Interventional Training: Letting Models Conduct Their Own Experiments

The one-sentence version

The way scientists distinguish correlation from causation is by conducting experiments: actively changing one variable to see how others change. Currently, Large Language Models (LLMs) learn almost exclusively from "observational" data (text written by others) and have never personally conducted an experiment. The concept proposed here is: turn the scientific cycle of "propose a hypothesis → design an intervention → execute → update beliefs based on results" into a part of LLM training. The reward is not "getting the answer right," but rather "did your experiment make you more certain about the causal structure?"

Positioning

Research concept. Distinguishes between existing evidence, reasonable speculation, and hypotheses to be verified.

Layer 1: The Intuition

1.1 Ice Cream and Drowning

In the summer, ice cream sales and the number of drownings both rise. Looking only at the data, you might think eating ice cream causes drowning.

A scientist would think: If I force a city to stop selling ice cream, will drownings decrease? No, because both are caused by "hot weather."

This "forced change" is an intervention. Textual data is full of "ice cream and drowning" style correlations; the causal knowledge a model learns from it depends heavily on whether someone happened to write down the correct causal relationship.

1.2 Concept: The Model as a Scientist

Step	What the Model Does
Observation	Reads a batch of data or a description of a system
Hypothesis	Proposes a possible causal graph (who influences whom)
Experimental Design	Selects the intervention that best distinguishes between different hypotheses
Execution	Executes in a simulator, code sandbox, or real-world experimental environment
Update	Refines the causal graph based on the results

The training goal is to make the model increasingly proficient at clarifying causal structures using the minimum number of experiments.

Layer 2: Existing Evidence

Direction	Representative Work
Causal Discovery Algorithms	PC Algorithm (based on conditional independence tests), NOTEARS (2018, turning DAG constraints into continuous optimization); with intervention data, the identifiable causal structures become much more numerous
Optimal Experimental Design / Active Learning	Bayesian Optimal Experimental Design: selecting experiments that maximize "information gain"
LLM's Causal Shortcomings	Benchmarks like CLadder show that LLMs are significantly weaker at intervention and counterfactual questions compared to associative questions (see Post-training for World Models)
LLM as a Research Agent	"AI Scientist" systems that automatically design and execute experiments have emerged: Coscientist (2023), which autonomously plans synthesis experiments in automated chemistry labs, and systems that can automatically propose hypotheses, write code to run experiments, and write papers
Children are Natural Experimenters	Developmental psychology research (e.g., Gopnik et al.) shows that infants test causal hypotheses through selective "play," behaving similarly to Bayesian optimal experimental design

Layer 3: Specific Forms of the Concept

3.1 Reward: Information Gain

Assume the model's belief about the causal structure is a distribution $P(G)$ (where $G$ is a possible causal graph). After executing experiment $e$ and observing result $y$ , the belief is updated to $P(G \mid y, e)$ . The reward is set as the information gain:

$r(e) = H\big[P(G)\big] - H\big[P(G \mid y, e)\big],$

$H$ is entropy. A good experiment makes the model more certain about "what the causal structure of the world is."

This reward does not require a ground truth answer: there is no need for a human to tell the model the correct causal graph; the model only needs its beliefs to become more concentrated and its subsequent predictions to become more accurate (to prevent cheating by "falsely becoming certain," validation via subsequent prediction accuracy is required).

3.2 Environments from Easy to Hard (Reasonable Speculation)

Stage	Environment	Example
1	Synthetic Causal Systems	Small randomly generated causal graphs where the model can intervene on any variable
2	Code sandbox	An unfamiliar program: changing inputs and configurations to infer internal logic
3	Physical Simulation	Building blocks, inclined planes, circuits
4	Scientific Simulation	Chemical reactions, gene regulatory network simulators
5	Real Laboratories	Automated wet-lab platforms (high cost, most authentic data)

3.3 Why It Is Worth Doing

Data Efficiency: Actively selected experiments contain much more information than passively observed data;
Transferable Ability: "How to design an experiment" is a general skill that may be more valuable than "memorizing ten thousand causal facts";
Reducing Hallucination: A model accustomed to "testing before believing" may be less likely to mistake correlation for causation.

3.4 Hypotheses to be Verified

Can the "experimental strategies" learned in synthetic causal systems be transferred to real-world problems described in natural language?
Will information gain rewards be subject to speculation (e.g., the model intentionally becoming very uncertain first to "gain" a large amount of information gain)? What constraints are needed?
Will the LLM's linguistic priors help (by providing reasonable initial hypotheses) or cause trouble (by stubbornly adhering to incorrect correlations in the text)?
Computational cost: Since causal beliefs must be maintained and updated at every step, can this be approximated using the model's own hidden states?

3.5 Risks

Autonomous experimental agents in real laboratories involve safety issues (hazardous chemicals, biological experiments) and must have strict permissions and human oversight (see Agent Permission Design).

Summary

Layer	Key points
Problem	The model has only "seen" the world, never "moved" the world
Concept	Train the model to propose hypotheses, design interventions, and update based on results
Reward	Information gain regarding the causal structure, paired with subsequent prediction accuracy validation
Basis	Causal discovery, optimal experimental design, AI scientist systems, and children's experimental learning