PatchBench: Measuring Collateral Damage in Activation Patching
Factual evidence
What the source reports
Researchers introduced PatchBench to evaluate whether activation patching in LLMs causes collateral damage like over-refusal of benign prompts.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.CL — Computation and Language · RESEARCH
- Published by source
- 8 October 2026
- Collected by OneBench
- 9 Oct 2026, 03:01 UK
- Original headline
- PatchBench: Measuring Collateral Damage in Activation Patching ↗
Stored source excerpt
arXiv:2610.10276v1 Announce Type: cross Abstract: An LLM safety patch can pass a benchmark while still being a poor repair. This risk is especially acute for…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
LLM safety patches and jailbreak fixes may introduce over-refusal risks or fail on slight prompt variants in production applications.
Do what
Review model safety testing protocols with the risk management team to account for collateral over-refusal in patched models.