Research map / Large language models / Language, safety and interpretability
Nuance Matters: Probing Epistemic Consistency in Causal Reasoning
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.
Introduces causal epistemic consistency and three metrics (intensity ranking concordance, cross-group position, intra-group clustering) to test whether LLMs consistently rank the polarity and intensity of their own generated causal intermediates, evaluating 21 LLMs and internal token probabilities as an alternative to prompting.
Problem
LLMs may generate fine-grained causal intermediates (supporting or defeating the cause-effect link with varying intensity) but fail to consistently rank or differentiate their own outputs by polarity and intensity; prior causal reasoning work ignores these nuances.
Approach
Measures self-consistency between an LLM's generation of ordered supporting/defeating intermediates and its subsequent ranking of those same intermediates, using Kendall-tau concordance, a cross-group position metric, and a silhouette-based clustering metric; also tests conditional token probability of conjunction-linked causal sentences as a non-prompting alternative.
Key results
- GPT-4o achieves the best closed-source consistency scores (τ-all 0.637, CGP 0.964, IGC 0.978), while LLaMA3-70B is the best open-source model (τ-all 0.586, CGP 0.887, IGC 0.923).
- Intensity consistency (τ-A, τ-D) is markedly harder than polarity and clustering; several small open models (Gemma-2B, LLaMA2-7B) score near random on τ metrics.
- Larger models within a family generally score higher, and coordinating/conjunctive-adverb connectives ("so", "therefore") outperform subordinate ones ("because", "since", "as") in the token-probability setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use internal token probabilities instead of prompting for ranking | Gemma and LLaMA2-7B/13B · δ-CAUSAL test | causal epistemic consistency | Helped |
| Use internal token probabilities instead of prompting for ranking | Open-source models overall · δ-CAUSAL test | causal epistemic consistency | Mixed |
| Use coordinating conjunction “so” or conjunctive adverbs “therefore,” “thus,” or “hence” rather than subordinate conjunctions | Internal token probability ranking · δ-CAUSAL test | causal epistemic consistency | Helped |
| Increase model size | Gemma, LLaMA2, Phi-3, and LLaMA3 · δ-CAUSAL test | causal epistemic consistency | Helped |
| Compare prompting with internal conditional token probability | Open-source LLMs · δ-CAUSAL test | τ-all, CGP, and IGC | Mixed |
Surprising findings
- Subordinate conjunctions such as "because" and "since" underperform in token-probability causal scoring, which the authors attribute to poor alignment with natural LLM sentence patterns.
- Gemma and mid-sized LLaMA2 models (7B, 13B) perform better with internal token probabilities than with prompting.
Limitations
- Closed-source models do not expose arbitrary-token log-probabilities, so the token-probability analysis is restricted to open-source models.
- Evaluation relies on a single foundational dataset (δ-CAUSAL test set).
Experimental setup
Datasets: δ-CAUSAL test set (1,970 defeasible cause-effect pairs)
Benchmarks: δ-CAUSAL
Metrics: Kendall tau (τ-A, τ-D, τ-all) cross-group position (CGP) intra-group clustering (IGC)
Baselines: Random
Builds on: δ-CAUSAL silhouette score
Closest papers by meaning
- Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? (NeurIPS 2024 · similarity 0.91)
- COLD: Causal reasOning in cLosed Daily activities (NeurIPS 2024 · similarity 0.89)
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.89)
- Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.89)
- Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.89)
- Compositional Causal Reasoning Evaluation in Language Models (ICML 2025 · similarity 0.88)
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.