Reinforcement learning (RL) has emerged as a powerful paradigm for improving the reasoning capabilities of large language models (LLMs). Despite its success, RL faces fundamental challenges, including low sample efficiency and a strong dependence on the quality of the base model: while some models improve rapidly with limited RL updates, others require substantial training data to achieve meaningful gains. Recent studies suggest that the patterns of thinking tokens play a critical role in RL performance, and that supervised fine-tuning (SFT) on datasets exhibiting desirable reasoning patterns can reduce reliance on base models and better prepare LLMs for RL. However, how to automatically discover such patterns across tasks remains unclear. In this work, we describe thinking token patterns with reasoning primitives and argue that initializing LLMs with diverse, high-quality primitives is crucial for stable and efficient RL training. We propose Tailor, a pipeline that automatically discovers such reasoning primitives and curates SFT datasets to prepare LLMs for RL. Extensive experiments on mathematical and logical reasoning benchmarks demonstrate that Tailor consistently improves downstream RL performance, outperforming strong baselines, including methods with expert domain knowledge.
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
Prompt injection attacks pose serious security risks across a wide range of real-world applications. While receiving increasing attention, the community faces a critical gap: the lack of a unified platform for prompt injection evaluation. This makes it challenging to reliably compare defenses, understand their true robustness under diverse attacks, or assess how well they generalize across tasks and benchmarks. For instance, many defenses initially reported as effective were later found to exhibit limited robustness on diverse datasets and attacks. To bridge this gap, we introduce PIArena, a unified and extensible platform for prompt injection evaluation that enables users to easily integrate state-of-the-art attacks and defenses and evaluate them across a variety of existing and new benchmarks. We also design a dynamic strategy-based attack that adaptively optimizes injected prompts based on defense feedback. Through comprehensive evaluation using PIArena, we uncover critical limitations of state-of-the-art defenses: limited generalizability across tasks, vulnerability to adaptive attacks, and fundamental challenges when an injected task aligns with the target task. The code and datasets are available at https://github.com/sleeepeer/PIArena.
Diagnosing Spatial Consistency across Perspectives and Viewpoints in Large Vision-Language Models
PDF ↗Consistent reasoning about 3D spatial relations across changing viewpoints is fundamental for Embodied AI agents operating in dynamic environments. While Large Vision-Language Models (LVLMs) have advanced multimodal perception, their ability to maintain spatial consistency across diverse perspectives remains underexplored. Existing benchmarks primarily assess spatial capabilities from a static, single-view, and egocentric perspective, failing to capture the dynamic nature of real-world spatial cognition.To address this gap, we introduce SCOPE (Spatial COnsistency across PErspectives and Viewpoints), a comprehensive benchmark designed to rigorously diagnose spatial reasoning capabilities. Grounded in human cognitive theories of dual spatial representations, SCOPE discretizes the 360^\circ field into multiview scenarios to systematically evaluate both allocentric and egocentric reasoning capabilities. Our dataset comprises 20.1K spatial VQA pairs derived from high-quality 3D environments. Through an extensive evaluation of 26 state-of-the-art LVLMs, we identify two fundamental limitations that prevent consistent spatial understanding across viewpoints.We hope SCOPE facilitates the diagnosis of spatial reasoning, serving as a stepping stone toward reliable embodied action.
Alexandria: A Multi-Domain Dialectal Arabic Machine Translation Dataset for Culturally Inclusive and Linguistically Diverse LLMs
PDF ↗Arabic is a highly diglossic language where most daily communication occurs in regional dialects rather than Modern Standard Arabic (MSA). Despite this, machine translation (MT) systems often generalize poorly to dialectal input, limiting their utility for millions of speakers. We introduce Alexandria, a large-scale, community-driven, human-translated dataset designed to bridge this gap. Alexandria covers 13 Arab countries and 11 high-impact domains, including health, education, and agriculture. Unlike previous resources, Alexandria provides unprecedented granularity by associating contributions with city-of-origin metadata, capturing authentic local varieties beyond coarse regional labels. The dataset consists of parallel English-Dialectal Arabic multi-turn conversational scenarios annotated with speaker-addressee gender configurations, enabling the study of gender-conditioned variation in dialectal use. Comprising 107K total turns, Alexandria serves as both a training resource and as a rigorous benchmark for evaluating MT and Large Language Models (LLMs). Our automatic and human evaluation benchmarks the current capabilities of Arabic-aware LLMs in translating across diverse Arabic dialects and sub-dialects while exposing significant persistent challenges.The Alexandria dataset, the creation prompts, the translation and revision guidelines, and the evaluation code are publicly available in the following repository: https://github.com/UBC-NLP/Alexandria
Model editing provides a promising mechanism for updating large language models (LLMs) without expensive retraining. Existing approaches, particularly locate-and-edit methods based on least-squares optimization, aim to introduce targeted knowledge changes while preserving pre-trained behavior. In this work, we show that this objective is fundamentally fragile under standard single-edit evaluation protocols. We first develop a unified theoretical framework that characterizes activation-based editing as a constrained intervention on intermediate representations. Within this framework, we demonstrate that least-squares edits cannot, in general, isolate target updates from unrelated activations, giving rise to unavoidable interference that accumulates with successive edits. Crucially, this degradation can remain undetected in single-edit settings when assessed using conventional success and locality metrics. To expose such hidden instabilities, we introduce an uncertainty-based evaluation protocol that combines structured semantic perturbations with uncertainty quantification based on Sampling with Perturbation for UQ. By measuring edit-induced growth in aleatoric and epistemic uncertainty, our method reveals local knowledge conflicts that are invisible to existing benchmarks. Extensive experiments across multiple models, datasets, and editing algorithms show that both least-squares and other parameter-update-based methods consistently increase post-edit uncertainty. Together, our results suggest that current evaluation practices substantially overestimate the reliability of single-edit model editing, and that uncertainty-based diagnostics are necessary for assessing edit stability.
Large language models (LLMs) in retrieval-augmented generation systems can still produce hallucinations, generating content that is unsupported or contradicted by the source texts and undermines reliability. Recent work addressed this problem by training span-level hallucination detectors using reinforcement learning (RL) and chain-of-thought (CoT) reasoning. In this work, we show through error analysis that incorrect predictions by existing reasoning-based detectors are strongly associated with CoT processes that lack explicit grounding in source evidence, particularly when verification steps do not quote or verify claims against the retrieved documents. This behaviour contrasts with human verification practices in benchmarks such as RAGTruth, where evidence quotation is a prerequisite for determining hallucinated spans. Motivated by this observation, we propose an evidence-grounded RL framework, namely RLSeek, to explicitly enforce active evidence seeking during CoT reasoning by requiring quotation of relevant source segments at each verification step. Experiments on the RAGTruth and NewsSum dataset demonstrate consistent improvements in hallucination span detection performance, with limited additional reasoning overhead and improved robustness in out-of-domain settings.
StratMem-Bench: Evaluating Strategic Memory Use in Virtual Character Conversation Beyond Factual Recall
PDF ↗Achieving realistic human-like conversation for virtual characters requires not only a simple memorization and recall of past events, but also the strategic utilization of memory to meet factual needs and social engagement. Current memory utilization relevant (e.g., memory-augmented generation, long-term dialogue, and etc.) benchmarks overlook this nuance, treating memory primarily as a static repository of facts rather than a dynamic resource to be strategically deployed in dialogues. To address this gap, we design StratMem-Bench, a new benchmark to evaluate strategic memory use in character-centric dialogues. This dataset comprises 657 instances where virtual characters must navigate heterogeneous memory pools containing required, supportive, and irrelevant memories. We also propose a framework with different evaluation metrics including Strict Memory Compliance, Memory Integration Quality, Proactive Enrichment Score and Conditional Irrelevance Rate, and to evaluate strategic memory use capabilities of virtual characters. Experiments on StratMem-Bench which leverage the state-of-the-art large language models as virtual characters show that all models perform well at distinguishing between required and irrelevant memories, but struggle once supportive memories are introduced into the decision process.
Evaluating the factual consistency of abstractive text summarization remains a significant challenge, particularly for long documents, where conventional metrics struggle with input length limitations and long-range dependencies. In this work, we systematically evaluate the reliability of six widely used reference-free factuality metrics, originally proposed for short-form summarization, in the long-document setting. We probe metric robustness through seven factuality-preserving perturbations applied to summaries, namely paraphrasing, simplification, synonym replacement, logically equivalent negations, vocabulary reduction, compression, and source text insertion, and further analyze their sensitivity to retrieval context and claim information density. Across three long-form benchmark datasets spanning science fiction, legal, and scientific domains, our results reveal that existing short-form metrics produce inconsistent scores for semantically equivalent summaries and exhibit declining reliability for information-dense claims whose content is semantically similar to many parts of the source document. While expanding the retrieval context improves stability in some domains, no metric consistently maintains factual alignment under long-context conditions. Finally, our results highlight concrete directions for improving factuality evaluation, including multi-span reasoning, context-aware calibration, and training on meaning-preserving variations to enhance robustness in long-form summarization.
Grounded claim factuality checking is important for large language model (LLM) applications such as retrieval-augmented generation, as it helps users assess the correctness of generated outputs. Existing metrics using entailment classifiers require dataset-specific threshold tuning, while LLM-based approaches often use direct prompting, which underutilises the reasoning capabilities of LLMs. We address this by formulating grounded claim factuality checking as a true/false reading comprehension task and prompting LLMs with explicit test-taking strategies for efficient reasoning. Our method reduces token usage by over 80% compared to unguided open-ended reasoning, and achieves competitive performance to more expensive alternatives across two factuality benchmarks, setting a new state of the art on one. To further reduce inference cost, we train small language models (SLMs) to replace LLMs in the checking pipeline. Using supervised fine-tuning (SFT) and a self-revision mechanism, the SLMs learn to improve their factuality judgements. Experimental results show that the resulting SLMs perform on par with strong baselines, combining low inference costs with generating supporting rationales to support interpretability. Code and datasets will be released upon acceptance.
We propose UniVocal, a unified framework that implicitly infers vocal modes from text context to pioneer Speech-Singing Code-Switching (SCS) Synthesis—a task where transitions are autonomously driven by textual semantics, akin to seamless human language blending. Unlike single-mode generation or systems relying on switching-control tags, our proposed UniVocal implicitly infers vocal modes solely from text context. To achieve this, we employ a data-efficient two-stage curriculum learning strategy that progressively trains a competitive TTS system to acquire the desired SCS capability. Addressing data scarcity, we introduce a scalable pipeline to synthesize diverse code-switching data that is both semantically and acoustically natural, alongside a new multi-scenario benchmark, SCSBench. To address limitations of semantic tokenizers in capturing acoustic details, we also introduce refined cent token and Chain-of-Thought (CoT) generation for planning prosody before content generation, effectively enhancing empathetic speech generation and singing melody. Experimental results demonstrate that UniVocal achieves state-of-the-art performance on SCSBench while maintaining competitive performance on regular speech and singing tasks. Audio samples are available at https://project-univocal-demo.github.io/demo/. The code and dataset are released at https://github.com/FunAudioLLM/FunResearch/tree/main/UniVocal.
Recent advances in large language models (LLMs) and text-aware graph learning have increased interest in reasoning over text-attributed graphs(TAGs). In many real-world settings, such graphs are inherently heterogeneous, with most existing benchmarks remaining largely homogeneous in structure. As a result, the lack of large-scale benchmarks for heterogeneous text-attributed graphs has hindered systematic evaluation and fair comparison of existing methods. In this work, we introduce CITE - **C**atalytic **I**nformation **T**extual **E**ntities Graph, the first and largest heterogeneous text-attributed citation graph benchmark for catalytic materials. CITE contains over 438K nodes and 1.2M edges spanning four node types and four relation types, with rich node-level textual information. We establish standardized evaluation protocols for node classification and link prediction, and conduct ablation studies to assess the impact of graph heterogeneity and textual attributes. Using CITE, we benchmark four classes of learning paradigms, including homogeneous graph models, heterogeneous graph models, LLM-centric models, and hybrid LLM–graph models. By providing a large-scale heterogeneous text-attributed benchmark together with standardized evaluation protocols and comprehensive baselines, CITE enables systematic assessment across diverse modeling paradigms and offers new insights into text-aware and LLM-enhanced graph learning. The dataset, codebase and evaluation suite are publicly available.
Large Language Models (LLMs) have demonstrated strong capabilities as autonomous agents through tool use, planning, and decision-making abilities, leading to their widespread adoption across diverse tasks. As task complexity grows, multi-agent LLM systems are increasingly used to solve problems collaboratively. However, safety and security of these systems remains largely under-explored. Existing benchmarks and datasets predominantly focus on single-agent settings, failing to capture the unique vulnerabilities of multi-agent dynamics and co-ordination. To address this gap, we introduce \textbf{T}hreats and \textbf{A}ttacks in \textbf{M}ulti-\textbf{A}gent \textbf{S}ystems (\textbf{TAMAS}), a benchmark designed to evaluate the robustness and safety of multi-agent LLM systems. TAMAS includes five distinct scenarios comprising 300 adversarial instances across six attack types and 211 tools, along with 100 harmless tasks. We assess system performance across ten backbone LLMs and three agent interaction configurations from Autogen and CrewAI frameworks, highlighting critical challenges and failure modes in current multi-agent deployments. Furthermore, we introduce Effective Robustness Score (ERS) to assess the tradeoff between safety and task effectiveness of these frameworks. Our findings show that multi-agent systems are highly vulnerable to adversarial attacks, underscoring the urgent need for stronger defenses. TAMAS provides a foundation for systematically studying and improving the safety of multi-agent LLM systems. Code and dataset is available at https://github.com/microsoft/TAMAS.
In large language model (LLM) agents, reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating memory. However, coherent reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate reasoning trajectories, we propose Self-Audited Verified Reasoning (SAVeR), a novel framework that enforces verification over internal belief states within the agent before action commitment, achieving faithful reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve reasoning faithfulness, we perform adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six benchmark datasets demonstrate that our approach consistently improves reasoning faithfulness while preserving competitive end-task performance.
SwissGov-RSD: A Human-annotated, Cross-lingual Benchmark for Token-level Recognition of Semantic Differences Between Related Documents
PDF ↗Recognizing semantic differences across documents is crucial for text generation evaluation and content alignment, especially in cross-lingual settings. However, as a standalone task, it has received little attention. We address this by introducing SwissGov-RSD, the first naturalistic, document-level, cross-lingual dataset for semantic difference recognition. It encompasses a total of 224 multi-parallel documents in English–German, English–French, and English–Italian with token-level difference annotations by human annotators.We evaluate a variety of open-source and closed-source large language models as well as encoder models across different fine-tuning settings on this new benchmark. Our results show that current automatic approaches perform poorly compared to their performance on monolingual, sentence-level, and synthetic benchmarks, revealing a considerable gap for both LLMs and encoder models.
Vision-language model-based mobile agents have gained the ability to understand complex instructions and mobile screenshots, benefiting from reinforcement learning paradigms like Group Relative Policy Optimization (GRPO). However, existing approaches centers on offline training or local action-level rewards often trap agents in local optima, hindering effective exploration and error correction with the environment. Crucially, we find that directly applying task-level rewards often leads to convergence difficulties due to the sparse nature of GUI interactions. To address these challenges, we present Mobile-R1, a systematic training recipe that bridges atomic action execution and strategic task completion. We propose a hierarchical curriculum consisting of three stages: (1) format alignment for reasoning structure, (2) on-policy exploration with verifiable action feedback to ground basic execution, and (3) multi-turn task-level training with realistic environment to unlock exploration and self-correction. This hierarchical strategy effectively bootstraps the agent, significantly enhancing its capability for exploration and self-correction (the “Eureka” moments). Furthermore, addressing the critical scarcity of diverse GUI data in non-English ecosystems, we contribute a comprehensive Chinese mobile dataset covering 28 applications with 24,521 high-quality manual annotations, and establish a rigorous benchmark with 500 trajectories. We will open source all resources, including the dataset, benchmark, model weight, and codes: https://mobile-r1.github.io/Mobile-R1/.
Entity alignment (EA) aims to identify entities across different knowledge graphs (KGs) that refer to the same real-world object and plays a critical role in knowledge fusion and integration. Traditional EA methods mainly rely on knowledge representation learning, but their performance is often limited under noisy or sparsely supervised scenarios. Recently, large language models (LLMs) have been introduced to EA and achieved notable improvements by leveraging rich semantic knowledge. However, existing LLM-based EA approaches typically treat LLMs as black-box decision makers, resulting in limited interpretability, and the direct use of large-scale triples substantially increases inference cost. To address these challenges, we propose EA-Agent, a reasoning-driven agent for EA. EA-Agent formulates EA as a structured reasoning process with multi-step planning and execution, enabling interpretable alignment decisions. Within this process, it introduces attribute and relation triple selectors to filter redundant triples before feeding them into the LLM, effectively addressing efficiency challenges. Experimental results on three benchmark datasets demonstrate that EA-Agent consistently outperforms existing EA methods and achieves state-of-the-art performance. The source code is available at https://anonymous.4open.science/r/EA-Agent-5696.
Existing jamming attacks on Retrieval-Augmented Generation (RAG) systems typically induce explicit refusals or denial-of-service behaviors, which are conspicuous and easy to detect. In this work, we formalize a subtler availability threat, termed soft failure, which degrades system utility by inducing fluent and coherent yet non-informative responses rather than overt failures. We propose Deceptive Evolutionary Jamming Attack (DEJA), an automated black-box attack framework that generates adversarial documents to trigger such soft failures by exploiting safety-aligned behaviors of large language models. DEJA employs an evolutionary optimization process guided by a fine-grained Answer Utility Score (AUS), computed via an LLM-based evaluator, to systematically undermine the certainty of answers while maintaining high retrieval success.Extensive experiments across multiple RAG configurations and benchmark datasets show that DEJA consistently drives responses toward low-utility soft failures and that the resulting adversarial documents maintain high stealth and effectiveness, proving resilient against common mitigation strategies including perplexity-based detection and input perturbations.
PersonalityDBench: A Dataset for Personality Disorders - from Modeling to Controlled Generation
PDF ↗Personality disorders (PDs) are a complex class of mental health (MH) conditions characterized by persistent patterns of cognition, behavior, and emotional regulation that deviate from cultural norms. While social media has become a valuable resource for MH research, NLP has largely focused on more prevalent conditions (e.g., depression), leaving PDs underexplored. In this work, we introduce PersonalityDBench, a large-scale, clinically grounded dataset that supports multidimensional study of personality pathology, and standardized, reproducible evaluation of LLM steering toward clinically grounded behavioral targets. The dataset comprises two parts: (1) PRISMA and (2) PersonaDSteering. (1) PRISMA (PeRsonality dISorder MAnifestations) is a clinically annotated collection of social media content spanning the full spectrum of PDs. It links clinically validated diagnostic criteria and dimensional trait frameworks with computational annotation and analysis methods to support fine-grained, multidimensional study of how PDs manifests in naturalistic, free-form language. Building on PRISMA, (2) PersonaDSteering is a benchmark for LLM steering evaluation that operationalizes clinically grounded PD profiles into structured behavioral elicitation tasks, enabling multidimensional steerability assessment beyond single-behavior settings and supporting PD-consistent persona construction for simulated patient generation. This dataset may have application in the study and modeling of PD and powering personality-specific text generation for adaptive, personalized chat systems.
From 1,000,000 Users to Every User: Scaling Up Personalized Preference for User-level Alignment
PDF ↗Large language models (LLMs) have traditionally been aligned through one-size-fits-all approaches that assume uniform human preferences, fundamentally overlooking the diversity in user values and needs. This paper introduces a comprehensive framework for scalable personalized alignment of LLMs. We establish a systematic preference space characterizing psychological and behavioral dimensions, alongside diverse persona representations for robust preference inference in real-world scenarios. Building upon this foundation, we introduce AlignX, a large-scale dataset of over 1.3 million personalized preference examples, and develop two complementary alignment approaches: in-context alignment directly conditioning on persona representations and preference-bridged alignment modeling intermediate preference distributions. Extensive experiments demonstrate substantial improvements over existing methods, with an average 17.06% accuracy gain across four benchmarks while exhibiting a strong adaptation capability to novel preferences, robustness to limited user data, and precise preference controllability. These results validate our approach toward user-adaptive AI systems.
Large language models (LLMs) have achieved remarkable performance across diverse tasks, largely driven by large-scale pretraining. However, this data abundance introduces test data contamination, where benchmark datasets overlap with pretraining corpora, undermining the reliability of model evaluation by confounding memorization with genuine generalization. To mitigate this issue, existing training data detectors attempt to identify clean (unseen) samples from contaminated test sets, but often suffer from residual contamination due to the black-box nature of LLMs. As a result, contaminated data may be mistakenly retained, leading to unreliable evaluation.To address this challenge, we propose FTD (FDR-controlled Training Data detection), a principled framework that detects and filters contaminated evaluation data while providing a statistical guarantee: the proportion of contaminated samples mistakenly retained as clean, the false discovery rate (FDR), is provably controlled below a user-specified threshold. FTD combines multiple complementary detectors via an adaptive weighting strategy, and we theoretically show it achieves high statistical power under valid FDR control. Extensive experiments on real-world benchmarks demonstrate that FTD significantly reduces residual contamination compared to existing methods while preserving evaluation consistency.