ICE: Intervention-Consistent Explanation Evaluation

Official implementation of "ICE: Intervention-Consistent Explanation Evaluation with Statistical Grounding for LLMs" (EMNLP 2026 Findings).

Key Finding

Faithfulness is operator-dependent: switching the intervention operator crosses the positive-evidence threshold in 18% of configurations (5/28), with gaps up to 44 percentage points. Nearly one-third of English deletion configurations are anti-faithful (18/56, worse than random), invisible without randomized baselines.

What ICE Does

ICE evaluates whether model explanations (attention, gradient) identify tokens that are genuinely more important than random tokens, under multiple intervention operators.

Input: "a gorgeous, witty, seductive movie"   (positive sentiment)
Rationale: {gorgeous, seductive}

               Deletion    Retrieval Infill
NSR:           0.79        0.39
Win Rate:      92%         68%
Verdict:       Faithful    Weakly Faithful   <-- operator changes conclusion

Installation

pip install ice-faithfulness

Quick Start

from ice import ICEEvaluator, ICEConfig

config = ICEConfig(
    k_values=[0.2],          # top 20% tokens as rationale
    n_permutations=50,        # random baselines
    operators="lite",         # "lite" (delete+retrieval) or "full"
)

evaluator = ICEEvaluator(model, tokenizer, config)
results = evaluator.evaluate_dataset(dataset, extractor="attention")

print(f"Win Rate: {results.win_rate:.1%}")
print(f"Effect Size (d_null): {results.effect_size:.2f}")

Causal LMs: restrict scoring to label tokens

For prompted classification with causal LMs (Llama, Mistral, Qwen, ...), restrict prediction and scoring to the task's label tokens via candidate_token_ids. Without this, argmax runs over the full vocabulary and most examples are filtered out by the confidence threshold.

# For sentiment classification with causal LMs
pos_id = tokenizer.encode(" positive", add_special_tokens=False)[0]
neg_id = tokenizer.encode(" negative", add_special_tokens=False)[0]

config = ICEConfig(
    k_values=[0.2],
    operators="lite",
    candidate_token_ids=[neg_id, pos_id],  # restrict to task labels
)

Core Modules

Evaluation Scripts

# English benchmarks (7 models x 4 tasks)
python scripts_for_repo/run_ice_llm.py --model meta-llama/Llama-3.2-3B-Instruct --dataset sst2

# Multilingual (6 languages)
python scripts_for_repo/run_ice_multilingual.py --model Qwen/Qwen2.5-7B-Instruct --languages french german turkish arabic

# Operator comparison
python scripts_for_repo/run_ice_llm_retrieval.py --model mistralai/Mistral-7B-Instruct-v0.3 --dataset esnli

Key Metrics

Results Summary

Evaluated on 7 LLMs (1.5B-8B), 4 English tasks, 6 non-English languages, 2 attribution methods:

• Operator gaps reach 44 pp (e.g., Llama-3.2 e-SNLI: 86% deletion vs 43% retrieval)
• Anti-faithfulness in nearly 1/3 of English deletion configurations (18/56; gradient selects function words)
• No correlation between faithfulness and human plausibility (|r| < 0.04)
• Cross-lingual faithfulness spans 12%-83%, not predicted by tokenization alone

Citation

@inproceedings{basu2026ice,
    title={ICE: Intervention-Consistent Explanation Evaluation with Statistical Grounding for LLMs},
    author={Basu, Abhinaba and Chakraborty, Pavan},
    booktitle={Findings of the Association for Computational Linguistics: EMNLP 2026},
    year={2026}
}

License

MIT