AI Paper Atlas

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

Faithful Bi-Directional Model Steering via Distribution Matching and Distributed Interchange Interventions

Yuntai Bao, Xuhong Zhang, Jintao Chen, Ge Su, Yuxiang Cai, Hao Peng, SUN Bing, Haiqin Weng, Liu Yan, Jianwei Yin

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 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
use JSD distribution matching instead of reverse KL (ablation)CDAS · AxBench Concept500 (Gemma-2-2B L20, 2 concepts)steering score: 0.840 -> 0.860Improved
use JSD distribution matching instead of forward KL (ablation)CDAS · AxBench Concept500 (Gemma-2-2B L20, 2 concepts)steering score: 0.560 -> 0.860Improved
tune the steering factor at inference instead of unit factor (ablation)CDAS · AxBench Gemma-2-2B L10steering score: 0.121 -> 0.631Improved
tune the steering factor at inference instead of unit factor (ablation)CDAS · AxBench Gemma-2-9B L20steering score: 0.238 -> 0.992Improved
scale model from Gemma-2-2B to Gemma-2-9B (layer 20)CDAS · AxBench Concept500steering score: 0.608 -> 0.992Improved
use CDAS instead of LoReFTGemma-2-9B L20 · AxBench Concept500steering score: 0.777 -> 0.992Improved
use CDAS instead of RePSGemma-2-9B L31 · AxBench Concept500steering score: 0.624 -> 0.518Hurt
move intervention from layer 10 to layer 20CDAS · AxBench Gemma-2-2Bsteering score: 0.631 -> 0.608No clear effect
move intervention from layer 10 to layer 20RePS · AxBench Gemma-2-2Bsteering score: 0.756 -> 0.606Hurt
use bi-directional CDAS instead of DAS objectiveLlama-3.1-8B · Refusal override suppression harmonic meanharmonic mean (%): 2 -> 95Improved
use CDAS instead of RePSPhi-3.5-mini · Refusal override suppression harmonic meanharmonic mean (%): 91 -> 46Hurt
use CDAS instead of RePSLlama-3.1-8B · MMLU change when overriding refusalaccuracy change (%): -35.57 -> 0.20Improved

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