PaperlayerView exact revision on arXiv

Source-linked AI summary

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists

Junlin Yang, Dylan Zhang, Xiangchen Song, Qirun Dai, Xiao Liu, Yuen Chen, Aniket Vashishtha, Jing Shi, Chenhao Tan, Hao Peng

arXiv:2605.26029v2cs.AIcs.CL

TL;DR

LLM causal-reasoning benchmarks often test question answering rather than whether models recover mechanisms from interventions. CausaLab evaluates interactive discovery in hidden SCMs and finds that prediction can remain accurate while mechanism recovery is incomplete, with observation-conditioned online intervention best balancing prediction and graph recovery.

  • Problem

    Existing causal benchmarks often reduce causal reasoning to question-answering or classification, leaving mechanism-based evaluation of LLMs limited.

  • Method

    CausaLab evaluates agents on hidden-SCM transfer episodes, scoring both held-out prediction and stepwise causal commitments converted into SCM artifacts.

  • Results

    Agents’ endpoint accuracy and mechanism fidelity move separately, while observation-conditioned online intervention best balances prediction and graph recovery.

  • Takeaways & Limitations

    CausaLab separates predictive success from causal understanding and exposes limits in LLM agents’ experimental mechanism recovery.

  • Takeaways & Limitations

    The benchmark covers synthetic 3–7 node SCMs with mostly linear mechanisms and does not represent the full range of real scientific causal structures.

Abstract

from arXiv · show

We introduce CausaLab, a scalable environment for evaluating interactive causal discovery by LLM agents. Unlike prior evaluations, CausaLab evaluates both whether an agent can solve a problem using causal evidence and whether its answer is grounded in a faithful recovered causal mechanism. Each episode places an agent in a synthetic laboratory: it receives prior measurement records, intervenes on a manipulator crystal, and predicts the resonance frequency of a held-out reactor crystal governed by the same mechanism. The hidden data-generating process is a randomly sampled structural causal model (SCM), so success requires recovering both a causal graph and structural equations rather than recalling prior knowledge. Experiments show a persistent gap between prediction and mechanism recovery: in the purely observational 6-node setting, GPT-5.2-high reaches 92% task accuracy but only 0.471 all-edge $F_1$. Mixed observation-intervention strategies improve structural fidelity, while pure intervention remains difficult even for strong agents. We identify premature stopping as a major weakness and show that consistency verification mitigates it. CausaLab therefore separates predictive success from causal understanding and exposes current LLM agents' limits as experimental causal reasoners.

1. Introduction

CausaLab evaluates LLM agents as interactive causal discoverers, testing both prediction from interventions and faithfulness of the recovered mechanism. It addresses limitations of static causal benchmarks by scoring agents’ evidence, hypotheses, experiments, actions, and mechanism trajectories.

  • Motivation: Causal reasoning matters because scientific, medical, and policy decisions depend on responses to interventions, while measuring causal reasoning remains challenging for LLMs.
  • Motivation: Existing benchmarks typically reduce causal reasoning to question-answering or classification over graphs, datasets, or narratives.
  • CausaLab: CausaLab places agents in episodes where they use prior records and interventions on one crystal to predict the held-out frequency of another crystal generated by a shared data-generating process.
  • CausaLab: A domain-specific language records evidence, graph and equation hypotheses, planned experiments, and actions, enabling evaluation of final predictions and trajectory-level mechanism faithfulness.
  • Key findings: Experiments show that correct predictions often do not reflect correct mechanism discovery, with endpoint accuracy and mechanism fidelity moving separately across multiple controls.

2. Background and Related Work

CausaLab frames causal reasoning as intervention- and counterfactual-based understanding formalized by structural causal models, rather than associational prediction alone. It extends interactive causal-discovery evaluation by testing whether agents recover mechanisms that transfer across related systems, while scoring both task utility and mechanism faithfulness.

  • Causal reasoning and SCMs: Structural causal models represent causal assumptions with directed graphs and structural equations, enabling evaluation of recovered mechanisms beyond reactor-value prediction.In CausaLab, the hidden SCM is both the episode ground truth and the evaluation target.
  • Prior LLM evaluations: Most LLM causal evaluations are static, covering textual questions, given-graph reasoning, cause–effect classification, formal inference, and causal-prior uses.Related uses include edge scoring, causal ordering, and query-efficient discovery.
  • Interactive environments: Interactive causal evaluations include scientific-discovery environments, budgeted graph-discovery games, causal games, intervention planners, and ReAct-style agent scaffolds.These settings evaluate sequential interaction rather than only one-shot answers.
  • CausaLab’s evaluation target: Auto-Bench primarily tests hidden-DAG discovery through interaction, whereas CausaLab tests whether a discovered mechanism transfers to a held-out reactor generated by the same SCM.CausaLab exposes graph, frequency-equation, and coefficient commitments through per-step DSL hypotheses.
  • CausaLab’s evaluation target: CausaLab combines explicit SCM recovery with sequential experiment design and scores both final-task utility and recovered-mechanism faithfulness using known SCMs and logged interaction traces.Each episode provides a ground-truth SCM and records the agent’s interaction trace.

3. The Construction of CausaLab

CausaLab constructs episodes as hidden-SCM transfer problems in which agents use finite-budget experiments on a manipulator crystal to predict a held-out reactor frequency and submit a causal mechanism hypothesis. The benchmark separates task success from mechanism recovery by evaluating both frequency prediction and fidelity to the ground-truth graph and structural equations.

  • Task formulation: Each episode hides the causal graph, structural equations, and coefficients, while providing prior records, a manipulator crystal for intervention, and a reactor crystal with held-out frequency.The two crystals share the same sampled SCM, requiring the agent to infer and transfer the common mechanism.
  • Structural causal model: The hidden SCM is instantiated as a DAG over observable properties and frequency, with optional unobserved disturbances resampled after interventions and propagated downstream.The disturbance is neither named nor directly observable or controllable; agents see only its downstream effects.
  • Interaction protocol: Agents repeatedly alternate between choosing a controllable manipulator-property intervention, observing the resulting measurements, and revising a DSL causal hypothesis.The reactor crystal is observed but not intervened on, and the DSL encodes the proposed graph, frequency equation, and coefficients.
  • Intervention semantics: Shift-style interventions replace a controllable variable’s base term while retaining incoming parent contributions, unlike hard do(X=v) interventions that sever upstream influence.This semantics models laboratory control that shifts a baseline while preserving upstream dependencies across sequential interventions.
  • Evaluation targets: Evaluation measures reactor-frequency accuracy separately from mechanism recovery using graph precision, recall, F1, SHD, and coefficient F1 against the ground-truth SCM.SHD counts missing, extra, and reversed directed edges, with lower values indicating closer graph recovery.

4. A DSL for Causal Trajectories

CausaLab introduces a domain-specific language that records causal commitments at every interaction step and converts them into parsable SCM artifacts. This enables trajectory-level scoring of recovered graphs and mechanisms, rather than relying only on final-answer accuracy.

  • Motivation: The DSL addresses the inability of final-answer accuracy to distinguish guessing from transferable mechanism discovery.It records per-step causal commitments and converts hypotheses into SCM artifacts for trajectory-level scoring.
  • DSL records: Each interaction record contains Memory, Thought, accumulated Past data, Hypothesis, and planned Experiment fields.Only the Hypothesis is scored as a causal artifact, specifying hypothesized edges, a frequency structural equation, and associated coefficients.
  • Parsable hypotheses: A fixed schema makes each hypothesis parsable through typed directed edges, a frequency structural equation, and numeric equation coefficients.A deterministic parser converts valid hypotheses into candidate graphs and target mechanisms, producing a trajectory of graph-mechanism pairs.
  • Trajectory scoring: Trajectory scoring evaluates each committed mechanism using the same graph, root, and coefficient metrics used for final evaluation.This avoids relying solely on the agent’s final numeric answer.

5. Experiments

CausaLab’s experiments separate endpoint prediction from faithful SCM recovery across graph sizes, interaction regimes, mechanisms, and model families. They show that mixed online intervention improves graph fidelity, larger models help unevenly, and premature commitment drives many failures.

  • Setup: The main suite evaluates four models on 3–7-node graph families, scoring reactor-frequency prediction separately from graph, parent, root, edge, and coefficient recovery.The reactor crystal’s hidden frequency is the target in every episode, while mechanism metrics are computed against the full episode SCM.
  • Prediction versus mechanism recovery: 48%→24%: replacing a linear mechanism with a hard-quadratic one sharply reduces GPT-5-mini accuracy, while root-node F1 rises from 0.559 to 0.829 and frequency-weight F1 falls from 0.589 to 0.251.The matched comparison fixes 50 four-node topologies and changes only the functional form, showing that graph recovery can persist despite mechanism-fitting failure.
  • Interaction regimes: Mixed observation-conditioned intervention consistently recovers more faithful graphs than pure observation at comparable or better task accuracy on 4-node and 6-node families.Observation narrows the hypothesis space, while agent-chosen interventions provide more faithful graph recovery.
  • Interaction regimes: 90% versus 48% on 4-node graphs and 44% versus 24% on 6-node graphs: Golden intervention traces improve task accuracy but reduce all-edge F1 relative to main-suite baselines.The result indicates that offline intervention chains act mostly as stronger observations rather than substituting for online experimental choice.
  • Model and scale effects: GPT-5.2-high is strongest overall, achieving the best endpoint accuracy and lowest directed all-edge SHD at every graph size, while scaling mainly improves mediated structure and quantitative mechanism fitting.Thinking improves Qwen3.5’s structure recovery over most graph sizes but does not close its gap to the strongest GPT model.
  • Failure analysis: Many failures reflect early overconfident commitment: successful runs usually produce hypotheses consistent with collected data, whereas failed runs stop before checking whether their SCM explains observed outcomes.Across 4-node and 5-node experiments, both successful and failed runs leave about half of the intervention budget unused, and additional budget yields only modest gains.

6. Discussion and Conclusion

CausaLab is a controlled stress test for interactive causal discovery within a deliberately limited synthetic scope, rather than a broad claim about causal reasoning in arbitrary real-world systems.

  • Scope and limitations: CausaLab evaluates interactive causal discovery in synthetic structural causal models with 3–7 nodes.The benchmark’s scope is explicitly limited to synthetic 3–7 node SCMs.
  • Scope and limitations: The benchmark primarily uses linear mechanisms, with one quadratic family, and covers a limited set of model families.These design choices constrain how broadly results should be generalized.
  • Scope and limitations: Its interventions are shift-style adjustments through the laboratory interface rather than perfect hard-do operations.The intervention setting does not implement idealized hard interventions.

7. Potential Risks

CausaLab is a synthetic benchmark and should not be treated as evidence that agents are ready for real scientific, medical, or policy deployment. Its main risk is overgeneralizing success on templated SCM tasks to high-stakes settings with real interventions and domain constraints.

  • Deployment risks: CausaLab’s synthetic results do not establish readiness for real scientific, medical, or policy deployment.The benchmark’s findings should not be interpreted as deployment evidence.
  • Deployment risks: The main risk is overgeneralizing success on templated SCM tasks to high-stakes settings with real interventions and domain constraints.Performance on synthetic causal mechanisms may not transfer to consequential real-world environments.

8. Limitations

CausaLab is a controlled synthetic benchmark whose current suites cover only a limited range of causal mechanisms and experimental settings. Its findings may differ across models, interfaces, decoding policies, exploration budgets, and trajectory-level analyses.

  • Benchmark scope: CausaLab’s 3–7-node suites use mostly linear mechanisms plus one hard-quadratic extension, limiting coverage of real-world causal structures and functional forms.The benchmark does not span the full range of latent variables or measurement processes found in real scientific domains.
  • Evaluation scope: The experiments cover a finite set of LLM agents, prompts, and interaction budgets, so performance may differ under other configurations.Potentially influential alternatives include other models, tool interfaces, decoding policies, and longer exploration budgets.
  • Evaluation scope: The main analyses emphasize final predictions and hypotheses rather than every possible trajectory-level diagnostic.This leaves trajectory-level behavior less fully characterized within the reported analyses.

A Appendix 16

Appendix 16 documents the benchmark, SCM and hidden-disturbance setup, artifacts, prompts, DSL implementation, visualizations, robustness controls, scaling analyses, and early-commitment diagnostics.

  • Appendix 16: The appendix covers benchmark setup and causal-discovery context, SCM and hidden-disturbance details, and artifact and implementation details.
  • Appendix 16: It includes prompt templates, DSL implementation details, and trajectory-level DSL visualization.
  • Appendix 16: It reports mechanism robustness and perturbation controls alongside intervention-trace controls.
  • Appendix 16: It addresses model-family and graph-size scaling, observation–intervention scaling, and early-commitment diagnostics.

A. Appendix · A.1. Benchmark Setup and Causal-Discovery Context

The appendix specifies CausaLab’s benchmark budgets and graph-family coverage, while situating the benchmark within classical causal-discovery research. It identifies CausaLab as closest to the shift-intervention regime, based on the supplied context.

  • A.1. Benchmark Setup and Causal-Discovery Context: The benchmark uses a fixed budget policy except in explicit observation–intervention scaling suites.The passage does not specify the exception’s altered policy beyond naming the scaling suites.
  • A.1. Benchmark Setup and Causal-Discovery Context: For a k-node graph, the observation budget is 2.This is the stated benchmark budget allocation for observations.
  • A.1. Benchmark Setup and Causal-Discovery Context: For a k-node graph, the intervention budget is 4(k −1).The intervention budget is given as a function of graph size.
  • A.1. Benchmark Setup and Causal-Discovery Context: The graph families span 3–7 node SCMs.This defines the range of structural causal model sizes used by the benchmark.
  • A.1. Benchmark Setup and Causal-Discovery Context: Table 1 summarizes the topology distribution used in the main experiments.The supplied passage identifies Table 1 as the summary of topology frequencies or categories.
  • A.1. Benchmark Setup and Causal-Discovery Context: Classical causal discovery learns causal structure from observational, interventional, or shifted data.The cited literature includes constraint- and score-based discovery, interventional Markov equivalence, multiple contexts, and unknown shift interventions.
  • A.1. Benchmark Setup and Causal-Discovery Context: CausaLab is closest to the shift-intervention regime of classical causal-discovery literature.The supplied passage explicitly makes this comparison but is truncated before stating the full distinction.

A.2. SCM and Hidden-Disturbance Details … A.11. Early-Commitment Diagnostics

CausaLab formalizes interactive causal discovery as recovery of a hidden SCM through observations and constrained interventions, with auditable hypothesis trajectories. Its supplementary analyses show that endpoint prediction and faithful mechanism recovery diverge, and that mixed interaction is especially important for graph recovery.

  • A.2. SCM and Hidden-Disturbance Details: CausaLab episodes instantiate SCMs with endogenous observables and frequency, exogenous disturbances, and structural equations mapping parent variables to values.Hidden disturbances may be resampled after interventions and applied as fixed-weight shifts to designated observable variables.
  • A.3. Artifact and Implementation Details: The benchmark extends DiscoveryWorld with new synthetic SCM generation, intervention mechanics, DSL traces, and scoring artifacts.Its laboratory variables are generated from fixed templates and contain no demographic attributes.
  • A.4. Prompt Templates: The two-stage prompts separately govern iterative hypothesis–experiment interaction and reactor activation, requiring causal equations rather than curve-fitting.Agents must preserve accumulated evidence and hypotheses, use interventions when available, and provide a complete frequency equation before reactor operation.
  • A.6. Trajectory-Level DSL Visualization: The DSL records make evolving causal hypotheses auditable by exposing ground-truth and agent graphs alongside recovery metrics over interaction sequences.Figure 7 presents this trajectory-level view rather than inferring mechanism recovery only from the final reactor prediction.
  • A.7. Mechanism Robustness and Perturbation Controls: Hidden-noise and target-family perturbations distinguish mechanism robustness from task performance: ordinary hidden perturbations mainly reduce graph F1, whereas frequency-target perturbations sharply reduce task accuracy.The hard-quadratic comparison similarly localizes degradation to recovered frequency mechanisms rather than root discovery.
  • A.8. Intervention-Trace Controls: Low-MEC intervention traces greatly improve endpoint prediction without producing matching gains in all-edge recovery.The Golden follow-up therefore tests whether better intervention traces alone suffice for faithful SCM recovery.
  • A.9. Model Family and Graph-Size Scaling: High task success does not guarantee causal disambiguation: even on simple 3–4 node settings, final IMEC remains above one, while model-family and graph-size sweeps expose capability gaps.The appendix combines GPT and Qwen results across 3–7 node graphs and reports that high accuracy need not identify a singleton SCM.
  • A.10. Observation–Intervention Scaling: Mixed observation–intervention budgets consistently lift explicit graph recovery, whereas observation-only budgets often improve endpoint prediction without certifying full causal recovery.Even GPT-5.2-high’s strong observation-only endpoint performance leaves graph metrics dependent on mixed interaction.
Loading 2605.26029v2…