AI
When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 04:55:20 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 11:18:24 UTC.
Briefings
SAKIKO is a framework for auditing internal LLM interventions by formalizing representation repair via directional error discovery. It evaluates seven LLMs on two datasets, revealing that while some interventions yield net gains, they often corrupt baseline decisions. The study highlights that behavioral movement does not equal repair, as interventions achieving +55 net gain corrupt over half of the baseline-corrected decisions. Key findings include formal declines for Qwen3-4B and Gemma-2-9B due to finite-sample uncertainty, and a distinction between pre-execution action selection and post-execution outcome auditing.
Key points
- SAKIKO formalizes representation repair via directional error discovery and router-conditioned intervention.
- Interventions achieving +55 net gain corrupt over half of the baseline-corrected decisions in seven LLMs.
Context and limitations
Evaluation is limited to seven LLMs on When2Call and MetaTool datasets; results are based on channel-keyed interventions and sealed evaluations with no matched random directions.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 04:55:20 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeideqfjrybtlxhyfyumgjqzjcflnz5ffxiuc43bwrnbqb33gilygzy