AI Paper Atlas

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

CircuitPrint: Mechanistic Circuit Fingerprints for Large Language Models

Zhenxiong Yan, Suhang Yao, Yu Liu, Wenqiang Jin

ICML 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 circuit-based fingerprint probes instead of behavioral adversarial prompts (ProFLingo, TRAP)LLM fingerprinting · Gemma-2-2B derivativesFSR: 0.93 vs 0.45 and 0.33 on gemma-2-2b-itImproved
use CircuitPrint instead of CTCC invasive watermarkingLLM fingerprinting · Gemma-2-2B derivativesFSR: 0.93 vs 0.95 on gemma-2-2b-itNo clear effect
increase fine-tuning along the training trajectory (ablation)CircuitPrint · Gemma-2-2B fine-tuning checkpointsFSR: initial drop then stabilizationMixed
reduce the victim's merge ratio under Task Arithmetic with DARE (ablation)CircuitPrint · model mergingFSRHurt
raise sampling temperature above 1.0 (ablation)CircuitPrint verification · Gemma-2-2BTPRHurt
vary top-p between 0.55 and 1.0 (ablation)CircuitPrint verification · Gemma-2-2BTPRNo clear effect
replace direct supernode intervention with adversarial GCG prefix (ablation)CircuitPrint probe · Gemma-2-2B and Llama-3.2-1Bcollapse token agreementNo clear effect
intervene on random nodes instead of supernodes (ablation)circuit interventions · derivatives of Gemma-2-2B and Llama-3.2-1Bcollapse token agreement: above 85% for supernodesHurt
test on strict and relaxed near-clone models trained from scratch (ablation)CircuitPrint · Pythia-410Mmean FSR: 0.85 -> 0.07 (strict), 0.05 (relaxed)Hurt
increase suppression strength of target supernodes (ablation)derivative models · collapse token agreementagreement ratioImproved

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