Ryo Kamoi

Yusen Zhang, Nan Zhang, Yixin Liu et al.

People from different social and demographic groups express diverse perspectives and conflicting opinions on a broad set of topics such as product reviews, healthcare, law, and politics. A fair summary should provide a comprehensive coverage of diverse perspectives without underrepresenting certain groups. However, current work in summarization metrics and Large Language Models (LLMs) evaluation has not explored fair abstractive summarization. In this paper, we systematically investigate fair abstractive summarization for user-generated data. We first formally define fairness in abstractive summarization as not underrepresenting perspectives of any groups of people, and we propose four reference-free automatic metrics by measuring the differences between target and source perspectives. We evaluate nine LLMs, including three GPT models, four LLaMA models, PaLM 2, and Claude, on six datasets collected from social media, online reviews, and recorded transcripts. Experiments show that both the model-generated and the human-written reference summaries suffer from low fairness. We conduct a comprehensive analysis of the common factors influencing fairness and propose three simple but effective methods to alleviate unfair summarization. Our dataset and code are available at https://github.com/psunlpgroup/FairSumm.

Ryo Kamoi, Tanya Goyal, Juan Diego Rodriguez et al.

Textual entailment models are increasingly applied in settings like fact-checking, presupposition verification in question answering, or summary evaluation. However, these represent a significant domain shift from existing entailment datasets, and models underperform as a result. We propose WiCE, a new fine-grained textual entailment dataset built on natural claim and evidence pairs extracted from Wikipedia. In addition to standard claim-level entailment, WiCE provides entailment judgments over sub-sentence units of the claim, and a minimal subset of evidence sentences that support each subclaim. To support this, we propose an automatic claim decomposition strategy using GPT-3.5 which we show is also effective at improving entailment models' performance on multiple datasets at test time. Finally, we show that real claims in our dataset involve challenging verification and retrieval problems that existing models fail to address.

Ryo Kamoi, Tanya Goyal, Greg Durrett

Despite recent progress in abstractive summarization, models often generate summaries with factual errors. Numerous approaches to detect these errors have been proposed, the most popular of which are question answering (QA)-based factuality metrics. These have been shown to work well at predicting summary-level factuality and have potential to localize errors within summaries, but this latter capability has not been systematically evaluated in past research. In this paper, we conduct the first such analysis and find that, contrary to our expectations, QA-based frameworks fail to correctly identify error spans in generated summaries and are outperformed by trivial exact match baselines. Our analysis reveals a major reason for such poor localization: questions generated by the QG module often inherit errors from non-factual summaries which are then propagated further into downstream modules. Moreover, even human-in-the-loop question generation cannot easily offset these problems. Our experiments conclusively show that there exist fundamental issues with localization using the QA framework which cannot be fixed solely by stronger QA and QG models.

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.

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.

Ryo Kamoi, Yusen Zhang, Sarkar Snigdha Sarathi Das et al.

Large Vision Language Models (LVLMs) have achieved remarkable performance in various vision-language tasks. However, it is still unclear how accurately LVLMs can perceive visual information in images. In particular, the capability of LVLMs to perceive geometric information, such as shape, angle, and size, remains insufficiently analyzed, although the perception of these properties is crucial for tasks that require a detailed visual understanding. In this work, we introduce VisOnlyQA, a dataset for evaluating the geometric perception of LVLMs, and reveal that LVLMs often cannot accurately perceive basic geometric information in images, while human performance is nearly perfect. VisOnlyQA consists of 12 tasks that directly ask about geometric information in geometric shapes, charts, chemical structures, and 3D shapes. Our experiments highlight the following findings: (i) State-of-the-art LVLMs struggle with basic geometric perception. 23 LVLMs we evaluate, including GPT-4o and Gemini 2.5 Pro, work poorly on VisOnlyQA. (ii) Additional training data does not resolve this issue. Fine-tuning on the training set of VisOnlyQA is not always effective, even for in-distribution tasks. (iii) LLM may be the bottleneck. LVLMs using stronger LLMs exhibit better geometric perception on VisOnlyQA, while it does not require complex reasoning, suggesting that the way LVLMs process information from visual encoders is a bottleneck. The datasets, code, and model responses are provided at https://github.com/psunlpgroup/VisOnlyQA.

Jihyun Janice Ahn, Ryo Kamoi, Berk Atil et al.

LLMs often generate seemingly valid answers to flawed or ill-posed inputs. This is not due to missing knowledge: under discriminative prompting, the same models can mostly identify such issues, yet fail to reflect this in standard generative responses. This reveals a fundamental know-act gap between discriminative recognition and generative behavior. Prior work largely characterizes this issue in narrow settings, such as math word problems or question answering, with limited focus on how to integrate these two modes. In this work, we present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty scientific questions. We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection (validate vs. answer) with content generation, preventing discriminative knowledge from being utilized. To address this, we propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision. Through self-distillation, the model unifies discriminative judgment and generative reasoning within a single backbone. Empirically, DeIllusionLLM substantially reduces answer-despite-error failures under natural prompting while maintaining general reasoning performance, demonstrating that self-distillation is an effective and scalable solution for bridging the discriminative-generative know-act gap

Renze Lou, Hanzi Xu, Sijia Wang et al.

Numerous studies have assessed the proficiency of AI systems, particularly large language models (LLMs), in facilitating everyday tasks such as email writing, question answering, and creative content generation. However, researchers face unique challenges and opportunities in leveraging LLMs for their own work, such as brainstorming research ideas, designing experiments, and writing or reviewing papers. In this study, we introduce AAAR-1.0, a benchmark dataset designed to evaluate LLM performance in three fundamental, expertise-intensive research tasks: (i) EquationInference, assessing the correctness of equations based on the contextual information in paper submissions; (ii) ExperimentDesign, designing experiments to validate research ideas and solutions; (iii) PaperWeakness, identifying weaknesses in paper submissions; and (iv) REVIEWCRITIQUE, identifying each segment in human reviews is deficient or not. AAAR-1.0 differs from prior benchmarks in two key ways: first, it is explicitly research-oriented, with tasks requiring deep domain expertise; second, it is researcher-oriented, mirroring the primary activities that researchers engage in on a daily basis. An evaluation of both open-source and proprietary LLMs reveals their potential as well as limitations in conducting sophisticated research tasks. We will keep iterating AAAR-1.0 to new versions.

Bowen Jiang, Taiwei Shi, Ryo Kamoi et al.

This paper introduces OMAR: One Model, All Roles, a reinforcement learning framework that enables AI to develop social intelligence through multi-turn, multi-agent conversational self-play. Unlike traditional paradigms that rely on static, single-turn optimizations, OMAR allows a single model to role-play all participants in a conversation simultaneously, learning to achieve long-term goals and complex social norms directly from dynamic social interaction. To ensure training stability across long dialogues, we implement a hierarchical advantage estimation that calculates turn-level and token-level advantages. Evaluations in the SOTOPIA social environment and Werewolf strategy games show that our trained models develop fine-grained, emergent social intelligence, such as empathy, persuasion, and compromise seeking, demonstrating the effectiveness of learning collaboration even under competitive scenarios. While we identify practical challenges like reward hacking, our results show that rich social intelligence can emerge without human supervision. We hope this work incentivizes further research on AI social intelligence in group conversations.

Sarkar Snigdha Sarathi Das, Ryo Kamoi, Bo Pang et al.

The effectiveness of large language models (LLMs) is closely tied to the design of prompts, making prompt optimization essential for enhancing their performance across a wide range of tasks. Many existing approaches to automating prompt engineering rely exclusively on textual feedback, refining prompts based solely on inference errors identified by large, computationally expensive LLMs. Unfortunately, smaller models struggle to generate high-quality feedback, resulting in complete dependence on large LLM judgment. Moreover, these methods fail to leverage more direct and finer-grained information, such as gradients, due to operating purely in text space. To this end, we introduce GReaTer, a novel prompt optimization technique that directly incorporates gradient information over task-specific reasoning. By utilizing task loss gradients, GReaTer enables self-optimization of prompts for open-source, lightweight language models without the need for costly closed-source LLMs. This allows high-performance prompt optimization without dependence on massive LLMs, closing the gap between smaller models and the sophisticated reasoning often needed for prompt refinement. Extensive evaluations across diverse reasoning tasks including BBH, GSM8k, and FOLIO demonstrate that GReaTer consistently outperforms previous state-of-the-art prompt optimization methods, even those reliant on powerful LLMs. Additionally, GReaTer-optimized prompts frequently exhibit better transferability and, in some cases, boost task performance to levels comparable to or surpassing those achieved by larger language models, highlighting the effectiveness of prompt optimization guided by gradients over reasoning. Code of GReaTer is available at https://github.com/psunlpgroup/GreaTer.

Ryo Kamoi, Ameya Godbole, Longqi Yang et al.

Simulating human conversations using large language models (LLMs) has emerged as a scalable methodology for modeling human social interaction. However, simulating human conversations is challenging because they inherently involve inconsistent and uncollaborative behaviors, such as misunderstandings and interruptions. Analysis comparing inconsistent and uncollaborative behaviors in human- and LLM-generated conversations remains limited, although reproducing these behaviors is integral to simulating human-like and complex social interaction. In this work, we introduce CoCoEval, an evaluation framework that analyzes LLM-simulated conversations by detecting 10 types of inconsistent and uncollaborative behaviors at the turn level using an LLM-as-a-Judge. Using CoCoEval, we evaluate GPT-4.1, GPT-5.1, and Claude Opus 4 by comparing the frequencies of detected behaviors in conversations simulated by each model and in human conversations across academic, business, and governmental meetings, as well as debates. Our analysis shows that (1) under vanilla prompting, LLM-simulated conversations exhibit far fewer inconsistent and uncollaborative behaviors than human conversations; (2) prompt engineering does not provide reliable control over these behaviors, as our results show that different prompts lead to their under- or overproduction; and (3) supervised fine-tuning on human conversations can lead LLMs to overproduce a narrow set of behaviors, such as repetition. Our findings highlight the difficulty of simulating human conversations, raising concerns about the use of LLMs as a proxy for human social interaction.

Ryo Kamoi, Yusen Zhang, Nan Zhang et al.

Process Reward Models (PRMs), which provide step-level feedback on reasoning traces generated by Large Language Models (LLMs), are receiving increasing attention. However, two key research gaps remain: creating PRM training data requires costly human annotation to label accurate step-level errors, and existing PRMs are limited to math reasoning domains. In response to these gaps, this paper aims to enable automatic synthesis of accurate PRM training data and the generalization of PRMs to diverse reasoning tasks beyond math reasoning. We propose FoVer, an approach to synthesize PRM training data with accurate step-level error labels automatically annotated by formal verification tools, such as Z3 and Isabelle. To show the practical effectiveness of FoVer, we synthesize a training dataset by annotating step-level error labels on LLM responses to formal logic and theorem proving tasks, without relying on human annotation. While FoVer creates training data with symbolic tasks compatible with formal verification, our experiments show that PRMs trained on our dataset exhibit cross-task generalization, enabling a single PRM to effectively perform verification across diverse reasoning tasks. Specifically, LLM-based PRMs trained with FoVer significantly outperform PRMs based on the original LLMs and achieve competitive or superior results compared to state-of-the-art PRMs, as measured by step-level verification on ProcessBench and Best-of-K performance across 12 reasoning benchmarks, including MATH, AIME, ANLI, MMLU, and BBH. The dataset and code are in the supplementary material and will be made public. The datasets, models, and code are provided at https://github.com/psunlpgroup/FoVer.

Jihyun Janice Ahn, Ryo Kamoi, Lu Cheng et al.

Mainstream LLM research has primarily focused on enhancing their generative capabilities. However, even the most advanced LLMs experience uncertainty in their outputs, often producing varied results on different runs or when faced with minor changes in input, despite no substantial change in content. Given multiple responses from the same LLM to the same input, we advocate leveraging the LLMs' discriminative capability to reduce this generative uncertainty, aiding in identifying the correct answers. Specifically, we propose and analyze three discriminative prompts: direct, inverse, and hybrid, to explore the potential of both closed-source and open-source LLMs in self-improving their generative performance on two benchmark datasets. Our insights reveal which discriminative prompt is most promising and when to use it. To our knowledge, this is the first work to systematically analyze LLMs' discriminative capacity to address generative uncertainty.

Yusen Zhang, Wenliang Zheng, Aashrith Madasu et al.

High-resolution image (HRI) understanding aims to process images with a large number of pixels, such as pathological images and agricultural aerial images, both of which can exceed 1 million pixels. Vision Large Language Models (VLMs) can allegedly handle HRIs, however, there is a lack of a comprehensive benchmark for VLMs to evaluate HRI understanding. To address this gap, we introduce HRScene, a novel unified benchmark for HRI understanding with rich scenes. HRScene incorporates 25 real-world datasets and 2 synthetic diagnostic datasets with resolutions ranging from 1,024 $\times$ 1,024 to 35,503 $\times$ 26,627. HRScene is collected and re-annotated by 10 graduate-level annotators, covering 25 scenarios, ranging from microscopic to radiology images, street views, long-range pictures, and telescope images. It includes HRIs of real-world objects, scanned documents, and composite multi-image. The two diagnostic evaluation datasets are synthesized by combining the target image with the gold answer and distracting images in different orders, assessing how well models utilize regions in HRI. We conduct extensive experiments involving 28 VLMs, including Gemini 2.0 Flash and GPT-4o. Experiments on HRScene show that current VLMs achieve an average accuracy of around 50% on real-world tasks, revealing significant gaps in HRI understanding. Results on synthetic datasets reveal that VLMs struggle to effectively utilize HRI regions, showing significant Regional Divergence and lost-in-middle, shedding light on future research.

Ryo Kamoi, Yusen Zhang, Nan Zhang et al.

Self-correction is an approach to improving responses from large language models (LLMs) by refining the responses using LLMs during inference. Prior work has proposed various self-correction frameworks using different sources of feedback, including self-evaluation and external feedback. However, there is still no consensus on the question of when LLMs can correct their own mistakes, as recent studies also report negative results. In this work, we critically survey broad papers and discuss the conditions required for successful self-correction. We first find that prior studies often do not define their research questions in detail and involve impractical frameworks or unfair evaluations that over-evaluate self-correction. To tackle these issues, we categorize research questions in self-correction research and provide a checklist for designing appropriate experiments. Our critical survey based on the newly categorized research questions shows that (1) no prior work demonstrates successful self-correction with feedback from prompted LLMs, except for studies in tasks that are exceptionally suited for self-correction, (2) self-correction works well in tasks that can use reliable external feedback, and (3) large-scale fine-tuning enables self-correction.

Ryo Kamoi, Kei Kobayashi

The Mahalanobis distance-based confidence score, a recently proposed anomaly detection method for pre-trained neural classifiers, achieves state-of-the-art performance on both out-of-distribution (OoD) and adversarial examples detection. This work analyzes why this method exhibits such strong performance in practical settings while imposing an implausible assumption; namely, that class conditional distributions of pre-trained features have tied covariance. Although the Mahalanobis distance-based method is claimed to be motivated by classification prediction confidence, we find that its superior performance stems from information not useful for classification. This suggests that the reason the Mahalanobis confidence score works so well is mistaken, and makes use of different information from ODIN, another popular OoD detection method based on prediction confidence. This perspective motivates us to combine these two methods, and the combined detector exhibits improved performance and robustness. These findings provide insight into the behavior of neural classifiers in response to anomalous inputs.

Ryo Kamoi, Kei Kobayashi

Recent work has shown that deep generative models assign higher likelihood to out-of-distribution inputs than to training data. We show that a factor underlying this phenomenon is a mismatch between the nature of the prior distribution and that of the data distribution, a problem found in widely used deep generative models such as VAEs and Glow. While a typical choice for a prior distribution is a standard Gaussian distribution, properties of distributions of real data sets may not be consistent with a unimodal prior distribution. This paper focuses on the relationship between the choice of a prior distribution and the likelihoods assigned to out-of-distribution inputs. We propose the use of a mixture distribution as a prior to make likelihoods assigned by deep generative models sensitive to out-of-distribution inputs. Furthermore, we explain the theoretical advantages of adopting a mixture distribution as the prior, and we present experimental results to support our claims. Finally, we demonstrate that a mixture prior lowers the out-of-distribution likelihood with respect to two pairs of real image data sets: Fashion-MNIST vs. MNIST and CIFAR10 vs. SVHN.