Large Reasoning Models (LRMs) have achieved remarkable performance across diverse domains, yet their decision-making under conflicting objectives remains insufficiently understood. This work investigates how LRMs respond to harmful queries when confronted with two categories of conflicts: internal conflicts that pit alignment values against each other and dilemmas, which impose mutually contradictory choices, including sacrificial, duress, agent-centered, and social forms. Using over 1,300 prompts across five benchmarks, we evaluate three representative LRMs - Llama-3.1-Nemotron-8B, QwQ-32B, and DeepSeek R1 - and find that conflicts significantly increase attack success rates, even under single-round non-narrative queries without sophisticated auto-attack techniques. Our findings reveal through layerwise and neuron-level analyses that safety-related and functional representations shift and overlap under conflict, interfering with safety-aligned behavior. This study highlights the need for deeper alignment strategies to ensure the robustness and trustworthiness of next-generation reasoning models. Our code is available at https://github.com/DataArcTech/ConflictHarm. Warning: This paper contains inappropriate, offensive and harmful content.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
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.
Despite strong performance on code generation tasks, it remains unclear whether large language models (LLMs) genuinely reason about code execution. Existing code reasoning benchmarks primarily evaluate final output correctness under a single canonical implementation, leaving two critical aspects underexplored: (1) whether LLMs predictions are consistent to functionally equivalent implementations, and (2) whether LLMs can accurately reason about intermediate execution states. We introduce CoRE, a Code Reasoning benchmark that evaluates code reasoning through implementation invariance and process transparency. Extensive evaluations on eight frontier LLMs reveal two fundamental limitations. First, models exhibit a substantial robustness gap, with performance varying significantly across equivalent implementations. Second, we observe superficial execution, where models arrive at correct final outputs without correctly reasoning about intermediate execution states. Together, these findings demonstrate that output-only evaluations are insufficient for assessing code reasoning and position CoRE as a necessary benchmark for evaluating robust and faithful code reasoning.
While LLM watermarking is essential for machine- generated content identification, existing paraphrase-based attacks struggle to balance watermark removal efficacy with text quality. We propose TSAPA, a training-free evolutionary framework that models watermark removal as a constrained multi-objective optimization problem. By leveraging genetic algorithms to navigate the Pareto front, TSAPA utilizes a Pseudo-Log-Likelihood (PLL)-guided mutation to precisely target and modify watermark-carrying tokens. Experiments on Qwen3 series (1.7B/8B/32B) across multiple watermark schemes show that TSAPA achieves over 90% attack success rate (ASR) while maintaining high text semantic fidelity, significantly outperforming baselines methods. This work exposes critical vulnerabilities in current watermarks and provides a new perspective for robust evaluation.
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.
Large language models (LLMs) offer a promising way to simulate human survey responses, potentially reducing the cost of large-scale data collection. However, existing zero-shot methods suffer from prompt sensitivity and low accuracy, while conventional fine-tuning approaches mostly fit the training set distributions and struggle to produce results more accurate than the training set itself, which deviates from the original goal of using LLMs to simulate survey responses. Building on this observation, we introduce Distribution Shift Alignment (DSA), a two-stage fine-tuning method that aligns both the output distributions and the distribution shifts across different backgrounds. By learning how these distributions change rather than fitting training data, DSA can provide results substantially closer to the true distribution than the training data. Empirically, DSA consistently outperforms other methods on five public survey datasets. We further conduct a comprehensive comparison covering accuracy, robustness, and data savings. DSA reduces the required real data by 53.48-69.12%, demonstrating its effectiveness and efficiency in survey simulation.
AlphaQuanter: An End-to-End Tool-Augmented Agentic Reinforcement Learning Framework for Stock Trading
PDF ↗While Large Language Model (LLM) agents show promise in automated trading, they still face critical limitations. Prominent multi-agent frameworks often suffer from inefficiency, produce inconsistent signals, and lack the end-to-end optimization required to learn a coherent strategy from market feedback. To address this, we introduce **AlphaQuanter**, a single-agent framework that uses reinforcement learning (RL) to learn a dynamic policy over a transparent, tool-augmented decision workflow, which empowers a single agent to *autonomously orchestrate tools* and *proactively acquire information* on demand, establishing a transparent reasoning process. Extensive experiments demonstrate that AlphaQuanter achieves state-of-the-art performance on key financial metrics. Besides, human evaluation shows the learned reasoning patterns reveal more faithful and coherent tool-usage behaviors, providing steps toward verifiable LLM-driven trading. Our code and data can be found at https://github.com/horizon-llm/AlphaQuanter.
LLMs Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions
PDF ↗Previous research has shown that LLMs finetuned on incorrect completions within narrow domains (\textit{e.g.}, insecure code or incorrect medical advice) can become broadly misaligned to exhibit harmful behaviors, which is called \textit{emergent misalignment}. In this work, we investigate whether this phenomenon can extend beyond safety behaviors to a broader spectrum of dishonesty and deception under high-stakes scenarios (\textit{e.g.}, lying under pressure and deceptive behavior). To explore this, we finetune open-sourced LLMs on misaligned completions across diverse domains. Experimental results demonstrate that LLMs show broadly misaligned behavior in dishonesty. Additionally, we further explore this phenomenon in a downstream combined finetuning setting, and find that introducing as little as 1% of misalignment data into a standard downstream task is sufficient to decrease honest behavior over 20%. Furthermore, we consider a more practical human-AI interaction environment where we simulate both benign and biased users to interact with the assistant LLM. Furthermore, we simulate both benign and biased users to interact with the assistant LLM, producing 20k trajectories for self-training in a more practical human-AI interaction environment. Notably, we find that the assistant model can be misaligned unintentionally to exacerbate its dishonesty with only 10% biased user population. In summary, we extend the study of \textit{emergent misalignment} to the domain of dishonesty under high-stakes scenarios, and highlight that this risk arises not only through direct finetuning, but also in downstream mixture tasks and human-AI interactions.
How Language Models Conflate Logical Validity with Plausibility: A Representational Analysis of Content Effects
PDF ↗Both humans and large language models (LLMs) exhibit content effects: biases in which the plausibility of the semantic content of a reasoning problem influences judgments regarding its logical validity. While this phenomenon in humans is best explained by the dual-process theory of reasoning, the mechanisms behind content effects in LLMs remain unclear. In this work, we address this issue by investigating how LLMs encode the concepts of validity and plausibility within their internal representations. We show that both concepts are linearly represented and strongly aligned in representational geometry, leading models to conflate plausibility with validity. Using steering vectors, we demonstrate that plausibility vectors can causally bias validity judgements, and vice versa, and that the degree of alignment between these two concepts predicts the magnitude of behavioral content effects across models. Finally, we construct debiasing vectors that disentangle these concepts, reducing content effects and improving reasoning accuracy. Our findings advance understanding of how abstract logical concepts are represented in LLMs and highlight representational interventions as a path toward more logical systems.
Query-focused table summarization requires generating natural language summaries of tabular data conditioned on a user query, enabling users to access insights beyond fact retrieval. Existing approaches face key limitations: table-to-text models require costly fine-tuning and struggle with complex reasoning, prompt-based LLM methods suffer from token-limit and efficiency issues while exposing sensitive data, and prior agentic pipelines often rely on decomposition, planning, or manual templates that lack robustness and scalability. To mitigate these issues, we introduce an agentic workflow, FACTS, a Fast, Accurate, and Privacy-Compliant Table Summarization approach via Offline Template Generation. FACTS produces offline templates, consisting of SQL queries and Jinja2 templates, which can be rendered into natural language summaries and are reusable across multiple tables sharing the same schema. It enables fast summarization through reusable offline templates, accurate outputs with executable SQL queries, and privacy compliance by sending only table schemas to LLMs. Evaluations on widely-used benchmarks show that FACTS consistently outperforms baseline methods, establishing it as a practical solution for real-world query-focused table summarization. Our code is available at https://github.com/BorealisAI/FACTS.
Large language models (LLMs) are increasingly capable of generating personalized, persuasive text at scale, raising new questions about bias and fairness in automated communication. This paper presents the first systematic analysis of how LLMs behave when tasked with demographic-conditioned targeted messaging. We introduce a controlled evaluation framework using three leading models: GPT-4o, Llama-3.3, and Mistral-Large-2.1, across two generation settings: **Standalone Generation**, which isolates intrinsic demographic effects, and **Context-Rich Generation**, which incorporates thematic and regional context to emulate realistic targeting. We evaluate generated messages along three dimensions: *lexical content*, *language style*, and *persuasive framing*. We instantiate this framework on ***climate communication*** and find consistent age- and gender-based asymmetries across models: male- and youth-targeted messages tend to emphasize more *assertive* and *progressive* framing, while female- and senior-targeted messages more often reflect *warmth*, *care*, and *traditional* themes. Contextual prompts systematically amplify these disparities, with *persuasion* scores being higher for male-targeted messages, while age-related differences vary across models. Our findings demonstrate how demographic stereotypes can surface and intensify in LLM-generated targeted communication, underscoring the need for bias-aware generation pipelines and transparent auditing frameworks that explicitly account for demographic conditioning in socially sensitive applications.
Zero-shot text-to-speech models can clone a speaker’s timbre from a short reference audio, but they also strongly inherit the speaking style present in the reference. As a result, synthesizing speech with a desired style often requires carefully selecting reference audio, which is impractical when only limited or mismatched references are available. While recent controllable TTS methods attempt to address this issue, they typically rely on absolute style targets and discrete textual prompts, and therefore do not support continuous and reference-relative style control. We propose ReStyle-TTS, a framework that enables continuous and reference-relative style control in zero-shot TTS. Our key insight is that effective style control requires first reducing the model’s implicit dependence on reference style before introducing explicit control mechanisms. To this end, we introduce Decoupled Classifier-Free Guidance (DCFG), which independently controls text and reference guidance, reducing reliance on reference style while preserving text fidelity. On top of this, we apply style-specific LoRAs together with Orthogonal LoRA Fusion to enable continuous and disentangled multi-attribute control, and introduce a Timbre Consistency Optimization module to mitigate timbre drift caused by weakened reference guidance. Experiments show that ReStyle-TTS enables user-friendly, continuous, and relative control over pitch, energy, and multiple emotions while maintaining intelligibility and speaker timbre, and performs robustly in challenging mismatched reference–target style scenarios. Code and data are available in supplementary materials.
Audio deepfakes pose a significant security threat, yet current state-of-the-art (SOTA) detection systems do not generalize well to realistic in-the-wild deepfakes. We introduce a novel In-Context Learning paradigm with comparison-guidance for Audio Deepfake detection (ICLAD). The framework enables the use of audio language models (ALMs) for training-free generalization to unseen deepfakes and provides rich textual explanations on the detection outcome. At the core of ICLAD is a pairwise comparative reasoning strategy that guides the ALM to discover and filter hallucinations and deepfake-irrelevant acoustic attributes. The ALM works alongside a specialized deepfake detector, whereby a routing mechanism feeds out-of-distribution samples to the ALM. On in-the-wild datasets, ICLAD improves macro F1 over the specialized detector, with up to 2× relative improvement. Further analysis demonstrates the flexibility of ICLAD and its potential for deployment on recent open-source ALMs.
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.
Long-horizon agents operate over extended sequences of reasoning and actions, but this inevitably accumulates context noise, resulting in excessive computational cost and information overload. Existing approaches commonly rely on fixed, rule-based summarization strategies (e.g., summarizing every few steps), which are inflexible, lack generalization, and often introduce irreversible information loss. We propose Self-Sum, a framework that empowers agents to autonomously decide when and what to summarize by modeling summarization as a first-class internal cognitive action, unified with external environmental actions within a multi-turn decision-making process. Specifically, we introduce a two-stage training recipe consisting of (i) a cold-start supervised fine-tuning stage that bootstraps summarization behavior, and (ii) a lightweight, summarization-aware reinforcement learning stage that refines summarization timing and content while discouraging unnecessary summaries. Experiments on multiple long-horizon benchmarks show that Self-Sum consistently outperforms no-summarization and rule-based baselines, with particularly strong gains in generalization. Analysis further reveals that Self-Sum learns to summarize sparsely at meaningful moments and preserves task-relevant information, highlighting the importance of jointly learning when and what to summarize for robust long-horizon agent behavior.
Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse
PDF ↗Supervised Semantic Differential (SSD) is a mixed quantitative–interpretive method that models how text meaning varies with continuous individual-difference variables by estimating a semantic gradient in an embedding space and interpreting its poles through clustering and text retrieval. SSD applies PCA before regression, but currently no systematic method exists for choosing the number of retained components, introducing avoidable researcher degrees of freedom in the analysis pipeline. We propose a PCA sweep procedure that treats dimensionality selection as a joint criterion over representation capacity, gradient interpretability, and stability across nearby values of K. We illustrate the method on a corpus of short posts about artificial intelligence written by Prolific participants who also completed Admiration and Rivalry narcissism scales. The sweep yields a stable, interpretable Admiration-related gradient contrasting optimistic, collaborative framings of AI with distrustful and derisive discourse, while no robust alignment emerges for Rivalry. We also show that a counterfactual using a high-PCA dimension solution heuristic produces diffuse, weakly structured clusters instead, reinforcing the value of the sweep-based choice of K. The case study shows how the PCA sweep constrains researcher degrees of freedom while preserving SSD’s interpretive aims, supporting transparent and psychologically meaningful analyses of connotative meaning.
Maintaining a stable persona is central to sustained spoken role-playing, yet when an agent breaks character, current evaluations often do not isolate which component caused the failure, making fixes slow and ad hoc.We propose PED (Persona-Emotion Decoupling), a diagnostic evaluation framework that decomposes persona expression into two observable routes: what the agent says (text) and how it sounds (speech).PED operationalizes the affective slice of persona expression by projecting transcripts and audio into a shared affective measurement space for route-comparable, reference-based analyses of separability, drift, failures, and coupling.We demonstrate PED via two worked instantiations spanning an end-to-end Speech LLM and a cascaded LLM+TTS pipeline under a fixed dialogue protocol.Within this setting, PED surfaces four recurring diagnostic signatures:(i) route-level separability is bounded by reference overlap and can differ sharply across architectures,(ii) text-route drift is stress-linked and tends toward a neutral-heavy region,(iii) text-audio consistency is weakly coupled, yielding route-asymmetric failures,and (iv) audio-route structure can be materially shaped by an explicit intermediate style cue in cascaded pipelines.Overall, PED reframes holistic "voice+character" grading as turn-level, fault-localizing signals for faster debugging and iteration.
Do Multimodal RAG Systems Leak Data? A Comprehensive Evaluation of Membership Inference and Image Caption Retrieval Attacks
PDF ↗The growing adoption of multimodal Retrieval-Augmented Generation (mRAG) pipelines for vision-centric tasks (e.g., visual QA) introduces important privacy challenges. In particular, while mRAG provides a practical capability to connect private datasets and improve model performance, it risks the leakage of private information from these datasets. In this paper, we perform an empirical study to analyze the privacy risks inherent in the mRAG pipeline observed through standard model prompting. Specifically, we implement a case study that attempts to determine whether a visual asset (e.g., image) is included in the mRAG, and, if present, to leak the metadata (e.g., caption) related to it.Our findings highlight the need for privacy-preserving mechanisms and motivate future research on mRAG privacy. Our code is published online: https://github.com/aliwister/mrag-attack-eval.