Research map / Large language models / Reasoning and chain of thought
CellDuality: Unlocking Biological Reasoning in LLMs with Self-Supervised RLVR
Yuhang Chen, Zhen Tan, Ruichen Zhang, Mufan Qiu, Tianlong Chen
Read the paperCheck your idea against 137,217 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 CellDuality, a self-supervised RLVR approach for unlocking open-ended biological reasoning in LLMs, presented at ICLR 2026.
Problem
Existing LLMs handle predictive biology tasks like cell type annotation but struggle with open-ended, mechanistic biological reasoning; RLVR typically needs verifiable answers that are scarce in biology.
Approach
Uses self-supervised signals to create verifiable rewards for RL training of LLMs on biological reasoning, avoiding the need for hand-labeled verifiable answers. Details beyond the title and truncated abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-supervised duality RL after SFT (ablation) | CellDuality (Llama-3.2-3B) · sci-Plex3 perturbation response (OOD) | scFID: 0.045 -> 0.038 | Improved |
| add self-supervised duality RL after SFT (ablation) | CellDuality · Bastidas-Ponce-2019 annotation (OOD) | accuracy: 72.87 -> 88.45 | Improved |
| add self-supervised duality RL after SFT (ablation) | CellDuality · Segerstolpe-2016 annotation | accuracy: 98.76 -> 99.81 | Improved |
| add self-supervised duality RL after SFT (ablation) | CellDuality · Drug sensitivity GSE110894 (OOD) | accuracy: 96.45 -> 96.12 | Hurt |
| add self-supervised duality RL after SFT (ablation) | CellDuality · Drug sensitivity GSE149383 | accuracy: 98.91 -> 99.12 | No clear effect |
| add self-supervised duality RL after SFT (ablation) | CellDuality · Conditional cell generation | GW distance: 63.87 -> 61.45 | Improved |
| use ground-truth reward RL instead of duality reward (ablation) | CellDuality · sci-Plex3 | scFID: 0.038 -> 0.025 | Improved |
| use ground-truth reward RL instead of duality reward (ablation) | CellDuality · He-2020-Liver | macro F1: 95.41 -> 94.85 | Hurt |
| use ground-truth reward RL instead of duality reward (ablation) | CellDuality · He-2020-Liver | accuracy: 96.34 -> 97.21 | Improved |
| use CellDuality instead of C2S-Scale SFT | Llama-3.2-3B · sci-Plex3 | scFID: 0.02 -> 0.038 | Hurt |
Experimental setup
Builds on: RLVR
Closest papers by meaning
- CellVerse: Do Large Language Models Really Understand Cell Biology? (NeurIPS 2025 · similarity 0.86)
- LaSeR: Reinforcement Learning with Last-Token Self-Rewarding (ICLR 2026 · similarity 0.86)
- How Post-Training Shapes Biological Reasoning Models (NeurIPS 2026 · similarity 0.85)
- Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing (NeurIPS 2026 · similarity 0.85)
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (NeurIPS 2025 · similarity 0.85)
- BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks (NeurIPS 2026 · similarity 0.85)
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.
