AI Paper Atlas

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

Rethinking LLM Fine-Tuning via Weight Space Reparameterization: Preserving Safety during Downstream Adaptation

Min-Seong Kim, Jongbok Won, Jeesup Park, Yonghee Choi, Dong-Jun Han

NeurIPS 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a weight-space reparameterization for LLM fine-tuning that aims to preserve safety alignment while adapting to downstream tasks.

Problem

Fine-tuning LLMs on downstream tasks weakens previously aligned safety behavior, creating a trade-off between task adaptation and safety preservation; safety information is distributed across many dimensions in the original parameter space.

Approach

Reparameterizes model weights into a different space so that safety-relevant information is handled during downstream adaptation, instead of identifying safety neurons, layers, or update directions in the original parameter space.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
reparameterize weights into a safety-conditioned basis before freezing important coefficients (ablation)freeze-and-adapt fine-tuning · Llama-2-7B GSM8Kjailbreak ASR (%): 9.17 -> 6.90Improved
reparameterized freeze at 10% vs original-space freeze at 10% (ablation)safety-tuned Llama-2-7B · jailbreak averagejailbreak ASR (%): 33.95 -> 14.67Improved
increase original-space freeze ratio from 10% to 50% (ablation)full fine-tuning · Llama-2-7Bjailbreak ASR (%): 33.95 -> 18.43Improved
WSR-Tune versus full-parameter FTLlama-2-7B-Chat · GSM8Kjailbreak ASR (%): 20.78 -> 6.90Improved
WSR-Tune versus full-parameter FTLlama-2-13B-Chat · GSM8Kaccuracy (%): 45.94 -> 49.58Improved
WSR-Tune versus full-parameter FTLlama-3.1-8B-Instruct · MATHaccuracy (%): 12.12 -> 13.70Improved
WSR-Tune versus full-parameter FTLlama-2-7B-Chat · GSM8Kaccuracy (%): 41.17 -> 38.99Hurt
add WSR-Tune to SN-TuneSN-Tune · Llama-2-7B-Chat GSM8Kjailbreak ASR (%): 27.18 -> 23.07Improved
add WSR-Tune to SafeInstrSafeInstr · Llama-2-13B-Chat GSM8Kjailbreak ASR (%): 5.82 -> 2.93Improved
add WSR-Tune to SEALSEAL · Llama-2-13B-Chat GSM8Kjailbreak ASR (%): 10.28 -> 1.00Improved
raise freeze ratio from 10% to 20% (ablation)WSR-Tune · Llama-2-7B-Chat GSM8Kjailbreak ASR (%): 6.90 -> 5.59Improved
vary freeze ratio 1% to 20% (ablation)WSR-Tune · Llama-2-7B-Chat GSM8Kaccuracy (%): 38.82 to 39.88No 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