AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models

Parth Asawa, Alan Zhu, Abigail O'Neill, Matei Zaharia, Alex Dimakis, Joseph E. Gonzalez

ICML 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

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
train the advisor with GRPO (ablation)Advisor Models · SWE Agent Efficiencyresolve rate (%): 52.6 -> 61.2Improved
train the advisor with GRPO (ablation)Advisor Models · MTOBchrF: 33.2 -> 45.8Improved
train the advisor with GRPO (ablation)Advisor Models · RuleArena Taxesaccuracy (%): 52.0 -> 76.8Improved
use learned per-instance advisor instead of static GEPA prompt optimizationAdvisor Models · MTOBchrF: 33.0 -> 45.8Improved
use learned advisor instead of GEPAAdvisor Models · RuleArena Taxesaccuracy (%): 58.4 -> 76.8Improved
transfer advisor trained on GPT-4.1 mini to GPT-5.2Advisor Models · RuleArena Taxesaccuracy (%): 67.2 -> 85.6Improved
transfer advisor trained on GPT-4o mini to GPT-5Advisor Models · MTOBchrF: 46.2 -> 47.8Improved
transfer advisor trained on Gemini 2.5 Flash to Gemini 3 ProAdvisor Models · SWE Agent Efficiencyaverage steps: 31.7 -> 26.3Improved
transfer advisor across model families to Claude 4.5 SonnetAdvisor Models · Review Lengthreward: 0.433 -> 0.836Improved
use weak instead of strong advisor initialization prompt (ablation)Advisor Models · Review Lengthreward: 0.941 -> 0.749 at 5 epochs; 0.941 -> 0.941 at 20 epochsMixed
use Advisor Models instead of direct RL on the student weightsAdvisor Models · MTOB (Qwen2.5-7B student)chrF: 42.38 -> 38.98Hurt
train advisor on MATH-500 with GPT-4o mini studentAdvisor Models · MATH-500accuracyNo 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