Research map / Large language models / Instruction and fine-tuning
On the Shelf Life of Finetuned LLM-Judges: Future Proofing, Backward Compatibility, and Question Generalization
Janvijay Singh, Austin Xu, Yilun Zhou, Yefan Zhou, Dilek Hakkani-Tür, Shafiq Joty
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 |
|---|---|---|---|
| evaluate weak-response-trained judges on newer, stronger generator responses | SFT, DPO, SFT+DPO judges (Llama-3.1-8B, Ministral-8B, Mistral-24B) · DeepScaleR | consistent accuracy (FutureProof) | Hurt |
| retrain the judge on strong responses instead of weak ones (ablation) | fine-tuned judges · DeepScaleR, strong-response evaluation | consistent accuracy (RefreshAdvantage): up to +7.6 points (DPO, Mistral-24B) | Improved |
| use DPO rather than SFT when retraining on strong responses (ablation) | fine-tuned judges · DeepScaleR | RefreshAdvantage: SFT minimal gain; DPO up to 7.6 points | Improved |
| add SFT loss on top of DPO (ablation) | judge training · DeepScaleR | RefreshAdvantage | Mixed |
| scale judge backbone from 8B to 24B (DPO) (ablation) | retrained judges · DeepScaleR | RefreshAdvantage: 4.3 -> 7.6 | Improved |
| evaluate strong-response-trained judges on older weak responses | DPO and SFT judges · DeepScaleR | BackCompatibility | Mixed |
| continually train weak-trained DPO judge on strong responses (ablation) | DPO judges · DeepScaleR | FutureProof / RefreshAdvantage | Improved |
| continual training versus strong-only training, backward compatibility (ablation) | DPO judges · DeepScaleR | BackCompatibility | Mixed |
| evaluate on questions unseen during training instead of seen questions with new responses | fine-tuned judges · DeepScaleR | consistent accuracy (QuestionGen) | Hurt |
| use SFT instead of DPO for question generalization (ablation) | fine-tuned judges · DeepScaleR | QuestionGen drop | Improved |
| move from math (DeepScaleR) to knowledge-heavy MMLU-Pro | judge shelf-life metrics · MMLU-Pro | FutureProof magnitude | Improved |
| retrain on strong responses with SFT+DPO, 24B vs 8B backbone (ablation) | judges · MMLU-Pro | RefreshAdvantage: 3.2 -> 4.4 | Improved |
Closest papers by meaning
- Evaluating Judges as Evaluators: The JETTS Benchmark of LLM-as-Judges as Test-Time Scaling Evaluators (ICML 2025 · similarity 0.91)
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.90)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.90)
- Representation Consistency for Accurate and Coherent LLM Answer Aggregation (NeurIPS 2025 · similarity 0.89)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.88)
- Rethinking Reward Models for Multi-Domain Test-Time Scaling (NeurIPS 2026 · similarity 0.88)
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.
