论文检索

Chenning Xu, Mao Zheng, Mingyu Zheng, Mingyang Song

Podcast script generation requires LLMs to synthesize structured, context-grounded dialogue from diverse inputs, yet systematic evaluation resources for this task remain limited. To bridge this gap, we introduce PodBench, a benchmark comprising 800 samples with inputs up to 21K tokens and complex multi-speaker instructions. We propose a multifaceted evaluation framework that integrates quantitative constraints with LLM-based quality assessment. Extensive experiments reveal that while proprietary models generally excel, open-source models equipped with explicit reasoning demonstrate superior robustness in handling long contexts and multi-speaker coordination compared to standard baselines. However, our analysis uncovers a persistent divergence where high instruction following does not guarantee high content substance. PodBench offers a reproducible testbed to address these challenges in long-form, audio-centric script generation.

Hieu Man, Van-Cuong Pham, Nghia Trung Ngo, Franck Dernoncourt, Thien Huu Nguyen

Learning robust representations of authorial style is crucial for authorship attribution and AI-generated text detection. However, existing methods often struggle with content-style entanglement, where models learn spurious correlations between authors’ writing styles and topics, leading to poor generalization across domains. To address this challenge, we propose Explainable Authorship Variational Autoencoder (EAVAE), a novel framework that explicitly disentangles style from content through architectural separation-by-design. EAVAE first pretrains style encoders using supervised contrastive learning on diverse authorship data, then finetunes with a Variational Autoencoder (VEA) architecture using separate encoders for style and content representations. Disentanglement is enforced through a novel discriminator that not only distinguishes whether pairs of style/content representations belong to the same or different authors/content sources, but also generates natural language explanation for their decision, simultaneously mitigating confounding information and enhancing interpretability. Extensive experiments demonstrate the effectiveness of EAVAE. On authorship attribution, we achieve state-of-the-art performance on various datasets, including Amazon Reviews, PAN21, and HRS. For AI-generated text detection, EAVAE excels in few-shot learning over the M4 dataset.

Md Asiful Islam, Mihai Surdeanu

We propose a lightweight explainable guardrail (LEG) method to detect unsafe prompts. LEG uses a multi-task learning architecture to jointly learn a prompt classifier and an explanation classifier, where the latter labels prompt words that explain the safe/unsafe overall decision. LEG is trained on synthetic explanation data, which is generated using a novel strategy that counteracts the confirmation biases of LLMs. Lastly, LEG’s training process uses a novel loss that captures global explanation signals as a weak supervision and combines cross-entropy and focal losses with uncertainty-based weighting. LEG obtains equivalent or better performance than the state-of-the-art for both prompt classification and explainability, both in-domain and out-of-domain on three datasets, despite the fact that its model size is considerably smaller than current approaches.

Shuo Yang, Caren Han, Xueqi Ma, Yan Li, Mohammad Reza Ghasemi Madani, Eduard Hovy

LLM-based agents depend on effective tool-use policies to solve complex tasks, yet optimizing these policies remains challenging dueto delayed supervision and the difficulty ofcredit assignment in long-horizon trajectories.Existing optimization approaches tend to beeither monolithic, which are prone to entangling behaviors, or single-aspect, which ignorecross-module error propagation. To addressthese limitations, we propose EVOTOOL, a self-evolving framework that optimizes a modulartool-use policy via a gradient-free evolutionary paradigm. EVOTOOL decomposes agent’stool-use policy into four modules, includingPlanner, Selector, Caller, and Synthesizer, anditeratively improves them through three mechanisms. Trajectory-Grounded Blame Attribution uses diagnostic traces to localize failuresto a specific module. Feedback-Guided Targeted Mutation then edits only that modulevia natural-language critique. Diversity-AwarePopulation Selection preserves complementarycandidates to ensure solution diversity. Acrossfour diverse benchmarks, EVOTOOL outperforms strong baselines by over 5 points on bothGPT-4.1 and Qwen3-8B, while achieving superior efficiency and transferability.

Bingkang Shi, Jen-tse Huang, Luo Long, Tianyu Zong, Hongzhu Yi, Yuanxiang Wang, Songlin Hu, Xiaodan Zhang, Zhongjiang Yao

Large Language Models (LLMs) have increasingly enhanced or replaced traditional Non-Player Characters (NPCs) in video games. However, these LLM-based NPCs inherit underlying social biases (e.g., race or class), posing fairness risks during in-game interactions. To address the limited exploration of this issue, we introduce FairGamer, the first benchmark to evaluate social biases across three interaction patterns: transaction, cooperation, and competition. FairGamer assesses four bias types, including class, race, age, and nationality, across 12 distinct evaluation tasks using a novel metric, FairMCV. Our evaluation of seven frontier LLMs reveals that: (1) models exhibit biased decision-making, with Grok-4-Fast demonstrating the highest bias (average FairMCV = 76.9%); and (2) larger LLMs display more severe social biases, suggesting that increased model capacity inadvertently amplifies these biases. We release FairGamer at https://github.com/BingkangShi/FairGamer to facilitate future research on NPC fairness.

Fengxian Ji, Jingpu Yang, Zirui Song, Lang Gao, Junhong Liang, Zhenhao Chen, Jinghui Zhang, Xiuying Chen

Recent image generation and editing models demonstrate robust adherence to instructions and high visual quality on academic benchmarks.However, their performance on paid, real-world design projects remains uncertain. We introduce ServImage, a benchmark that explicitly correlates model outputs with economic value in commercial design projects. ServImage consists of (i) ServImageBench: a dataset of 1.07k paid commercial design tasks and 2.05k designer deliverables totaling over $295k, covering portrait, product, and digital content, along with 33k candidate images and 33k human annotations.(ii) ServImageScore: an integrated scoring system that combines three quality dimensions: baseline requirements fulfilment, visual execution quality, and commercial necessity satisfaction. These three dimensions are designed to characterize the factors that drive human payment decisions and indicate whether an image is commercially acceptable.(iii) ServImageModel: under this scoring system, we propose a payment prediction model trained on the human-annotated candidate images, achieving 82.00% accuracy in predicting human payment decisions and producing calibrated payment probabilities.ServImage provides a comprehensive foundation for assessing the commercial viability of image generation models and offers a scalable resource for future research on economically grounded vision systems Github.

Ruifeng Yuan, Chaohao Yuan, David Dai, Yu Rong, Hong Cheng, Hou Pong Chan, Chenghao Xiao

Recent retrieval-augmented generation (RAG) approaches have demonstrated strong capability in handling complex queries, yet current research overlooks a critical challenge: different retrievers require fundamentally different query formulation strategies for optimal performance. In this work, we present the first systematic analysis of how LLMs can learn to adapt their query formulation strategies for different retrievers via reinforcement learning (RL). Our empirical study reveals that RL effectively teaches an LLM to tailor its queries to specific retriever characteristics. We discover that different retrievers exhibit surprisingly distinct optimal query styles (e.g., descriptive vs. question-like), suggesting strategies learned for one retriever ineffective for another. We further show that performance can be enhanced by incorporating retriever-specific human guidance and by scaling model size. To facilitate learning over multi-retrieval-step trajectories, we introduce a branching-based rollout technique that improves training stability. Our work provides the first empirical evidence and actionable insights for building truly retriever-aware RAG systems. Code and resources are available at https://github.com/LCO-Embedding/Envs-aware-Information-Retrieval.

Jikun Wan, Chen Gong, Guohong Fu

Multimodal emotion cause analysis in conversation aims to identify the causes of emotions by leveraging multimodal information. Existing studies mainly formulate this problem as either utterance-level emotion cause extraction, which provides clear cause localization but limited explanation, or multimodal emotion cause generation, which offers fine-grained explanations but lacks explicit traceability to source utterances. Moreover, existing datasets rely heavily on human judgment and lack well-defined structured theoretical guidance, leading to subjective and inconsistent annotations. To address these issues, we introduce joint Multimodal Emotion Cause Extraction and Summarization in conversation (MECES), a new task that simultaneously extracts emotion cause utterances and generates cause summaries, enabling both precise localization and interpretable explanations of emotion cause. We further construct a MECES dataset guided by the Activating Events–Beliefs–Consequences theory from psychology. This dataset consists of 5,787 emotion utterances annotated with causes, comprising 12,231 emotion-cause pairs and 6,040 cause summaries. We also propose an effective end-to-end joint learning approach for MECES task, establishing strong benchmark results for this newly introduced task and dataset.

Kwangwook Seo, Dongha Lee

Recent approaches in personalized reward modeling have primarily focused on leveraging user interaction history to align model judgments with individual preferences. However, existing approaches largely treat user context as a static or implicit conditioning signal, failing to capture the dynamic and multi-faceted nature of human judgment. In this paper, we propose P-Check, a novel personalized reward modeling framework, designed to train a plug-and-play checklist generator that synthesizes dynamic evaluation criteria for guiding the reward prediction. To better align these checklists with personalized nuances, we introduce Preference-Contrastive Criterion Weighting, a training strategy that assigns saliency scores to criteria based on their discriminative power for personalized judgment. We conduct extensive experiments and demonstrate that P-Check not only improves reward accuracy but also enhances downstream personalized generation, and remains robust in OOD scenarios.

Wei-Chieh Huang, Henry Peng Zou, Yaozu Wu, Dongyuan Li, Yankai Chen, Weizhi Zhang, Yangning Li, Angelo Zangari, Jizhou Guo, Chunyu Miao 等

Deep research frameworks have shown promising capabilities in synthesizing comprehensive reports from web sources. While deep research possesses significant potential to address complex issues through planning and research cycles, existing frameworks are deficient in sufficient evaluation procedures and stage-specific protections. They typically treat evaluation as exact match accuracy of question-answering, but overlook crucial aspects of report quality such as credibility, coherence, breadth, depth, and safety. This oversight may result in hazardous or malicious sources being integrated into the final report. To address this, we introduce DeepResearchGuard, a framework featuring four-stage safeguards with open-domain evaluation, and DRSafeBench, a novel stage-wise safety benchmark. Evaluating across GPT-4o, o4-mini, Gemini-2.5-flash, DeepSeek-v3, and GPT-5, DeepResearchGuard improves defense success rates by an absolute 16.53% while reducing over-refusal rates to approximately 6%. Through extensive experiments, we show that DeepResearchGuard enables comprehensive open-domain evaluation and stage-aware defenses that effectively block harmful content propagation, while systematically improving report quality without excessive over-refusal rates.

Weiyang Huang, Xuefeng Bai, Kehai Chen, Xinyang Chen, Yibin Chen, Weili Guan, Min Zhang

Large Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive "overthinking", generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure. SAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (SLOW, NORMAL, FAST, SKIP). It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy. Code is available at https://github.com/byxw13/SAT_Code.

Robin Vujanic, Thomas Rückstieß

We present a knowledge distillation framework for text embedding models. A key distinguishing feature is that our distilled models are compatible with their teacher, enabling flexible asymmetric architectures where documents are encoded with the larger teacher model, while queries use smaller student models. We also show that our models automatically inherit MRL and robustness to output quantization whenever these properties are present in the teacher model, without explicitly training for them. To demonstrate the effectiveness of our framework we publish leaf-ir, a 23M parameters information retrieval oriented model that, besides being teacher-compatibile, sets a new state-of-the-art (SOTA) on BEIR, ranking no.1 on the public leaderboard for models of its size. Asymmetric mode further increases its retrieval performance. Our scheme is however not restricted to information retrieval. We demonstrate its wider applicability by synthesizing the multi-task leaf-mt model. This also sets a new SOTA, achieving no.1 on the public MTEB v2 (English) leaderboard for models of its size. Our technique is applicable to black-box models, requires no judgments nor hard negatives, and training can be conducted using small batch sizes. Thus, dataset and training infrastructure requirements for our framework are modest. We make our models publicly available under a permissive Apache 2.0 license.

Yiming Yao, Jianwei Niu, Bin Dai, Tao Ren

Recent breakthroughs in Transformer-based large models, have driven widespread tasks, yet their reliance on centralized cloud deployment raises significant privacy risks due to sensitive data exposure. While edge-based collaborative inference offers a privacy-preserving alternative, existing methods face critical limitations: static model partitioning cannot adapt to dynamic edge resource fluctuations, and rigid multi-head attention handling overlooks semantic-critical prioritization and parallelism. We propose EdgeFormer, a latency-aware framework for distributed Transformer inference in resource-constrained edge networks. EdgeFormer dynamically allocates model blocks across devices via efficiency-storage trade-off optimization and introduces collaborative Multi-Head Attention (cMHA), which distributes semantic-critical attention heads across devices while pruning redundant ones under real-time constraints. We further develop LiScore, a composite metric integrating attention diversity and latency costs, alongside a similarity-based retrieval method to reduce recomputation overhead. Extensive experiments demonstrate that EdgeFormer achieves up to 2.01 \\times inference acceleration over state-of-the-art baselines with \\leq1.06% accuracy loss, maintaining robustness under varying edge conditions.

Kunal Samanta, Faisal Tareque Shohan, Amine Trabelsi, Richard Khoury

Multi-party social dialogue remains underexplored in the literature,in part due to the difficulty and cost of evaluation. As a result,recent work on synthetic dialogue generation often relies on automatedmetrics and LLM-as-a-Judge frameworks, despite limited evidence thatsuch judges reflect human preferences in social settings. In this work,we introduce a lightweight and controllable multi-party dialoguegeneration framework (MPOD) as an experimental instrument forstudying generation and evaluation in social interaction. Using thisframework, we conduct human evaluations of open-domain multi-partydialogue simulation and directly compare human judgments againststate-of-the-art LLM judges. Across 319 pairwise comparisons, weobserve near-random agreement between humans and automated judges(Cohen’s \kappa \approx 0.11), driven by systematic behaviorsincluding extreme tie aversion and strong sensitivity toassistant-style verbosity. Crucially, human–human inter-annotatoragreement (\kappa = 0.29) is substantially higher than human–LLMagreement. To isolate themechanism underlying this misalignment, we introduce a controlledTransplant Ablation, showing that LLM judges consistentlyprefer conversations containing a single proprietary, assistant-styleagent. Additional stress tests show that judges prefer GPT-styleconversations even when utterance order is randomly shuffled,indicating insensitivity to conversational structure and coherence.Our findings provide controlled evidence that currentinstruction-tuned LLM judges do not reliably reflect human preferences for naturalness, engagingness, and overall quality in multi-party social dialogue, calling into question their widespreaduse for validating synthetic conversational data.

Junpeng Liu, Jiuyi Li, Kaiyu Huang, Bo Jin, Degen Huang, Hui Xiong

Current large language models (LLMs) often exhibit performance imbalances between dominant languages (e.g., English) and non-dominant ones due to the skewed distribution of pretraining data. A common strategy to address this issue is to enhance cross-lingual alignment, thereby facilitating non-dominant language processing. However, existing methods typically rely on additional training objectives or language-specific parameters, which increase training complexity and cost. In this work, we propose a selective bidirectional language projection framework that enables efficient multilingual alignment and language shift using the intrinsic parameters. Specifically, we first identify the layers most sensitive to language projection between non-dominant and dominant languages through neuron activation analysis. We then perform sequential language projection within the selected layers by mapping non-dominant representations into the dominant language space and reverting them before generation. The bidirectional projection benefits the subsequent instruction tuning in non-dominant languages. Experiments on seven benchmarks demonstrate that our method remarkably enhances the performance of non-dominant languages. Further analyses indicate that our method learns better internal representations and exhibits strong generalization capabilities.

Rui Li, Shuang Cao

Inference-time search can substantially improve LLM reasoning when tasks admit deterministic verification, but existing methods largely refine single trajectories and lack a reliable mechanism for composing partial solutions across candidates. We propose contract-checked graph editing: represent each candidate as an interface-typed reasoning DAG and validate every nontrivial edit with a deterministic structural gate (acyclicity, namespace closure, schema validity, terminal constraints) before invoking the verifier. The gate certifies runnability only and emits auditable rejection reasons; semantic correctness is determined solely by the verifier. Instantiated in Genetic Inference Search (GIS) with Qwen2.5-32B-Instruct under strictly matched token budgets (8K tokens), contract-checked grafting increases verifier-runnable recombination from 41.2% to 92.8% and improves accuracy over rStar (+6.1 on MATH, +9.1 on MATH L5) while using 42% fewer verifier calls. The same operators transfer across outer loops (beam, best-first, MCTS) and to structured generation and code, outperforming execution-guided beam search on Spider (+2.8) and improving multi-file code generation on HumanEval-MF (+9.2).

Rohit Gupta, Jayakrishnan Unnikrishnan, Fan Fei, Sheng Liu, Son Tran, Mubarak Shah

Video Large Language Models (VideoLLMs) excel at video understanding tasks where outputs are textual, such as Video Question Answering and Video Captioning. However, they underperform specialized embedding-based models in Retrieval tasks, such as Text-toVideo Retrieval and Moment Retrieval. We introduce ViLL-E (Video-LLM-Embed), a unified VideoLLM architecture endowed with a novel embedding generation mechanism that allows the model to "think longer" for complex videos and stop early for easy ones. We train this model with a three-stage training methodology combining generative and contrastive learning: initial large-scale pre-training with video-caption pairs; followed by continual training on a smaller, detailed-caption dataset; and concluding with task-specific fine-tuning on a novel multi-task dataset covering Video QA, Temporal Localization, Video Retrieval, and Video-Text Matching. Our model significantly improves temporal localization (on avg. 7% over other VideoLLMs) and video retrieval (up to 4% over dual encoder models), achieving performance comparable to state-of-the-art specialized embedding models while remaining competitive on VideoQA tasks. Furthermore, our joint contrastive-generative training unlocks new zero-shot capabilities, significantly outperforming state-of-the-art methods in composed video retrieval (+5% over SotA) and retrieval from long text (+2% over SotA).

Iseo Kim, Eunjin Hong, Juae Kim

This study investigates how different moral conditions influence the general problem-solving capabilities of Large Language Models (LLMs). We aim to explore whether the role of morality as a cognitive function operating in human decision-making can be extended to LLMs. Specifically, we define distinct moral stages based on Kohlberg’s theory of moral development and design prompts to elicit model responses aligned with each corresponding condition. The validity of this alignment is verified using the Defining Issues Test, a human evaluation tool. Subsequently, models reflecting the characteristics of each condition are evaluated using the MMLU benchmark, which requires general problem-solving abilities across various domains. Experimental results show that different moral perspectives lead to changes in the model’s decision-making during general reasoning, reflected in both responses and internal representations. Notably, conditions grounded in more advanced moral stages tend to elicit more thoughtful and reflective problem-solving behavior, which is often associated with enhanced performance. Our study attempts to broaden the concept of LLM morality, which has traditionally only been considered in ethical judgment scenarios. Furthermore, it emphasizes that morality is not merely a means of ensuring safety but a crucial factor that shapes a model’s behavior and thought processes.

Ziqiao Xi, Shuang Liang, Qi Liu, Jiaqing Zhang, Letian Peng, Fang Nan, Meshal Nayim, Tianhui Zhang, Rishika Mundada, Lianhui Qin 等

To close the gap between LLM-based agents and humans in planning and reasoning, agents need large-scale, diverse environments for continuous learning—yet building such environments is itself prohibitively expensive. We present C-World, an environment creation system that enables users to build agent environments on demand. We define a complete agent environment through four components: an Action Space of 5,571 format-unified tools across 204 common applications, a Task Distribution engine that synthesizes long-horizon workflows with wild constraints, a Transition Function implemented as a state controller that injects realistic failures and perturbations, and a Reward Signal combining verifiable metrics with LLM-based judgment. C-World operates in two modes: a realistic mode grounded in live API execution, and a synthesized mode powered by the World Engine, which approximates tool behavior without live service access, enabling scalable environment creation—including environments for domains and tools that do not yet exist in the real world. Evaluation of nine state-of-the-art LLMs reveals that planning ability is uniformly strong but execution remains the bottleneck, and that constraint following—not tool invocation—is the dominant failure mode. The World Engine achieves Spearman \rho = 0.883 ranking correlation with real execution, and fine-tuning on just 1,170 C-World trajectories outperforms baselines trained on 119k samples, demonstrating C-World’s dual value as a rigorous evaluation environment and a scalable data engine. Our code and data are available at https://ziqiao-git.github.io/C-World/.

Sijia Luo, Xiaokang Zhang, Yuxuan Hu, Bohan Zhang, Ke Wang, Jinbo Su, Mengshu Sun, Lei Liang, Jing Zhang

Reinforcement Learning (RL) has become essential for eliciting complex reasoning capabilities in Large Language Models (LLMs). However, the substantial memory overhead of storing Key-Value (KV) caches during long-horizon rollouts acts as a critical bottleneck, often prohibiting efficient training on limited hardware. While existing KV compression techniques offer a remedy for inference, directly applying them to RL training induces a severe policy mismatch, leading to catastrophic performance collapse. To address this, we introduce Sparse-RL, which empowers stable RL training under sparse rollouts. We show that instability arises from a fundamental policy mismatch among the dense old policy, the sparse sampler policy, and the learner policy. To mitigate this issue, Sparse-RL incorporates Sparsity-Aware Rejection Sampling and Importance-based Reweighting to correct the off-policy bias introduced by compression-induced information loss. Experimental results show that Sparse-RL reduces rollout overhead compared to dense baselines while preserving the performance. Furthermore, Sparse-RL inherently implements sparsity-aware training, significantly enhancing model robustness during sparse inference deployment.