AI Paper Atlas

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

Efficient Transcoder Adaptation for Fine-Tuned Models: Revealing Medical Reasoning Mechanisms in Large Language Models

Zhouxing Tan, Hanlin Xue, Yulong Wan, Ruochong Xiong, Xu Chu, Xiang Li, Junfei Liu

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 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
fine-tune only MLP layers with frozen attention (FTFA) instead of full-parameter SFTGemma-2-2b · Medical diagnosis benchmarkaccuracy: 50.06 -> 52.27Improved
fine-tune only MLP layers with frozen attention (FTFA) instead of full-parameter SFTGemma-2-2b · Medical diagnosis benchmarkBERTScore: 54.18 -> 56.36Improved
apply FTFA fine-tuning relative to the untuned base modelGemma-2-2b · Medical diagnosis benchmarkaccuracy: 50.79 -> 52.27Improved
use base-model pretrained transcoders directly on the fine-tuned model instead of PAT adaptation (ablation)FTFA Gemma2-2b replacement model · Top-1 accuracy of replacement modelTop-1 accuracy: nearly zero Top-1 accuracyHurt
adapt transcoders with PAT instead of training from scratchGemma2-2b transcoders (16k) · Transcoder training costtraining tokens: 1B -> 1MImproved
adapt transcoders with PAT instead of training from scratchGemma2-2b transcoders (16k) · Transcoder training costH100 hours: 210 -> 0.2Improved

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