Agent-RLVR, a framework that makes RLVR effective in challenging agentic settings, with an initial focus on software engineering tasks. Inspired by human pedagogy, Agent-RLVR introduces agent guidance, a mechanism that actively steers the agent towards successful trajectories by leveraging diverse informational cues. These cues, ranging from high-level strategic plans to dynamic feedback on the agent's errors and environmental interactions, emulate a teacher's guidance, enabling the agent to navigate difficult solution spaces and promotes active self-improvement via additional environment exploration. In the Agent-RLVR training loop, agents first attempt to solve tasks to produce initial trajectories, which are then validated by unit tests and supplemented with agent guidance. Agents then reattempt with guidance, and the agent policy is updated with RLVR based on the rewards of these guided trajectories. Agent-RLVR elevates the pass@1 performance of Qwen-2.5-72B-Instruct from 9.4% to 22.4% on SWE-Bench Verified. We find that our guidance-augmented RLVR data is additionally useful for test-time reward model training, shown by further boosting pass@1 to 27.8%. Agent-RLVR lays the groundwork for training agents with RLVR in complex, real-world environments where conventional RL methods struggle.

Xiang Deng, Jeff Da, Edwin Pan et al.

We introduce SWE-Bench Pro, a substantially more challenging benchmark that builds upon the best practices of SWE-BENCH [25], but is explicitly designed to capture realistic, complex, enterprise-level problems beyond the scope of SWE-BENCH. SWE-BENCH PRO contains 1,865 problems sourced from a diverse set of 41 actively maintained repositories spanning business applications, B2B services, and developer tools. The benchmark is partitioned into a public set with open access to problems sourced from 11 repositories, a held-out set of 12 repositories and a commercial set of 18 proprietary repositories where we have formal partnership agreements with early-stage startups. Problems in the held-out and the commercial set are not publicly accessible, but we release results on the commercial set. Our benchmark features long-horizon tasks that may require hours to days for a professional software engineer to complete, often involving patches across multiple files and substantial code modifications. All tasks are human-verified and augmented with sufficient context to ensure resolvability. To better understand these limitations, we cluster the failure modes observed in the collected agent trajectories for a clearer characterization of the error patterns exhibited by current models. Overall, SWE-BENCH PRO provides a contamination-resistant testbed that more faithfully captures the complexity and diversity of real-world software development, advancing the pursuit of truly autonomous software engineering agents at a professional level.

Jeff Da, Ronan Le Bras, Ximing Lu et al.

Recently, commonsense knowledge models - pretrained language models (LM) fine-tuned on knowledge graph (KG) tuples - showed that considerable amounts of commonsense knowledge can be encoded in the parameters of large language models. However, as parallel studies show that LMs are poor hypothesizers of declarative commonsense relationships on their own, it remains unclear whether this knowledge is learned during pretraining or from fine-tuning on KG examples. To investigate this question, we train commonsense knowledge models in few-shot settings to study the emergence of their commonsense representation abilities. Our results show that commonsense knowledge models can rapidly adapt from limited examples, indicating that KG fine-tuning serves to learn an interface to encoded knowledge learned during pretraining. Importantly, our analysis of absolute, angular, and distributional parameter changes during few-shot fine-tuning provides novel insights into how this interface is learned.

Jeff Da, Maxwell Forbes, Rowan Zellers et al.

Multimodal disinformation, from 'deepfakes' to simple edits that deceive, is an important societal problem. Yet at the same time, the vast majority of media edits are harmless -- such as a filtered vacation photo. The difference between this example, and harmful edits that spread disinformation, is one of intent. Recognizing and describing this intent is a major challenge for today's AI systems. We present the task of Edited Media Understanding, requiring models to answer open-ended questions that capture the intent and implications of an image edit. We introduce a dataset for our task, EMU, with 48k question-answer pairs written in rich natural language. We evaluate a wide variety of vision-and-language models for our task, and introduce a new model PELICAN, which builds upon recent progress in pretrained multimodal representations. Our model obtains promising results on our dataset, with humans rating its answers as accurate 40.35% of the time. At the same time, there is still much work to be done -- humans prefer human-annotated captions 93.56% of the time -- and we provide analysis that highlights areas for further progress.

Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras et al.

Recent years have brought about a renewed interest in commonsense representation and reasoning in the field of natural language understanding. The development of new commonsense knowledge graphs (CSKG) has been central to these advances as their diverse facts can be used and referenced by machine learning models for tackling new and challenging tasks. At the same time, there remain questions about the quality and coverage of these resources due to the massive scale required to comprehensively encompass general commonsense knowledge. In this work, we posit that manually constructed CSKGs will never achieve the coverage necessary to be applicable in all situations encountered by NLP agents. Therefore, we propose a new evaluation framework for testing the utility of KGs based on how effectively implicit knowledge representations can be learned from them. With this new goal, we propose ATOMIC 2020, a new CSKG of general-purpose commonsense knowledge containing knowledge that is not readily available in pretrained language models. We evaluate its properties in comparison with other leading CSKGs, performing the first large-scale pairwise study of commonsense knowledge resources. Next, we show that ATOMIC 2020 is better suited for training knowledge models that can generate accurate, representative knowledge for new, unseen entities and events. Finally, through human evaluation, we show that the few-shot performance of GPT-3 (175B parameters), while impressive, remains ~12 absolute points lower than a BART-based knowledge model trained on ATOMIC 2020 despite using over 430x fewer parameters.

Jeff Da

We introduce a simple yet effective method of integrating contextual embeddings with commonsense graph embeddings, dubbed BERT Infused Graphs: Matching Over Other embeDdings. First, we introduce a preprocessing method to improve the speed of querying knowledge bases. Then, we develop a method of creating knowledge embeddings from each knowledge base. We introduce a method of aligning tokens between two misaligned tokenization methods. Finally, we contribute a method of contextualizing BERT after combining with knowledge base embeddings. We also show BERTs tendency to correct lower accuracy question types. Our model achieves a higher accuracy than BERT, and we score fifth on the official leaderboard of the shared task and score the highest without any additional language model pretraining.

Jeff Da, Jungo Kasai

Pretrained deep contextual representations have advanced the state-of-the-art on various commonsense NLP tasks, but we lack a concrete understanding of the capability of these models. Thus, we investigate and challenge several aspects of BERT's commonsense representation abilities. First, we probe BERT's ability to classify various object attributes, demonstrating that BERT shows a strong ability in encoding various commonsense features in its embedding space, but is still deficient in many areas. Next, we show that, by augmenting BERT's pretraining data with additional data related to the deficient attributes, we are able to improve performance on a downstream commonsense reasoning task while using a minimal amount of data. Finally, we develop a method of fine-tuning knowledge graphs embeddings alongside BERT and show the continued importance of explicit knowledge graphs.

Saadia Gabriel, Antoine Bosselut, Jeff Da et al.

We introduce a general framework for abstractive summarization with factual consistency and distinct modeling of the narrative flow in an output summary. Our work addresses current limitations of models for abstractive summarization that often hallucinate information or generate summaries with coherence issues. To generate abstractive summaries with factual consistency and narrative flow, we propose Cooperative Generator -- Discriminator Networks (Co-opNet), a novel transformer-based framework where a generator works with a discriminator architecture to compose coherent long-form summaries. We explore four different discriminator objectives which each capture a different aspect of coherence, including whether salient spans of generated abstracts are hallucinated or appear in the input context, and the likelihood of sentence adjacency in generated abstracts. We measure the ability of Co-opNet to learn these objectives with arXiv scientific papers, using the abstracts as a proxy for gold long-form scientific article summaries. Empirical results from automatic and human evaluations demonstrate that Co-opNet learns to summarize with considerably improved global coherence compared to competitive baselines.