Research map / Large language models / Language, safety and interpretability

Causal Prompting: Debiasing Large Language Model Prompting Based on Front-Door Adjustment

AAAI 2025 · Large language models › Language, safety and interpretability · approach: In-context learning analysis · 18 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 Causal Prompting, which applies front-door adjustment with CoT as mediator, combining CoT-SC clustering and an NWGM-style ICL intervention, with a contrastively fine-tuned BERT encoder to debias LLM prompting.

Problem

LLM prompting methods (ICL, CoT) suffer from biases such as reliance on spurious text spans, and standard debiasing (data augmentation, reweighting, counterfactual/back-door adjustment) is hard to apply to prompting because the confounder is unobservable and logits are often inaccessible.

Approach

Uses a structural causal model with the chain-of-thought as mediator and applies front-door adjustment, decomposing P(A|do(X)) into CoT-SC-style cluster probabilities P(r|do(X)) and ICL demonstration-based interventions P(A|do(r)), so no logits or confounder values are needed; a BERT encoder is contrastively aligned to the LLM's CoT semantics.

Key results

• On LLaMA3, Causal Prompting reaches 87.95% on GSM8K vs 82.41% for CoT-SC, and 62.76% on MATH vs 56.61%.
• On GPT-3.5, Causal Prompting reaches 85.44% on GSM8K vs 80.21% for CoT-SC and 58.20 EM on HotpotQA vs 41.60.
• On the LLaMA3 adversarial (OOD) splits, Causal Prompting improves ABSA Adv to 78.69 vs 73.53 for CoT-SC.

Measured effects reported in the paper

Surprising findings

• Applying front-door adjustment to all samples (s=1.0) degrades NLU performance, attributed to wrong/correct CoT demonstrations with opposite labels causing answer flips.
• On the original ABSA split, Causal Prompting underperforms CoT-SC on LLaMA3 (79.78 vs 80.56), while gaining on the adversarial split.

Limitations

• Assumes a single confounder and ignores confounding of the CoT with other variables, so new biases from generated CoTs are not addressed.
• Substantially higher inference cost than CoT-SC (multiple CoTs, clustering, and K*T further queries).
• Evaluated on only seven datasets and three LLMs; datasets were subsampled and MATH restricted to algebra.

Experimental setup

Datasets: GSM8K MATH HotpotQA MuSiQue SemEval2014-Laptop MNLI FEVER Symmetric FEVER TextFlint

Benchmarks: GSM8K MATH HotpotQA MuSiQue ABSA NLI FEVER

Metrics: accuracy exact match F1

Baselines: Standard ICL CoT CoT-SC

Builds on: chain-of-thought prompting self-consistency front-door adjustment normalized weighted geometric mean approximation InfoNCE contrastive learning KATE

Closest papers by meaning

• Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.90)
• Evaluating the Efficacy of Prompting Techniques for Debiasing Language Model Outputs (Student Abstract) (AAAI 2024 · similarity 0.89)
• Eliciting Causal Abilities in Large Language Models for Reasoning Tasks (AAAI 2025 · similarity 0.88)
• Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales? (NeurIPS 2024 · similarity 0.88)
• Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.87)
• Rethinking the Development of Large Language Models from the Causal Perspective: A Legal Text Prediction Case Study (AAAI 2024 · similarity 0.86)