Skip to content

A causal world engine for experimental intelligence

Give experimental intelligence its own world engine

Static benchmarks ask what a model knows. ChemWorld asks how it experiments when the answer is hidden.

ChemWorld is a replayable causal world-model environment. Under partial observability, finite budgets, and operational constraints, agents choose operations and measurements, form hypotheses, and revise strategies from evidence. Hidden kinetics, phase behavior, and process rules can change while the public task remains stable, so memorizing one optimal recipe is not enough.

ChemWorld is not a universal real-reaction predictor. It is a research environment for making experimental decision making scalable, comparable, and falsifiable.

Replay-verified trajectories Causal world shifts BO · RL · LLM · World Models

Current evidence

Two flagship tasks now have formal descriptive no-dossier campaigns and paired opaque / nominal / misindexed anonymous-material experiments. Correct information has confirmed positive value in electrochemistry and an inconclusive effect in crystallization. A targeted wrong prior changes early behavior in both tasks, but neither passes the joint recovery rule.

A five-task, five-world development extension is also complete. The shared Codex strategy is above the best classical method mean in electrochemistry and the new reaction–distillation task, but below it in crystallization, partition, and continuous flow. All 15 task contracts are executable; two tasks have formal results and three more now have development comparisons. Historical RC28 Gate A has a stale current-source binding, and Participant Gates B–E remain unexecuted. ChemWorld is therefore a candidate research environment with benchmark_ready=false, not a completed leaderboard release.

Inspect exact results and status →

Why a world engine

Real chemical experiments are slow, costly, and risk-bearing. Static datasets test knowledge and prediction, but not whether an agent selects an informative experiment, interprets failure, manages resources, or adapts when its model is wrong.

Static chemistry benchmark ChemWorld
Answer a given question Decide what experiment to do next
One-shot input and output Repeated observation and action
Fixed data and rules Intervenable hidden world rules
Error lowers a score Error consumes budget and changes state

The central experiment

The same public task can run under different rate laws, reaction topologies, constitutive relations, or equipment boundaries. Agents are not given a world label. They must use experiments to detect which assumptions still hold and recover when the rules change.

Three agent tracks

  • Campaign Design: choose the next complete experiment—BO, safe BO, active learning, recipe-level LLMs.
  • Procedure Execution: choose the next operation—hierarchical RL, state machines, operation-level LLMs.
  • Process Control: choose bounded equipment setpoints and process-control actions—SAC, MPC, system identification, world-model control. This is not a claim of universal high-frequency continuous control.

World-model adaptation cuts across all three: infer the current world from history and recover quickly after a shift.

Start here

Goal Page
Understand the research thesis Why ChemWorld
Define experimental intelligence Experimental Intelligence
Understand changing worlds Causal Worlds
Read the evaluation design Benchmark
Inspect current evidence Research Findings
Understand the real-world path Real-world Bridge

The complete API and local setup reference currently lives in the Chinese technical documentation.

Research status: benchmark candidate. Engine and replay controls are operational; formal cross-method adaptation, private evaluation, and external bridging remain incomplete. RC28 Gate A passed on its frozen source, but its current source binding is stale and current benchmark_ready=false pending recertification.