Research map / Large language models / Agents, code and math
Language Agents Meet Causality -- Bridging LLMs and Causal World Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Integrates causal representation learning (BISCUIT-based causal world model with language-grounded actions and state descriptions) with LLM planning via a RAP-style search, outperforming an LLaMA 3-based RAP baseline on causal inference and planning in GridWorld and iTHOR.
Problem
LLMs' common-sense causal knowledge is often incomplete or inapplicable to specific environments, limiting their causal reasoning and planning in interactive settings.
Approach
Learn a causal world model with BISCUIT-style CRL in which actions are encoded from natural-language descriptions and latent causal variables are decoded back into text, so an LLM can query the causal simulator during tree-search planning.
Key results
- Causal world model reaches 0.758 accuracy at 8-step GridWorld inference versus 0.005 for the baseline LM.
- On iTHOR 2-step planning, success rate 0.58 for the causal model vs 0.25 for the baseline LM.
- On GridWorld 8-step planning, success rate 0.42 for the causal model vs 0.06 for the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a causal world model alongside the LLaMA 3 planning agent instead of a baseline LM world model | LLaMA 3 planning agent · iTHOR, 2-step planning | success rate: 0.25 -> 0.58 | Helped |
| Use a causal world model alongside the LLaMA 3 planning agent instead of a baseline LM world model | LLaMA 3 planning agent · iTHOR, 4-step planning | success rate: 0.11 -> 0.44 | Helped |
| Use a causal world model alongside the LLaMA 3 planning agent instead of a baseline LM world model | LLaMA 3 planning agent · GridWorld, 2-step planning | success rate: 0.20 -> 0.95 | Helped |
| Use a causal world model alongside the LLaMA 3 planning agent instead of a baseline LM world model | LLaMA 3 planning agent · GridWorld, 4-step planning | success rate: 0.11 -> 0.73 | Helped |
| Use a causal world model alongside the LLaMA 3 planning agent instead of a baseline LM world model | LLaMA 3 planning agent · GridWorld, 6-step planning | success rate: 0.08 -> 0.46 | Helped |
| Use a causal world model alongside the LLaMA 3 planning agent instead of a baseline LM world model | LLaMA 3 planning agent · GridWorld, 8-step planning | success rate: 0.06 -> 0.42 | Helped |
| Use the causal world model instead of the baseline LM | Causal inference model · iTHOR, 1-step causal inference | accuracy: 0.482 -> 0.824 | Helped |
| Use the causal world model instead of the baseline LM | Causal inference model · iTHOR, 4-step causal inference | accuracy: 0.110 -> 0.630 | Helped |
| Use the causal world model instead of the baseline LM | Causal inference model · GridWorld, 8-step causal inference | accuracy: 0.005 -> 0.758 | Helped |
| Use text-based action representations rather than coordinate-based representations | Causal representation learning in GridWorld · GridWorld, 1.5% training data | R²: 0.548±0.022 -> 0.603±0.003 | Helped |
| Use hybrid coordinate-and-text action representations rather than coordinate-based representations | Causal representation learning in GridWorld · GridWorld, 0.5% training data | R²: 0.366±0.000 -> 0.433±0.001 | Helped |
| Use text-based action representations rather than coordinate-based representations | Causal representation learning in GridWorld · GridWorld, 0.3% training data | R²: 0.392 -> 0.374±0.000 | Hurt |
Surprising findings
- Text-based action encodings match or beat coordinate-based encodings in low-data regimes for causal variable recovery.
- The baseline LM performs worse on GridWorld than on iTHOR, which the authors attribute to lack of visual input.
Limitations
- Experiments limited to two simple environments (8x8 GridWorld and a static iTHOR kitchen).
- Requires a small set of labeled causal variables to map latents to language, and a rule-based state description generator.
- Weak performance on PutObject and PickupObject in iTHOR because positions are modeled as independent binary-like variables.
Experimental setup
Datasets: GridWorld synthetic trajectories iTHOR kitchen trajectories (FloorPlan10)
Benchmarks: GridWorld (custom) iTHOR (custom)
Metrics: R^2 for latent-causal alignment N-step causal inference accuracy planning success rate average steps
Baselines: RAP with LLaMA 3 (8B) world model
Builds on: BISCUIT Reasoning via Planning (RAP) RealNVP CoordConv
Closest papers by meaning
- Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? (NeurIPS 2024 · similarity 0.92)
- COLD: Causal reasOning in cLosed Daily activities (NeurIPS 2024 · similarity 0.89)
- Do LLMs Build World Representations? Probing Through the Lens of State Abstraction (NeurIPS 2024 · similarity 0.88)
- Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.87)
- The Abstraction Gap in Vision-Language Causal Reasoning (ICML 2026 · similarity 0.87)
- Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL (NeurIPS 2025 · similarity 0.87)
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.