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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| train the advisor with GRPO (ablation) | Advisor Models · SWE Agent Efficiency | resolve rate (%): 52.6 -> 61.2 | Improved |
| train the advisor with GRPO (ablation) | Advisor Models · MTOB | chrF: 33.2 -> 45.8 | Improved |
| train the advisor with GRPO (ablation) | Advisor Models · RuleArena Taxes | accuracy (%): 52.0 -> 76.8 | Improved |
| use learned per-instance advisor instead of static GEPA prompt optimization | Advisor Models · MTOB | chrF: 33.0 -> 45.8 | Improved |
| use learned advisor instead of GEPA | Advisor Models · RuleArena Taxes | accuracy (%): 58.4 -> 76.8 | Improved |
| transfer advisor trained on GPT-4.1 mini to GPT-5.2 | Advisor Models · RuleArena Taxes | accuracy (%): 67.2 -> 85.6 | Improved |
| transfer advisor trained on GPT-4o mini to GPT-5 | Advisor Models · MTOB | chrF: 46.2 -> 47.8 | Improved |
| transfer advisor trained on Gemini 2.5 Flash to Gemini 3 Pro | Advisor Models · SWE Agent Efficiency | average steps: 31.7 -> 26.3 | Improved |
| transfer advisor across model families to Claude 4.5 Sonnet | Advisor Models · Review Length | reward: 0.433 -> 0.836 | Improved |
| use weak instead of strong advisor initialization prompt (ablation) | Advisor Models · Review Length | reward: 0.941 -> 0.749 at 5 epochs; 0.941 -> 0.941 at 20 epochs | Mixed |
| use Advisor Models instead of direct RL on the student weights | Advisor Models · MTOB (Qwen2.5-7B student) | chrF: 42.38 -> 38.98 | Hurt |
| train advisor on MATH-500 with GPT-4o mini student | Advisor Models · MATH-500 | accuracy | No clear effect |
Closest papers by meaning
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.85)
- Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs (NeurIPS 2025 · similarity 0.85)
- InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.84)
- The Best Instruction-Tuning Data are Those That Fit (NeurIPS 2025 · similarity 0.84)
- It Helps to Take a Second Opinion: Teaching Smaller LLMs To Deliberate Mutually via Selective Rationale Optimisation (ICLR 2025 · similarity 0.84)
- Use Your INSTINCT: INSTruction optimization for LLMs usIng Neural bandits Coupled with Transformers (ICML 2024 · similarity 0.83)
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.
