Source-linked AI summary
EvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation
Qing Zhao, Haowei Li, Weijian Deng, Pengxu Wei, Liang Lin
TL;DR
Scientific agents often lack explicit, revisable beliefs when new evidence challenges their conclusions. EvoSCM represents those beliefs as evolving populations of structural causal models and, on DiscoverPhysics, consistently improves discovery while enabling transfer across model architectures.
Problem
Existing scientific agents often leave hypotheses distributed across free-form reasoning, records, and summaries, making belief revision under contradictory evidence unreliable.
Method
EvoSCM maintains competing SCM hypotheses and evolves them through abduction, discriminative intervention, falsifiable prediction, discrepancy-based correction, structural revision, and validation.
Results
EvoSCM consistently outperforms baseline agents on DiscoverPhysics, with faster convergence, lower prediction error, fewer experiment episodes, and transferable evolved SCMs.
Takeaways & Limitations
Evolved SCMs provide explicit, portable causal knowledge for reasoning, prediction, experiment planning, and generalization to unseen interventions.
Takeaways & Limitations
Revised hypotheses must remain valid DAGs with well-defined, mutually compatible structural equations and mechanisms.
Abstract
from arXiv · showhide
Scientific agents must learn not only how to reason, but also what to believe. However, existing LLM agents typically express scientific hypotheses in free-form text, leaving their beliefs implicit and difficult to test or revise. We introduce EvoSCM, which equips scientific agents with explicit structural causal models that evolve as new experimental evidence is collected. EvoSCM maintains a population of competing SCM hypotheses, each encoding a candidate causal explanation of the environment, and evolves them through a closed discovery loop. In each round, the agent abduces latent mechanisms from accumulated evidence, designs discriminative interventions, and commits to falsifiable predictions that it tests through experimentation. Discrepancies between prediction and observation are inductively distilled into correction rules that revise the causal structures and mechanisms of each hypothesis, and the agent then deductively validates the revised population against accumulated evidence and structural consistency to guide the next round. We evaluate EvoSCM on DiscoverPhysics, a benchmark requiring agents to uncover the hidden dynamics of noncanonical physical worlds through experimentation. EvoSCM consistently improves scientific discovery over baselines, yielding more accurate explanations and predictions while making more effective use of experimental interactions.
1 INTRODUCTION
EvoSCM addresses the difficulty of revising implicit scientific beliefs by representing them as persistent structural causal models. On DiscoverPhysics, it improves discovery performance, convergence speed, prediction error, and transferability compared with baseline agents.
- Motivation: Existing scientific agents struggle to revise beliefs when new evidence contradicts earlier conclusions because their knowledge is distributed across free-form traces, records, and summaries.This makes failed assumptions and appropriate revisions difficult to identify.
- Approach: EvoSCM makes structural causal models the explicit, persistent epistemic state of a scientific agent.It maintains competing candidate explanations rather than leaving beliefs implicit in reasoning artifacts.
- Approach: Its closed loop alternates causal experiment design with SCM evolution, using interventions, falsifiable predictions, discrepancy-based correction rules, and structural validation.The process revises causal structures, latent variables, mechanisms, and parameters before the next round.
- Results: EvoSCM consistently outperforms baseline agents on DiscoverPhysics while requiring fewer experiment episodes and supports direct transfer of evolved SCMs across base models.The benchmark tests recovery of hidden dynamics in noncanonical physical environments through experimentation.
2 EVOSCM: SELF-EVOLVING CAUSAL MODELS
EvoSCM treats scientific discovery as sequential causal-model recovery with a population of competing SCM hypotheses. It repeatedly designs discriminative interventions, compares predictions with observations, distills correction rules, revises hypotheses, and validates the resulting population.
- Problem formulation: Scientific discovery is formalized as recovering an unknown causal model through interventions and observed outcomes over a finite experimental budget.The agent knows observable variables but must recover latent variables, causal structure, mechanisms, and parameters.
- SCM representation: Each SCM hypothesis combines a causal graph with structural equations and parameters, forming a falsifiable explanation whose interventional predictions can be tested.Prediction discrepancies can be traced to structural or parametric commitments.
- Population-based epistemic state: EvoSCM maintains a diverse population of competing hypotheses to avoid premature commitment and identify interventions where candidate predictions diverge.Population disagreement directly guides comparative experiment design.
- Causal experiment design: For each hypothesis, the agent abducts latent states from accumulated evidence, selects a maximally discriminative intervention, and commits to a hypothesis-specific prediction before experimentation.The sequence follows abduction, action, and prediction within a comparative experimental loop.
- SCM hypothesis revision: After observing outcomes, EvoSCM compares them with predictions, distills recurring failures into correction rules, and applies targeted structural or mechanistic revisions.Supported edits include adding or removing edges and latent variables and updating mechanisms and parameters.
1. Evidence Validation. The agent derives predictions from ˜Ht+1
EvoSCM validates revised causal hypotheses against accumulated evidence and structural consistency before retaining them, then uses surviving SCMs to predict unseen interventions. Its explicit models are inspectable, portable, and transferable across architectures.
- Evidence validation: Revised hypotheses must explain all accumulated interventions and observations; candidates resolving a new discrepancy while contradicting earlier evidence are rejected.
- Evidence validation: Consistency checks require a valid DAG, well-defined structural equations, and mutually compatible mechanisms.
- Population evolution: Hypotheses passing both checks form the evolved population used with updated evidence to seed the next experiment-design iteration.
- Inference: For unseen interventions, each surviving hypothesis derives predictions through its structural equations, and population predictions can be aggregated into a consensus outcome.
- Inference: The do(·) operator supports predictions for unseen interventions by simulating causal consequences through discovered graphs and mechanisms rather than stored input–output associations.
- Model portability: Evolved SCMs make qualitative and quantitative claims inspectable while remaining portable for new agents, future experiment planning, and cross-architecture transfer.
3 EXPERIMENT
EvoSCM is evaluated on DiscoverPhysics, an interactive benchmark requiring agents to uncover hidden dynamics through active experimentation. It consistently outperforms baselines in discovery quality, prediction accuracy, efficiency, and cross-model transfer.
- Benchmark: DiscoverPhysics requires agents to formulate hypotheses, design interventions, and iteratively refine physical-law understanding from observational feedback.
- Scientific discovery performance: EvoSCM consistently outperforms direct baselines across metrics and backbones, including mechanism explanation quality, prediction error, pass rates, and interaction efficiency.
- Scientific discovery performance: On GPT-5.5, EvoSCM raises mechanism explanation from 0.5164 to 0.7509 and reduces normalized MSE from 2.83 × 10−2 to 2.77 × 10−4.
- Cross-model SCM transferability: Qwen3.6-35B-A3B rises from a 0.00% pass rate to 63.64% pass@5 and a 0.740 explanation score when equipped with the GPT-5.6-Sol SCM.
- Cross-model SCM transferability: Evolved SCMs transfer across architectures, enabling a smaller model to approach frontier-model zero-shot discovery performance.
4 CONCLUSION
EvoSCM makes structural causal models the persistent epistemic state of scientific agents, turning free-form reasoning into a structured, falsifiable, and revisable process. Experiments show stronger discovery with fewer episodes and portable causal models across architectures.
- EvoSCM addresses the gap between improving how agents reason and revising what they believe by maintaining and evolving competing SCM hypotheses.
- Its closed loop combines abduction, intervention, induction, and deduction to make scientific reasoning structured, falsifiable, and cumulatively revisable.
- DiscoverPhysics evaluations show superior forward prediction and mechanism alignment with significantly fewer experiment episodes.
- The evolved causal models support zero-shot transfer across different underlying architectures.