CORTEXA
← Browse
arxivcs.AIcs.CR2026-07-11

When Are Sparse Feature Interventions Actually Localized? Matched Evaluation for SAE-Based Safety Control

Daming Luo

We evaluate when sparse autoencoder (SAE) features act as localized control handles for safety-relevant behavior. This question is difficult because apparent success can arise from weak interventions, mismatched baselines, model robustness, or degenerate outputs that automated safety judges mark as unsafe without representing meaningful harmful compliance. We introduce a matched coherence-gated evaluation protocol for runtime safety interventions: methods are compared at matched target-effect points, and the primary target metric counts harmful compliance only when an output is both judge-unsafe and coherent. Applying this protocol to three prompt splits on Gemma-2-9B-it with a Gemma Scope layer-20 residual SAE, we find that SAE feature ablation has a narrow useful regime. SAE top800 reaches a low-to-mid target effect with lower total perturbation and competitive utility, but SAE top1600 loses utility relative to a matched dense refusal-direction baseline, and SAE top3200 primarily induces coherence collapse. Human audit confirms that coherence gating removes unsafe-only artifacts, and feature diagnostics show that the useful regime is driven by a stable head of refusal-aligned features whose activation separation decays rapidly with rank. These results argue that SAE-based safety interventions should be evaluated as regime-dependent control mechanisms rather than assumed to be uniformly localized.

View free PDFSource page

Related papers

arxivcs.LGcs.AIcs.CR2026-06-26

TRACE: Trajectory-Based Safety Patch Learning for LLM Post-Training Realignment

Changyue Li, Jiaming He, Youliang Yuan, Jialin Wu, Boxi Yu, Zhicong Huang, et al.

Fine-Tuning-as-a-Service (FTaaS) platforms let users train large language models (LLMs) on customized tasks, but this pipeline could erode models' safety alignment. In practice, service providers need to recover models' safety without re-running full alignment, or destroying the…

View free PDFSource page
arxivcs.CRcs.AIcs.CL2026-07-20

JailMeter: An Evidence-Based Evaluation Framework for Jailbreak Attacks on Large Language Models

Qingjia Huang, Jingyu Zhang, Jianguo Wu, Yakai Li, Weijuan Zhang, Yankai Rong, et al.

The assessment of jailbreak attacks against large language models currently suffers from inconsistent evaluation criteria and methods, leading to unreliable estimates of attack success rates. We propose JailMeter, an evidence-based evaluation framework designed to more faithfully…

View free PDFSource page
arxivcs.CRcs.AIcs.IRcs.LGcs.MA2026-07-08

Who Broke the System? Failure Localization in LLM-Based Multi-Agent Systems

Yufei Xia, Anjun Gao, Yueyang Quan, Zhuqing Liu, Minghong Fang

Large language model (LLM) based multi-agent systems enable complex problem solving through coordinated reasoning and action, but their distributed structure also introduces new challenges in diagnosing system-level failures. When an execution fails, identifying which agent is re…

View free PDFSource page
arxivcs.AIcs.CRcs.MA2026-07-08

Operational Reframing and Approval-Framed Delegation in Multi-Agent LLM Safety

Lifei Liu, Haoran Yu, Xiaochong Jiang, Su Wang, Pin Qian, Yihang Chen

Safety evaluations of multi-agent LLM systems often compare a direct prompt with a planner-executor pipeline and report the difference as a single "pipeline effect." We argue that this aggregate is difficult to interpret because it conflates three mechanisms: harmful intent may b…

View free PDFSource page
arxivcs.CRcs.AI2026-07-22

GPE: Evaluating Robust Evidence Aggregation for Fact Verification under Controllable GEO-Style Poisoning

Zhaoqi Wang, Zijian Zhang, Xiaomei Yuan, Pengtao Kou, Jiamou Liu, Zhen Li, et al.

Large language models increasingly use search tools to retrieve up-to-date information, introducing a new attack surface in which retrieved documents can be manipulated. This risk is amplified by the development of generative engine optimization, which can make selected content m…

View free PDFSource page