# Externalized CPDAG Summaries Improve LLM Causal Deduction

arXiv: https://arxiv.org/abs/2609.31071 · HTML: https://mlsanity.com/paper/2609.31071/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 23 of its 30 references that are in the map.
- First posted on arXiv 25 Sep 2026.

## Builds on (most cited first, 23 of 23)

- [ReAct: Synergizing Reasoning and Acting in Language Models](https://mlsanity.com/paper/2210.03629/) (2022)
- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [Toolformer: Language Models Can Teach Themselves to Use Tools](https://mlsanity.com/paper/2302.04761/) (2023)
- [Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks](https://mlsanity.com/paper/2211.12588/) (2022)
- [PAL: Program-aided Language Models](https://mlsanity.com/paper/2211.10435/) (2022)
- [Causal Reasoning and Large Language Models: Opening a New Frontier for Causality](https://mlsanity.com/paper/2305.00050/) (2023)
- [Efficient Guided Generation for Large Language Models](https://mlsanity.com/paper/2307.09702/) (2023)
- [TALM: Tool Augmented Language Models](https://mlsanity.com/paper/2205.12255/) (2022)
- [Causal Inference and Causal Explanation with Background Knowledge](https://mlsanity.com/paper/1302.4972/) (2013)
- [Can Large Language Models Infer Causation from Correlation?](https://mlsanity.com/paper/2306.05836/) (2023)
- [Can large language models build causal graphs?](https://mlsanity.com/paper/2303.05279/) (2023)
- [Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?](https://mlsanity.com/paper/2506.21215/) (2025)
- [Causal Inference with Large Language Model: A Survey](https://mlsanity.com/paper/2409.09822/) (2024)
- [Can Foundation Models Talk Causality?](https://mlsanity.com/paper/2206.10591/) (2022)
- [Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning](https://mlsanity.com/paper/2506.09853/) (2025)
- [A Critical Review of Causal Reasoning Benchmarks for Large Language Models](https://mlsanity.com/paper/2407.08029/) (2024)
- [C2P: Featuring Large Language Models with Causal Reasoning](https://mlsanity.com/paper/2407.18069/) (2024)
- [Prompting Strategies for Enabling Large Language Models to Infer Causation from Correlation](https://mlsanity.com/paper/2412.13952/) (2024)
- [C2-Faith: Benchmarking LLM Judges for Causal and Coverage Faithfulness in Chain-of-Thought Reasoning](https://mlsanity.com/paper/2603.05167/) (2026)
- [Causal Reasoning in Pieces: Modular In-Context Learning for Causal Discovery](https://mlsanity.com/paper/2507.23488/) (2025)
- [Structured Thinking Matters: Improving LLMs Generalization in Causal Inference Tasks](https://mlsanity.com/paper/2505.18034/) (2025)
- [Bypassing the Rationale: Causal Auditing of Implicit Reasoning in Language Models](https://mlsanity.com/paper/2602.03994/) (2026)
- [CLEAR-3K: Assessing Causal Explanatory Capabilities in Language Models](https://mlsanity.com/paper/2506.17180/) (2025)

## Abstract (arXiv)

Corr2Cause asks whether a causal claim holds in every DAG compatible with observed correlations and conditional independencies. We frame this as latent-object reasoning: the label is defined by a CPDAG query, but free-form chain-of-thought often collapses the Markov-equivalence-class problem into local pattern matching. We propose Structured Thinking, a two-turn pipeline that first externalizes a typed, schema-constrained CPDAG summary and then answers against that graph state. On the Corr2Cause full test, Structured Thinking raises Qwen3.5-27B from 73.0 to 86.4 F1(Yes) over a strong PC-instruction baseline in the primary paired run (+13.4 pp; McNemar p=2.4× 10-6; bootstrap 95% CI [+8.4, +18.6]); across three full-ID seeds, the mean gain is +8.1 ± 5.3 pp. A PC-scaffolded two-turn prose control reaches only 67.6 F1, indicating that a detailed PC scaffold plus a schema-free prose intermediate is not sufficient. The same pattern holds on Qwen3.6-27B, Paraphrase-OOD, and GPT-5.4-mini. Scrambling the emitted CPDAG costs 12.0 pp F1, and a full-split audit shows close agreement with the reference CPDAG (ID skeleton F1 0.960; exact match 75.9%). These results support a bounded design principle: externalize the latent object that defines the label, constrain its form, and test whether downstream answers use it.