Research map / Probabilistic and scientific modeling / Causal discovery
Externalized CPDAG Summaries Improve LLM Causal Deduction
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 Structured Thinking, a two-turn pipeline where the model first emits a schema-constrained CPDAG summary and then answers Corr2Cause causal deduction queries against it, improving over a strong PC-instruction direct-answer baseline.
Problem
Corr2Cause requires deciding whether a causal claim holds in every DAG in the Markov equivalence class consistent with given correlations and conditional independencies; free-form chain-of-thought poorly tracks this latent CPDAG object.
Approach
Externalize the latent CPDAG as a typed, schema-constrained object (via constrained decoding or tool calling) in Turn 1, then answer the hypothesis by querying that materialized graph in Turn 2, with no fine-tuning or symbolic solver.
Key results
- On Corr2Cause ID, Structured Thinking raises Qwen3.5-27B F1(Yes) from 73.01 to 86.36 over BaselineEnhanced (+13.35 pp; McNemar p=2.4e-6); across three full-ID seeds mean gain is +8.08 pp.
- A PC-scaffolded two-turn prose control reaches only 67.55 F1(Yes) on Qwen3.5-27B ID.
- Scrambling the full emitted CPDAG costs 12.0 pp F1; the audit shows skeleton F1 0.960 and exact CPDAG match 75.9% on ID.
Surprising findings
- Removing the PC-algorithm prompt collapses performance to 36.2 F1, far below baseline, showing causal-discovery content is required.
- Thinking tokens are harmful at 9B scale while helpful at 27B; strict tool forcing on GPT-5.4-mini compressed reasoning and dropped F1 by 10.7 pp.
Limitations
- Graph-consistency violations occur in about 18-21% of outputs; schema does not enforce acyclicity or Meek closure.
- Most cross-model rows, Paraphrase-OOD, and controls are single-seed; the gains are limited to CPDAG-style deduction.
- Ablations are one-factor removals and bundle schema, field names, and query instructions, so mechanism is not fully decomposed.
Experimental setup
Datasets: Corr2Cause test split (ID, n=1162) Paraphrase-OOD (n=2246)
Benchmarks: Corr2Cause ID test Paraphrase-OOD larger-graph stress tests (Random n=7, ALARM n=7, ASIA n=8)
Metrics: F1 (Yes) accuracy precision recall skeleton F1 v-structure F1 directed-edge F1 exact CPDAG match
Baselines: BaselineEnhanced (PC-instruction one-turn) TwoTurn-Prose-PC SingleTurn-StructPrompt
Builds on: Corr2Cause PC algorithm Meek rules xgrammar
Closest papers by meaning
- Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.87)
- Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.86)
- Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning (NeurIPS 2025 · similarity 0.85)
- Causal Abstraction Inference under Lossy Representations (ICML 2025 · similarity 0.85)
- What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoT (ICML 2026 · similarity 0.85)
- Thought Branches: Interpreting LLM Reasoning Requires Resampling (ICLR 2026 · 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.