arXiv subjects

Yu-Hang Wu

Publications and source records attributed to Yu-Hang Wu.

4 recordsLinked to original sources

arXiv2026

Does Deeper Reasoning Compromise Alignment? Revealing and Mitigating of Alignment Collapse in Large Reasoning Models

The emergence of Chain-of-Thought (CoT) has established a robust foundation for Large Reasoning Models (LRMs). While deep reasoning is widely believed to enhance safety alignment, the stability of alignment mechanisms under extended reasoning remains underexplored. This paper challenges the prevailing view by revealing a critical vulnerability: Deep Reasoning May Induce Alignment Collapse. To rigorously quantify this phenomenon, we propose the Alignment Loss Rate (ALR) metric. Our experiments demonstrate that as reasoning depth increases, ALR rises significantly, indicating a severe degradation in model robustness against external perturbations. Capitalizing on this instability, a novel jailbreaking paradigm, Reasoning Trap (RT), is proposed. RT induces the model into extended reasoning to amplify the impact of adversarial attacks, leading to a sharp decline in safety capabilities. To elucidate the mechanism behind this collapse, we identify Attention Dilution as the root cause, arising from the competition for attention between the extended reasoning process and the original input. To mitigate this, Reasoning Residual Alignment (RRA), a lightweight defense strategy that dynamically re-emphasizes the input via residual connections integrated with the reasoning process.

cs.AI↗

arXiv2026

Freeze Deep, Train Shallow: Interpretable Layer Allocation for Continued Pre-Training

Selective layer-wise updates are essential for low-cost continued pre-training of Large Language Models (LLMs), yet determining which layers to freeze or train remains an empirical black-box problem due to the lack of interpretable guidance. To address this issue, we propose LayerTracer, an architecture-agnostic diagnostic framework that reveals the evolution patterns of layer-wise representations and stability by locating task execution positions and quantifying layer sensitivity. Analysis results reveal that deep layers act as critical regions for task execution and maintain high stability against disruptive updates. Guided by this finding, we conduct three controlled continued pre-training trials to compare diverse freeze-train strategies, demonstrating that training shallow layers while freezing deep layers consistently outperforms full-parameter fine-tuning and the opposite allocation on both C-Eval and CMMLU benchmarks. We further present a hybrid model case study, which validates that placing high-quality pre-trained modules in deep layers effectively preserves inherent knowledge of the model. This work delivers a low-cost and interpretable solution for resource-constrained teams, offering actionable guidance for layer-wise parameter allocation in continued pre-training and hybrid model construction.

cs.CL↗

arXiv2025

Positivity bounds in scalar-QED EFT at one-loop level

Understanding the implication of positivity bounds on loop-generated dim-8 operator coefficients is a nontrivial task, as these bounds only strictly hold when all the contributions are included in the dispersion relation up to a certain loop order in the UV theory. As a step towards more realistic gauge theories such as the Standard Model, in this paper we study the positivity bounds in the Scalar QED Effective Field Theory (EFT) from the scalar-photon scattering ($γϕ\to γϕ$) and the photon-photon scattering ($γγ\to γγ$), derived from the dispersion relation of the full one-loop EFT amplitudes. Assuming the UV theory is weakly coupled and all heavy particles have spin $\leq1$, the leading dim-8 interaction for both amplitudes are generated at the one-loop level in the UV theory. Gauge invariance imposes strong constraints on the loop structures, while potential IR divergences also require careful treatments. Our findings reveal that, for $γϕ\to γϕ$, while the tree-level bound does not necessarily hold, the one-loop $β$-function of the corresponding coefficient always tends to restore the tree-level bound in the IR, unless its actual loop order in the UV theory is further suppressed. For $γγ\to γγ$, on the other hand, the tree-level positivity bound is still robust at the one-loop level in the UV theory. These findings are verified in two example UV models with a heavy scalar extension. Importantly, the bounds on the $β$-functions that we obtain should be considered as an accidental feature at one loop, rather than a fundamental property of the theory.

hep-ph↗

arXiv2025

Sugar-Coated Poison: Benign Generation Unlocks LLM Jailbreaking

With the increasingly deep integration of large language models (LLMs) across diverse domains, the effectiveness of their safety mechanisms is encountering severe challenges. Currently, jailbreak attacks based on prompt engineering have become a major safety threat. However, existing methods primarily rely on black-box manipulation of prompt templates, resulting in poor interpretability and limited generalization. To break through the bottleneck, this study first introduces the concept of Defense Threshold Decay (DTD), revealing the potential safety impact caused by LLMs' benign generation: as benign content generation in LLMs increases, the model's focus on input instructions progressively diminishes. Building on this insight, we propose the Sugar-Coated Poison (SCP) attack paradigm, which uses a "semantic reversal" strategy to craft benign inputs that are opposite in meaning to malicious intent. This strategy induces the models to generate extensive benign content, thereby enabling adversarial reasoning to bypass safety mechanisms. Experiments show that SCP outperforms existing baselines. Remarkably, it achieves an average attack success rate of 87.23% across six LLMs. For defense, we propose Part-of-Speech Defense (POSD), leveraging verb-noun dependencies for syntactic analysis to enhance safety of LLMs while preserving their generalization ability.

cs.CR↗