Research map / Multi-agent systems and games / LLM agents
Cycle-of-Science: Reliable Reasoning through Counterfactual Verification for Agent Decision Making
Ruojie Zhang, Wencheng Zhu, Peiyuan Jiang, dayong zhu
Read the paperCheck your idea against 137,217 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 Cycle-of-Science, a scientific-method-inspired framework where LLM agents autonomously explore potential causal relationships to improve decision-making reliability via counterfactual verification.
Problem
LLM agents learn passively from similar past experiences, which confounds correlation with causality and hurts robust decision-making.
Approach
Inspired by the Scientific Method, the agent cycles through exploring potential causal relations and verifying them through counterfactual checks, rather than learning passively from similar experiences.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Hypothesis-Experiment-Validation scientific loop (SFT+CPO) (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 56.41 -> 67.23 | Improved |
| add the scientific loop (SFT+CPO) (ablation) | CoS Llama-3-8B · ScienceWorld (unseen) | success rate (%): 55.23 -> 65.35 | Improved |
| add the scientific loop (SFT+CPO) (ablation) | CoS Gemma-7B · ScienceWorld (unseen) | success rate (%): 51.17 -> 60.47 | Improved |
| add CPO after SFT instead of SFT only (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 62.34 -> 67.23 | Improved |
| use CPO only instead of SFT only (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 62.34 -> 65.18 | Improved |
| increase hypotheses from N=1 to N=3 (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 58.47 -> 67.23 | Improved |
| increase hypotheses from N=3 to N=5 (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 67.23 -> 65.74 | Hurt |
| use adaptive threshold instead of best fixed threshold delta=0.5 (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 65.20 -> 67.23 | Improved |
| add intervention-validation filtering (vs no filtering) (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 59.34 -> 67.23 | Improved |
| raise fixed threshold from 0.3 to 0.7 (ablation) | CoS Mistral-7B · ScienceWorld (unseen) | success rate (%): 65.47 -> 62.48 | Hurt |
| add self-play trajectories (5K human only to 1:9 mix, 50K total) (ablation) | CoS · ALFWorld | success rate (%): 72.18 -> 80.51 | Improved |
| increase self-play data from 1:9 to 1:19 (ablation) | CoS · ALFWorld | success rate (%): 80.51 -> 79.94 | No clear effect |
Closest papers by meaning
- ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.91)
- Thought Branches: Interpreting LLM Reasoning Requires Resampling (ICLR 2026 · similarity 0.88)
- Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making (ICML 2025 · similarity 0.88)
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.87)
- Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning (NeurIPS 2025 · similarity 0.87)
- Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? (NeurIPS 2024 · similarity 0.87)
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.
