Reward models (RMs) play a crucial role in reinforcement learning from human feedback (RLHF), aligning model behavior with human preferences. However, existing benchmarks for reward models show a weak correlation with the performance of optimized policies, suggesting that they fail to accurately assess the true capabilities of RMs. To bridge this gap, we explore several evaluation designs through the lens of reward overoptimization, i.e., a phenomenon that captures both how well the reward model aligns with human preferences and the dynamics of the learning signal it provides to the policy. The results highlight three key findings on how to construct a reliable benchmark: (i) it is important to minimize differences between chosen and rejected responses beyond correctness, (ii) evaluating reward models requires multiple comparisons across a wide range of chosen and rejected responses, and (iii) given that reward models encounter responses with diverse representations, responses should be sourced from a variety of models. However, we also observe that a extremely high correlation with degree of overoptimization leads to comparatively lower correlation with certain downstream performance. Thus, when designing a benchmark, it is desirable to use the degree of overoptimization as a useful tool, rather than the end goal.

Large language models (LLMs) have recently gained much attention in building autonomous agents. However, performance of current LLM-based web agents in long-horizon tasks is far from optimal, often yielding errors such as repeatedly buying a non-refundable flight ticket. By contrast, humans can avoid such an irreversible mistake, as we have an awareness of the potential outcomes (e.g., losing money) of our actions, also known as the "world model". Motivated by this, our study first starts with preliminary analyses, confirming the absence of world models in current LLMs (e.g., GPT-4o, Claude-3.5-Sonnet, etc.). Then, we present a World-model-augmented (WMA) web agent, which simulates the outcomes of its actions for better decision-making. To overcome the challenges in training LLMs as world models predicting next observations, such as repeated elements across observations and long HTML inputs, we propose a transition-focused observation abstraction, where the prediction objectives are free-form natural language descriptions exclusively highlighting important state differences between time steps. Experiments on WebArena and Mind2Web show that our world models improve agents' policy selection without training and demonstrate our agents' cost- and time-efficiency compared to recent tree-search-based agents.

Seungone Kim, Juyoung Suk, Ji Yong Cho, Shayne Longpre, Chaeeun Kim, Dongkeun Yoon, Guijin Son, Yejin Cho, Sheikh Shafayat, Jinheon Baek, Sue Hyun Park, Hyeonbin Hwang, Jinkyung Jo, Hyowon Cho, Haebin Shin, Seongyun Lee, Hanseok Oh, Noah Lee, Namgyu Ho, Se June Joo, Miyoung Ko, Yoonjoo Lee, Hyungjoo Chae, Jamin Shin, Joel Jang, Seonghyeon Ye, Bill Yuchen Lin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, Minjoon Seo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Kai Tzu-iunn Ong, Namyoung Kim, Minju Gwak, Hyungjoo Chae, Taeyoon Kwon, Yohan Jo, Seung-won Hwang, Dongha Lee, Jinyoung Yeo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Web navigation is a unique domain that can automate many repetitive real-life tasks and is challenging as it requires long-horizon sequential decision making beyond typical multimodal large language model (MLLM) tasks. Yet, specialized reward models for web navigation that can be utilized during both training and test-time have been absent until now. Despite the importance of speed and cost-effectiveness, prior works have utilized MLLMs as reward models, which poses significant constraints for real-world deployment. To address this, in this work, we propose the first process reward model (PRM) called Web-Shepherd which could assess web navigation trajectories in a step-level. To achieve this, we first construct the WebPRM Collection, a large-scale dataset with 40K step-level preference pairs and annotated checklists spanning diverse domains and difficulty levels. Next, we also introduce the WebRewardBench, the first meta-evaluation benchmark for evaluating PRMs. In our experiments, we observe that our Web-Shepherd achieves about 30 points better accuracy compared to using GPT-4o on WebRewardBench. Furthermore, when testing on WebArena-lite by using GPT-4o-mini as the policy and Web-Shepherd as the verifier, we achieve 10.9 points better performance, in 10x less cost compared to using GPT-4o-mini as the verifier. Our model, dataset, and code are publicly available at https://github.com/kyle8581/Web-Shepherd.

Recent approaches in domain-specific named entity recognition (NER), such as biomedical NER, have shown remarkable advances.However, they still lack of faithfulness, producing erroneous predictions.We assume that knowledge of entities can be useful in verifying the correctness of the predictions.Despite the usefulness of knowledge, resolving such errors with knowledge is nontrivial, since the knowledge itself does not directly indicate the groundtruth label.To this end, we propose VER-IFINER, a post-hoc verification framework that identifies errors from existing NER methods using knowledge and revises them into more faithful predictions.Our framework leverages the reasoning abilities of large language models to adequately ground on knowledge and the contextual information in the verification process.We validate effectiveness of VERIFINER through extensive experiments on biomedical datasets.The results suggest that VERIFINER can successfully verify errors from existing models as a model-agnostic approach.Further analyses on out-of-domain and low-resource settings show the usefulness of VERIFINER on real-world applications.1

Hyungjoo Chae, Yeonghyeon Kim, Seungone Kim, Kai Tzu-iunn Ong, Beong-woo Kwak, Moohyeon Kim, Sunghwan Kim, Taeyoon Kwon, Jiwan Chung, Youngjae Yu, Jinyoung Yeo. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Hyungjoo Chae, Taeyoon Kwon, Seungjun Moon, Yongho Song, Dongjin Kang, Kai Tzu-iunn Ong, Beong-woo Kwak, Seonghyeon Bae, Seung-won Hwang, Jinyoung Yeo. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Recent studies have investigated utilizing Knowledge Graphs (KGs) to enhance Question Answering (QA) performance of Large Language Models (LLMs), yet structured KG verbalization remains challenging.Existing methods, such as triple-form or free-form textual conversion of triple-form facts, encounter several issues.These include reduced evidence density due to duplicated entities or relationships, and reduced evidence clarity due to an inability to emphasize crucial evidence.To address these issues, we propose EFSUM, an Evidence-focused Fact Summarization framework for enhanced QA with knowledge-augmented LLMs.We optimize an open-source LLM as a fact summarizer through distillation and preference alignment.Our extensive experiments show that EFSUM improves LLM's zero-shot QA performance, and it is possible to ensure both the helpfulness and faithfulness of the summary.

In this paper, we propose Tutoring bot, a generative chatbot trained on a large scale of tutor-student conversations for English-language learning. To mimic a human tutor's behavior in language education, the tutor bot leverages diverse educational instructions and grounds to each instruction as additional input context for the tutor response generation. As a single instruction generally involves multiple dialogue turns to give the student sufficient speaking practice, the tutor bot is required to monitor and capture when the current instruction should be kept or switched to the next instruction. For that, the tutor bot is learned to not only generate responses but also infer its teaching action and progress on the current conversation simultaneously by a multi-task learning scheme. Our Tutoring bot is deployed under a non-commercial use license at https://tutoringai.com.

Hyungjoo Chae, Yongho Song, Kai Ong, Taeyoon Kwon, Minjin Kim, Youngjae Yu, Dongha Lee, Dongyeop Kang, Jinyoung Yeo. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

In this paper, we propose to leverage the unique characteristics of dialogues sharing commonsense knowledge across participants, to resolve the difficulties in summarizing them. We present SICK, a framework that uses commonsense inferences as additional context. Compared to previous work that solely relies on the input dialogue, SICK uses an external knowledge model to generate a rich set of commonsense inferences and selects the most probable one with a similarity-based selection method. Built upon SICK, SICK++ utilizes commonsense as supervision, where the task of generating commonsense inferences is added upon summarizing the dialogue in a multi-task learning setting. Experimental results show that with injected commonsense knowledge, our framework generates more informative and consistent summaries than existing methods.