论文检索

Guido Ivetta, Pietro Palombini, Sofía Martinelli, Marcos J Gomez, M Emilia Echeveste, Sunipa Dev, Vinodkumar Prabhakaran, Luciana Benotti

The evaluation of societal biases in NLP models is critically hindered by a geo-cultural gap. This leaves regions such as Latin America severely underserved, making it impossible to adequately assess or mitigate the perpetuation of harmful regional stereotypes in language technologies. This paper presents LACES, a stereotype association dataset, for 15 Latin American countries. This dataset includes 4,789 stereotype associations[The de-identified dataset can be accessed via GitHub], manually created and annotated by 83 participants. The dataset was developed through targeted community partnerships across Latin America. Additionally, in this paper, we propose a novel adaptive data collection methodology that uniquely integrates the sourcing of new stereotype entries and the validation of existing data within a single, unified workflow. This approach results in a resource with more unique stereotypes than previous static collection methods, enabling a more efficient stereotype collection. The paper further supports the quality of LACES by demonstrating reduced efficacy of debiasing methods on this dataset in comparison to existing popular stereotype benchmarks.Content Warning: This research involves the study of social biases. Consequently, the paper contains examples of discriminatory language and stereotypes that may be sensitive or upsetting to readers. These examples are included for the purpose of scientific analysis and do not reflect the views of the authors.

Kaiser Sun, Fan Bai, Mark Dredze

Large language models (LLMs) rely on both contextual knowledge and parametric memory, yet these sources can conflict. Prior analysis largely focused on contextual question answering, suggesting that models tend to favor parametric knowledge under conflict, but this setting assumes that tasks should always rely on the provided passage. It therefore remains unclear how LLMs behave when tasks demand different kinds and degrees of knowledge utilization. We address this gap with a model-agnostic diagnostic framework that holds underlying knowledge constant while injecting controlled conflicts across tasks with varying knowledge requirements. Evaluating representative open-source LLMs, we find that: (1) performance degradation under conflict correlates with a task’s knowledge reliance rather than conflict plausibility alone; (2) strategies such as explanatory rationales or reiteration increase context reliance, helping context-only tasks but harming those that require parametric knowledge; and (3) these behaviors bias model-based evaluation, raising concerns about the reliability of LLMs as judges. Together, our findings show that context–memory conflict is fundamentally task-dependent and motivate task-aware approaches to balancing context and memory in LLM deployment and evaluation.

He Wang, Yu Gu, Fangfang Li, Zhigang Wang, Zhenghao Liu, Ning Wang, Xiaohua Li, Ge Yu

The autoregressive inference in large language models requires repeated computation across transformer layers. While caching intermediate key-value (KV) pairs eliminates redundancy, it introduces severe memory overhead, particularly in long-context settings. Most existing cache compression methods operate solely on either quantization or eviction, based on importance estimation of cached data. However, they are limited by coarse compression choices and inaccurate importance assessment, leading to suboptimal inference quality. To address this, we propose HqeKV, a hybrid compression framework built on both quantization and eviction, offering finer-grained compression options that adapt smoothly to the varying importance of cached KV pairs. An integrated optimizer automatically selects the best compression action for each cached element, maximizing quality while insulating end-users from tedious low-level tuning details. We further design a joint K–V importance metric to provide more accurate importance assessment results so that the optimizer can make smarter decisions. Additionally, HqeKV supports flexible conversion policies across multiple quantization precision levels, to further reduce quality degradation. Extensive experiments show that HqeKV improves output quality under the same memory constraints, outperforming state-of-the-art alternatives. Code is available at https://github.com/skywclouds/HqeKV.

Tongxi Wang

Large language models (LLMs) excel across many tasks, yet inference is still dominated by strictly token-by-token autoregression. Existing acceleration methods largely patch this pipeline and miss core human-reading ingredients: content-adaptive foresight, chunk-structure-aware compute allocation, and train–test consistency for preview/skimming. We propose the Fovea–Block–Skip Transformer (FBS), which injects a causal, trainable loop into Transformers via Parafovea-Attention Window (PAW), Chunk-Head (CH), and Skip-Gate (SG). Across diverse benchmarks, FBS improves the quality-efficiency trade-off without increasing parameters, and ablations show the three modules are complementary.

Yangzhuo Li, Shengpeng Ji, Yifu Chen, Tianle Liang, Haoyu Yang, Junboli, Jun Fang, Lin Li, Qingyang Hong

Integrating explicit Chain-of-Thought (CoT) into end-to-end spoken dialogue models enhances intelligence but incurs prohibitive latency. While the "Thinking-while-Talking" paradigm alleviates this delay, it fundamentally compromises block atomicity, severing the logical connection between interleaved thought and speech. To address this, we present Dual-Reasoner, employing a Streaming Masking Mechanism underpinned by our Dual-Think-30k dataset to guarantee uninterrupted audio streaming. Crucially, to strictly align the fragmented thinking blocks to service speech generation, we introduce the Atomic-Consistency Restoration framework. To secure comprehensive capabilities in high-difficulty reasoning, this mechanism utilizes a quadruple-constraint system to reconstruct logical atomicity, ensuring that "think" chunks act as a rigorous anchor for "talk" outputs. Experimental results demonstrate that Dual-Reasoner achieves comprehensive reasoning enhancements within ultra-low latency constraints: it elevates the VoiceBench score from 67.24 to 73.41 over the baseline, while significantly reducing the Time-to-First-Audio (TTFA) from 20.35s to 3.65s and the Real-Time Factor (RTF) from 7.04 to 1.05.

Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao

Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO often rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address these limitations, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.

Zequn Xie, Guijin Luo, Chuxin Wang, Sihang Cai, Tao Jin, Zhou Zhao, Yixuan Tang

Text-based person anomaly search retrieves specific behavioral events from surveillance archives using natural-language queries. Although recent pose-aware methods align geometric structures well, they face a fundamental Pose-Semantic Gap: semantically different actions can share similar skeletal geometries. While Multimodal Large Language Models (MLLMs) can reduce this ambiguity, using them for large-scale retrieval is computationally prohibitive. We propose the Structure-Semantic Decoupled Cascade (SSDC) framework, which decouples retrieval into two stages: (1) Structure-Aware Coarse Retrieval, where a lightweight model quickly filters candidates by skeletal similarity; and (2) Detective Squad Interaction, a multi-agent semantic verification module. The squad consists of a Detective for fast binary filtering, an Analyst for evidence extraction, and a Writer for semantic synthesis. Finally, we re-rank candidates by fusing the synthesized captions with structural priors. Experiments on the PAB benchmark show that SSDC achieves state-of-the-art performance by balancing efficiency and semantic reasoning.

Yuxuan Jiang, Dawei Li, Francis Ferraro

While Large Reasoning Models (LRMs) excel at complex tasks via long Chain-of-Thought (CoT) reasoning, their outputs are often excessively verbose, leading to inefficiency. This problem is amplified when the student’s long-form reasoning mismatches the concise outputs of smaller teacher models—common in LLM distillation to avoid using costly large teachers. To address this issue, we propose Distilled Reasoning Pruning (DRP), a hybrid framework that combines inference-time pruning with tuning-based distillation. DRP leverages a teacher model to perform mathematical problem-solving skill-aware step decomposition and pruning, then distills the refined reasoning paths into a student model, enabling efficient and accurate reasoning. Across challenging math datasets, DRP significantly reduces token usage without sacrificing accuracy—for instance, cutting tokens on GSM8K from 917 to 328 while improving accuracy from 91.7% to 94.1%, and reducing AIME tokens by 43% with no performance drop. Further analysis shows that aligning training CoT structure with the student’s capacity is key to effective knowledge transfer.

Qianyu Wang, Xiaoman Wang, Yuanyuan Liang, Xinyuan Li, Yunshi Lan

Grammatical error correction (GEC) systems are usually trained and evaluated on GEC benchmarks, but their performance often drops sharply once the surrounding context is slightly perturbed or extended. This indicates that the existing GEC models usually fail to understand the error patterns in the varying contexts. In this paper, we thoroughly investigate the counterfactuals for GEC tasks, where the subtle changes to the contexts could lead to the label flipping issue. We address this robustness gap by viewing contextual variation through the lens of counterfactual data. We propose CoCoGEC, a counterfactual generation framework that creates copies of training instances with error-irrelevant contexts altered. Our framework systematically generates counterfactuals by (1) generating intra- and inter-sentence counterfactuals that maintain the error patterns as well as syntax of the original instances by altering the word-level and sentence-level contexts; (2) revising the generated counterfactuals by selecting the instances with flipped labels and high GEC Mutual Information (MI) coefficient. Extensive experiments show that our method substantially improves the stability of GEC models, outperforming a set of data augmentation baselines. Particularly, it could achieve absolute F_{0.5} gains of +9.9, +11.3, and +20.8 points on the perturbed BEA-19*,CoNLL-14*, and TEM-8* data set.Our code is released at https://github.com/Quinnok/CoCoGEC.

Panuthep Tasawong, Jian Gang Ngui, Alham Fikri Aji, Trevor Cohn, Peerat Limkonchotiwat

Safeguard models help large language models (LLMs) detect and block harmful content, but most evaluations remain English-centric and overlook linguistic and cultural diversity. Existing multilingual safety benchmarks often rely on machine-translated English data, which fails to capture nuances in low-resource languages. Southeast Asian (SEA) languages are underrepresented despite the region’s linguistic diversity and unique safety concerns, from culturally sensitive political speech to region-specific misinformation. Addressing these gaps requires benchmarks that are natively authored to reflect local norms and harm scenarios. We introduce SEA-SafeguardBench, the first human-verified safety benchmark for SEA, covering eight languages, 21,640 samples, across three subsets: general, in-the-wild, and content generation. The experimental results from our benchmark demonstrate that even state-of-the-art LLMs and guardrails are challenged by SEA cultural and harm scenarios and underperform when compared to English texts.

Minh Duc Bui, Xenia Heilmann, Mattia Cerrato, Manuel Mager, Katharina von der Wense

Prior work evaluates code generation bias primarily through simple conditional statements, which represent only a narrow slice of real-world programming and reveal solely overt, explicitly encoded bias. We demonstrate that this approach dramatically underestimates real-world bias by examining a more realistic task: generating machine learning (ML) pipelines. Testing both code-specialized and general-instruction large language models, we find that ML pipelines exhibit substantially greater bias than simple conditionals across all conditions: standard generation, with varying prompt-based mitigation strategies, varying numbers of attributes, and different ML pipeline difficulty levels. Even attribute selection alone, the simplest pipeline difficulty, shows higher bias compared to conditionals, demonstrating that ML pipelines inherently amplify bias beyond what isolated conditionals reveal. Critically, we uncover a stark asymmetry: models maintain equivalent bias detection performance on both simple conditionals and ML pipelines, revealing that models recognize bias equally well in both contexts yet generate significantly more biased code in ML pipelines. These findings challenge simple conditionals as valid proxies for bias evaluation and suggest current benchmarks mischaracterize model safety in practical deployment contexts.

Xiaobin Tian, Shuai Yuan, Muyun Ding, Haonan Chen, Xiaoxi Jiang

Large Reasoning Models (LRMs) benefit from generating intermediate reasoning steps, enabling more reliable and interpretable decision-making. While outcome-based supervision has proven effective for LRMs across diverse tasks, it focuses solely on final answers and cannot guarantee high-quality intermediate reasoning. In contrast, existing process supervision is largely limited to verifiable domains such as mathematics or code, where intermediate steps can be explicitly checked, restricting its applicability to open-ended reasoning tasks. To address these limitations, we propose Rubrics-in-Thinking Reinforcement Learning (RiT), the first framework to introduce thinking-rubric supervision into intermediate reasoning. RiT automatically generates fine-grained rubrics and integrates them into a reward function via gated fusion with outcome-based rewards, guiding models to reason in a coherent and task-aligned manner, improving both intermediate steps and the final response. Experiments on reasoning-intensive and open-ended benchmarks demonstrate that RiT consistently outperforms outcome-only RL baselines.

Xinjie Xu, Yongqi Fan, Shuang-shuang Chen, Qi Ye, Weibin Guo, Xinxuan Hu

Accurate International Classification of Diseases (ICD) coding is crucial for hospital management and healthcare data governance. In clinical practice, straightforward cases can often be matched directly to ICD codes via diagnostic text, establishing retrieval-based methods as the baseline. More advanced approaches leverage large language models to rerank these results. However, real-world coding scenarios are typically more complex, demanding reasoning that goes beyond superficial descriptions. For instance, it involves synthesizing key information such as disease subtype, anatomical location, and complications from complex progress notes to accurately identify the primary diagnosis. However, a comprehensive evaluation framework for ICD coding based on complete EMRs is still lacking. To address these challenges, we constructed the Code4Detail dataset, which comprises 560 real clinical records covering 434 common diseases across 19 core chapters of ICD-10. To systematically explore the capability boundaries of large language models under different paradigms, we further propose the Travel on the ICD Tree (ToT-ICD) evaluation framework. Unlike the conventional retrieval-recall approach, ToT-ICD treats ICD coding as a structured exploration process across a hierarchical taxonomy. We design an agentic workflow that integrates similarity retrieval, path-guided navigation, and dynamic backtracking, enabling logical reasoning and decision-making under coding rules.

Amin Rakhsha, Thomas Hehn, Pietro Mazzaglia, Fabio Valerio Massoli, Arash Behboodi, Tribhuvanesh Orekondy

Large language models can perform well on many isolated tasks, yet they continue to struggle on multi-turn, long-horizon agentic problems that require skills such as planning, state tracking, and long context processing. In this work, we aim to better understand the relative importance of advancing these underlying capabilities for success on such tasks. We develop an oracle counterfactual framework for multi-turn problems that asks: how would an agent perform if it could leverage an oracle to perfectly execute a specific skill? The change in the agent’s performance due to this oracle assistance allows us to measure the criticality of that skill in the future advancement of AI agents. We introduce a suite of procedurally generated, game-like tasks with tunable complexity. These controlled environments allow us to provide precise oracle interventions, such as perfect planning or flawless state tracking, and make it possible to isolate the contribution of each oracle without confounding effects present in real-world benchmarks. Our results show that while some interventions (e.g., planning) consistently improve performance across settings, the usefulness of other skills is dependent on the properties of the environment and language model. Our work sheds light on the challenges of multi-turn agentic environments to guide the future efforts in the development of AI agents and language models.

Heng Yu, Rui Li, Qi Liu, Wenjun Feng, Junfeng Kang, Yi Zhan

Retrieval-augmented generation (RAG) effectively extends the knowledge boundaries of large language models (LLMs) for complex tasks, yet current paradigms typically optimize for an interleaving of reasoning and retrieval, where models fail to critically evaluate retrieved information against the target question. Most existing methods rely on sparse outcome-based rewards, failing to provide explicit supervision for the internal reasoning process or to diagnose information inadequacy. To address this, we propose Eval-RAR, an Evaluation-driven Retrieval-Augmented Reasoning framework. Eval-RAR introduces a "Search-then-Evaluate" paradigm where the model performs explicit self-evaluation after each search step, generating a rationale to either identify sufficient evidence or specify missing information to guide subsequent queries. To optimize this process, we employ reinforcement learning with a fine-grained evaluation reward, providing intermediate feedback that encourages the model to track core entities and maintain logical consistency. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate that Eval-RAR outperforms existing methods.

Junyan Cheng, Ankit Srivastava, Jessie Zeng, Milenko Drinic, Jack W. Stokes

We introduce Apeiron, a scalable and extensible framework for addressing *amorphous* user demands through autonomous, full-lifecycle application synthesis. Apeiron models the unstructured app development process as a heuristic optimization problem combining (i) a Computer-Use Agent (CUA) evaluator that simulates personas and demands, (ii) an *Activity Tracer* that grounds feedback in code-level interaction traces, and (iii) a *Locality Controller* that constrains changes during continuous integration and delivery (CI/CD). Furthermore, we introduce an innovative data generation approach using CUA-as-a-Judge to tackle data scarcity. Across 300 app scenarios, 2,400 personas, and 46,338 demands, Apeiron outperformed baselines by 10.7% in CUA ratings and 27.8% in user-demand task scores. The optimization process enhances task scores by 64.7%, and the tracer contributes a 25.1% gain. In CI/CD, Apeiron effectively restores 96.9% of the pre-shift mean CUA rating in one optimization step with <30% code changes in response to 30% demand shifts. Finally, a user study (N=18) shows that our CUA ratings strongly correlate with human judgment (Spearman’s \rho=0.685) and that users prefer Apeiron-synthesized apps over baselines.

Shouqing Yang, Qi Zhang, Yuhang Yang, Ruikang Xu, Yuwei Hou, Zhulin Jia, Lirong Gao, Haobo Wang, Jinglei Chen, Jiexiang Wang 等

Retrieval-augmented generation (RAG) has become a widely adopted paradigm for realistic financial analysis over financial documents. However, existing benchmarks fail to capture realistic financial analysis settings that involve cross-document retrieval, multi-page evidence integration, and diverse analytical tasks. To address this gap, we introduce \textbf{FinMRAGBench}, a comprehensive multi-modal financial RAG benchmark in which most questions require retrieving evidence scattered across multiple pages and documents, constructed from large-scale real-world annual reports and comprising 887 expert-verified QA pairs spanning five representative financial analysis tasks. Moreover, we introduce \textbf{FinMRAGAgent}, an agent trained on high-quality agentic trajectories following the reasoning-and-acting (ReAct) paradigm, capable of dynamic tool invocation and multi-step financial analysis. Our extensive experiments show that current multi-modal RAG systems still struggle with incomplete retrieval and complex financial reasoning. In contrast, FinMRAGAgent achieves the strongest overall performance across all models, demonstrating that our structured reasoning approach significantly enhances multi-modal RAG in realistic financial scenarios. The code and data are available at https://github.com/sqyangit/FinMRAGBench.

Jianfei Wu, Zhichun Wang, Zhensheng Wang, Zhiyu He

While Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, their potential for purpose-driven exploration in dynamic geo-spatial environments remains under-investigated. Existing Geo-Spatial Question Answering (GSQA) benchmarks predominantly focus on static retrieval, failing to capture the complexity of real-world planning that involves dynamic user locations and compound constraints. To bridge this gap, we introduce EVGeoQA, a novel benchmark built upon Electric Vehicle (EV) charging scenarios that features a distinct location-anchored and dual-objective design. Specifically, each query in EVGeoQA is explicitly bound to a user’s real-time coordinate and integrates the dual objectives of a charging necessity and a co-located activity preference. To systematically assess models in such complex settings, we further propose GeoRover, a general evaluation framework based on a tool-augmented agent architecture to evaluate the LLMs’ capacity for dynamic, multi-objective exploration. Our experiments reveal that while LLMs successfully utilize tools to address sub-tasks, they struggle with long-range spatial exploration. Notably, we observe an emergent capability: LLMs can summarize historical exploration trajectories to enhance exploration efficiency. These findings establish EVGeoQA as a challenging testbed for future geo-spatial intelligence. The dataset and prompts are available at https://github.com/Hapluckyy/EVGeoQA/.

Peng Liao, Peijia Zheng, Lingbo Li, Shangsong Liang, Lin Chen

Offline preference optimization methods, such as Direct Preference Optimization (DPO), offer significant advantages in aligning Large Language Models (LLMs) with human values. However, achieving optimal performance with these methods typically involves additional hyperparameter tuning, resulting in substantial time overhead. Although prior work has proposed a range of improvements, these methods remain limited in effectiveness and have not fully eliminated reliance on hyperparameter tuning. In this work, we introduce RMiPO, a lightweight and efficient framework for offline preference optimization. RMiPO leverages intrinsic **R**esponse-level **M**utual **i**nformation for **P**reference **O**ptimization with hyperparameter modulation, dynamically decoupling preference contributions at negligible additional computational cost. Extensive experimental results demonstrate that RMiPO achieves consistently superior performance over existing methods while reducing training overhead by more than 15%. Our code is available at https://github.com/liavonpenn/rmipo.

Yuxiang Chai, Shunye Tang, Han Xiao, Weifeng Lin, Hanhao Li, Jiayu Zhang, Liang Liu, Pengxiang Zhao, Guangyi Liu, Guozhi Wang 等

The advancement of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has catalyzed the development of mobile graphic user interface (GUI) AI agents, which is designed to autonomously perform tasks on mobile devices. However, a significant gap persists in mobile GUI agent evaluation, where existing benchmarks predominantly rely on either static frame assessments such as AndroidControl or offline static apps such as AndroidWorld and thus fail to capture agent performance in dynamic, real-world online mobile apps. To address this gap, we present Android Agent Arena (A3), a novel "essential-state" based procedural evaluation system for mobile GUI agents. A3 introduces a benchmark of 100 tasks derived from 20 widely-used, dynamic online apps across 20 categories from the Google Play Store, ensuring evaluation comprehension. A3 also presents a novel "essential-state" based procedural evaluation method that leverages MLLMs as reward models to progressively verify task completion and process achievement. This evaluation approach address the limitations of traditional function based evaluation methods on online dynamic apps. Furthermore, A3 includes a toolkit to streamline Android device interaction, reset online environment and apps and facilitate data collection from both human and agent demonstrations. The complete A3 system, including the benchmark and tools, will be publicly released to provide a robust foundation for future research and development in mobile GUI agents.