Research map / Large language models / Agents, code and math
Discovery of the Hidden World with Large Language 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.
Proposes COAT, a framework where LLMs propose and annotate high-level factors from unstructured data while causal discovery (FCI) finds Markov blanket structure and feeds back hard-to-explain samples to refine factor proposals.
Problem
Causal discovery usually needs well-defined high-level variables given by experts, which are often unavailable for unstructured data such as reviews, clinical notes, and images.
Approach
Use LLMs to propose and annotate high-level factors from raw observations, run a causal discovery algorithm (FCI) over the factors to recover the target's Markov blanket, and construct feedback from poorly explained sample clusters to prompt new factor proposals in later rounds.
Key results
- On AppleGastronome with GPT-4, COAT reaches F1 0.85 for factor proposal versus 0.49 for META and 0.72 for DATA.
- Causal relation extraction on AppleGastronome: COAT with GPT-4o achieves SHD 1.00 and F1 0.89 versus 5.33 SHD and 0.62 F1 for pairwise LLM reasoning.
- On Neuropathic with GPT-4, COAT attains accuracy 0.96 and F1 0.80 on recovering causal ancestors, outperforming META and DATA baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add causal-discovery feedback across rounds instead of using DATA's single round (ablation) | GPT 4 factor proposal on AppleGastronome · AppleGastronome | F1: 0.72±0.04 -> 0.85±0.11 | Helped |
| Use COAT rather than direct pairwise LLM causal-relation reasoning | GPT 4O causal relation recovery · AppleGastronome | F1: 0.62±0.07 -> 0.89±0.02 | Helped |
| Use COAT rather than direct pairwise LLM causal-relation reasoning | GPT 4O causal relation recovery · AppleGastronome | SHD: 5.33±2.62 -> 1.00±0.00 | Helped |
| Add causal-discovery feedback across rounds instead of using DATA's single round (ablation) | GPT 4 factor proposal on Neuropathic · Neuropathic | F1: 0.50 -> 0.80 | Helped |
| Use COAT rather than DATA+COT | GPT 4 factor proposal on Neuropathic · Neuropathic | Accuracy: 0.81 -> 0.96 | Helped |
Surprising findings
- Directly prompting LLMs to reason about causal relations produces many false-positive edges, while COAT's FCI-based recovery is much more accurate.
- Even the weakest backbone (LLaMA2-7B) in COAT beats baselines that use stronger LLMs on Neuropathic.
Limitations
- Relies on FCI assumptions (faithfulness, no selection bias, sufficient sample size) and FCI cannot orient some edges.
- LLM-proposed factors may be multicollinear or redundant, requiring PCA-based filtering.
- Evaluated mainly on a single synthetic benchmark (AppleGastronome) generated by GPT-4 and focused on one target variable.
Experimental setup
Datasets: AppleGastronome Neuropathic brain tumor MRI NYT stock news NOAA climate reanalysis (ENSO)
Benchmarks: AppleGastronome Neuropathic brain tumor stock news ENSO
Metrics: recall precision F1 SHD SID accuracy MB/NMB/OT counts
Baselines: META DATA DATA+COT PAIRWISE
Builds on: FCI Kiciman et al. LLM causal discovery chain-of-thought prompting
Closest papers by meaning
- Revealing Multimodal Causality with Large Language Models (NeurIPS 2025 · similarity 0.87)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.86)
- Smoke and Mirrors in Causal Downstream Tasks (NeurIPS 2024 · similarity 0.86)
- COLD: Causal reasOning in cLosed Daily activities (NeurIPS 2024 · similarity 0.85)
- Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.85)
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.85)
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.