arXiv subjects

Shi Feng

Publications and source records attributed to Shi Feng.

At least 73 records · Page 4Linked to original sources

arXiv2026

Defending Large Language Models Against Jailbreak Attacks via In-Decoding Safety-Awareness Probing

Large language models (LLMs) have achieved impressive performance across natural language tasks and are increasingly deployed in real-world applications. Despite extensive safety alignment efforts, recent studies show that such alignment is often shallow and remains vulnerable to jailbreak attacks. Existing defense mechanisms, including decoding-based constraints and post-hoc content detectors, struggle against sophisticated jailbreaks, often intervening robust detection or excessively degrading model utility. In this work, we examine the decoding process of LLMs and make a key observation: even when successfully jailbroken, models internally exhibit latent safety-related signals during generation. However, these signals are overridden by the model's drive for fluent continuation, preventing timely self-correction or refusal. Building on this observation, we propose a simple yet effective approach that explicitly surfaces and leverages these latent safety signals for early detection of unsafe content during decoding. Experiments across diverse jailbreak attacks demonstrate that our approach significantly enhances safety, while maintaining low over-refusal rates on benign inputs and preserving response quality. Our results suggest that activating intrinsic safety-awareness during decoding offers a promising and complementary direction for defending against jailbreak attacks. Code is available at: https://github.com/zyz13590/SafeProbing.

cs.AI↗

arXiv2026

Unsupervised Elicitation of Language Models

To steer pretrained language models for downstream tasks, today's post-training paradigm relies on humans to specify desired behaviors. However, for models with superhuman capabilities, it is difficult or impossible to get high-quality human supervision. To address this challenge, we introduce a new unsupervised algorithm, Internal Coherence Maximization (ICM), to fine-tune pretrained language models on their own generated labels, \emph{without external supervision}. On GSM8k-verification, TruthfulQA, and Alpaca reward modeling tasks, our method matches the performance of training on golden labels and outperforms training on crowdsourced human supervision. On tasks where LMs' capabilities are strongly superhuman, our method can elicit those capabilities significantly better than training on human labels. Finally, we show that our method can improve the training of frontier LMs: we use our method to train an unsupervised reward model and use reinforcement learning to train a Claude 4 Sonnet-based assistant. The resulting assistant matches its counterpart trained on production-grade human labels on average, with higher scores on chat and safety yet lower scores on math and coding.

cs.CL↗

arXiv2026

Variance bounds in product measures without exponential tails

We establish analogs of Cheeger's inequality for probability measures with heavy tails. As one of the principal applications, suppose $λ> 3$ and define the (Pareto) probability measure $μ_λ$ on $[1,\infty)$ by $dμ_λ(x) = (λ- 1) x^{-λ}$. Let $μ_λ^n$ denote the product measure of $μ_λ$ on $\mathbb{R}^n$. Then, for any $1$-Lipschitz function (with respect to the Euclidean distance) $f : \mathbb{R}^n \to \mathbb{R}$, we obtain the variance bound $\operatorname{Var}_{μ_λ^n}(f) \le C(λ)\, n^{\frac{2}{λ- 1}}$, where $C(λ)$ is an explicit constant depending only on $λ$. This improves upon the existing bound $\operatorname{Var}_{μ_λ^n}(f) = O(n)$ derived from the Efron--Stein inequality. Moreover, this bound is asymptotically tight when considering the $1$-Lipschitz function $f(x) = |x|_{\infty}$ corresponding to the $L^{\infty}$ norm. In probabilistic terms, suppose $X_1, \dots, X_n$ are i.i.d.\ random variables with distribution $μ_λ$. Then, for any $1$-Lipschitz function $f$, we have $\operatorname{Var}(f(X_1, \dots, X_n)) \le C'(λ)\operatorname{Var}(\max\{X_1, \dots, X_n\}) = Θ\!\left(n^{\frac{2}{λ- 1}}\right)$, where $C'(λ)$ is another explicit constant depending only on $λ$.

math.PR↗

arXiv2026

MoLAN: A Unified Modality-Aware Noise Dynamic Editing Framework for Multimodal Sentiment Analysis

Multimodal Sentiment Analysis aims to integrate information from various modalities, such as audio, visual, and text, to make complementary predictions. However, it often struggles with irrelevant or misleading visual and auditory information. Most existing approaches typically treat the entire modality information (e.g., a whole image, audio segment, or text paragraph) as an independent unit for feature enhancement or denoising. They often suppress the redundant and noise information at the risk of losing critical information. To address this challenge, we propose MoLAN, a unified ModaLity-aware noise dynAmic editiNg framework. Specifically, MoLAN performs modality-aware blocking by dividing the features of each modality into multiple blocks. Each block is then dynamically assigned a distinct denoising strength based on its noise level and semantic relevance, enabling fine-grained noise suppression while preserving essential multimodal information. Notably, MoLAN is a unified and flexible framework that can be seamlessly integrated into a wide range of multimodal models. Building upon this framework, we further introduce MoLAN+, a new multimodal sentiment analysis approach. Experiments across five models and four datasets demonstrate the broad effectiveness of the MoLAN framework. Extensive evaluations show that MoLAN+ achieves the state-of-the-art performance. The code is publicly available at https://github.com/betterfly123/MoLAN-Framework.

cs.LG↗

arXiv2026

Analyzing Residential Speeding Using Connected Vehicle Data: A Case Study in Charlottesville, VA Area

This study uses connected vehicle data to analyze speeding behavior on residential roads. A scalable pipeline processes trajectory data and supplements missing speed limits to generate summaries at OpenStreetMap's way ID level. The findings reveal a highly skewed distribution of both aggressive and reckless speeding. Based on a case study of Charlottesville, VA's connected vehicle data on residential roads, we found that 38% of segments had at least one instance of aggressive speeding, and 20% had at least one instance of reckless speeding. In addition, night time speeding is 27 times more prevalent than day time, and extreme violations on specific road segments highlight how severe the issue can be. Several segments rank among the top 10 for both aggressive and reckless speedings, indicating that there exist high-risk residential roads. These findings support the need for both spatial and behavioral interventions. The analysis provides a rich foundation for policy and planning, offering a valuable complement to traditional enforcement and planning tools. In conclusion, this framework sets the foundation for future applications in traffic safety analytics, demonstrating the growing potential of telematics data to inform safer, more livable communities.

stat.AP↗

arXiv2026

ES4R: Speech Encoding Based on Prepositive Affective Modeling for Empathetic Response Generation

Empathetic speech dialogue requires not only understanding linguistic content but also perceiving rich paralinguistic information such as prosody, tone, and emotional intensity for affective understandings. Existing speech-to-speech large language models either rely on ASR transcription or use encoders to extract latent representations, often weakening affective information and contextual coherence in multi-turn dialogues. To address this, we propose \textbf{ES4R}, a framework for speech-based empathetic response generation. Our core innovation lies in explicitly modeling structured affective context before speech encoding, rather than relying on implicit learning by the encoder or explicit emotion supervision. Specifically, we introduce a dual-level attention mechanism to capture turn-level affective states and dialogue-level affective dynamics. The resulting affective representations are then integrated with textual semantics through speech-guided cross-modal attention to generate empathetic responses. For speech output, we employ energy-based strategy selection and style fusion to achieve empathetic speech synthesis. ES4R consistently outperforms strong baselines in both automatic and human evaluations and remains robust across different LLM backbones.

eess.AS↗

arXiv2026

SAD: A Large-Scale Strategic Argumentative Dialogue Dataset

Argumentation generation has attracted substantial research interest due to its central role in human reasoning and decision-making. However, most existing argumentative corpora focus on non-interactive, single-turn settings, either generating arguments from a given topic or refuting an existing argument. In practice, however, argumentation is often realized as multi-turn dialogue, where speakers defend their stances and employ diverse argumentative strategies to strengthen persuasiveness. To support deeper modeling of argumentation dialogue, we present the first large-scale \textbf{S}trategic \textbf{A}rgumentative \textbf{D}ialogue dataset, SAD, consisting of 392,822 examples. Grounded in argumentation theories, we annotate each utterance with five strategy types, allowing multiple strategies per utterance. Unlike prior datasets, SAD requires models to generate contextually appropriate arguments conditioned on the dialogue history, a specified stance on the topic, and targeted argumentation strategies. We further benchmark a range of pretrained generative models on SAD and present in-depth analysis of strategy usage patterns in argumentation.

cs.CL↗

arXiv2026

High-Rank Structured Modulation for Parameter-Efficient Fine-Tuning

As the number of model parameters increases, parameter-efficient fine-tuning (PEFT) has become the go-to choice for tailoring pre-trained large language models. Low-rank Adaptation (LoRA) uses a low-rank update method to simulate full parameter fine-tuning, which is widely used to reduce resource requirements. However, decreasing the rank encounters challenges with limited representational capacity when compared to full parameter fine-tuning. We present \textbf{SMoA}, a high-rank \textbf{S}tructured \textbf{MO}dulation \textbf{A}dapter that uses fewer trainable parameters while maintaining a higher rank, thereby improving the model's representational capacity and offering improved performance potential. The core idea is to freeze the original pretrained weights and selectively amplify or suppress important features of the original weights across multiple subspaces. The subspace mechanism provides an efficient way to increase the capacity and complexity of a model. We conduct both theoretical analyses and empirical studies on various tasks. Experiment results show that SMoA outperforms LoRA and its variants on 10 tasks, with extensive ablation studies validating its effectiveness.

cs.CL↗

arXiv2026

PlaM: Training-Free Plateau-Guided Model Merging for Better Visual Grounding in MLLMs

Multimodal Large Language Models (MLLMs) rely on strong linguistic reasoning inherited from their base language models. However, multimodal instruction fine-tuning paradoxically degrades this text's reasoning capability, undermining multimodal performance. To address this issue, we propose a training-free framework to mitigate this degradation. Through layer-wise vision token masking, we reveal a common three-stage pattern in multimodal large language models: early-modal separation, mid-modal alignment, and late-modal degradation. By analyzing the behavior of MLLMs at different stages, we propose a plateau-guided model merging method that selectively injects base language model parameters into MLLMs. Experimental results based on five MLLMs on nine benchmarks demonstrate the effectiveness of our method. Attention-based analysis further reveals that merging shifts attention from diffuse, scattered patterns to focused localization on task-relevant visual regions. Our repository is on https://github.com/wzj1718/PlaM.

cs.CL↗

arXiv2026

CIRAG: Construction-Integration Retrieval and Adaptive Generation for Multi-hop Question Answering

Triple-based Iterative Retrieval-Augmented Generation (iRAG) mitigates document-level noise for multi-hop question answering. However, existing methods still face limitations: (i) greedy single-path expansion, which propagates early errors and fails to capture parallel evidence from different reasoning branches, and (ii) granularity-demand mismatch, where a single evidence representation struggles to balance noise control with contextual sufficiency. In this paper, we propose the Construction-Integration Retrieval and Adaptive Generation model, CIRAG. It introduces an Iterative Construction-Integration module that constructs candidate triples and history-conditionally integrates them to distill core triples and generate the next-hop query. This module mitigates the greedy trap by preserving multiple plausible evidence chains. Besides, we propose an Adaptive Cascaded Multi-Granularity Generation module that progressively expands contextual evidence based on the problem requirements, from triples to supporting sentences and full passages. Moreover, we introduce Trajectory Distillation, which distills the teacher model's integration policy into a lightweight student, enabling efficient and reliable long-horizon reasoning. Extensive experiments demonstrate that CIRAG achieves superior performance compared to existing iRAG methods.

cs.CL↗

arXiv2026

Spin-orbit coupling controlled two-dimensional magnetism in chromium trihalides

CrX$_3$ (X = Cl, Br, I) have the same crystal structure and Hamiltonian but different ligand spin-orbit coupling (SOC) constant $λ_X$, providing excellent material platform exploring for exotic two-dimensional (2D) spin orders. Their microscopic mechanism underlying 2D spin physics remain unestablished, along with experimental corroboration of Kitaev exchange interaction, central to realizing topological quantum spin liquids. Finding direct evidence for Kitaev interaction and determining its value has been an essential but formidable challenge in Kitaev physics. Here we report the direct Kitaev interaction signature in magnetic anisotropy measured by ferromagnetic resonance (FMR) spectroscopy. We present measured values of Heisenberg $J$, Kitaev $K$, and off-diagonal symmetric $Γ$ exchange interactions in CrX$_3$ determined using FMR and exact diagonalization. $K$ and $Γ$ exhibit dominant dependencies on $λ_X$, indicating its central role in 2D magnetism. Our study provides a foundation for designing 2D magnetic materials exhibiting novel behaviors by tuning intrinsic material parameters such as SOC.

cond-mat.str-el↗

arXiv2026

SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning

Existing fraud detection methods predominantly rely on transcribed text, suffering from ASR errors and missing crucial acoustic cues like vocal tone and environmental context. This limits their effectiveness against complex deceptive strategies. To address these challenges, we first propose \textbf{SAFE-QAQ}, an end-to-end comprehensive framework for audio-based slow-thinking fraud detection. First, the SAFE-QAQ framework eliminates the impact of transcription errors on detection performance. Secondly, we propose rule-based slow-thinking reward mechanisms that systematically guide the system to identify fraud-indicative patterns by accurately capturing fine-grained audio details, through hierarchical reasoning processes. Besides, our framework introduces a dynamic risk assessment framework during live calls, enabling early detection and prevention of fraud. Experiments on the TeleAntiFraud-Bench demonstrate that SAFE-QAQ achieves dramatic improvements over existing methods in multiple key dimensions, including accuracy, inference efficiency, and real-time processing capabilities. Currently deployed and analyzing over 70,000 calls daily, SAFE-QAQ effectively automates complex fraud detection, reducing human workload and financial losses. Code: https://anonymous.4open.science/r/SAFE-QAQ.

cs.SD↗

arXiv2025

Gaming the Answer Matcher: Examining the Impact of Text Manipulation on Automated Judgment

Automated answer matching, which leverages LLMs to evaluate free-text responses by comparing them to a reference answer, shows substantial promise as a scalable and aligned alternative to human evaluation. However, its reliability requires robustness against strategic attacks such as guesswork or verbosity that may artificially inflate scores without improving actual correctness. In this work, we systematically investigate whether such tactics deceive answer matching models by prompting examinee models to: (1) generate verbose responses, (2) provide multiple answers when unconfident, and (3) embed conflicting answers with the correct answer near the start of their response. Our results show that these manipulations do not increase scores and often reduce them. Additionally, binary scoring (which requires a matcher to answer with a definitive "correct" or "incorrect") is more robust to attacks than continuous scoring (which requires a matcher to determine partial correctness). These findings show that answer matching is generally robust to inexpensive text manipulation and is a viable alternative to traditional LLM-as-a-judge or human evaluation when reference answers are available.

cs.CL↗

arXiv2025

Emergent Persuasion: Will LLMs Persuade Without Being Prompted?

With the wide-scale adoption of conversational AI systems, AI are now able to exert unprecedented influence on human opinion and beliefs. Recent work has shown that many Large Language Models (LLMs) comply with requests to persuade users into harmful beliefs or actions when prompted and that model persuasiveness increases with model scale. However, this prior work looked at persuasion from the threat model of $\textit{misuse}$ (i.e., a bad actor asking an LLM to persuade). In this paper, we instead aim to answer the following question: Under what circumstances would models persuade $\textit{without being explicitly prompted}$, which would shape how concerned we should be about such emergent persuasion risks. To achieve this, we study unprompted persuasion under two scenarios: (i) when the model is steered (through internal activation steering) along persona traits, and (ii) when the model is supervised-finetuned (SFT) to exhibit the same traits. We showed that steering towards traits, both related to persuasion and unrelated, does not reliably increase models' tendency to persuade unprompted, however, SFT does. Moreover, SFT on general persuasion datasets containing solely benign topics admits a model that has a higher propensity to persuade on controversial and harmful topics--showing that emergent harmful persuasion can arise and should be studied further.

cs.AI↗

arXiv2025

Do LLM Evaluators Prefer Themselves for a Reason?

Large language models (LLMs) are increasingly used as automatic evaluators in applications such as benchmarking, reward modeling, and self-refinement. Prior work highlights a potential self-preference bias where LLMs favor their own generated responses, a tendency often intensifying with model size and capability. This raises a critical question: Is self-preference harmful, or does it simply reflect the genuinely higher-quality outputs of stronger models? Answering this has been difficult as prior works mostly relied on subjective tasks that lack an objective ground truth, meaning that either preference can be reasonably justified. To address this ambiguity, we investigate self-preference using verifiable benchmarks (mathematical reasoning, factual knowledge, code generation) that allow objective ground-truth assessment. This enables us to distinguish harmful (favoring objectively worse responses) from legitimate (favoring genuinely superior ones) self-preference. Our large-scale experiments across 7 model families reveal three key findings: (1) While stronger models exhibit greater self-preference, much of this preference aligns with objectively superior performance, indicating stronger models prefer themselves mostly legitimately. (2) Harmful self-preference persists when evaluator models err as generators, and stronger models display more pronounced harmful self-preference bias when they do err. This suggests stronger models struggle more to recognize when they are wrong. (3) Inference-time scaling strategies, such as generating a long Chain-of-Thought before evaluation, effectively reduce the harmful self-preference. Additionally, we experiment with LMArena and show that our findings extend beyond verifiable benchmarks to real-world, subjective domains. These results provide a more nuanced understanding of LLM-based evaluation and practical insights for improving its reliability.

cs.CL↗

arXiv2025

Triangular $J_1$-$J_2$ Heisenberg Antiferromagnet in a Magnetic Field

The behavior of the paradigmatic $J_1-J_2$ triangular lattice Heisenberg antiferromagnet in a magnetic field remains unsettled despite decades of study. We map out the phase diagram using three complementary approaches, including self-consistent nonlinear spin-wave theory, density-matrix renormalization group, and variational Monte Carlo. This combined analysis resolves the competition among different field-induced magnetic orders and magnetization plateaux across the classically frustrated parameter range. In particular, there is a finite range in the parameter regime around $J_2/J_1=\frac{1}{8}$ in which i) upon the application of the external field, the gapless quantum spin liquid acquires a finite density of monopoles, and ii) by further increasing the field, two plateaux are clearly obtained at $m=\frac{1}{3}$ and $m=\frac{1}{2}$. We discuss the experimental importance of the consecutive magnetization plateaux transitions as a signature of an underlying quantum spin-liquid phase.

cond-mat.str-el↗

arXiv2025

Mitigating Self-Preference by Authorship Obfuscation

Language models (LMs) judges are widely used to evaluate the quality of LM outputs. Despite many advantages, LM judges display concerning biases that can impair their integrity in evaluations. One such bias is self-preference: LM judges preferring their own answers over those produced by other LMs or humans. The bias is hard to eliminate as frontier LM judges can distinguish their own outputs from those of others, even when the evaluation candidates are not labeled with their sources. In this paper, we investigate strategies to mitigate self-preference by reducing the LM judges' ability to recognize their own outputs. We apply black-box perturbations to evaluation candidates in pairwise comparison to obfuscate the authorship and reduce self-recognition. We find that perturbations as simple as synonym replacement for a few words predictably reduce self-preference. However, we also uncover fundamental challenges to eliminating the bias: when we extrapolate our perturbations to a more complete neutralization of stylistic differences between the evaluation candidates, self-preference recovers. Our findings suggest that self-recognition and self-preference can happen on many semantic levels, and complete mitigation remains challenging despite promising initial results.

cs.CL↗

arXiv2025

AI Debate Aids Assessment of Controversial Claims

As AI grows more powerful, it will increasingly shape how we understand the world. But with this influence comes the risk of amplifying misinformation and deepening social divides-especially on consequential topics where factual accuracy directly impacts well-being. Scalable Oversight aims to ensure AI systems remain truthful even when their capabilities exceed those of their evaluators. Yet when humans serve as evaluators, their own beliefs and biases can impair judgment. We study whether AI debate can guide biased judges toward the truth by having two AI systems debate opposing sides of controversial factuality claims on COVID-19 and climate change where people hold strong prior beliefs. We conduct two studies. Study I recruits human judges with either mainstream or skeptical beliefs who evaluate claims through two protocols: debate (interaction with two AI advisors arguing opposing sides) or consultancy (interaction with a single AI advisor). Study II uses AI judges with and without human-like personas to evaluate the same protocols. In Study I, debate consistently improves human judgment accuracy and confidence calibration, outperforming consultancy by 4-10% across COVID-19 and climate change claims. The improvement is most significant for judges with mainstream beliefs (up to +15.2% accuracy on COVID-19 claims), though debate also helps skeptical judges who initially misjudge claims move toward accurate views (+4.7% accuracy). In Study II, AI judges with human-like personas achieve even higher accuracy (78.5%) than human judges (70.1%) and default AI judges without personas (69.8%), suggesting their potential for supervising frontier AI models. These findings highlight AI debate as a promising path toward scalable, bias-resilient oversight in contested domains.

cs.CL↗