Research map / Large language models / Instruction and fine-tuning
Learning from Scoring Disagreements: Contrastive Error Mining for Efficient and Robust LLM-based Assessment
Lei Chen, Tengteng Cheng, BoYu Gao, Zitao Liu, Weiqi Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-4B · SciEntsbank Unseen Answers | ACC: 0.7593 -> 0.7685 | Improved |
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-4B · SciEntsbank Unseen Questions | ACC: 0.6248 -> 0.6968 | Improved |
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-4B · SciEntsbank Unseen Domains | QWK: 0.6183 -> 0.6309 | Improved |
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-4B · Beetle Unseen Answers | ACC: 0.7677 -> 0.7995 | Improved |
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-4B · Beetle Unseen Questions | QWK: 0.6549 -> 0.6554 | Improved |
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-1.7B · Mohler Unseen Answers | MAE: 0.4216 -> 0.4019 | Improved |
| add contrastive hard-sample LoRA tuning after full fine-tuning | Qwen3-1.7B · Mohler Unseen Questions | MAE: 0.5116 -> 0.5109 | No clear effect |
| full fine-tuning instead of few-shot prompting | Qwen3-4B · SciEntsbank Unseen Answers | ACC: 0.5241 -> 0.7593 | Improved |
| apply CEM-FT across five Qwen base models | Qwen3-0.6B to Qwen3-8B · Mohler (average over tasks) | TAA: improves by 1.78% to 2.98% | Improved |
| raise beta from 1 to 2 (alpha=1) in hard-sample threshold (ablation) | CEM-FT · SciEntsbank | ACC: 0.6891 -> 0.7003 | Improved |
| raise beta from 1 to 2 (alpha=1) in hard-sample threshold (ablation) | CEM-FT · Beetle | ACC: 0.7311 -> 0.7447 | Improved |
| raise beta from 1 to 1.5 (alpha=1) in hard-sample threshold (ablation) | CEM-FT · Mohler | MAE: 0.4584 -> 0.4564 | Improved |
Closest papers by meaning
- Leveraging Online Olympiad-Level Math Problems for LLMs Training and Contamination-Resistant Evaluation (ICML 2025 · similarity 0.85)
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.85)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.85)
- Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals (ICML 2026 · similarity 0.85)
- Adaptive Testing for LLM Evaluation: A Psychometric Alternative to Static Benchmarks (ICML 2026 · similarity 0.84)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 2025 · similarity 0.84)
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.
