AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Beyond I’m Sorry, I Can’t: Dissecting Large-Language-Model Refusal

Nirmalendu Prakash, Yeo Wei Jie, Amir Abdullah, Ranjan Satapathy, Erik Cambria, Roy Ka-Wei Lee

AAAI 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs · 2 citations (OpenAlex)

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
ablate random active features instead of jailbreak-critical features (ablation)SAE feature ablation · D_harmfuljailbreak count: 0 samples on LLaMA, 5 samples on GemmaHurt
keep silent (zero-activation) critical features in the ablation set (ablation)Stage 2 feature set · D_harmfuljailbreak success: drops to 372 (LLaMA) and 103 (Gemma) samples when inactive features removedImproved
rank Stage 3 features with factorization machine instead of linear probe (ablation)Stage 3 feature discovery · D_harmful, LLaMAjailbroken samples: 101 samples for linear probe, significantly fewer than FMImproved
FM embedding size 5 vs 20 (ablation)Factorization machine · Feature embeddingsembedding diversity: cosine similarity > 0.9 at size 5Improved
FM embedding size 50 instead of 20 (ablation)Factorization machine · Feature embeddingsembedding quality / training costNo clear effect
add Stage 3 FM-found features to Stage 2 setRefusal-feature pipeline · D_harmful ASRASR: 0.33 -> 0.31 (Gemma), 0.70 -> 0.57 (LLaMA)Hurt
ablate only Stage 3 features (ablation)Stage 3 features · D_harmfuljailbreak rate: 330/372 (89%) on LLaMA, 8 (3%) on GemmaMixed
use SAE refusal-feature ablation instead of single refusal direction ablationLLaMA-3.1-8B-IT · ALPACA CE lossCE loss: 1.78 -> 2.17Hurt
use features active on at least k tokens (Phi-3 Mini approach) instead of the proposed pipelineBaseline feature search · D_harmfulmax ASR: 0.1 Gemma, 0.07 LLaMAHurt
exclude layers 1 and 2 features from candidate set (ablation)Stage 1 · Candidate selectionresponse qualityImproved
increase K beyond 2000 in Stage 3 (ablation)LLaMA · Stage 3 top-K ablationresponse qualityHurt
clamp redundant features to high values on BOS tokenGemma-2-2B-IT · benign promptsrefusal behaviourImproved

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