Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models

• URL: http://arxiv.org/abs/2602.02600v2
• Date: Sat, 07 Feb 2026 19:58:16 GMT
• Title: Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models
• Abstract summary: Diffusion language models (DLMs) have recently emerged as a promising alternative to autoregressive (AR) models.<n>We present a fundamental analytical framework for step-wise refusal dynamics, enabling comparison between AR and diffusion sampling.<n>We introduce the Step-Wise Refusal Internal Dynamics (SRI) signal, which supports interpretability and improved safety for both AR and DLMs.
• Score: 5.578295626701434
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Diffusion language models (DLMs) have recently emerged as a promising alternative to autoregressive (AR) models, offering parallel decoding and controllable sampling dynamics while achieving competitive generation quality at scale. Despite this progress, the role of sampling mechanisms in shaping refusal behavior and jailbreak robustness remains poorly understood. In this work, we present a fundamental analytical framework for step-wise refusal dynamics, enabling comparison between AR and diffusion sampling. Our analysis reveals that the sampling strategy itself plays a central role in safety behavior, as a factor distinct from the underlying learned representations. Motivated by this analysis, we introduce the Step-Wise Refusal Internal Dynamics (SRI) signal, which supports interpretability and improved safety for both AR and DLMs. We demonstrate that the geometric structure of SRI captures internal recovery dynamics, and identifies anomalous behavior in harmful generations as cases of \emph{incomplete internal recovery} that are not observable at the text level. This structure enables lightweight inference-time detectors that generalize to unseen attacks while matching or outperforming existing defenses with over $100\times$ lower inference overhead.

Related papers

• Learning to Watch: Active Video Anomaly Understanding via Interleaved Policy Optimization [56.150184676745205]
  Video anomaly understanding (VAU) relies on sparse, context-dependent cues.<n>We propose $Anomtext-$, a closed-loop framework that reconceptualizes video understanding as an active sequential decision-making process.<n>Inspired by human video-reviewing behavior, this framework unifies internal cognitive reasoning and strategic evidence acquisition into an interleaved policy.
  arXiv  Detail & Related papers  (2026-07-01T08:41:49Z)
• Spatiotemporal Hidden-State Dynamics as a Signature of Internal Reasoning in Large Language Models [0.0]
  Internal representations carry correctness-related signals, but their coarse aggregations may obscure the token and layer structure underlying reasoning.<n>We investigate hidden-state transitions across decoding steps and layers, and identify a pattern in large reasoning models (LRMs)<n>We formalize this characteristic as Stemporal Amplitude of Latent Transition (StALT), a statistic trajectory that summarizes temporal changes between adjacent tokens weighted by within-token layer saliency.
  arXiv  Detail & Related papers  (2026-05-03T12:46:41Z)
• Silencing the Guardrails: Inference-Time Jailbreaking via Dynamic Contextual Representation Ablation [16.17685258866394]
  We propose Contextual Representation Ablation (CRA), a novel inference-time intervention framework designed to dynamically silence model guardrails.<n>CRA identifies and suppresses refusal-inducing activation patterns during decoding without requiring expensive parameter updates or training.<n> Empirical evaluation across multiple safety-aligned open-source LLMs demonstrates that CRA significantly outperforms baselines.
  arXiv  Detail & Related papers  (2026-04-09T05:35:56Z)
• Dynamic analysis enhances issue resolution [53.50448142467294]
  DAIRA (Dynamic Analysis-enhanced Issue Resolution Agent) is an automated repair framework that embeds dynamic analysis into the agent's reasoning cycle.<n> Driven by a Test Tracing-Driven methodology, DAIRA utilizes lightweight monitors to extract critical runtime data.<n>When powered by Gemini 3 Flash Preview, DAIRA establishes a new state-of-the-art (SOTA) performance, achieving a 79.4% resolution rate on the SWE-bench Verified dataset.
  arXiv  Detail & Related papers  (2026-03-23T14:48:54Z)
• Adaptive Guidance for Retrieval-Augmented Masked Diffusion Models [60.25003813232697]
  We propose Adaptive Retrieval-Augmented Masked Diffusion (ARAM) as a training-free adaptive guidance framework for Masked Diffusion Models (MDMs)<n>ARAM calibrates the guidance scale during denoising according to the Signal-to-Noise Ratio (SNR) of the distributional shift induced by retrieved context.<n>Experiments on multiple knowledge-intensive QA benchmarks show that ARAM improves overall QA performance over competitive RAG baselines.
  arXiv  Detail & Related papers  (2026-03-18T12:54:50Z)
• Are Reasoning LLMs Robust to Interventions on Their Chain-of-Thought? [79.86483056611105]
  Reasoning LLMs generate step-by-step chains of thought before giving an answer.<n>How robust are these reasoning traces to disruptions that occur within them?<n>We introduce a controlled evaluation framework that perturbs a model's own CoT at fixed timesteps.
  arXiv  Detail & Related papers  (2026-02-07T10:02:58Z)
• Lost in the Noise: How Reasoning Models Fail with Contextual Distractors [57.31788955167306]
  Recent advances in reasoning models and agentic AI systems have led to an increased reliance on diverse external information.<n>We introduce NoisyBench, a comprehensive benchmark that systematically evaluates model robustness across 11 datasets in RAG, reasoning, alignment, and tool-use tasks.<n>Our evaluation reveals a catastrophic performance drop of up to 80% in state-of-the-art models when faced with contextual distractors.
  arXiv  Detail & Related papers  (2026-01-12T05:43:51Z)
• Dynamical Systems Analysis Reveals Functional Regimes in Large Language Models [0.8694591156258423]
  Large language models perform text generation through high-dimensional internal dynamics.<n>Most interpretability approaches emphasise static representations or causal interventions, leaving temporal structure largely unexplored.<n>We discuss a composite dynamical metric, computed from activation time-series during autoregressive generation.
  arXiv  Detail & Related papers  (2026-01-11T21:57:52Z)
• Explainability-Guided Defense: Attribution-Aware Model Refinement Against Adversarial Data Attacks [6.573058520271728]
  We identify a connection between interpretability and robustness that can be directly leveraged during training.<n>We introduce an attribution-guided refinement framework that transforms Local Interpretable Model-Agnostic Explanations into an active training signal.
  arXiv  Detail & Related papers  (2026-01-02T19:36:03Z)
• Improving Deepfake Detection with Reinforcement Learning-Based Adaptive Data Augmentation [60.04281435591454]
  CRDA (Curriculum Reinforcement-Learning Data Augmentation) is a novel framework guiding detectors to progressively master multi-domain forgery features.<n>Central to our approach is integrating reinforcement learning and causal inference.<n>Our method significantly improves detector generalizability, outperforming SOTA methods across multiple cross-domain datasets.
  arXiv  Detail & Related papers  (2025-11-10T12:45:52Z)
• Drift No More? Context Equilibria in Multi-Turn LLM Interactions [58.69551510148673]
  contexts drift is the gradual divergence of a model's outputs from goal-consistent behavior across turns.<n>Unlike single-turn errors, drift unfolds temporally and is poorly captured by static evaluation metrics.<n>We show that multi-turn drift can be understood as a controllable equilibrium phenomenon rather than as inevitable decay.
  arXiv  Detail & Related papers  (2025-10-09T04:48:49Z)
• Retrieval is Not Enough: Enhancing RAG Reasoning through Test-Time Critique and Optimization [58.390885294401066]
  Retrieval-augmented generation (RAG) has become a widely adopted paradigm for enabling knowledge-grounded large language models (LLMs)<n>RAG pipelines often fail to ensure that model reasoning remains consistent with the evidence retrieved, leading to factual inconsistencies or unsupported conclusions.<n>We propose AlignRAG, a novel iterative framework grounded in Critique-Driven Alignment (CDA)<n>We introduce AlignRAG-auto, an autonomous variant that dynamically terminates refinement, removing the need to pre-specify the number of critique iterations.
  arXiv  Detail & Related papers  (2025-04-21T04:56:47Z)
• Model Hemorrhage and the Robustness Limits of Large Language Models [119.46442117681147]
  Large language models (LLMs) demonstrate strong performance across natural language processing tasks, yet undergo significant performance degradation when modified for deployment.<n>We define this phenomenon as model hemorrhage - performance decline caused by parameter alterations and architectural changes.
  arXiv  Detail & Related papers  (2025-03-31T10:16:03Z)
• Towards More Robust Retrieval-Augmented Generation: Evaluating RAG Under Adversarial Poisoning Attacks [45.07581174558107]
  Retrieval-Augmented Generation (RAG) systems have emerged as a promising solution to mitigate hallucinations.<n>RAG systems are vulnerable to adversarial poisoning attacks, where malicious passages injected into the retrieval corpus can mislead models into producing factually incorrect outputs.<n>We present a rigorously controlled empirical study of how RAG systems behave under such attacks and how their robustness can be improved.
  arXiv  Detail & Related papers  (2024-12-21T17:31:52Z)
• Reinforcement Learning under Latent Dynamics: Toward Statistical and Algorithmic Modularity [51.40558987254471]
  Real-world applications of reinforcement learning often involve environments where agents operate on complex, high-dimensional observations. This paper addresses the question of reinforcement learning under $textitgeneral$ latent dynamics from a statistical and algorithmic perspective.
  arXiv  Detail & Related papers  (2024-10-23T14:22:49Z)
• Sycophancy in Vision-Language Models: A Systematic Analysis and an Inference-Time Mitigation Framework [18.54098084470481]
  We analyze sycophancy across vision-language benchmarks and propose an inference-time mitigation framework.<n>Our framework effectively mitigates sycophancy across all evaluated models, while maintaining performance on neutral prompts.
  arXiv  Detail & Related papers  (2024-08-21T01:03:21Z)
• Robust VAEs via Generating Process of Noise Augmented Data [9.366139389037489]
  This paper introduces a novel framework that enhances robustness by regularizing the latent space divergence between original and noise-augmented data. Our empirical evaluations demonstrate that this approach, termed Robust Augmented Variational Auto-ENcoder (RAVEN), yields superior performance in resisting adversarial inputs.
  arXiv  Detail & Related papers  (2024-07-26T09:55:34Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.