Research map / Label-efficient and robust learning / Recognition and long-tail
DC-W2S: Dual-Consensus Weak-to-Strong Training for Reliable Process Reward Modeling in Biological Reasoning
Chi-Min Chan, Ehsan Hajiramezanali, Xiner Li, Edward De Brouwer, Carl Edwards, Wei Xue, Sirui Han, Yike Guo, Gabriele Scalia
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 |
|---|---|---|---|
| aggregate multiple weak label sources (LLM judges plus MC rollouts) instead of a single source (ablation) | DC-W2S PRM · PerturbQA, 4 cell lines | average F1 (BoN N=8): 66.9% -> 68.5% | Improved |
| add DC-W2S distribution-balanced instance sampling (100k) instead of full 351k set (ablation) | PRM training · PerturbQA RPE1 (OOD) | F1 (%): 64.0 -> 64.5 (vs Full Set multi-label) | Improved |
| train only on P1 labels (about 62% of steps) (ablation) | PRM masking · PerturbQA RPE1 (OOD) | F1 (%): approx 64% F1 with about 62% of labels | No clear effect |
| train only on low-reliability regimes P3 or P4 (ablation) | PRM masking · PerturbQA RPE1 (OOD) | F1 (%): Only P3 57.52; Only P4 52.40 vs Only P1 63.40 | Hurt |
| add P3 steps on top of P1 anchors (ablation) | PRM masking · PerturbQA RPE1 (OOD) | F1 (%): 63.40 -> 64.66 | Improved |
| add P2 steps on top of P1 anchors (ablation) | PRM masking · PerturbQA RPE1 (OOD) | F1 (%): 63.40 -> 64.16 | Improved |
| add P4 steps on top of P1 anchors (ablation) | PRM masking · PerturbQA RPE1 (OOD) | F1 (%): 63.40 -> 64.23 | No clear effect |
| train on 100k DC-W2S subset instead of full weak-label set (ablation) | PRM · BioReason KEGG | weighted F1 (%): 87.40 -> 88.89 (CellProfiler), 89.08 (ESM) | Improved |
| use 100k DC-W2S subset instead of the full set with rbio1 policy (ablation) | PRM · RPE1 (OOD) rbio1 average | F1 (%): 70.59 -> 68.43 (CellProfiler), 69.73 (ESM) | Hurt |
| use PRM best-of-N instead of greedy decoding with question-only prompts | Qwen3-4B policy · RPE1 (OOD) Only Query | average F1 (%): 27.25 -> 52.56 | Improved |
| use PRM best-of-N instead of greedy decoding with rbio1 policy | rbio1 · RPE1 (OOD) | average F1 (%): 56.61 -> 70.59 | Improved |
| use biologically grounded embeddings (ESM, CellProfiler) for neighborhood consensus rather than semantic embeddings (ablation) | DC-W2S · RPE1-DE | F1 gain from P3 | Improved |
Closest papers by meaning
- Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models (ICLR 2026 · similarity 0.87)
- Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning (ICLR 2025 · similarity 0.87)
- P2S: Probabilistic Process Supervision for General-Domain Reasoning Question Answering (AAAI 2026 · similarity 0.86)
- From Mathematical Reasoning to Code: Generalization of Process Reward Models in Test-Time Scaling (AAAI 2026 · similarity 0.86)
- Rethinking Reward Models for Multi-Domain Test-Time Scaling (NeurIPS 2026 · similarity 0.86)
- SCAN: Self-Denoising Monte Carlo Annotation for Robust Process Reward Learning (NeurIPS 2025 · similarity 0.86)
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.
