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

Mitigating Hallucinations in Large Language Models via Causal Reasoning

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 1 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 CDCR-SFT, an SFT framework that teaches LLMs to build variable-level causal DAGs and reason over them, plus the 25,368-sample CausalDR dataset built from CLADDER with DeepSeek-R1, reporting gains on CLADDER, WIQA and HaluEval.

Problem

LLMs produce logically inconsistent hallucinations, and existing reasoning methods (CoT, ToT, GoT) model token or semantic dependencies rather than variable-level causal structure, so they cannot represent conditional independencies or causal identification.

Approach

Fine-tune LLMs with LoRA SFT to first output an explicit variable-level causal DAG and then produce a reasoning trace over that graph, using DAG-annotated data generated by DeepSeek-R1 from CLADDER and augmented by node/edge permutation.

Key results

• On CLADDER with Llama-3.1-8B, CDCR-SFT reaches 95.33% overall accuracy vs 72.88% for the best baseline (CoT-SC), per the paper's table.
• On HaluEval with Llama-3.1-8B, overall accuracy rises to 54.93% vs 51.73% for the best baseline (CausalCoT); the abstract cites roughly 10% improvement.
• Ablation: CDCR-SFT-Ablated (QA-only fine-tuning) drops below the baseline on WIQA and HaluEval for Llama-3.1-8B, while full CDCR-SFT improves on both.

Surprising findings

• Fine-tuning on causal QA pairs without DAG construction lowers WIQA and HaluEval accuracy versus the best prompting baseline.
• DeepSeek-R1-Distill-Llama-8B gains only 1.3-ish points on overall CLADDER-style DAG quality (8.97 to 9.17), suggesting the base distillation already yields good graphs.

Limitations

• Training and evaluation both rely on CLADDER-derived, synthetic-style causal stories, limiting generality.
• Hallucination evaluation on HaluEval is indirect; no hallucination-focused supervision is used, and absolute HaluEval accuracies remain near 50%.
• DAG quality is scored by GPT-4o-mini rather than a deterministic or human metric.

Experimental setup

Datasets: CLADDER Alpaca WIQA HaluEval

Benchmarks: CLADDER WIQA HaluEval

Metrics: accuracy causal DAG node score causal DAG edge score causal DAG structural score

Baselines: CoT CoT-SC CausalCoT ToT GoT DeepSeek-R1-Distill-Llama-8B

Builds on: Chain-of-Thought Tree-of-Thoughts Graph-of-Thoughts CLADDER CausalCoT LoRA

Closest papers by meaning

• Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.91)
• Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.91)
• Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? (NeurIPS 2024 · similarity 0.90)
• Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.90)
• COPO: Causal-Oriented Policy Optimization for Hallucinations of MLLMs (CVPR 2026 · similarity 0.90)
• Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-Based Retrofitting (AAAI 2024 · similarity 0.90)