Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognition

• URL: http://arxiv.org/abs/2607.27240v1
• Date: Sun, 26 Jul 2026 19:10:34 GMT
• Title: Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognition
• Abstract summary: We use two Gemma-3-1B-IT finetunes on two complementary safety objectives: CARES harm-level classification and WildJailbreak adversarial refusal.<n>We merge the two fine-tunes using Linear, SLERP, TIES, and DARE-TIES, and evaluate the merged models on classification accuracy, attack resistance, and benign compliance.
• Score: 4.800831341795478
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Model merging is often used to combine capabilities from separately fine-tuned models without additional training, but it is unclear whether standard merging methods preserve multiple safety-relevant behaviors simultaneously. We study this question through a controlled case study using two Gemma-3-1B-IT finetunes on two complementary safety objectives: CARES harm-level classification and WildJailbreak adversarial refusal. We merge the two fine-tunes using Linear, SLERP, TIES, and DARE-TIES, and evaluate the merged models on classification accuracy, attack resistance, and benign compliance. Across all four methods, attack resistance transfers significantly more than classification accuracy: merged models retain 81-85% jailbreak refusal rates while CARES accuracy falls to at most 12.9%. Weight-space measurements suggest that this asymmetry is not caused by strongly opposing task-vector directions: the two task vectors are nearly orthogonal (cosine similarity 0.011). Instead, the refusal fine-tune induces consistently larger per-layer task-vector magnitudes, causing magnitude-sensitive methods to favor refusal updates. These results show that standard model merging can collapse safety recognition into broad refusal when safety-relevant task vectors differ substantially in scale.

Related papers

• Detecting Safety Training Modification in Language Models via Activation Analysis [0.0]
  We introduce AMS, a tool that detects modifications to safety training in language models by measuring the geometric structure of safety-relevant concepts in activation space.<n>We validate AMS across 14 model configurations spanning 4 architecture families.<n>Mechanistic analysis identifies a four-class taxonomy of safety-training modifications distinguished by activation-space signature.
  arXiv  Detail & Related papers  (2026-08-06T03:57:38Z)
• Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs [54.96236442484317]
  We show that hazardous knowledge can be retained in the model and behaviorally gated by a privileged control token.<n>Our method, Token Inoculation, introduces a binding-and-branching approach.
  arXiv  Detail & Related papers  (2026-07-21T02:15:39Z)
• The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs [1.4323566945483497]
  We introduce Contrastive Logit Steering ( CLS), a zero-optimization framework that isolates the "refusal direction"<n> CLS operates directly on the output distribution, serving as a diagnostic probe for alignment.<n>Our experiments on 7 model families reveal that safety implementation is architecturally deterministic.
  arXiv  Detail & Related papers  (2026-06-21T22:04:48Z)
• What Accuracy and Gradient Cosine Miss: Evaluating Feedback Alignment via Scale Stability, Reference Validity, and Depth Utility [48.10132234701036]
  We propose a diagnostic evaluation protocol based on three checks -- scale stability, reference validity, and depth utility -- together with per-layer rather than aggregate cosine reporting.<n>Across multiple architectures and methods, our protocol identifies all failures with wide calibration margins.
  arXiv  Detail & Related papers  (2026-06-19T06:04:17Z)
• From Refusal Geometry to Safety Geometry: Harmfulness--Refusal Coupling under Dynamic Adversarial Fine-Tuning [14.6508023458559]
  Existing evaluations do not reveal whether a model has learned to recognize harmfulness.<n>We study this question with a dual safety-geometry protocol that measures harmfulness carriers, refusal carriers, and their coupling.
  arXiv  Detail & Related papers  (2026-06-15T07:50:00Z)
• Building a Foundational Guardrail for General Agentic Systems via Synthetic Data [76.18834864749606]
  LLM agents can plan multi-step tasks, intervening at the planning stage-before any action is executed-is often the safest way to prevent harm.<n>Existing guardrails mostly operate post-execution, which is difficult to scale and leaves little room for controllable supervision at the plan level.<n>We introduce AuraGen, a controllable engine that synthesizes benign trajectories, injects category-labeled risks with difficulty, and filters outputs via an automated reward model.
  arXiv  Detail & Related papers  (2025-10-10T18:42:32Z)
• DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models [50.21378052667732]
  We conduct an in-depth analysis of dLLM vulnerabilities to jailbreak attacks across two distinct dimensions: intra-step and inter-step dynamics.<n>We propose DiffuGuard, a training-free defense framework that addresses vulnerabilities through a dual-stage approach.
  arXiv  Detail & Related papers  (2025-09-29T05:17:10Z)
• The Rogue Scalpel: Activation Steering Compromises LLM Safety [11.402179030703188]
  Activation steering is a technique for controlling LLM behavior by adding semantically meaningful vectors directly into a model's hidden states during inference.<n>We demonstrate the opposite: steering systematically breaks model alignment safeguards, making it comply with harmful requests.
  arXiv  Detail & Related papers  (2025-09-26T08:49:47Z)
• Indiscriminate Disruption of Conditional Inference on Multivariate Gaussians [60.22542847840578]
  Despite advances in adversarial machine learning, inference for Gaussian models in the presence of an adversary is notably understudied. We consider a self-interested attacker who wishes to disrupt a decisionmaker's conditional inference and subsequent actions by corrupting a set of evidentiary variables. To avoid detection, the attacker also desires the attack to appear plausible wherein plausibility is determined by the density of the corrupted evidence.
  arXiv  Detail & Related papers  (2024-11-21T17:46:55Z)
• Improving Adversarial Robustness to Sensitivity and Invariance Attacks with Deep Metric Learning [80.21709045433096]
  A standard method in adversarial robustness assumes a framework to defend against samples crafted by minimally perturbing a sample. We use metric learning to frame adversarial regularization as an optimal transport problem. Our preliminary results indicate that regularizing over invariant perturbations in our framework improves both invariant and sensitivity defense.
  arXiv  Detail & Related papers  (2022-11-04T13:54:02Z)

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.