AI Paper Atlas

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

At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization

Praneet Suresh, Jack Stanley, Sonia Joseph, Luca Scimeca, Danilo Bzdok

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs · 0 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
add typos to 5% of MMLU prompt wordsLlama 3.1 8B · MMLUaccuracy: 66.7% -> 51.01%Hurt
add typos to 5% of MMLU prompt wordsGPT-4o mini · MMLUaccuracy: 82.10% -> 74.85%Hurt
add typos to 5% of MMLU prompt wordsGPT-5-thinking-nano · MMLUaccuracy: 91.26% -> 86.45%Hurt
increase typo rate from clean to every wordGPT-2 25M · TinyStories layer 6number of active SAE concepts: nearly 30% more featuresHurt
fine-tune on high-energy-score samples instead of lowest decile (ablation)GPT-2 25M · TinyStories typos (layer 6)validation loss: over 3% lower final lossImproved
select fine-tuning samples by SAE energy score instead of entropy or Mahalanobis distance (ablation)GPT-2 25M · TinyStories typosvalidation loss: lower final loss at every decileImproved
align SAE features of successful jailbreaks to unsuccessful centroid via LoRALlama 3.1 8B · WildJailbreak held-outjailbreak success rate: 46% -> 7%Improved
SAE-informed LoRA jailbreak alignmentLlama 3.1 8B · MMLUaccuracy: 0.09% lowerNo clear effect
SAE-informed LoRA jailbreak alignmentLlama 3.1 8B · OR-Benchover-refusal rate: 3.5% -> 3.0%No clear effect
align SAE features of ASR transcripts to clean counterparts via LoRALlama 3.1 8B · Spoken-SQuADexact match: 49.45 -> 58.33Improved
same SAE-aligned adapters applied to clean textLlama 3.1 8B · SQuADexact match: 59.97 -> 67.88Improved
choose fine-tuning samples from middle energy deciles in middle layers (ablation)GPT-2 25M · TinyStories typosvalidation loss: U-shaped loss across decilesMixed

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