The paper investigates why large language models lose safety alignment after benign fine‑tuning. It argues that safety alignment relies on a low‑rank, output‑routing geometry that becomes flatter during fine‑tuning, and that after only 100 benign examples this routing is sharpened in output‑side MLPs, leading to fragile safety while general performance remains relatively intact. Techniques like LoRA and ASAM can delay this collapse by reducing output‑side sharpness, but their effectiveness diminishes with larger fine‑tuning scales.
arXiv:2608. 12821v1 Announce Type: new Abstract: Large language models (LLMs) remain vulnerable to harmful requests and jailbreak attacks.
The paper introduces RASET, a router‑agnostic safety‑critical expert tuning framework for Mixture‑of‑Experts (MoE) large language models. RASET identifies a small subset of experts that are responsible for safety enforcement and applies parameter‑efficient tuning only to those experts, preserving the model’s intrinsic routing behavior. Experiments on five open‑weight MoE backbones show that RASET achieves a high safety‑bypass yield, outperforming existing baselines by a significant margin.
arXiv:2608. 11583v1 Announce Type: new Abstract: Safety alignment in large language models is often treated as a distributed property of the entire network, yet its practical brittleness suggests that refusal behavior may be concentrated in a smaller set of parameters.
arXiv:2506. 08473v4 Announce Type: replace Abstract: Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures.
arXiv:2607. 08883v1 Announce Type: new Abstract: Behavioral alignment in large language models often masks fragile internal safety representations.
CLEAR is a conditional safety adaptation framework that employs a lightweight hidden‑state gate to continuously control the activation strength of a safety low‑rank adapter. It aims to reduce harmful completions while preserving the performance of the frozen backbone on benign prompts. Experiments on safety and utility benchmarks, including HarmBench and GSM8K, show that CLEAR significantly lowers HarmBench ASR and improves utility compared to globally applied safety tuning methods such as SFT or standard LoRA.
arXiv:2606. 04168v1 Announce Type: new Abstract: Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens.
arXiv:2605.01913v2 Announce Type: replace-cross Abstract: Fine-tuning safety-aligned language models for downstream tasks often leads to substantial degradation of refusal behavior, making models vul...
arXiv:2606. 22686v2 Announce Type: replace-cross Abstract: Modern Large Language Models (LLMs) rely on extensive safety alignment, yet the mechanistic basis of refusal remains opaque.
arXiv:2508.20766v2 Announce Type: replace-cross Abstract: Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent researc...
The paper addresses the problem of large language models (LLMs) over-refusing to comply with benign safety-related instructions. It identifies that a small number of hypersensitive safety heads in transformer attention misfire on hard-safe prompts, causing abnormal attention entanglement and high-entropy routing conflicts that lead to refusals. The authors propose Semantic Routing Calibration (SRC), a lightweight, training‑free inference framework that dynamically suppresses these problematic heads and fuses logits to restore trustworthy reasoning while preserving intrinsic safety performance.
The paper addresses the problem of over‑refusal in safety‑aligned large language models, where benign instructions are incorrectly rejected. It identifies that a small set of hypersensitive safety heads in transformer attention misfire on hard‑safe prompts, causing abnormal attention entanglement and high‑entropy routing conflicts that block necessary attention to target entities. To mitigate this, the authors propose Semantic Routing Calibration (SRC), a lightweight, training‑free inference framework that dynamically suppresses these hypersensitive heads and fuses logits from dual branches to restore trustworthy reasoning while preserving intrinsic safety performance.