ICE: Intervention-Consistent Explanation Evaluation with Statistical Grounding for LLMs
This addresses the challenge of reliably assessing explanation faithfulness in LLMs for researchers and practitioners, though it is incremental as it builds on existing intervention-based methods.
The paper tackled the problem of evaluating whether explanations faithfully reflect a model's reasoning by introducing ICE, a framework that uses multiple intervention operators and statistical tests to compare explanations against random baselines, finding that faithfulness varies significantly with operators and shows no correlation with human plausibility.
Evaluating whether explanations faithfully reflect a model's reasoning remains an open problem. Existing benchmarks use single interventions without statistical testing, making it impossible to distinguish genuine faithfulness from chance-level performance. We introduce ICE (Intervention-Consistent Explanation), a framework that compares explanations against matched random baselines via randomization tests under multiple intervention operators, yielding win rates with confidence intervals. Evaluating 7 LLMs across 4 English tasks, 6 non-English languages, and 2 attribution methods, we find that faithfulness is operator-dependent: operator gaps reach up to 44 percentage points, with deletion typically inflating estimates on short text but the pattern reversing on long text, suggesting that faithfulness should be interpreted comparatively across intervention operators rather than as a single score. Randomized baselines reveal anti-faithfulness in one-third of configurations, and faithfulness shows zero correlation with human plausibility (|r| < 0.04). Multilingual evaluation reveals dramatic model-language interactions not explained by tokenization alone. We release the ICE framework and ICEBench benchmark.