Research map / Large language models / Instruction and fine-tuning
Towards Unifying Evaluation of Counterfactual Explanations: Leveraging Large Language Models for Human-Centric Assessments
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.
Builds CounterEval, a 30-scenario counterfactual explanation dataset rated by 206 humans on 8 metrics, and fine-tunes Llama 3 models (and zero-shot GPT-4) to predict averaged or individual ratings in three classes.
Problem
Evaluation of counterfactual explanations relies on fragmented quantitative metrics and costly, non-comparable user studies that do not capture human perspectives on explanatory qualities.
Approach
Collect human ratings of diverse counterfactual scenarios across eight explanatory qualities and fine-tune LLMs with QLoRA to predict low/medium/high human judgments, both averaged and per individual participant.
Key results
- Fine-tuned Llama 3 70B Instruct and Llama 3.1 8B reach 85% accuracy on the metric split (3-class prediction).
- Zero-shot GPT-4 reaches 63% accuracy on the metric split, the best zero-shot result.
- Fine-tuning Llama 3 70B on individual participants reaches roughly 87-90% accuracy for three of four participants, but only 66% for one.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on averaged human ratings | Llama 3 70B Instruct · metric split | accuracy: 0.57 -> 0.85 | Helped |
| fine-tune on averaged human ratings | Llama 3 70B Instruct · question split | accuracy: 0.59 -> 0.81 | Helped |
| fine-tune on averaged human ratings | Llama 3 70B Instruct · Trust, question split | accuracy: 0.50 -> 0.50 | No clear effect |
| Use the prompt with all metric definitions instead of the baseline prompt (ablation) | LLM counterfactual evaluator · prompt comparison across four models | accuracy: 0.54 -> 0.54 | No clear effect |
| Add two prompt examples to the prompt with all metric definitions (ablation) | LLM counterfactual evaluator · prompt comparison across four models | accuracy: 0.54 -> 0.51 | Hurt |
| Add all metric definitions to the baseline prompt (ablation) | Llama 3 8B Instruct · preliminary prompt evaluation | accuracy: 0.56 -> 0.63 | Helped |
| Add two examples to the prompt with all metric definitions (ablation) | Llama 3 70B Instruct · preliminary prompt evaluation | accuracy: 0.70 -> 0.75 | Helped |
| Add all metric definitions to the baseline prompt (ablation) | Llama 2 7B Chat · preliminary prompt evaluation | accuracy: 0.46 -> 0.44 | Hurt |
| Add two examples to the prompt with all metric definitions (ablation) | Llama 2 7B Chat · preliminary prompt evaluation | accuracy: 0.44 -> 0.37 | Hurt |
| Fine-tune on one participant's responses | Llama 3 70B Instruct · individual preference evaluation, participant B | accuracy: 0.58 -> 0.66 | Helped |
| Fine-tune on one participant's responses | Llama 3 70B Instruct · individual preference evaluation, participant C | accuracy: 0.69 -> 0.90 | Helped |
Surprising findings
- Feasibility and Trust remained hard to predict even after fine-tuning, attributed to the difficulty of judging whether categorical changes would achieve the outcome.
- Averaged across prompts, the baseline and definition-rich system prompts performed equally (0.54), and adding examples lowered average accuracy (0.51).
Limitations
- Dataset is small: only 30 unique counterfactual explanations.
- Per-participant experiment covers only 4 participants and 30 explanations, so conclusions are tentative.
- Ethical risks of bias and gaming the evaluator when optimizing explanations against the model.
Experimental setup
Datasets: CounterEval (own human-rated counterfactual scenarios)
Benchmarks: CounterEval
Metrics: accuracy per-metric accuracy confusion matrix
Baselines: GPT-4 Mistral-7B Instruct Llama 2 7B Chat Llama 3 8B Instruct
Builds on: QLoRA trl SFTTrainer completion-only collator
Closest papers by meaning
- Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations (ICML 2024 · similarity 0.91)
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 2024 · similarity 0.90)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.89)
- What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models (CVPR 2024 · similarity 0.88)
- Generating Likely Counterfactuals Using Sum-Product Networks (ICLR 2025 · similarity 0.87)
- Counterfactual Metarules for Local and Global Recourse (ICML 2024 · similarity 0.87)
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.