Faithful by Design: Improving Large Language Model Rationales through Counterfactual Consistency Verification
Bibliographic record
Abstract
Rationale generation in large language models faces critical challenges in ensuring faithfulness and interpretability, limiting their deployment in high-stakes applications requiring transparent reasoning. This study introduces a novel counterfactual verification framework designed to enhance rationale quality across diverse reasoning tasks and model architectures. We evaluate our approach using three state-of-the-art models—GPT- 4o, Claude-3.5-Sonnet, and LLaMA-3.3-70B—across benchmark datasets representing natural language inference (e-SNLI), multi-hop question answering (HotpotQA), and reading comprehension (MultiRC). The evaluation uses four metrics: counterfactual consistency scores, deletion-AUC for sufficiency analysis, perturbation robustness, and human simulatability assessments. Results show improvements in counterfactual consistency (median improvement = 0.086, p ¡ 0.001, Wilcoxon signed-rank test, N = 6,000), with effect sizes ranging from small-to-medium for computational metrics to very large for human performance measures. Human simulatability studies show a 14-point improvement in prediction accuracy (Cohen’s d = 3.50, p ¡ 0.001). Optimal rationale length is 15-35 tokens. Deletion curves indicate better information organization and sufficiency. GPT-4o outperforms other models, with Claude-3.5-Sonnet balancing efficiency and quality. Counterfactual verification enhances AI explainability by improving computational faithfulness and human comprehension, providing a reliable basis for trustworthy reasoning systems in critical applications.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.018 | 0.095 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.003 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.003 |
| Scholarly communication | 0.004 | 0.006 |
| Open science | 0.004 | 0.005 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".