AI Paper Atlas

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

COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics

Kartik Sharma, Rakshit Trivedi

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 4 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
steer with COLD-FD (finite-difference learning dynamics) instead of unsteered baseCOLD-FD · CAA coordinate-ais, Qwen-2.5-7B-Instructselection accuracy: 0.02 -> 0.98Improved
use COLD-FD on Llama-2-7b-chat coordinate-aisCOLD-FD · CAA coordinate-ais, Llama-2-7b-chatselection accuracy: 0.28 -> 0.90Improved
use COLD-FD on myopic-reward (Llama-2-7b-chat, pairwise)COLD-FD · CAA myopic-rewardselection accuracy: 0.76 -> 0.60Hurt
use COLD-Kernel (unit kernel) instead of base on QwenCOLD-Kernel · CAA, Qwen-2.5-7B-Instructselection accuracy: 0.32 -> 0.32No clear effect
use COLD-Kernel with positive-only examples on Llama-2-7b-hfCOLD-Kernel · CAA coordinate-ais, Llama-2-7b-hfselection accuracy: 0.52 -> 0.90Improved
use COLD-FD on other model families (Gemma, Mistral)COLD-FD · CAA hallucinationselection accuracy: 0.64 -> 0.70 (Gemma); 0.62 -> 0.88 (Mistral)Improved
use COLD-Kernel for pluralistic alignment instead of baseCOLD-Kernel · OpinionsQA, Black respondentsKL divergence: 2.43 -> 0.86Improved
use COLD-FD for pluralistic alignmentCOLD-FD · OpinionsQA, RepublicansKL divergence: 2.38 -> 2.40No clear effect
use COLD-Kernel instead of ReFT for runtimeCOLD-Kernel · CAA pairwise runtimeruntime: 109.57 -> 39.29Improved
increase number of in-context examples (ablation)COLD-FD · CAA tasks, Llama-2-7b-hfselection accuracyNo 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