Research map / Large language models / Language, safety and interpretability
Narrow Finetuning Leaves Clearly Readable Traces in the Activation Differences
Julian Minder, Clément Dumas, Stewart Slocum, Helena Casademunt, Cameron Holmes, Robert West, Neel Nanda
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 changed | Where | Measured | Effect |
|---|---|---|---|
| give interpretability agent ADL results (Patchscope, Logit Lens, steering) versus blackbox only | gpt-5 interpretability agent · finetuning objective identification (grade >= 2) | % organisms: 39 (blackbox i=50) -> 91 (ADL i=5) | Improved |
| give ADL agent zero extra model interactions (ablation) | ADL agent · grade >= 2 | % organisms: 91 -> 82 | No clear effect |
| increase interaction budget (ablation) | interpretability agent · agent grades | grade | Improved |
| switch agent LLM from GPT5 to Gemini2.5 Pro (ablation) | interpretability agent · agent grades | grade | No clear effect |
| compare base model instead of chat model against finetuned chat model (ablation) | ADL on SDF organisms · agent grade | grade | No clear effect |
| remove bias direction (causal ablation) on finetuning data (ablation) | finetuned model · finetuning data CE loss | cross-entropy | Hurt |
| remove bias direction on pretraining data (ablation) | finetuned model · pretraining data CE loss | cross-entropy | Mixed |
| mix unrelated pretraining data into SDF finetuning (ablation) | SDF model organisms · bias traces (steering, token relevance) | similarity / relevance: 1:0.1 ratio already significant reduction | Improved |
| mix unrelated data at 1:0.1 ratio (ablation) | Llama 3.2 1B SDF · False Fact Alignment | FFA score | Hurt |
| reduce number of finetuning samples (ablation) | SDF organisms · bias and fact alignment | bias / FFA | Mixed |
| concept ablation during finetuning (CAFT) (ablation) | SDF organisms · bias traces | bias | No clear effect |
| apply ADL to broad chat finetuning | chat-tuned models · trace detectability | detectability | Hurt |
Closest papers by meaning
- When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.91)
- Mechanistically analyzing the effects of fine-tuning on procedurally defined tasks (ICLR 2024 · similarity 0.89)
- Exploring Activation Patterns of Parameters in Language Models (AAAI 2025 · similarity 0.89)
- TuCo: Measuring the Contribution of Fine-Tuning to Individual Responses of LLMs (ICML 2025 · similarity 0.89)
- An Emulator for Fine-tuning Large Language Models using Small Language Models (ICLR 2024 · similarity 0.88)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 2025 · similarity 0.88)
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.
