EvoSCM

Preprint

EvoSCM Scientific Belief Revision Through Causal Model Evolution and Experimentation

Qing Zhao1Haowei Li1Weijian Deng2 Sibei Yang1Pengxu Wei1Liang Lin1

1Sun Yat-sen University

2SIGS, Tsinghua University

Paper arXiv Code BibTeX
PhysicsChemistry & MaterialsBiology↓

Abstract

Scientific discovery depends on the ability to form hypotheses, test them through experiments, and revise them when evidence disagrees. Existing LLM agents support this process by improving their reasoning or actions, but their scientific beliefs are often scattered across free-form reasoning and difficult to update coherently. This makes it difficult to identify what failed, what should change, and whether revisions remain consistent with prior evidence. We introduce EvoSCM, which represents scientific beliefs as a population of structural causal model (SCM) hypotheses that can be tested and revised across experiments. EvoSCM formulates scientific discovery as a closed loop in which causal hypotheses guide experimentation and experimental outcomes drive causal model evolution. Competing SCM hypotheses make falsifiable predictions and guide discriminative experiments that separate alternative explanations. When observations contradict these predictions, EvoSCM distills discrepancies into correction rules identifying which aspects of the hypotheses fail to explain the evidence. These rules guide revisions to causal dependencies, latent factors, mechanisms, and parameters. Revised hypotheses are validated against accumulated evidence and carried forward to guide subsequent experiments, allowing scientific beliefs to evolve cumulatively. We evaluate EvoSCM across physics, chemistry and materials, and biology. It consistently outperforms baseline agents and existing evolution methods, yielding more accurate explanations and predictions with more effective use of experimental budgets. The evolved SCMs also transfer across base models, suggesting reusable scientific knowledge beyond any single model's reasoning process.

Figure 1: Discovery performance and transferability of EvoSCM on DiscoverPhysics. (a) Faster Convergence & Lower Error: EvoSCM achieves lower prediction error while requiring fewer experimental episodes than the baseline agent. (b) Cross-Model Transferability: evolved SCMs can be transferred directly across different base models.

EvoSCM achieves efficient scientific discovery and evolves reusable scientific knowledge that transfers across base models.

EvoSCM: Causal Model Evolution

Figure 2: Overview of EvoSCM. EvoSCM maintains a population of competing SCM hypotheses and evolves it through a closed discovery loop: (1) Causal Experiment Design: the agent abductively interprets accumulated evidence, selects discriminative interventions, and commits to falsifiable predictions tested through experimentation; (2) SCM Hypothesis Revision: prediction–observation discrepancies are inductively distilled into correction rules that guide targeted revisions to causal structure and mechanisms, while revised hypotheses are deductively validated against accumulated evidence and structural consistency to form the next-round population.

Analysis

Figure 3: Experimentation and belief revision analysis on DiscoverPhysics. (a) Experiment Informativeness: EvoSCM designs experiments with higher information scores. The score peaks early when the hypothesis population is most diverse and declines as competing hypotheses converge, leaving less disagreement to exploit. (b) Scientific Belief Revision: EvoSCM achieves consistently larger posterior entropy reductions, confirming more effective belief revision.

EvoSCM uses scientific beliefs to guide informative experiments and experimental evidence to revise those beliefs more effectively.

Figure 4: SCM hypothesis evolution on the Yukawa world in DiscoverPhysics. EvoSCM progressively recovers the 2D Yukawa ground-truth law through physically meaningful intermediate states, rather than merely fitting parameters along a smooth refinement path. (1) Round 9: the agent starts from an embryonic Coulomb-like hypothesis (at ∝ 1/rt); (2) Round 30: a revision introduces a screening factor St = K1(rt/λ)/(2πλ), coinciding with the Yukawa screening mechanism, resolving intermediate-range overprediction; (3) Round 59: a further revision introduces a softened distance ρt = √(rt2 + ε2), coinciding with the finite core model, resolving residual short-range discrepancies.

EvoSCM evolves scientific hypotheses through physically meaningful intermediate states, rather than merely refining parameters.

Results

Physical Law Discovery

Table 1: Physical law discovery on DiscoverPhysics. EvoSCM consistently outperforms all baselines in explanation score, prediction error, success rate, and experimental efficiency on both GPT-5.4 and GPT-5.5.

BackboneMethodExplanation ↑norm MSE ↓pass@1 ↑pass@2 ↑pass@3 ↑pass@4 ↑pass@5 ↑Episodes ↓
GPT-5.4Baseline39.80%5.25e−11.86%3.88%5.60%7.14%9.09%1,045
GEPA35.64%5.83e−10.00%0.00%0.00%0.00%0.00%976
ReasoningBank36.55%1.84e05.45%10.00%13.64%16.36%18.18%908
ModelSMC28.55%1.28e01.82%3.64%5.45%7.27%9.09%901
PiEvo28.18%2.25e01.82%3.64%5.45%7.27%9.09%1,033
EvoSCM (Ours)65.30%2.34e−338.08%50.16%53.74%54.55%54.55%736
GPT-5.5Baseline51.64%2.83e−27.27%13.64%19.09%23.64%27.27%2,206
GEPA45.09%3.55e−29.09%14.55%17.27%18.18%18.18%2,062
ReasoningBank51.27%2.78e−210.91%19.09%24.55%27.27%27.27%1,797
ModelSMC43.27%1.83e−27.27%13.64%19.09%23.64%27.27%1,884
PiEvo48.18%5.01e−29.09%16.36%21.82%25.45%27.27%1,845
EvoSCM (Ours)75.09%2.77e−456.36%62.73%63.64%63.64%63.64%1,685

Citation

BibTeX
@article{zhao2026evoscm,
  title={{EvoSCM}: Scientific Belief Revision Through Causal Model Evolution and Experimentation},
  author={Zhao, Qing and Li, Haowei and Deng, Weijian and Yang, Sibei and Wei, Pengxu and Lin, Liang},
  journal={arXiv preprint arXiv:2609.01526},
  year={2026}
}