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

Export BriefConnect with Author
View PDF ↗
PDF Viewer
Open Full PDF

BUILDER'S SANDBOX

Build This Paper

Use an AI coding agent to implement this research.

OpenAI Codex
OpenAI CodexAI Agent

Lightweight coding agent in your terminal.

Claude Code
Claude CodeAI Agent

Agentic coding tool for terminal workflows.

AntiGravity IDE
AntiGravity IDEScaffolding

AI agent mindset installer and workflow scaffolder.

Cursor
CursorIDE

AI-first code editor built on VS Code.

VS Code
VS CodeIDE

Free, open-source editor by Microsoft.

Estimated $10K - $14K over 6-10 weeks.

See exactly what it costs to build this -- with 3 comparable funded startups.

7-day free trial. Cancel anytime.

Discover the researchers behind this paper and find similar experts.

7-day free trial. Cancel anytime.

References (53)

[1]
Angular Steering: Behavior Control via Rotation in Activation Space
2025Hieu M. Vu, Tan M. Nguyen
[2]
Learning without training: The implicit dynamics of in-context learning
2025Benoit Dherin, Michael Munn et al.
[3]
HyperSteer: Activation Steering at Scale with Hypernetworks
2025Jiuding Sun, Sidharth Baskaran et al.
[4]
Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
2025Yuxiao Qu, Matthew Y. R. Yang et al.
[5]
VITAL: A New Dataset for Benchmarking Pluralistic Alignment in Healthcare
2025Anudeex Shetty, Amin Beheshti et al.
[6]
STEER-ME: Assessing the Microeconomic Reasoning of Large Language Models
2025Narun K. Raman, Taylor Lundy et al.
[7]
Scaling Test-Time Compute Without Verification or RL is Suboptimal
2025Amrith Rajagopal Setlur, Nived Rajaraman et al.
[8]
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
2025Jonas Geiping, Sean McLeish et al.
[9]
s1: Simple test-time scaling
2025Niklas Muennighoff, Zitong Yang et al.
[10]
AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
2025Zhengxuan Wu, Aryaman Arora et al.
[11]
LatentQA: Teaching LLMs to Decode Activations Into Natural Language
2024Alexander Pan, Lijie Chen et al.
[12]
Benchmarking Distributional Alignment of Large Language Models
2024Nicole Meister, Carlos Guestrin et al.
[13]
Controlling Language and Diffusion Models by Transporting Activations
2024Pau Rodríguez López, Arno Blaas et al.
[14]
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
2024C. Snell, Jaehoon Lee et al.
[15]
PERSONA: A Reproducible Testbed for Pluralistic Alignment
2024Yuntao Bai, Andy Jones et al.
[16]
Learning Dynamics of LLM Finetuning
2024Yi Ren, Danica J. Sutherland
[17]
Refusal in Language Models Is Mediated by a Single Direction
2024Andy Arditi, Oscar Obeso et al.
[18]
HelpSteer2: Open-source dataset for training top-performing reward models
2024Zhilin Wang, Yi Dong et al.
[19]
Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization
2024Yuanpu Cao, Tianrong Zhang et al.
[20]
Context Steering: Controllable Personalization at Inference Time
2024J. He, Sashrika Pandey et al.

Showing 20 of 53 references

Founder's Pitch

"COLD-Steer enables efficient, training-free control of LLM behavior at inference time, requiring significantly fewer examples than existing methods."

LLM SteeringScore: 7View PDF ↗

Commercial Viability Breakdown

0-10 scale

High Potential

2/4 signals

5

Quick Build

4/4 signals

10

Series A Potential

2/4 signals

5

Sources used for this analysis

arXiv Paper

Full-text PDF analysis of the research paper

GitHub Repository

Code availability, stars, and contributor activity

Citation Network

Semantic Scholar citations and co-citation patterns

Community Predictions

Crowd-sourced unicorn probability assessments

Analysis model: GPT-4o · Last scored: 3/6/2026

Explore the full citation network and related research.

7-day free trial. Cancel anytime.

Understand the commercial significance and market impact.

7-day free trial. Cancel anytime.

Get detailed profiles of the research team.

7-day free trial. Cancel anytime.

Related Papers

Loading…