Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Passive learning of active causal strategies in agents and language models
Andrew Kyle Lampinen, Stephanie C.Y. Chan, Ishita Dasgupta, Andrew Joohun Nam, Jane X Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove active interventional exploration phase (skip or observe only) (ablation) | imitation agent · causal DAG environment | reward | Hurt |
| replace skipped exploration with observation-only phase (ablation) | imitation agent · causal DAG environment | reward | Improved |
| reduce training DAG structures from 1277 to 428 (ablation) | imitation agent · held-out causal DAG evaluation | generalization: 1277 -> 428 structures | Hurt |
| reduce training DAG structures from 1277 to 195 (ablation) | imitation agent · held-out causal DAG evaluation | generalization: 1277 -> 195 structures | Hurt |
| train on linear DAGs, evaluate on nonlinear DAGs (ablation) | imitation agent · causal DAG environment | fraction of optimal reward: over 75% of optimal | Mixed |
| add auxiliary language explanation prediction (ablation) | imitation agent · odd-one-out environments | learning and generalization | Improved |
| train to predict explanations focused on one feature on confounded data (ablation) | imitation agent · deconfounded odd-one-out evaluation | OOD choice following explained feature | Improved |
| include explanations and reasoning traces in few-shot prompt (ablation) | Chinchilla 70B · odd-one-out LM, held-out dimension | generalization | Improved |
| use only explanations or only reasoning (not both) in prompt (ablation) | Chinchilla 70B · odd-one-out LM | average performance | Mixed |
| make expert policy in examples switch to exploiting after finding correct dimension (ablation) | Chinchilla 70B · odd-one-out LM, texture held out | generalization | Hurt |
| train on subsets of variables with relevance cue | imitation agent · adaptive experimentation, n=10 variables | optimal exploration on held-out variable sets | Improved |
Closest papers by meaning
- Teaching Transformers Causal Reasoning through Axiomatic Training (ICML 2025 · similarity 0.89)
- ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.86)
- Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? (NeurIPS 2024 · similarity 0.86)
- Language Agents Meet Causality -- Bridging LLMs and Causal World Models (ICLR 2025 · similarity 0.86)
- Eliciting Causal Abilities in Large Language Models for Reasoning Tasks (AAAI 2025 · similarity 0.86)
- Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
