AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMs

Ziqian Zhong, Aditi Raghunathan

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use weight-difference singular directions for monitoring instead of BEATbackdoor detection on PPO trojans · PPO trojan3 TPRTPR: 31.95 -> 100.00Improved
use weight-difference directions instead of activation difference norm baselinebackdoor detection on PPO trojans · PPO trojan3 TPRTPR: 11.50 -> 100.00Improved
weight-based monitoring on LoRA mtba backdoorWeightWatch · LoRA mtba% prompts flagged: 27.27 (vs 100.00 on other triggers)Hurt
use different trigger words for mtba LoRA replicationWeightWatch · LoRA mtba% prompts flagged: 27.27 -> 52.53Improved
regenerate cleaner FFT backdoor datasetWeightWatch FFT backdoor models · FFT ctba% prompts flagged: 38 -> 100.00Improved
steer along weight directions with default calibration rangeZephyr-RMU · WMDP-Bioaccuracy: 31.2 -> 39.7Improved
widen steering range (drop top/bottom 0.1% calibration), FPR 0.94% -> 8.54% (ablation)Zephyr-RMU steering · WMDP-Bioaccuracy: 39.7 -> 56.2Improved
widen steering range (FPR 8.54%) (ablation)Zephyr-RMU steering · MMLUaccuracy: 57.1 -> 57.0No clear effect
steer along directions plus prompt prefixCircuit Breaker · HarmBenchattack success rate: 9.6 -> 82.08Improved
steer along weight directionsWHP unlearned model · knowledge recoveryqualitativeNo clear effect
more fine-tuning data on math in OLMo SFT/DPOdirection counts · math directions% directions: 15 (1.5%) directions despite 4.84%-7.74% math dataNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy