Research map / Large language models / Instruction and fine-tuning
InfiMed-ORBIT: Aligning LLMs on Open-Ended Complex Tasks via Rubric-Based Incremental Training
Pengkai Wang, Pengwei Liu, Qi Zuo, Zhijie Sang, Congkai Xie, Hongxia Yang
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 |
|---|---|---|---|
| train with ORBIT rubric-guided RL on 2k samples | Qwen3-4B-Instruct · HealthBench-Hard | total score: 7.0 -> 27.5 | Improved |
| scale ORBIT training data from 2k to 28k | InfiMed-ORBIT-4B · HealthBench-Hard | total score: 27.5 -> 33.6 -> 37.3 | Improved |
| add retrieval-augmented (human seed) rubric generation (ablation) | ORBIT · HealthBench-Hard | total score: 10.69 -> 17.89 | Improved |
| use multi-LLM synthetic contrast seeds instead of human seeds (ablation) | ORBIT · HealthBench-Hard (140 steps) | total score: 15.34 -> 15.60 | No clear effect |
| use a stronger training-time judge (4B self to 30B-A3B) (ablation) | ORBIT · HealthBench-Hard | total score: 13.64 -> 17.89 | Improved |
| add moderate sample pass@k filtering (0-0.75) (ablation) | ORBIT · HealthBench-Hard | total score: 20.2 -> 19.7 | No clear effect |
| add strict sample pass@k filtering (0-0.50) (ablation) | ORBIT · HealthBench-Hard | total score: 20.2 -> 14.5 | Hurt |
| add rubric-level pass@k filtering (0-0.75) (ablation) | ORBIT · HealthBench-Hard | total score: 20.2 -> 19.9 | No clear effect |
| add multi-stage restart training (ablation) | ORBIT 8k · HealthBench-Hard | total score: 25.9 -> 27.3 | Improved |
| use DeepSeek-R1 rubrics instead of GPT-5-Chat rubrics (ablation) | ORBIT · HealthBench-Hard (OSS-120B) | total score: 12.3 -> 20.2 | Improved |
| use rubric-guided RL instead of inference scaling K=8 to 40 | Qwen3-4B-Instruct · best-of-K score distribution | normalized score: ~0.65 -> beyond 0.85 | Improved |
| add ORBIT training | Qwen3-4B-Instruct · HealthBench-Hard communication quality | axis score: -8.2 | Hurt |
Closest papers by meaning
- Zhongjing: Enhancing the Chinese Medical Capabilities of Large Language Model through Expert Feedback and Real-World Multi-Turn Dialogue (AAAI 2024 · similarity 0.88)
- CareBot: A Pioneering Full-Process Open-Source Medical Language Model (AAAI 2025 · similarity 0.86)
- MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.85)
- Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning (ICLR 2026 · similarity 0.85)
- Step-wise Rubric Rewards for LLM Reasoning (NeurIPS 2026 · similarity 0.84)
- CLR-voyance : Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics (NeurIPS 2026 · 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.
