Large language models (LLMs) can generate fluent dialogue, but prior works lack situational grounding, dynamic strategy control, and evaluation aligned with clinical standards in motivational interviewing (MI). We introduce StoryMI, a multi-LLM agent framework for controllable MI dialogue generation, where questionnaire-based client profiles are expanded into situational stories that provide narrative context for the dialogue. Therapist and client agents generate MI-coded utterances guided by MI codes selected by the interaction agent, while an interaction agent dynamically coordinates exchanges to control MI strategies during a multi-turn conversation. We propose a two-level evaluation protocol: lexical metrics and MI-specific measures of macro-level counseling strategies, alongside LLM-as-judge and human expert assessments. We construct a dataset of 6K simulated MI dialogues grounded in 1K questionnaire-story pairs, covering 12 MI codes and 13 symptom domains, and benchmark six open- and closed-source LLMs. Our results show that situational grounding and macro-level control can improve MI adherence and clinical plausibility, demonstrating the effectiveness of a structured multi-agent workflow for psychotherapy dialogue generation. We provide code and data for reproducibility.
论文检索
输入标题、作者或关键词，从 12,319 篇学术成果中精准定位
EduMARS: Can Vision-Language Models Grade Like Teachers? Benchmarking Multimodal, Rubric-Based Assessment on Chinese K-12 Answers
PDF ↗Automated grading of student work is a critical application of AI in education. However, existing benchmarks fall short in evaluating models on realistic, cognitively demanding tasks. Most rely on synthetic, well-structured text inputs, overlooking the multimodal, error-prone, and often handwritten nature of real student responses, especially in K-12 settings. We introduce EduMARS, a multimodal benchmark designed for rubric-aligned evaluation of real Chinese K-12 student answers. The dataset contains over 4,500 authentic responses from high-stakes exams across eight subjects, featuring noisy handwriting,mixed-layout diagrams,mathematical expressions, and narrative reasoning. Each response is meticulously annotated by expert teachers using step-wise scoring rubrics, error classifications, and key-point mappings, providing fine-grained supervision aligned with real-world pedagogical practices. We evaluated existing SOTA MLLMs across the dimensions of final score and the reasoning process of grading, reveals a significant gap between existing SOTA MLLMs and human-level performance. To bridge this performance gap, we propose the Retrieval-Augmented Adaptive-Rubric Grading (RARG), enabling models to emulate expert grading logic by dynamically synthesizing case-specific evaluation schemas. RARG effectively enhances the performance and interpretability of various MLLMs on EduMARS, surpassing in-context learning and chain-of-thought.
One paradigm of language model (LM) fine-tuning relies on creating large training datasets, under the assumption that high quantity and diversity will enable models to generalize to novel tasks after post-training. In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive; worse, there is no guarantee that the resulting model will handle complex scenarios or generalize better. Moreover, existing techniques rarely assess whether a training sample provides novel information, resulting in unnecessary costs. In this work, we explore a new Test-Time Self-Improvement (TT-SI) algorithm to create more effective and generalizable agentic LMs on-the-fly. TT-SI can be summarized in three steps: (i) first it identifies the samples that model struggles with (self-awareness), (ii) then generates similar examples from detected uncertain samples (self-data augmentation), and (iii) uses these newly generated samples at test-time training (self-improvement). We further explore Test-Time Distillation (TT-D), which leverages a stronger supervisor for targeted data generation. Empirical evaluations across different agent benchmarks demonstrate that TT-SI improves the performance with +5.48% absolute accuracy gain on average across all benchmarks and surpasses other standard learning methods more efficiently. Our findings highlight the promise of TT-SI, demonstrating the potential of self-improvement algorithms at test-time as a new paradigm for building more capable agents toward self-evolution.
Evaluating the instruction-following (IF) capabilities of Multimodal Large Language Models (MLLMs) is essential for rigorously assessing how faithfully model outputs adhere to user-specified intentions. Nevertheless, existing benchmarks for evaluating MLLMs’ instruction-following capability primarily focus on verbal instructions in the textual modality. These limitations hinder a thorough analysis of instruction-following capabilities, as they overlook the implicit constraints embedded in the semantically rich visual modality. To address this gap, we introduce VC-IFEval, a new benchmark accompanied by a systematically constructed dataset that evaluates MLLMs’ instruction-following ability under multimodal settings. Our benchmark systematically incorporates vision-dependent constraints into instruction design, enabling a more rigorous and fine-grained assessment of how well MLLMs align their outputs with both visual input and textual instructions. Furthermore, by fine-tuning MLLMs on our dataset, we achieve substantial gains in visual instruction-following accuracy and adherence. Through extensive evaluation across representative MLLMs, we provide new insights into the strengths and limitations of current models.
PhysPRM: A Generative Process Reward Model with Fine-grained Diagnosis for Physics Problem Solving
PDF ↗Despite the remarkable progress of Large Language Models (LLMs) in abstract reasoning tasks, they continue to struggle with physics problem solving due to difficulties in decoding implicit constraints and maintaining physical consistency. To address these challenges, Process Reward Models (PRMs) have emerged as a promising approach to verify intermediate reasoning steps. Existing PRMs attempt to mitigate reasoning errors but typically rely on scalar scoring, which lacks the explanatory power necessary to diagnose complex physical misconceptions. In this work, we introduce PhysPRM, a Generative PRM that treats evaluation as a generative task to produce fine-grained diagnoses comprising critiques, final judgments, and specific error types. To facilitate this, we develop an automated data synthesis pipeline to construct PhysPRM30K, a comprehensive training dataset, and PhysProcessBench, a rigorously human-verified benchmark. By employing a two-stage training paradigm that integrates Supervised Fine-Tuning with Group Relative Policy Optimization, PhysPRM significantly enhances the physics reasoning capabilities of various LLMs. Extensive experiments demonstrate that PhysPRM improves performance across seven benchmarks in both Best-of-N and critique refinement strategies.
Modern large language models (LLMs) are typically trained and deployed using structured role tags (e.g. system, user, assistant, tool) that explicitly mark the source of each piece of context. While these tags are essential for instruction following and controllability, asymmetries in the training data associated with different role tags can potentially introduce inductive biases. In this paper, we study this phenomenon by formalizing user–assistant bias, defined as the tendency of an LLM to preferentially rely on information from either the user or assistant role when they provide incompatible information about the same entity in the context history. We introduce a task-agnostic benchmark UserAssist and evaluate such bias in 52 frontier models. We observe that most of the instruction-tuned models exhibit strong user bias, whereas base and reasoning models are close to neutral. Using controlled fine-tuning experiments, we isolate which post-training recipes drive the observed user–assistant bias. We find that human-preference alignment amplifies user bias, while reasoning fine-tuning reduces it. Finally, we show that user–assistant bias can be bidirectionally controlled via direct preference optimization (DPO) on UserAssist-train, and that the resulting bias reliably generalizes to two realistic multi-turn debate datasets spanning philosophical opinions and natural argumentative exchanges on factual/policy topics. These results reveal an underexplored consequence of role-tagged training and provide a principled framework to diagnose and control tag-induced biases in modern LLMs.
Large Language Models (LLMs) are commonly trained on multilingual corpora that include Greek, yet reliable evaluation benchmarks for Greek—particularly those based on authentic, native-sourced content—remain limited. Existing datasets are often machine-translated from English, failing to capture Greek linguistic and cultural characteristics. We introduce GreekMMLU, a native-sourced benchmark for massive multitask language understanding in Greek, comprising 21,805 multiple-choice questions across 45 subject areas, organized under a newly defined subject taxonomy and annotated with educational difficulty levels spanning primary to professional examinations. All questions are sourced or authored in Greek from academic, professional, and governmental exams. We publicly release 16,857 samples and reserve 4,948 samples for a private leaderboard to enable robust and contamination-resistant evaluation. Evaluations of over 80 open- and closed-source LLMs reveal substantial performance gaps between frontier and open-weight models, as well as between Greek-adapted models and general multilingual ones. Finally, we provide a systematic analysis of factors influencing performance—including model scale, adaptation, and prompting—and derive insights for improving LLM capabilities in Greek.
The advancement of Multimodal Emotion Recognition (MER) in Chinese is significantly hindered by the scarcity of high-quality, spontaneous dialogue datasets compared to their English counterparts. In this work, we introduce EmotionTalk, the first interactive Chinese multimodal dataset designed to capture the nuance of authentic emotional interplay. Collected from 19 professional actors, the dataset spans 23.6 hours of dyadic conversations across diverse scenarios. A key contribution of EmotionTalk is its multi-grained annotation system, which integrates standard categorical and dimensional labels with fine-grained emotional speaking style captions, enabling research into interpretable emotion analysis. We establish comprehensive benchmarks for emotion recognition and captioning tasks, verifying the dataset’s effectiveness and the necessity of multimodal fusion. EmotionTalk serves as a critical resource for bridging the gap in non-English affective computing and is publicly released for the research community.
PBEBench: A Multi-Step Programming by Examples Reasoning Benchmark inspired by Historical Linguistics
PDF ↗While many benchmarks evaluate the reasoning abilities of Large Language Models (LLMs), few isolate reasoning as a capability independent of domain knowledge. We introduce a new benchmark for inductive reasoning inspired by Sound Law Induction (SLI) in historical linguistics and formulated in a simple multi-step Programming by Example (PBE) framework. The task requires inducing a cascade of string rewrite programs that transform inputs into target outputs. We present PBEBench, a fully automated evaluation approach that generates such problems with controllable difficulty and ordering constraints, enabling scalable and contamination-resistant evaluation of sequential inductive reasoning. Using this approach, we construct three datasets that show a large gap between models that leverage test-time compute or long chain-of-thought reasoning and those that do not. Although recent models such as GPT-5 and gpt-oss-120b show promise, solve rates remain below 5% on hard PBEBench instances with long program cascades, even under computationally expensive scaling strategies. Finally, we show that PBEBench scores are more predictive of performance on real SLI than are other inductive reasoning benchmarks. We will release code and data to support further research.
Long-form text generation remains a significant challenge for large language models (LLMs), particularly in maintaining coherence, ensuring logical consistency, and preserving text quality as sequence length increases. To address these limitations, we propose \textit{SuperWriter}-Agent, an agent-based framework designed to enhance the quality and consistency of long-form text generation. \textit{SuperWriter}-Agent introduces explicit structured thinking-through planning and refinement stages—into the generation pipeline, guiding the model to follow a more deliberate and cognitively grounded process akin to that of a professional writer. Based on this framework, we construct a supervised fine-tuning dataset to train a 7B \textit{SuperWriter}-LM. We further develop a hierarchical Direct Preference Optimization (DPO) procedure that uses Monte Carlo Tree Search (MCTS) to propagate final quality assessments and optimize each generation step accordingly. Empirical results across diverse benchmarks demonstrate that \textit{SuperWriter}-LM achieves state-of-the-art performance, surpassing even larger-scale baseline models in both automatic evaluation and human evaluation. Furthermore, comprehensive ablation studies demonstrate the effectiveness of hierarchical DPO and underscore the value of incorporating structured thinking steps to improve the quality of long-form text generation.
Reward models (RMs) play a critical role in aligning large language models (LLMs) with human preferences. Yet in the domain of tool learning, the lack of RMs specifically designed for function-calling tasks has limited progress toward more capable agentic AI. We introduce ToolRM, a family of lightweight reward models tailored for general tool-use scenarios. To build these models, we propose a novel pipeline that constructs high-quality pairwise preference data using rule-based scoring and multidimensional sampling. This yields ToolPref-Pairwise-30K, a diverse, balanced, and challenging preference dataset that supports both generative and discriminative reward modeling. We also introduce TRBench_{BFCL}, a benchmark built on the agent evaluation suite BFCL to evaluate RMs on tool calling tasks. Trained on our constructed data, models from the Qwen3-4B/8B series achieve up to 17.94% higher accuracy, substantially outperforming frontier LLMs and RMs in pairwise reward judgments. Beyond training objectives, generative ToolRM generalizes to broader critique tasks, including Best-of-N sampling and self-correction. Experiments on ACEBench highlight its effectiveness and efficiency, enabling inference-time scaling while reducing output token usage by over 66%. Its support for downstream RL training further validates its practical utility. We release data to facilitate future research.
Large Language Models (LLMs) are challenged by generating hallucinations and factually incorrect responses, particularly in complex and specialized medical question answering (QA). Integrating knowledge graphs (KGs) through retrieval-augmented generation (RAG) methods has emerged as a promising direction. However, existing graph-based RAG methods heuristically retrieve and refine question-relevant subgraphs, potentially introducing redundant and noisy factual information that is difficult for LLMs to process, ultimately limiting reasoning capability. To incorporate a concise yet informative evidence subgraph, we propose an iterative medical QA framework. It optimizes graph-based RAG methods by selectively retrieving focused knowledge from KGs to construct a precise evidence subgraph and progressively pruning it utilizing structured feature representations. The targeted KG integration maintains coherent and reliable inference. Experiments on three medical QA benchmark datasets demonstrate that the framework achieves state-of-the-art performance against representative baseline competitors, highlighting the importance of efficient KG integration.
Process Reward Models (PRMs) have emerged as a promising approach for improving LLM reasoning capabilities by providing process supervision over reasoning traces. However, existing approaches for constructing PRM training data remain costly and noisy, as they typically rely on human annotation or sampling-based labeling methods that require repeated LLM calls. In this work, we propose FoVer, a framework that synthesizes PRM training data from formal reasoning tasks by annotating step-level error labels using formal verification tools such as Z3 and Isabelle. By leveraging formal verification, FoVer enables efficient and accurate PRM data construction without requiring human annotation or additional LLM calls. Using FoVer, we create PRM training data from formal logic and theorem proving tasks. Experiments on 12 reasoning benchmarks show that fine-tuning on our training data improves PRMs not only on math and logic reasoning tasks, which are informal variants of the training tasks, but also on NLI and BBH benchmarks, which differ substantially from the tasks used to construct the training data. These results demonstrate the practical effectiveness of FoVer, showing that PRM training data created using formal verification improves PRMs on informal reasoning tasks written in natural language. The datasets, models, and code are provided at https://github.com/psunlpgroup/FoVer.
Anonymizing textual documents is a highly context-sensitive problem: the appropriate balance between privacy protection and utility preservation varies with the data domain, privacy objectives, and downstream application. However, existing anonymization methods rely on static, manually designed strategies that lack the flexibility to adjust to diverse requirements and often fail to generalize across domains. We introduce adaptive text anonymization, a new task formulation in which anonymization strategies are automatically adapted to specific privacy–utility requirements. We propose a framework for task-specific prompt optimization that automatically constructs anonymization instructions for language models, enabling adaptation to different privacy goals, domains, and downstream usage patterns. To evaluate our approach, we present a benchmark spanning five datasets with diverse domains, privacy constraints, and utility objectives. Across all evaluated settings, our framework consistently achieves a better privacy–utility trade-off than existing baselines, while remaining computationally efficient and effective on open-source language models, with performance comparable to larger closed-source models. Additionally, we show that our method can discover novel anonymization strategies that explore different points along the privacy–utility trade-off frontier.
LIBERTy: A Causal Framework for Benchmarking Concept-Based Explanations of LLMs with Structural Counterfactuals
PDF ↗Concept-based explanations quantify how high-level concepts (e.g., gender or experience) influence model behavior, which is crucial for decision-makers in high-stakes domains. Recent work evaluates the faithfulness of such explanations by comparing them to reference causal effects estimated from counterfactuals. In practice, existing benchmarks rely on costly human-written counterfactuals that serves as imperfect proxy. To address this, we introduce a framework for constructing datasets containing structural counterfactual pairs: LIBERTy (LLM-based Interventional Benchmark for Explainability with Reference Targets). LIBERTy is grounded in explicitly defined Structured Causal Models (SCMs) of the text generation, interventions on a concept propagate through the SCM until an LLM generates the counterfactual. We introduce three datasets (disease detection, CV screening, and workplace violence prediction) together with a new evaluation metric, order-faithfulness. Using them, we evaluate a wide range of methods across five models and identify substantial headroom for improving concept-based explanations. LIBERTy also enables systematic analysis of model sensitivity to interventions: we find that proprietary LLMs show markedly reduced sensitivity to demographic concepts, likely due to post-training mitigation. Overall, LIBERTy provides a much-needed benchmark for developing faithful explainability methods.
PaperScope: A Multi-Modal Multi-Document Benchmark for Agentic Deep Research Across Massive Scientific Papers
PDF ↗Leveraging Multi-modal Large Language Models (MLLMs) to accelerate frontier scientific research is promising, yet how to rigorously evaluate such systems remains unclear. Existing benchmarks mainly focus on single-document understanding, whereas real scientific workflows require integrating evidence from multiple papers, including their text, tables, and figures. As a result, multi-modal, multi-document scientific reasoning remains underexplored and lacks systematic evaluation. To address this gap, we introduce PaperScope, a multi-modal multi-document benchmark designed for agentic deep research. PaperScope presents three advantages: (1) Structured scientific grounding. It is built on a knowledge graph of over 2,000 AI papers spanning three years, providing a structured foundation for research-oriented queries. (2) Semantically dense evidence construction. It integrates semantically related key information nodes and employs optimized random-walk article selector to sample thematically coherent paper sets, thereby ensuring adequate semantic density and task complexity. (3) Multi-task evaluation of scientific reasoning. It contains over 2,000 QA pairs across reasoning, retrieval, summarization, and problem solving, enabling evaluation of multi-step scientific reasoning. Experimental results show that even advanced systems such as OpenAI Deep Research and Tongyi Deep Research achieve limited scores on PaperScope, highlighting the difficulty of long-context retrieval and deep multi-source reasoning. PaperScope thus provides a rigorous benchmark alongside a scalable pipeline for constructing large-scale multi-modal, multi-source deep research datasets.
FinMaster: A Holistic Benchmark for Full-Pipeline Financial Management with Large Language Models
PDF ↗Financial management is high-stakes, where small errors can propagate into reporting deviations and costly downstream decisions, yet real-world workflows remain labor-intensive and fragmented, and existing automation supports only isolated steps rather than complete workflows. Large language models (LLMs) show promise in automating financial workflows, but current benchmarks lack domain-specific data, realistic workflow-level task design, and standardized workflow-level evaluation. To address these gaps, we present **FinMaster**, a benchmark for evaluating large language models on full financial management workflows spanning financial literacy, accounting, auditing, and consulting. **FinMaster** comprises three modules: *FinSim* generates synthetic datasets compliant with real-world accounting standards for diverse company types, enabling realistic evaluation without relying on proprietary financial records. *FinSuite* offers 183 tasks across core financial domains. *FinEval* provides a unified evaluation framework. Extensive experiments on state-of-the-art models including GPT-4o-mini, Claude-3.7-Sonnet, and DeepSeek-V3 reveal critical capability gaps in financial reasoning, with accuracy dropping from over 90% on basic tasks to 40% on complex scenarios requiring multi-step reasoning. This degradation reflects error propagation, where accuracy reaches 58% for single-metric calculations but decreases to 37% in multi-metric settings. **FinMaster** provides scalable and reproducible benchmarking for realistic end-to-end financial workflows, helping advance reliable deployment of LLMs in financial practice.
Large Language Models (LLMs) are increasingly deployed in human-centric applications, yet they often fail to provide substantive emotional support. While Reinforcement Learning (RL) has been utilized to enhance empathy of LLMs, existing reward models typically evaluate empathy from a single perspective, overlooking the inherently bidirectional interaction nature of empathy between the supporter and seeker as defined by Empathy Cycle theory. To address this limitation, we propose Psychology-grounded Empathetic Reward Modeling (PERM). PERM operationalizes empathy evaluation through a bidirectional decomposition: 1) Supporter perspective, assessing internal resonation and communicative expression; 2) Seeker perspective, evaluating emotional reception. Additionally, it incorporates a bystander perspective to monitor overall interaction quality. Extensive experiments on a widely-used emotional intelligence benchmark and an industrial daily conversation dataset demonstrate that PERM outperforms state-of-the-art baselines by over 10%. Furthermore, a blinded user study reveals a 70% preference for our approach, highlighting its efficacy in generating more empathetic responses.
Recent advances in AI and wearable devices, such as augmented-reality glasses, have made it possible to augment human memory by retrieving personal experiences in response to natural language queries. However, existing egocentric video datasets fall short in supporting the personalization and long-context reasoning required for episodic memory retrieval. To address these limitations, we introduce EgoMemory, a benchmark derived from Ego4D, enriched with 165,795 user-specific object annotations over 245 videos from 45 participants, yielding 639 distinct, human-curated, and evaluated queries for rich and individualized episodic memory retrieval. Leveraging this resource, we present EgoRetriever, a novel, training-free retrieval framework that combines Multimodal Large Language Models with reflective Chain-of-Thought prompting. Our approach enables interpretive inference of user intent and generates detailed target video descriptions by leveraging contextualized personal memory for video retrieval. Extensive experiments on three benchmarks, including EgoMemory, EgoCVR, and EgoLife, demonstrate that EgoRetriever consistently and substantially outperforms state-of-the-art baselines, highlighting its strong generalizability and practical potential for personalized, long-context egocentric video retrieval.
Nowadays, wearable devices can continuously lifelog ambient conversations, creating substantial opportunities for memory systems. However, existing benchmarks primarily focus on online one-on-one chatting or human-AI interactions, thus neglecting the unique demands of real-world scenarios. Given the scarcity of public lifelogging audio datasets, we propose a hierarchical synthesis framework to curate LifeDialBench, a novel benchmark comprising two complementary subsets: EgoMem, built on real-world egocentric videos, and LifeMem, constructed using simulated virtual community. Crucially, to address the issue of temporal leakage in traditional offline settings, we propose an Online Evaluation protocol that strictly adheres to temporal causality, ensuring systems are evaluated in a realistic streaming fashion. Our experimental results reveal a counterintuitive finding: current sophisticated memory systems fail to outperform a simple RAG-based baseline. This highlights the detrimental impact of over-designed structures and lossy compression in current approaches, emphasizing the necessity of high-fidelity context preservation for lifelog scenarios.