Research map / Large language models / Language, safety and interpretability
xFinder: Large Language Models as Automated Evaluators for Reliable Evaluation
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.
Proposes xFinder, a fine-tuned LLM evaluator that replaces RegEx answer extraction and judge models for objective-question evaluation, trained on the new KAF dataset.
Problem
RegEx-based answer extraction in LLM evaluation frameworks is error-prone and fine-tuned LLM judge models generalize poorly, making evaluation results unreliable.
Approach
Treat key answer extraction as a learned task: fine-tune small LLMs (0.5B–8B) on a newly built dataset of LLM responses with human-verified key answers, so they extract answers from free-form responses and feed a deterministic match for judgment.
Key results
- On the KAF generalization set, xFinder-qwen1505 reaches 93.42% overall extraction accuracy versus at most 74.38% for RegEx (OpenCompass), and xFinder-llama38it reaches 95.18%.
- xFinder-llama38it achieves 97.61% judgment accuracy, versus 88.70% for OpenCompass, 84.20% for GPT-4 and 78.13% for JudgeLM-33B.
- xFinder-qwen1505 processes 200 items per task in about 10.67 s on average, compared with 534.69 s for JudgeLM-33B; its cost is about $0.02, 0.4% of GPT-4 API cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace RegEx extraction with xFinder-qwen1505 | LLM answer evaluation · KAF test set | extraction accuracy: 0.8113 -> 0.9688 | Helped |
| Replace RegEx extraction with xFinder-qwen1505 | LLM answer evaluation · KAF generalization set | extraction accuracy: 0.7438 -> 0.9342 | Helped |
| Use xFinder-llama38it instead of GPT-4 as an extractor | LLM answer evaluation · KAF generalization set | extraction accuracy: 0.6957 -> 0.9518 | Helped |
| Use xFinder-llama38it instead of GPT-4 as a judge | LLM response evaluation · KAF generalization set | judgment accuracy: 0.8420 -> 0.9761 | Helped |
| Add chain-of-thought prompting | GPT-4 judge · KAF generalization set | judgment accuracy: 0.8420 -> 0.8842 | Helped |
| Use xFinder-qwen1505 instead of xFinder-llama38it | xFinder extraction · KAF generalization set | extraction accuracy: 0.9518 -> 0.9342 | Hurt |
| Use xFinder-qwen1505 instead of xFinder-llama38it | xFinder judgment · KAF generalization set | judgment accuracy: 0.9761 -> 0.9748 | Hurt |
Surprising findings
- GPT-4 as extractor achieves only 69.57% extraction accuracy on the generalization set, below the OpenCompass RegEx baseline on the same set.
- Changing alphabet-option format to short-text format changes CommonsenseQA model rankings substantially, suggesting alphabet-option evaluation is unreliable.
Limitations
- Evaluated only on objective tasks with deterministic answers (MCQ and math), not open-ended generation.
- Fine-tuned on responses from a limited set of ~10 LLMs and 8 tasks; generalization to other models and tasks is only partially tested.
- Training labels for the training split are partly GPT-4 pre-annotated and human-rechecked, so label quality is not uniform across splits.
Experimental setup
Datasets: KAF training set KAF test set KAF generalization set AgNews Amazon ARC BoolQ DBPedia GSM8K MATH MMLU OpenbookQA
Benchmarks: KAF test set KAF generalization set real-world evaluation on 14 tasks
Metrics: extraction accuracy judgment accuracy accuracy gap inference time
Baselines: OpenCompass LM Eval Harness UltraEval PandaLM-7B JudgeLM-7B JudgeLM-33B GPT-4 as Judge
Builds on: QLoRA XTuner Chain of Thought self-consistency
Closest papers by meaning
- tinyBenchmarks: evaluating LLMs with fewer examples (ICML 2024 · similarity 0.88)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.88)
- RocketEval: Efficient automated LLM evaluation via grading checklist (ICLR 2025 · similarity 0.88)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.88)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.88)
- PiCO: Peer Review in LLMs based on Consistency Optimization (ICLR 2025 · 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.