Research map / Reinforcement learning / RL from feedback
Agentic Monte Carlo: Simulating Reinforcement Learning for Black-Box Agents
Dae Yon Hwang, Raunaq Suri, Valentin Villecroze, Anthony L. Caterini, Jesse C. Cresswell, Noël Vouitsis, Brendan Leigh Ross
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 |
|---|---|---|---|
| use learned value-function resampling (AMC) instead of Best-of-15 | ReAct agent, Llama-3.2-11B · WebShop | score: 0.562 -> 0.625 | Improved |
| use AMC instead of Best-of-15 | ReAct agent, GPT-4.1-mini · WebShop | score: 0.403 -> 0.488 | Improved |
| use AMC instead of Best-of-15 | ReAct agent, Llama-3.2-11B · TextCraft | score: 0.296 -> 0.543 | Improved |
| use AMC instead of Best-of-15 | ReAct agent, GPT-5.1 · TextCraft | score: 0.889 -> 0.790 | Hurt |
| use AMC instead of Best-of-15 | ReflAct agent, Llama-3.1-8B · SciWorld | score: 0.347 -> 0.376 | Improved |
| add FoA heuristic-prompt value function SMC over Best-of-15 | Llama-3.2-11B · WebShop | score: 0.562 -> 0.580 | No clear effect |
| add a trained value function (vs prompted zero-shot value) (ablation) | AMC, Llama-3.2-11B · WebShop | score: 0.556 -> 0.625 | Improved |
| add a trained value function (vs prompted zero-shot value) (ablation) | AMC, GPT-5.1 · SciWorld | score: 0.506 -> 0.597 | Improved |
| add a trained value function (vs prompted zero-shot value) (ablation) | AMC, GPT-5.1 · TextCraft | score: 0.815 -> 0.790 | Mixed |
| train value function on trajectories from the actual policy rather than a surrogate (ablation) | AMC, GPT-4.1-mini · TextCraft | score: 0.790 -> 0.852 | Improved |
| train value function on trajectories from the actual policy rather than a surrogate (ablation) | AMC, GPT-4.1-mini · WebShop | score: 0.467 -> 0.488 | Improved |
| pair cheap GPT-4.1-mini with AMC instead of GPT-5.1 Best-of-15 | cost comparison · SciWorld | score and cost: 0.533 -> 0.673; cost 0.18 -> 0.06 | Improved |
Closest papers by meaning
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.89)
- Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.89)
- A Black-box Approach for Non-stationary Multi-agent Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems (NeurIPS 2026 · similarity 0.88)
- Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.88)
- Foundations of Multi-Agent Learning in Dynamic Environments: Where Reinforcement Learning Meets Strategic Decision-Making (AAAI 2025 · similarity 0.88)
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.
