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

Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs

NeurIPS 2025 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes M-Pilot, a lightweight white-box LLM controller (LLaMA-3-8B-Instruct) trained with iterative DPO to generate intermediate guidance prompts that steer black-box LLMs (gpt-4o-mini, gpt-3.5-turbo) on personalization, reasoning, and planning tasks.

Problem

Black-box LLMs lack the ability to be improved via parameter access for complex long-horizon tasks such as reasoning, planning, and personalization; existing ICL and best-of-N adapter methods rely on heuristic prompts or the generator's own candidate quality.

Approach

Treat the black-box LLM as an environment and train a small white-box controller to emit intermediate guidance (decompositions, plans, user summaries) via multi-turn interaction, with correctness feedback forming preference pairs optimized by iterative DPO (self-improving with previous iteration as reference).

Key results

• On ALFWorld, M-Pilot with gpt-3.5-turbo reaches 96.27% overall success rate vs. 88.06% for AdaPlanner.
• On GSM8K with gpt-3.5-turbo, M-Pilot reaches 93.1% accuracy vs. 86.4% for PAL Self-Debug.
• On LaMP-2N, M-Pilot with gpt-4o-mini reaches 0.823 accuracy vs. 0.792 for RAG (k=4); BLEU on LaMP-4 improves to 4.298 vs. 2.953.

Measured effects reported in the paper

Surprising findings

• A controller trained with gpt-4o-mini as environment transfers plug-and-play to gpt-3.5-turbo and gemini-1.5-flash without retraining.

Limitations

• Potential misuse for jailbreaking black-box LLMs and extracting sensitive information, as noted by the authors.
• Privacy risks from storing interaction data and retrieved user histories in personalization tasks.

Experimental setup

Datasets: LaMP GSM8K ALFWorld MATH500

Benchmarks: LaMP GSM8K GSM-HARD ALFWorld MATH500

Metrics: accuracy F1 MAE RMSE ROUGE-1 ROUGE-L BLEU success rate

Baselines: RAG PAG CoT Least-to-Most PAL PAL Self-Debug BUTLER ReAct Reflexion AdaPlanner Fermi ORPO

Builds on: DPO STaR SimPO ReAct AdaPlanner

Closest papers by meaning

• Enhancing Decision-Making of Large Language Models via Actor-Critic (ICML 2025 · similarity 0.87)
• Cross-model Control: Improving Multiple Large Language Models in One-time Training (NeurIPS 2024 · similarity 0.87)
• FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents (ICML 2026 · similarity 0.87)
• Controlling Large Language Model with Latent Action (ICML 2025 · similarity 0.86)
• Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL (NeurIPS 2025 · similarity 0.86)
• InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.86)