Research map / Large language models / Reasoning and chain of thought

Grounding Multi-Hop Reasoning in Structural Causal Models via Group Relative Policy Optimization

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

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 SCM-GRPO, an SCM-inspired structural dependency reasoning framework for multi-hop fact verification, trained with distillation-based SFT followed by GRPO using a composite correctness/structure/length reward, evaluated on HoVer and EX-FEVER.

Problem

Multi-hop fact verification with LLMs suffers from hallucinations and fractured logical chains; existing CoT methods do not explicitly model directed dependencies between evidence and claims, and unconstrained structured reasoning chains can degrade accuracy.

Approach

Represent reasoning as an SCM-inspired DAG of exogenous evidence variables, endogenous intermediate conclusions and structural functions (used as a structural abstraction, not causal inference), distill it into a student via SFT from a teacher, then apply GRPO with a composite correctness, evidence-vs-inference structure, and length reward to control chain complexity.

Key results

• SCM-GRPO reaches 71.02% average accuracy across HoVer (2/3/4-hop) and EX-FEVER (2/3-hop), the best among listed methods, versus 69.58% for GPT-3.5 Direct Prediction.
• Under the same Qwen3-8B backbone, DP and ProgramFC reproduced at 58.41% and 64.31% average; SCM-GRPO outperforms both.
• Ablation on Qwen3-8B: SFT 73.33% vs SFT+GRPO 75.42% (ablation table) and removing the length reward drops accuracy to 64.33%.

Surprising findings

• Adding SCM-CoT without RL reduces accuracy for several backbones (e.g., Qwen3-14B on HoVer 2-hop drops from 57.48% to 47.22%).
• CPO degrades SFT performance (73.33% to 71.91%).

Limitations

• Dependence on quality of automatically constructed teacher-distilled structured data.
• Generalization to ultra-long reasoning chains (beyond 4 hops) unverified.
• Reported ablation numbers are inconsistent across tables and comparisons with GPT-3.5 baselines are not same-backbone.

Experimental setup

Datasets: HoVer EX-FEVER FEVER

Benchmarks: HoVer EX-FEVER

Metrics: accuracy

Baselines: Direct Prediction ProgramFC FOLK RAG Decompose-Verify QACheck Search-o1 Qwen3-8B SFT CPO

Builds on: GRPO ProgramFC Chain-of-Thought

Closest papers by meaning

• MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · similarity 0.86)
• Causal Walk: Debiasing Multi-Hop Fact Verification with Front-Door Adjustment (AAAI 2024 · similarity 0.86)
• Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace (NeurIPS 2026 · similarity 0.86)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.86)
• Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models (ICLR 2026 · similarity 0.86)
• Verifying Chain-of-Thought Reasoning via Its Computational Graph (ICLR 2026 · similarity 0.86)