Arman Cohan

David Wadden, Kyle Lo, Bailey Kuehl et al. · allen-ai, uw

While research on scientific claim verification has led to the development of powerful systems that appear to approach human performance, these approaches have yet to be tested in a realistic setting against large corpora of scientific literature. Moving to this open-domain evaluation setting, however, poses unique challenges; in particular, it is infeasible to exhaustively annotate all evidence documents. In this work, we present SciFact-Open, a new test collection designed to evaluate the performance of scientific claim verification systems on a corpus of 500K research abstracts. Drawing upon pooling techniques from information retrieval, we collect evidence for scientific claims by pooling and annotating the top predictions of four state-of-the-art scientific claim verification models. We find that systems developed on smaller corpora struggle to generalize to SciFact-Open, exhibiting performance drops of at least 15 F1. In addition, analysis of the evidence in SciFact-Open reveals interesting phenomena likely to appear when claim verification systems are deployed in practice, e.g., cases where the evidence supports only a special case of the claim. Our dataset is available at https://github.com/dwadden/scifact-open.

Rodney Kinney, Chloe Anastasiades, Russell Authur et al. · allen-ai, microsoft-research

The volume of scientific output is creating an urgent need for automated tools to help scientists keep up with developments in their field. Semantic Scholar (S2) is an open data platform and website aimed at accelerating science by helping scholars discover and understand scientific literature. We combine public and proprietary data sources using state-of-the-art techniques for scholarly PDF content extraction and automatic knowledge graph construction to build the Semantic Scholar Academic Graph, the largest open scientific literature graph to-date, with 200M+ papers, 80M+ authors, 550M+ paper-authorship edges, and 2.4B+ citation edges. The graph includes advanced semantic features such as structurally parsed text, natural language summaries, and vector embeddings. In this paper, we describe the components of the S2 data processing pipeline and the associated APIs offered by the platform. We will update this living document to reflect changes as we add new data offerings and improve existing services.

Chunyuan Deng, Yilun Zhao, Xiangru Tang et al. · gatech

Recent observations have underscored a disparity between the inflated benchmark scores and the actual performance of LLMs, raising concerns about potential contamination of evaluation benchmarks. This issue is especially critical for closed-source models and certain open-source models where training data transparency is lacking. In this paper we study data contamination by proposing two methods tailored for both open-source and proprietary LLMs. We first introduce a retrieval-based system to explore potential overlaps between evaluation benchmarks and pretraining corpora. We further present a novel investigation protocol named \textbf{T}estset \textbf{S}lot Guessing (\textit{TS-Guessing}), applicable to both open and proprietary models. This approach entails masking a wrong answer in a multiple-choice question and prompting the model to fill in the gap. Additionally, it involves obscuring an unlikely word in an evaluation example and asking the model to produce it. We find that certain commercial LLMs could surprisingly guess the missing option in various test sets. Specifically, in the TruthfulQA benchmark, we find that LLMs exhibit notable performance improvement when provided with additional metadata in the benchmark. Further, in the MMLU benchmark, ChatGPT and GPT-4 demonstrated an exact match rate of 52\% and 57\%, respectively, in guessing the missing options in benchmark test data. We hope these results underscore the need for more robust evaluation methodologies and benchmarks in the field.

Xiangru Tang, Yuliang Liu, Zefan Cai et al. · pku

Despite Large Language Models (LLMs) like GPT-4 achieving impressive results in function-level code generation, they struggle with repository-scale code understanding (e.g., coming up with the right arguments for calling routines), requiring a deeper comprehension of complex file interactions. Also, recently, people have developed LLM agents that attempt to interact with repository code (e.g., compiling and evaluating its execution), prompting the need to evaluate their performance. These gaps have motivated our development of ML-Bench, a benchmark rooted in real-world programming applications that leverage existing code repositories to perform tasks. Addressing the need for LLMs to interpret long code contexts and translate instructions into precise, executable scripts, ML-Bench encompasses annotated 9,641 examples across 18 GitHub repositories, challenging LLMs to accommodate user-specified arguments and documentation intricacies effectively. To evaluate both LLMs and AI agents, two setups are employed: ML-LLM-Bench for assessing LLMs' text-to-code conversion within a predefined deployment environment, and ML-Agent-Bench for testing autonomous agents in an end-to-end task execution within a Linux sandbox environment. Our findings indicate that while GPT-4o leads with a Pass@5 rate surpassing 50%, there remains significant scope for improvement, highlighted by issues such as hallucinated outputs and difficulties with bash script generation. Notably, in the more demanding ML-Agent-Bench, GPT-4o achieves a 76.47% success rate, reflecting the efficacy of iterative action and feedback in complex task resolution. Our code, dataset, and models are available at https://github.com/gersteinlab/ML-bench.

Xiangru Tang, Anni Zou, Zhuosheng Zhang et al.

Large language models (LLMs), despite their remarkable progress across various general domains, encounter significant barriers in medicine and healthcare. This field faces unique challenges such as domain-specific terminologies and reasoning over specialized knowledge. To address these issues, we propose MedAgents, a novel multi-disciplinary collaboration framework for the medical domain. MedAgents leverages LLM-based agents in a role-playing setting that participate in a collaborative multi-round discussion, thereby enhancing LLM proficiency and reasoning capabilities. This training-free framework encompasses five critical steps: gathering domain experts, proposing individual analyses, summarising these analyses into a report, iterating over discussions until a consensus is reached, and ultimately making a decision. Our work focuses on the zero-shot setting, which is applicable in real-world scenarios. Experimental results on nine datasets (MedQA, MedMCQA, PubMedQA, and six subtasks from MMLU) establish that our proposed MedAgents framework excels at mining and harnessing the medical expertise within LLMs, as well as extending its reasoning abilities. Our code can be found at https://github.com/gersteinlab/MedAgents.

Kalpesh Krishna, Erin Bransom, Bailey Kuehl et al. · allen-ai, deepmind

While human evaluation remains best practice for accurately judging the faithfulness of automatically-generated summaries, few solutions exist to address the increased difficulty and workload when evaluating long-form summaries. Through a survey of 162 papers on long-form summarization, we first shed light on current human evaluation practices surrounding long-form summaries. We find that 73% of these papers do not perform any human evaluation on model-generated summaries, while other works face new difficulties that manifest when dealing with long documents (e.g., low inter-annotator agreement). Motivated by our survey, we present LongEval, a set of guidelines for human evaluation of faithfulness in long-form summaries that addresses the following challenges: (1) How can we achieve high inter-annotator agreement on faithfulness scores? (2) How can we minimize annotator workload while maintaining accurate faithfulness scores? and (3) Do humans benefit from automated alignment between summary and source snippets? We deploy LongEval in annotation studies on two long-form summarization datasets in different domains (SQuALITY and PubMed), and we find that switching to a finer granularity of judgment (e.g., clause-level) reduces inter-annotator variance in faithfulness scores (e.g., std-dev from 18.5 to 6.8). We also show that scores from a partial annotation of fine-grained units highly correlates with scores from a full annotation workload (0.89 Kendall's tau using 50% judgments). We release our human judgments, annotation templates, and our software as a Python library for future research.

Simeng Han, Hailey Schoelkopf, Yilun Zhao et al. · salesforce

Large language models (LLMs) have achieved remarkable performance on a variety of natural language understanding tasks. However, existing benchmarks are inadequate in measuring the complex logical reasoning capabilities of a model. We present FOLIO, a human-annotated, logically complex and diverse dataset for reasoning in natural language (NL), equipped with first-order logic (FOL) annotations. FOLIO consists of 1,430 examples (unique conclusions), each paired with one of 487 sets of premises used to deductively reason for the validity of each conclusion. The logical correctness of the premises and conclusions is ensured by their FOL annotations, which are automatically verified by an FOL inference engine. In addition to the main NL reasoning task, NL-FOL pairs in FOLIO constitute a new NL-FOL translation dataset. Our experiments on FOLIO systematically evaluate the FOL reasoning ability of supervised fine-tuning on medium-sized language models. For both NL reasoning and NL-FOL translation, we benchmark multiple state-of-the-art language models. Our results show that a subset of FOLIO presents a challenge for one of the most capable {Large Language Model (LLM)} publicly available, GPT-4.

Ansong Ni, Pengcheng Yin, Yilun Zhao et al. · salesforce

Recently, large language models (LLMs), especially those that are pretrained on code, have demonstrated strong capabilities in generating programs from natural language inputs in a few-shot or even zero-shot manner. Despite promising results, there is a notable lack of a comprehensive evaluation of these models language-to-code generation capabilities. Existing studies often focus on specific tasks, model architectures, or learning paradigms, leading to a fragmented understanding of the overall landscape. In this work, we present L2CEval, a systematic evaluation of the language-to-code generation capabilities of LLMs on 7 tasks across the domain spectrum of semantic parsing, math reasoning and Python programming, analyzing the factors that potentially affect their performance, such as model size, pretraining data, instruction tuning, and different prompting methods. In addition to assessing model performance, we measure confidence calibration for the models and conduct human evaluations of the output programs. This enables us to identify and analyze the typical failure modes across various tasks and models. L2CEval offers a comprehensive understanding of the capabilities and limitations of LLMs in language-to-code generation. We also release the evaluation framework and all model outputs, hoping to lay the groundwork for further future research in this domain.

Orion Weller, Kyle Lo, David Wadden et al. · allen-ai

Using large language models (LMs) for query or document expansion can improve generalization in information retrieval. However, it is unknown whether these techniques are universally beneficial or only effective in specific settings, such as for particular retrieval models, dataset domains, or query types. To answer this, we conduct the first comprehensive analysis of LM-based expansion. We find that there exists a strong negative correlation between retriever performance and gains from expansion: expansion improves scores for weaker models, but generally harms stronger models. We show this trend holds across a set of eleven expansion techniques, twelve datasets with diverse distribution shifts, and twenty-four retrieval models. Through qualitative error analysis, we hypothesize that although expansions provide extra information (potentially improving recall), they add additional noise that makes it difficult to discern between the top relevant documents (thus introducing false positives). Our results suggest the following recipe: use expansions for weaker models or when the target dataset significantly differs from training corpus in format; otherwise, avoid expansions to keep the relevance signal clear.

Dustin Wright, David Wadden, Kyle Lo et al. · allen-ai, uw

Automated scientific fact checking is difficult due to the complexity of scientific language and a lack of significant amounts of training data, as annotation requires domain expertise. To address this challenge, we propose scientific claim generation, the task of generating one or more atomic and verifiable claims from scientific sentences, and demonstrate its usefulness in zero-shot fact checking for biomedical claims. We propose CLAIMGEN-BART, a new supervised method for generating claims supported by the literature, as well as KBIN, a novel method for generating claim negations. Additionally, we adapt an existing unsupervised entity-centric method of claim generation to biomedical claims, which we call CLAIMGEN-ENTITY. Experiments on zero-shot fact checking demonstrate that both CLAIMGEN-ENTITY and CLAIMGEN-BART, coupled with KBIN, achieve up to 90% performance of fully supervised models trained on manually annotated claims and evidence. A rigorous evaluation study demonstrates significant improvement in generated claim and negation quality over existing baselines

Xiangru Tang, Yiming Zong, Jason Phang et al.

Despite the remarkable capabilities of Large Language Models (LLMs) like GPT-4, producing complex, structured tabular data remains challenging. Our study assesses LLMs' proficiency in structuring tables and introduces a novel fine-tuning method, cognizant of data structures, to bolster their performance. We unveil Struc-Bench, a comprehensive benchmark featuring prominent LLMs (GPT-NeoX-20B, GPT-3.5, GPT-4, and Vicuna), which spans text tables, HTML, and LaTeX formats. Our proposed FormatCoT aids in crafting format-specific instructions from the intended outputs to populate this benchmark. Addressing the gap in task-centered evaluation, we propose two innovative metrics, P-Score (Prompting Score) and H-Score (Heuristical Score), to more accurately gauge LLM performance. Our experiments show that applying our structure-aware fine-tuning to LLaMA-7B leads to substantial performance gains, outshining its LLM counterparts across most measures. In-depth error analysis and creating an ability map across six dimensions -- coverage, formatting, reasoning, comprehension, pragmatics, and hallucination -- highlight areas for future enhancements and suggest forthcoming research trajectories. Our code and models can be found at https://github.com/gersteinlab/Struc-Bench.

Jon Saad-Falcon, Amanpreet Singh, Luca Soldaini et al. · allen-ai, stanford

Real-world applications of neural language models often involve running many different models over the same corpus. The high computational cost of these runs has led to interest in techniques that can reuse the contextualized embeddings produced in previous runs to speed training and inference of future ones. We refer to this approach as embedding recycling (ER). While multiple ER techniques have been proposed, their practical effectiveness is still unknown because existing evaluations consider very few models and do not adequately account for overhead costs. We perform an extensive evaluation of ER across eight different models (17 to 900 million parameters) and fourteen tasks in English. We show how a simple ER technique that caches activations from an intermediate layer of a pretrained model, and learns task-specific adapters on the later layers, is broadly effective. For the best-performing baseline in our experiments (DeBERTa-v2 XL), adding a precomputed cache results in a >90% speedup during training and 87-91% speedup for inference, with negligible impact on accuracy. Our analysis reveals important areas of future work.

Yijie Zhou, Kejian Shi, Wencai Zhang et al.

Open-domain Multi-Document Summarization (ODMDS) is a critical tool for condensing vast arrays of documents into coherent, concise summaries. With a more inter-related document set, there does not necessarily exist a correct answer for the retrieval, making it hard to measure the retrieving performance. We propose a rule-based method to process query-based document summarization datasets into ODMDS datasets. Based on this method, we introduce a novel dataset, ODSum, a sophisticated case with its document index interdependent and often interrelated. We tackle ODMDS with the \textit{retrieve-then-summarize} method, and the performance of a list of retrievers and summarizers is investigated. Through extensive experiments, we identify variances in evaluation metrics and provide insights into their reliability. We also found that LLMs suffer great performance loss from retrieving errors. We further experimented methods to improve the performance as well as investigate their robustness against imperfect retrieval. We will release our data and code at https://github.com/yale-nlp/ODSum.

John Giorgi, Luca Soldaini, Bo Wang et al. · allen-ai, utoronto

Multi-document summarization (MDS) assumes a set of topic-related documents are provided as input. In practice, this document set is not always available; it would need to be retrieved given an information need, i.e. a question or topic statement, a setting we dub "open-domain" MDS. We study this more challenging setting by formalizing the task and bootstrapping it using existing datasets, retrievers and summarizers. Via extensive automatic and human evaluation, we determine: (1) state-of-the-art summarizers suffer large reductions in performance when applied to open-domain MDS, (2) additional training in the open-domain setting can reduce this sensitivity to imperfect retrieval, and (3) summarizers are insensitive to the retrieval of duplicate documents and the order of retrieved documents, but highly sensitive to other errors, like the retrieval of irrelevant documents. Based on our results, we provide practical guidelines to enable future work on open-domain MDS, e.g. how to choose the number of retrieved documents to summarize. Our results suggest that new retrieval and summarization methods and annotated resources for training and evaluation are necessary for further progress in the open-domain setting.

Simeng Han, Tianyu Liu, Chuhan Li et al.

LLMs approach logical and mathematical reasoning through natural or symbolic languages. While natural language offers human-accessible flexibility but suffers from ambiguity, symbolic reasoning provides precise, machine-executable inferences at the cost of strict domain constraints. We introduce HYBRIDMIND, an adaptive strategy that selects the optimal reasoning approach for each reasoning problem. Through extensive experiments, we evaluate both prompting-based approaches with state-of-the-art LLMs and fine-tuned open-source models. We find that fine-tuning LLaMA-3.1-8B-Instruct as a meta-selector outperforms GPT-4o's natural language reasoning by 4.4\% on FOLIO and 1.3\% on MATH. More notably, using GPT-3.5-turbo as a prompted meta-selector yields a 10\% improvement on FOLIO's challenging subset compared to GPT-4o. We will release our code and data to support future research.

Lin Fu, Zheyuan Yang, Yang Wang et al.

We introduce VideoKR, the first large-scale training corpus specifically designed to strengthen knowledge- and reasoning-intensive video understanding. It comprises 315K video reasoning examples over 145K newly collected, CC-licensed, expert-domain videos. We develop a human-in-the-loop, skill-oriented example generation pipeline that targets progressively deeper video reasoning capabilities while ensuring the difficulty, diversity, and reliability of both the examples and their CoT rationales. We also curate VideoKR-Eval, a new expert-annotated benchmark where questions require genuine video understanding and knowledge-intensive reasoning rather than textual shortcuts. Our experiments show that, under a standard SFT$\rightarrow$GRPO pipeline, models post-trained on VideoKR outperform prior post-training approaches on knowledge-intensive video reasoning while remaining competitive on general video reasoning, highlighting data design as a key driver of progress in video reasoning. We further conduct comprehensive ablations to isolate the contributions of VideoKR, providing actionable insights for future work.

Yixin Liu, Alexander R. Fabbri, Jiawen Chen et al.

While large language models (LLMs) can already achieve strong performance on standard generic summarization benchmarks, their performance on more complex summarization task settings is less studied. Therefore, we benchmark LLMs on instruction controllable text summarization, where the model input consists of both a source article and a natural language requirement for desired summary characteristics. To this end, we curate an evaluation-only dataset for this task setting and conduct human evaluations of five LLM-based systems to assess their instruction-following capabilities in controllable summarization. We then benchmark LLM-based automatic evaluation for this task with 4 different evaluation protocols and 11 LLMs, resulting in 40 evaluation methods. Our study reveals that instruction controllable text summarization remains a challenging task for LLMs, since (1) all LLMs evaluated still make factual and other types of errors in their summaries; (2) no LLM-based evaluation methods can achieve a strong alignment with human annotators when judging the quality of candidate summaries; (3) different LLMs show large performance gaps in summary generation and evaluation capabilities. We make our collected benchmark InstruSum publicly available to facilitate future research in this direction.

Areeb Gani, Asal Meskin, Gabrielle Kaili-May Liu et al.

Reliable uncertainty communication is critical to the trustworthiness of LLMs, yet faithful calibration (FC)--the alignment between models' intrinsic and (linguistically) expressed confidence--is a persistent failure mode. This challenge is key for large reasoning models (LRMs), whose extended reasoning traces are often interpreted by users as evidence of deliberation, competence, and confidence. Despite the importance of FC and wide usage of LRMs, the extent to which LRMs can faithfully express their confidence remains poorly understood. Moreover, the prevailing paradigm to measure FC does not generalize well to the long chain-of-thought outputs generated by LRMs, which tend to lack clear step boundaries, involve inconsistent step structure, and encode complex conditional dependencies throughout the trace--complicating estimation of intrinsic confidence. To address this challenge, we introduce a novel framework to systematically quantify FC of LRMs. Our framework analyzes linguistic decisiveness relative to three sources of internal uncertainty, based on token probabilities, hidden states, and sampled response consistency. We also devise a prefix-conditioned sampling approach to control for conditional and structural variation across traces. Applying our framework to a diverse suite of leading models, datasets, and prompts, we find that faithful confidence expression is a significant challenge for LRMs. Reasoning behaviors do not automatically translate to improved FC, and prompt interventions for non-reasoning models do not improve faithfulness in the reasoning setting. Different confidence estimators further produce divergent assessments of the same traces, revealing fragility in prior evaluation methodologies. Taken together, our work establishes FC as a distinct reliability and alignment target for LRMs, particularly as such systems are increasingly deployed in high-stakes contexts.

Amanpreet Singh, Mike D'Arcy, Arman Cohan et al.

Learned representations of scientific documents can serve as valuable input features for downstream tasks without further fine-tuning. However, existing benchmarks for evaluating these representations fail to capture the diversity of relevant tasks. In response, we introduce SciRepEval, the first comprehensive benchmark for training and evaluating scientific document representations. It includes 24 challenging and realistic tasks, 8 of which are new, across four formats: classification, regression, ranking and search. We then use this benchmark to study and improve the generalization ability of scientific document representation models. We show how state-of-the-art models like SPECTER and SciNCL struggle to generalize across the task formats, and that simple multi-task training fails to improve them. However, a new approach that learns multiple embeddings per document, each tailored to a different format, can improve performance. We experiment with task-format-specific control codes and adapters and find they outperform the existing single-embedding state-of-the-art by over 2 points absolute. We release the resulting family of multi-format models, called SPECTER2, for the community to use and build on.

Hyunji Lee, Luca Soldaini, Arman Cohan et al. · allen-ai

Information retrieval methods often rely on a single embedding model trained on large, general-domain datasets like MSMARCO. While this approach can produce a retriever with reasonable overall performance, they often underperform models trained on domain-specific data when testing on their respective domains. Prior work in information retrieval has tackled this through multi-task training, but the idea of routing over a mixture of domain-specific expert retrievers remains unexplored despite the popularity of such ideas in language model generation research. In this work, we introduce RouterRetriever, a retrieval model that leverages a mixture of domain-specific experts by using a routing mechanism to select the most appropriate expert for each query. RouterRetriever is lightweight and allows easy addition or removal of experts without additional training. Evaluation on the BEIR benchmark demonstrates that RouterRetriever outperforms both models trained on MSMARCO (+2.1 absolute nDCG@10) and multi-task models (+3.2). This is achieved by employing our routing mechanism, which surpasses other routing techniques (+1.8 on average) commonly used in language modeling. Furthermore, the benefit generalizes well to other datasets, even in the absence of a specific expert on the dataset. RouterRetriever is the first work to demonstrate the advantages of routing over a mixture of domain-specific expert embedding models as an alternative to a single, general-purpose embedding model, especially when retrieving from diverse, specialized domains.

Thong Nguyen, Andrew Yates, Ayah Zirikly et al.

Automated methods have been widely used to identify and analyze mental health conditions (e.g., depression) from various sources of information, including social media. Yet, deployment of such models in real-world healthcare applications faces challenges including poor out-of-domain generalization and lack of trust in black box models. In this work, we propose approaches for depression detection that are constrained to different degrees by the presence of symptoms described in PHQ9, a questionnaire used by clinicians in the depression screening process. In dataset-transfer experiments on three social media datasets, we find that grounding the model in PHQ9's symptoms substantially improves its ability to generalize to out-of-distribution data compared to a standard BERT-based approach. Furthermore, this approach can still perform competitively on in-domain data. These results and our qualitative analyses suggest that grounding model predictions in clinically-relevant symptoms can improve generalizability while producing a model that is easier to inspect.

Zonghai Yao, Benlu Wang, Yifan Zhang et al.

Large language models perform well on many medical QA benchmarks, but real clinical reasoning often requires integrating evidence across multiple images rather than interpreting a single view. We introduce MedThinkVQA, an expert-annotated benchmark for thinking with multiple images, where models must interpret each image, combine cross-view evidence, and answer diagnostic questions with intermediate supervision and step-level evaluation. The dataset contains 8,067 cases, including 720 test cases, with an average of 6.62 images per case, substantially denser than prior work, whose expert-level benchmarks use at most 1.43 images per case. On the test set, the best closed-source models, Claude-4.6-Opus, Gemini-3-Pro, and GPT-5.2-xhigh, reach only 57.2%, 55.3%, and 54.9% accuracy, while GPT-5-mini and GPT-5-nano reach 39.7% and 30.8%. Strong open-source models lag behind, led by Qwen3.5-397B-A17B at 52.2% and Qwen3.5-27B at 50.6%. Further analysis identifies grounded multi-image reasoning as the main bottleneck: models often fail to extract, align, and compose evidence across views before higher-level inference can help. Providing expert single-image cues and cross-image summaries improves performance, whereas replacing them with self-generated intermediates reduces accuracy. Step-level analysis shows that over 70% of errors arise from image reading and cross-view integration. Scaling results further show that additional inference-time computation helps only when visual grounding is already reliable; when early evidence extraction is weak, longer reasoning yields limited or unstable gains and can amplify misread cues. These results suggest that the key challenge is not reasoning length alone, but reliable mechanisms for grounding, aligning, and composing distributed evidence across real-world multimodal clinical inputs.

Santiago Gonzalez, Alireza Amiri Bavandpour, Peter Ye et al.

As Large Language Models (LLMs) saturate elementary benchmarks, the research frontier has shifted from generation to the reliability of automated evaluation. We demonstrate that standard "LLM-as-a-Judge" protocols suffer from a systematic Alignment Gap when applied to upper-undergraduate to early graduate level mathematics. To quantify this, we introduce QEDBench, the first large-scale dual-rubric alignment benchmark to systematically measure alignment with human experts on university-level math proofs by contrasting course-specific rubrics against expert common knowledge criteria. By deploying a dual-evaluation matrix (7 judges x 5 solvers) against 1,000+ hours of human evaluation, we reveal that certain frontier evaluators like Claude Opus 4.5, DeepSeek-V3, Qwen 2.5 Max, and Llama 4 Maverick exhibit significant positive bias (up to +0.18, +0.20, +0.30, +0.36 mean score inflation, respectively). Furthermore, we uncover a critical reasoning gap in the discrete domain: while Gemini 3.0 Pro achieves state-of-the-art performance (0.91 average human evaluation score), other reasoning models like GPT-5 Pro and Claude Sonnet 4.5 see their performance significantly degrade in discrete domains. Specifically, their average human evaluation scores drop to 0.72 and 0.63 in Discrete Math, and to 0.74 and 0.50 in Graph Theory. In addition to these research results, we also release QEDBench as a public benchmark for evaluating and improving AI judges. Our benchmark is publicly published at https://github.com/qqliu/Yale-QEDBench.

Yilun Zhao, Yitao Long, Hongjun Liu et al.

Recent LLMs have demonstrated remarkable performance in solving exam-like math word problems. However, the degree to which these numerical reasoning skills are effective in real-world scenarios, particularly in expert domains, is still largely unexplored. This paper introduces DocMath-Eval, a comprehensive benchmark specifically designed to evaluate the numerical reasoning capabilities of LLMs in the context of understanding and analyzing specialized documents containing both text and tables. We conduct an extensive evaluation of 48 LLMs with Chain-of-Thought and Program-of-Thought prompting methods, aiming to comprehensively assess the capabilities and limitations of existing LLMs in DocMath-Eval. We found that even the current best-performing system (i.e., GPT-4o) still significantly lags behind human experts in solving complex numerical reasoning problems grounded in long contexts. We believe that DocMath-Eval can serve as a valuable benchmark for evaluating LLMs' capabilities in solving challenging numerical reasoning problems within expert domains.

Yilun Zhao, Hongjun Liu, Yitao Long et al.

We introduce FinanceMath, a novel benchmark designed to evaluate LLMs' capabilities in solving knowledge-intensive math reasoning problems. Compared to prior works, this study features three core advancements. First, FinanceMath includes 1,200 problems with a hybrid of textual and tabular content. These problems require college-level knowledge in the finance domain for effective resolution. Second, we provide expert-annotated, detailed solution references in Python program format, ensuring a high-quality benchmark for LLM assessment. We also construct a finance-domain knowledge bank and investigate various knowledge integration strategies. Finally, we evaluate a wide spectrum of 44 LLMs with both Chain-of-Thought and Program-of-Thought prompting methods. Our experimental results reveal that the current best-performing system (i.e., GPT-4o) achieves only 60.9% accuracy using CoT prompting, leaving substantial room for improvement. Moreover, while augmenting LLMs with external knowledge can improve model performance (e.g., from 47.5% to 54.5% for Gemini-1.5-Pro), their accuracy remains significantly lower than the estimated human expert performance of 92%. We believe that FinanceMath can advance future research in the area of domain-specific knowledge retrieval and integration, particularly within the context of solving reasoning-intensive tasks.

Jinbiao Wei, Qianran Ma, Yilun Zhao et al.

We present OpenComputer, a verifier-grounded framework for constructing verifiable software worlds for computer-use agents. OpenComputer integrates four components: (1) app-specific state verifiers that expose structured inspection endpoints over real applications, (2) a self-evolving verification layer that improves verifier reliability using execution-grounded feedback, (3) a task-generation pipeline that synthesizes realistic and machine-checkable desktop tasks, and (4) an evaluation harness that records full trajectories and computes auditable partial-credit rewards. In its current form, OpenComputer covers 33 desktop applications and 1,000 finalized tasks spanning browsers, office tools, creative software, development environments, file managers, and communication applications. Experiments show that OpenComputer's hard-coded verifiers align more closely with human adjudication than LLM-as-judge evaluation, especially when success depends on fine-grained application state. Frontier agents struggle with end-to-end completion despite partial progress, and open-source models exhibit sharp drops from their OSWorld-Verified scores, exposing a persistent gap in robust computer automation.

Linyong Nan, Ellen Zhang, Weijin Zou et al.

This study introduces a new long-form database question answering dataset designed to evaluate how Large Language Models (LLMs) interact with a SQL interpreter. The task necessitates LLMs to strategically generate multiple SQL queries to retrieve sufficient data from a database, to reason with the acquired context, and to synthesize them into a comprehensive analytical narrative. Our findings highlight that this task poses great challenges even for the state-of-the-art GPT-4 model. We propose and evaluate two interaction strategies, and provide a fine-grained analysis of the individual stages within the interaction. A key discovery is the identification of two primary bottlenecks hindering effective interaction: the capacity for planning and the ability to generate multiple SQL queries. To address the challenge of accurately assessing answer quality, we introduce a multi-agent evaluation framework that simulates the academic peer-review process, enhancing the precision and reliability of our evaluations. This framework allows for a more nuanced understanding of the strengths and limitations of current LLMs in complex retrieval and reasoning tasks.

Yixin Liu, Pengfei Liu, Arman Cohan

Direct Preference Optimization (DPO) has become a widely used training method for the instruction fine-tuning of large language models (LLMs). In this work, we explore an under-investigated aspect of DPO - its dependency on the reference model or policy. Such reference policies, typically instantiated as the model to be further fine-tuned, are important since they can impose an upper limit on DPO's effectiveness. Therefore, we address three related research questions in this work. First, we explore the optimal strength of the KL divergence constraint in DPO, which penalizes deviations from the reference policy, and find that DPO is sensitive to this strength. Next, we examine the necessity of the KL-constraint from the reference policies in DPO by providing both theoretical and empirical comparisons between DPO and related learning objectives, demonstrating DPO's superiority in this controlled setting. Additionally, we investigate whether DPO benefits from stronger reference policies, finding that a stronger reference policy can lead to improved performance, but only when it is similar to the model being fine-tuned. Our findings highlight the confounding role of reference policies in DPO and offer insights for best practices, while also identifying open research questions for future studies.

Gabrielle Kaili-May Liu, Arman Cohan

LLMs' linguistically expressed confidence should faithfully reflect their intrinsic uncertainty. While recent work shows LLMs struggle to use epistemic markers (e.g., "it is likely...") in a human-aligned fashion, it remains unclear whether models can apply their own linguistic confidence framework to associate markers with specific confidence levels in a stable and generalizable way, and how contextual features impact this ability. We conduct the first systematic study of this question, formalizing _marker internal confidence_ (MIC) as the estimated intrinsic confidence a model associates with a specific epistemic marker in a given task domain. We present 7 metrics to evaluate the stability of MICs within and across distributions. Applying our analysis framework to diverse models and tasks, we find that LLMs remain faithfully miscalibrated even under model-centric interpretation of marker meanings, struggling to differentiate markers by internal confidence across distributions despite preserving a somewhat consistent ranking order across tasks. This supplies critical, complementary evidence to existing work toward a holistic understanding of faithful calibration in LLMs, emphasizing the need for more aligned and stable marker use to improve trustworthiness and reliability.

Jinu Lee, Kyoung-Woon On, Simeng Han et al.

Evaluating the quality of LLM-generated reasoning traces in expert domains (e.g., law) is essential for ensuring credibility and explainability, yet remains challenging due to the inherent complexity of such reasoning tasks. We introduce LEGIT (LEGal Issue Trees), a novel large-scale (24K instances) expert-level legal reasoning dataset with an emphasis on reasoning trace evaluation. We convert court judgments into hierarchical trees of opposing parties' arguments and the court's conclusions, which serve as rubrics for evaluating the issue coverage and correctness of the reasoning traces. We verify the reliability of these rubrics via human expert annotations and comparison with coarse, less informative rubrics. Using the LEGIT dataset, we show that (1) LLMs' legal reasoning ability is seriously affected by both legal issue coverage and correctness, and that (2) retrieval-augmented generation (RAG) and RL with rubrics bring complementary benefits for legal reasoning abilities, where RAG improves overall reasoning capability, whereas RL improves correctness albeit with reduced coverage.

Sihong Wu, Yiling Ma, Yilun Zhao et al.

Large language models (LLMs) are increasingly used across the scientific workflow, including to draft peer-review reports. However, many AI-generated reviews are superficial and insufficiently actionable, leaving authors without concrete, implementable guidance and motivating the gap this work addresses. We propose RbtAct, which targets actionable review feedback generation and places existing peer review rebuttal at the center of learning. Rebuttals show which reviewer comments led to concrete revisions or specific plans, and which were only defended. Building on this insight, we leverage rebuttal as implicit supervision to directly optimize a feedback generator for actionability. To support this objective, we propose a new task called perspective-conditioned segment-level review feedback generation, in which the model is required to produce a single focused comment based on the complete paper and a specified perspective such as experiments and writing. We also build a large dataset named RMR-75K that maps review segments to the rebuttal segments that address them, with perspective labels and impact categories that order author uptake. We then train the Llama-3.1-8B-Instruct model with supervised fine-tuning on review segments followed by preference optimization using rebuttal derived pairs. Experiments with human experts and LLM-as-a-judge show consistent gains in actionability and specificity over strong baselines while maintaining grounding and relevance.

Tianyu Yang, Sihong Wu, Yilun Zhao et al.

Multimodal Mathematical Reasoning (MMR) has recently attracted increasing attention for its capability to solve mathematical problems that involve both textual and visual modalities. However, current models still face significant challenges in real-world visual math tasks. They often misinterpret diagrams, fail to align mathematical symbols with visual evidence, and produce inconsistent reasoning steps. Moreover, existing evaluations mainly focus on checking final answers rather than verifying the correctness or executability of each intermediate step. To address these limitations, a growing body of recent research addresses these issues by integrating structured perception, explicit alignment, and verifiable reasoning within unified frameworks. To establish a clear roadmap for understanding and comparing different MMR approaches, we systematically study them around four fundamental questions: (1) What to extract from multimodal inputs, (2) How to represent and align textual and visual information, (3) How to perform the reasoning, and (4) How to evaluate the correctness of the overall reasoning process. Finally, we discuss open challenges and offer perspectives on promising directions for future research.

Jinbiao Wei, Yilun Zhao, Kangqi Ni et al.

End-to-end GUI agents for real desktop environments require large amounts of high-quality interaction data, yet collecting human demonstrations is expensive and existing synthetic pipelines often suffer from limited task diversity or noisy, goal-drifting trajectories. We present a trajectory expansion framework Anchor that bootstraps scalable desktop supervision from a small set of verified seed demonstrations. Starting from each seed, we identify branch points that correspond to meaningful state changes and propose new, state-grounded task variants conditioned on the current GUI context. An executing agent then follows the proposed instructions to generate new trajectories, while a verifier enforces task completion via state-aware checks and trajectory-level consistency. To improve supervision quality, we further apply task-conditioned step-level filtering to remove ungrounded actions and denoise post-branch segments to maintain coherent intent. Experiments on standard desktop benchmarks, OSWorld and WindowsAgentArena, show that models fine-tuned on our expanded corpus achieve consistent improvements over zero-shot agents and representative synthesis baselines, and generalize across applications and operating systems.

Tiansheng Hu, Yilun Zhao, Canyu Zhang et al.

Deep research agents have emerged as powerful systems for addressing complex queries. Meanwhile, LLM-based retrievers have demonstrated strong capability in following instructions or reasoning. This raises a critical question: can LLM-based retrievers effectively contribute to deep research agent workflows? To investigate this, we introduce SAGE, a benchmark for scientific literature retrieval comprising 1,200 queries across four scientific domains, with a 200,000 paper retrieval corpus. We evaluate six deep research agents and find that all systems struggle with reasoning-intensive retrieval. Using DR Tulu as backbone, we further compare BM25 and LLM-based retrievers (i.e., ReasonIR and gte-Qwen2-7B-instruct) as alternative search tools. Surprisingly, BM25 significantly outperforms LLM-based retrievers by approximately 30%, as existing agents generate keyword-oriented sub-queries. To improve performance, we propose a corpus-level test-time scaling framework that uses LLMs to augment documents with metadata and keywords, making retrieval easier for off-the-shelf retrievers. This yields 8% and 2% gains on short-form and open-ended questions, respectively.

Zhaojian Yu, Kaiyue Feng, Yilun Zhao et al.

Large language models have made significant progress in complex but easy-to-verify problems, yet they still struggle with discovering the unknown. In this paper, we present \textbf{AlphaResearch}, an autonomous research agent designed to discover new algorithms on open-ended problems. To synergize the feasibility and innovation of the discovery process, we construct a novel dual research environment by combining the execution-based verify and simulated real-world peer review environment. AlphaResearch discovers new algorithm by iteratively running the following steps: (1) propose new ideas (2) verify the ideas in the dual research environment (3) optimize the research proposals for better performance. To promote a transparent evaluation process, we construct \textbf{AlphaResearchComp}, a new evaluation benchmark that includes an eight open-ended algorithmic problems competition, with each problem carefully curated and verified through executable pipelines, objective metrics, and reproducibility checks. AlphaResearch gets a 2/8 win rate in head-to-head comparison with human researchers, demonstrate the possibility of accelerating algorithm discovery with LLMs. Notably, the algorithm discovered by AlphaResearch on the \emph{``packing circles''} problem achieves the best-of-known performance, surpassing the results of human researchers and strong baselines from recent work (e.g., AlphaEvolve). Additionally, we conduct a comprehensive analysis of the remaining challenges of the 6/8 failure cases, providing valuable insights for future research.

Andrew M. Bean, Ryan Othniel Kearns, Angelika Romanou et al.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

Ryo Kamoi, Sarkar Snigdha Sarathi Das, Renze Lou et al.

With Large Language Models (LLMs) being widely used across various tasks, detecting errors in their responses is increasingly crucial. However, little research has been conducted on error detection of LLM responses. Collecting error annotations on LLM responses is challenging due to the subjective nature of many NLP tasks, and thus previous research focuses on tasks of little practical value (e.g., word sorting) or limited error types (e.g., faithfulness in summarization). This work introduces ReaLMistake, the first error detection benchmark consisting of objective, realistic, and diverse errors made by LLMs. ReaLMistake contains three challenging and meaningful tasks that introduce objectively assessable errors in four categories (reasoning correctness, instruction-following, context-faithfulness, and parameterized knowledge), eliciting naturally observed and diverse errors in responses of GPT-4 and Llama 2 70B annotated by experts. We use ReaLMistake to evaluate error detectors based on 12 LLMs. Our findings show: 1) Top LLMs like GPT-4 and Claude 3 detect errors made by LLMs at very low recall, and all LLM-based error detectors perform much worse than humans. 2) Explanations by LLM-based error detectors lack reliability. 3) LLMs-based error detection is sensitive to small changes in prompts but remains challenging to improve. 4) Popular approaches to improving LLMs, including self-consistency and majority vote, do not improve the error detection performance. Our benchmark and code are provided at https://github.com/psunlpgroup/ReaLMistake.

Zhaoling Chen, Xiangru Tang, Gangda Deng et al.

Code localization--identifying precisely where in a codebase changes need to be made--is a fundamental yet challenging task in software maintenance. Existing approaches struggle to efficiently navigate complex codebases when identifying relevant code sections. The challenge lies in bridging natural language problem descriptions with the appropriate code elements, often requiring reasoning across hierarchical structures and multiple dependencies. We introduce LocAgent, a framework that addresses code localization through graph-based representation. By parsing codebases into directed heterogeneous graphs, LocAgent creates a lightweight representation that captures code structures (files, classes, functions) and their dependencies (imports, invocations, inheritance), enabling LLM agents to effectively search and locate relevant entities through powerful multi-hop reasoning. Experimental results on real-world benchmarks demonstrate that our approach significantly enhances accuracy in code localization. Notably, our method with the fine-tuned Qwen-2.5-Coder-Instruct-32B model achieves comparable results to SOTA proprietary models at greatly reduced cost (approximately 86% reduction), reaching up to 92.7% accuracy on file-level localization while improving downstream GitHub issue resolution success rates by 12% for multiple attempts (Pass@10). Our code is available at https://github.com/gersteinlab/LocAgent.

Xiangru Tang, Daniel Shao, Jiwoong Sohn et al.

Large Language Models (LLMs) have shown impressive performance on existing medical question-answering benchmarks. This high performance makes it increasingly difficult to meaningfully evaluate and differentiate advanced methods. We present MedAgentsBench, a benchmark that focuses on challenging medical questions requiring multi-step clinical reasoning, diagnosis formulation, and treatment planning-scenarios where current models still struggle despite their strong performance on standard tests. Drawing from seven established medical datasets, our benchmark addresses three key limitations in existing evaluations: (1) the prevalence of straightforward questions where even base models achieve high performance, (2) inconsistent sampling and evaluation protocols across studies, and (3) lack of systematic analysis of the interplay between performance, cost, and inference time. Through experiments with various base models and reasoning methods, we demonstrate that the latest thinking models, DeepSeek R1 and OpenAI o3, exhibit exceptional performance in complex medical reasoning tasks. Additionally, advanced search-based agent methods offer promising performance-to-cost ratios compared to traditional approaches. Our analysis reveals substantial performance gaps between model families on complex questions and identifies optimal model selections for different computational constraints. Our benchmark and evaluation framework are publicly available at https://github.com/gersteinlab/medagents-benchmark.

Xiangru Tang, Tianyu Hu, Muyang Ye et al.

Chemical reasoning usually involves complex, multi-step processes that demand precise calculations, where even minor errors can lead to cascading failures. Furthermore, large language models (LLMs) encounter difficulties handling domain-specific formulas, executing reasoning steps accurately, and integrating code effectively when tackling chemical reasoning tasks. To address these challenges, we present ChemAgent, a novel framework designed to improve the performance of LLMs through a dynamic, self-updating library. This library is developed by decomposing chemical tasks into sub-tasks and compiling these sub-tasks into a structured collection that can be referenced for future queries. Then, when presented with a new problem, ChemAgent retrieves and refines pertinent information from the library, which we call memory, facilitating effective task decomposition and the generation of solutions. Our method designs three types of memory and a library-enhanced reasoning component, enabling LLMs to improve over time through experience. Experimental results on four chemical reasoning datasets from SciBench demonstrate that ChemAgent achieves performance gains of up to 46% (GPT-4), significantly outperforming existing methods. Our findings suggest substantial potential for future applications, including tasks such as drug discovery and materials science. Our code can be found at https://github.com/gersteinlab/chemagent

Yukun Huang, Yixin Liu, Raghuveer Thirukovalluru et al.

To enhance Large Language Models' (LLMs) reliability, calibration is essential -- the model's assessed confidence scores should align with the actual likelihood of its responses being correct. However, current confidence elicitation methods and calibration metrics typically rely on a binary true/false assessment of response correctness. This approach does not apply to long-form generation, where an answer can be partially correct. Addressing this gap, we introduce a unified calibration framework, in which both the correctness of the LLMs' responses and their associated confidence levels are treated as distributions across a range of scores. Within this framework, we develop three metrics to precisely evaluate LLM calibration and further propose two confidence elicitation methods based on self-consistency and self-evaluation. Our experiments, which include long-form QA and summarization tasks, demonstrate that larger models don't necessarily guarantee better calibration, that calibration performance is found to be metric-dependent, and that self-consistency methods excel in factoid datasets. We also find that calibration can be enhanced through techniques such as fine-tuning, integrating relevant source documents, scaling the temperature, and combining self-consistency with self-evaluation. Lastly, we showcase a practical application of our system: selecting and cascading open-source models and ChatGPT to optimize correctness given a limited API budget. This research not only challenges existing notions of LLM calibration but also offers practical methodologies for improving trustworthiness in long-form generation.

Lorenzo Jaime Yu Flores, Heyuan Huang, Kejian Shi et al.

Text simplification has emerged as an increasingly useful application of AI for bridging the communication gap in specialized fields such as medicine, where the lexicon is often dominated by technical jargon and complex constructs. Despite notable progress, methods in medical simplification sometimes result in the generated text having lower quality and diversity. In this work, we explore ways to further improve the readability of text simplification in the medical domain. We propose (1) a new unlikelihood loss that encourages generation of simpler terms and (2) a reranked beam search decoding method that optimizes for simplicity, which achieve better performance on readability metrics on three datasets. This study's findings offer promising avenues for improving text simplification in the medical field.

Leyao Wang, Yanan He, Peng Chen et al.

Deep research agents increasingly automate complex information-seeking tasks, producing evidence-grounded reports via multi-step reasoning, tool use, and synthesis. Their growing role demands scalable, reliable evaluation, positioning LLM-as-judge as a supervision paradigm for assessing factual accuracy, evidence use, and reasoning quality. Yet the reliability of these judges for deep research agents remains poorly understood, posing a critical meta-evaluation problem: before deploying LLM judges to supervise research agents, we must first evaluate the judges themselves. Existing meta-evaluations fall short in two ways: (1) reliance on coarse, subjective human-preference agreement; (2) focus on instruction-following or verifiable tasks, leaving open-ended agent executions unexplored. To address these gaps, we introduce REFLECT (REliable Fine-grained LLM judge Evaluation via Controlled inTervention), a meta-evaluation benchmark targeting fine-grained failure detection in agentic environments. REFLECT defines a detailed taxonomy of process- and outcome-level failure modes, instantiated by performing controlled and localized interventions on quality-screened agent execution traces. This yields verifiable, comprehensive, and fine-grained instances for validating the judge models. Our experiments show that current LLM judges remain unreliable: even the best-performing models achieve overall accuracies below 55% across reasoning, tool-use, and report-quality failures, with especially poor performance on evidence verification. Together, our taxonomy and findings expose systematic judge limitations, reveal tradeoffs in cost and reliability, and offer actionable guidance for building more reliable evaluation pipelines for deep research agents.

Yixin Liu, Yue Yu, DiJia Su et al.

Reasoning LLMs-as-Judges, which can benefit from inference-time scaling, provide a promising path for extending the success of reasoning models to non-verifiable domains where the output correctness/quality cannot be directly checked. However, while reasoning judges have shown better performance on static evaluation benchmarks, their effectiveness in actual policy training has not been systematically examined. Therefore, we conduct a rigorous study to investigate the actual impact of non-reasoning and reasoning judges in reinforcement-learning-based LLM alignment. Our controlled synthetic setting, where a "gold-standard" judge (gpt-oss-120b) provides preference annotations to train smaller judges, reveals key differences between non-reasoning and reasoning judges: non-reasoning judges lead to reward hacking easily, while reasoning judges can lead to policies that achieve strong performance when evaluated by the gold-standard judge. Interestingly, we find that the reasoning-judge-trained policies achieve such strong performance by learning to generate highly effective adversarial outputs that can also score well on popular benchmarks such as Arena-Hard by deceiving other LLM-judges. Combined with our further analysis, our study highlights both important findings and room for improvements for applying (reasoning) LLM-judges in non-verifiable LLM post-training.

Balaji Dinesh Gangireddi, Aniketh Garikaparthi, Manasi Patwardhan et al.

Existing prompt-optimization techniques rely on local signals to update behavior, often neglecting broader and recurring patterns across tasks, leading to poor generalization; they further rely on full-prompt rewrites or unstructured merges, resulting in knowledge loss. These limitations are magnified in research-coding workflows, which involve heterogeneous repositories, underspecified environments, and weak feedback, where reproducing results from public codebases is an established evaluation regime. We introduce Reflective Evolving Research Engineer (REVERE), a framework that continuously learns from Global Training Context, recognizes recurring failure modes in cross-repository execution trajectories, distills them into reusable heuristics, and performs targeted edits across three configurable fields: the system prompt, a task-prompt template, and a cumulative cheatsheet. REVERE, via this reflective optimization framework, improves performance over prior state-of-the-art expert-crafted instructions on research coding tasks by 4.50% on SUPER, 3.51% on ResearchCodeBench, and 4.89% on ScienceAgentBench across their respective metrics. These results demonstrate that agents equipped with mechanisms for continual learning and global memory consolidation can meaningfully evolve their capabilities over time.

Kejian Shi, Yixin Liu, Peifeng Wang et al.

While Reinforcement Learning with Verifiable Rewards (RLVR) has shown strong effectiveness in reasoning tasks, it cannot be directly applied to non-verifiable domains lacking ground-truth verifiers, such as LLM alignment. In this work, we investigate whether reference-guided LLM-evaluators can bridge this gap by serving as soft "verifiers". First, we design evaluation protocols that enhance LLM-based evaluators for LLM alignment using reference outputs. Through comprehensive experiments, we show that a reference-guided approach substantially improves the accuracy of less capable LLM-judges using references from frontier models; stronger LLM-judges can also be enhanced by high-quality (i.e., human-written) references. Building on these improved judges, we demonstrate the utility of high-quality references in alignment tuning, where LLMs guided with references are used as judges to self-improve. We show that reference-guided self-improvement yields clear gains over both direct SFT on reference outputs and self-improvement with reference-free judges, achieving performance comparable to training with ArmoRM, a strong finetuned reward model. Specifically, our method achieves 73.1% and 58.7% on AlpacaEval and Arena-Hard with Llama-3-8B-Instruct, and 70.0% and 74.1% with Qwen2.5-7B, corresponding to average absolute gains of +20.2 / +17.1 points over SFT distillation and +5.3 / +3.6 points over reference-free self-improvement on AlpacaEval / Arena-Hard. These results highlight the potential of using reference-guided LLM-evaluators to enable effective LLM post-training in non-verifiable domains.

Gabrielle Kaili-May Liu, Bowen Shi, Avi Caciularu et al.

Multi-document (MD) processing is crucial for LLMs to handle real-world tasks such as summarization and question-answering across large sets of documents. While LLMs have improved at processing long inputs, MD contexts still present unique difficulties, including management of inter-document dependencies, redundancy, and incoherent structures. To address this challenge, we introduce MDCure, a scalable and effective instruction data generation framework to enhance the MD capabilities of LLMs without the computational cost of pre-training or reliance on human-annotated data. MDCure generates high-quality synthetic MD instruction data over sets of articles via targeted prompts. We also introduce MDCureRM, a cost-effective, MD-specific reward model to score and filter generated data based on their training utility for MD settings. MDCure is compatible with open- and closed-source models in addition to policy optimization methods such as PPO, enabling even small open-source models to surpass proprietary LLMs as strong generators of high-quality MD instruction data without further data filtering. With MDCure, we fine-tune a wide variety of LLMs up to 70B parameters in size from the FlanT5, Qwen2, and LLAMA3.1 model families. Extensive evaluations on a wide range of MD and long-context benchmarks spanning various tasks and domains show MDCure consistently improves performance over pre-trained baselines and base models by up to 75.1%. Our code, datasets, and models are available at https://github.com/yale-nlp/MDCure.

Yilun Zhao, Kaiyan Zhang, Tiansheng Hu et al. · allen-ai

We present SciArena, an open and collaborative platform for evaluating foundation models on scientific literature tasks. Unlike traditional benchmarks for scientific literature understanding and synthesis, SciArena engages the research community directly, following the Chatbot Arena evaluation approach of community voting on model comparisons. By leveraging collective intelligence, SciArena offers a community-driven evaluation of model performance on open-ended scientific tasks that demand literature-grounded, long-form responses. The platform currently supports 23 open-source and proprietary foundation models and has collected over 13,000 votes from trusted researchers across diverse scientific domains. We analyze the data collected so far and confirm that the submitted questions are diverse, aligned with real-world literature needs, and that participating researchers demonstrate strong self-consistency and inter-annotator agreement in their evaluations. We discuss the results and insights based on the model ranking leaderboard. To further promote research in building model-based automated evaluation systems for literature tasks, we release SciArena-Eval, a meta-evaluation benchmark based on our collected preference data. The benchmark measures the accuracy of models in judging answer quality by comparing their pairwise assessments with human votes. Our experiments highlight the benchmark's challenges and emphasize the need for more reliable automated evaluation methods.

Jaesung Tae, Hamish Ivison, Sachin Kumar et al. · uw

We introduce TESS 2, a general instruction-following diffusion language model that outperforms contemporary instruction-tuned diffusion models, as well as matches and sometimes exceeds strong autoregressive (AR) models. We train TESS 2 by first adapting a strong AR model via continued pretraining with the usual cross-entropy as diffusion loss, and then performing further instruction tuning. We find that adaptation training as well as the choice of the base model is crucial for training good instruction-following diffusion models. We further propose reward guidance, a novel and modular inference-time guidance procedure to align model outputs without needing to train the underlying model. Finally, we show that TESS 2 further improves with increased inference-time compute, highlighting the utility of diffusion LMs in having fine-grained controllability over the amount of compute used at inference time. Code and models are available at https://github.com/hamishivi/tess-2.

Shruti Singh, Nandan Sarkar, Arman Cohan

Scientific literature is typically dense, requiring significant background knowledge and deep comprehension for effective engagement. We introduce SciDQA, a new dataset for reading comprehension that challenges LLMs for a deep understanding of scientific articles, consisting of 2,937 QA pairs. Unlike other scientific QA datasets, SciDQA sources questions from peer reviews by domain experts and answers by paper authors, ensuring a thorough examination of the literature. We enhance the dataset's quality through a process that carefully filters out lower quality questions, decontextualizes the content, tracks the source document across different versions, and incorporates a bibliography for multi-document question-answering. Questions in SciDQA necessitate reasoning across figures, tables, equations, appendices, and supplementary materials, and require multi-document reasoning. We evaluate several open-source and proprietary LLMs across various configurations to explore their capabilities in generating relevant and factual responses. Our comprehensive evaluation, based on metrics for surface-level similarity and LLM judgements, highlights notable performance discrepancies. SciDQA represents a rigorously curated, naturally derived scientific QA dataset, designed to facilitate research on complex scientific text understanding.