Research map / Large language models / Agents, code and math
Towards Knowledge‑and‑Data‑Driven Organic Reaction Prediction: RAG‑Enhanced and Reasoning‑Powered Hybrid System with LLMs
Qingyu Wang, Xinyuan Cai, Xiang Cheng, YuZhe Gao, Bo XU
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 retrieved reaction type and similar cases (RAG) to the fine-tuned LLM (ablation) | Qwen3-32B · ORD test | Exact Match (%): 77.35 -> 83.13 | Improved |
| add MORGAN fingerprint similarity (FTS) reward to GRPO (ablation) | DeepSeek-R1-Distill-Qwen-32B GRPO · ORD test | Exact Match (%): 68.24 -> 56.83 | Hurt |
| add SFT cold-start before GRPO (ablation) | DeepSeek-R1-Distill-Qwen-32B · ORD test | Exact Match (%): 9.67 -> 68.24 | Improved |
| add GRPO after SFT (ablation) | DeepSeek-R1-Distill-Qwen-32B · ORD test | Exact Match (%): 59.93 -> 68.24 | Improved |
| add GRPO without SFT (ablation) | DeepSeek-R1-Distill-Qwen-32B · ORD test | Exact Match (%): 6.52 -> 9.67 | Improved |
| loosen similarity threshold M from 10 to 100 (ablation) | RAG retrieval · ORD test | Total Acc (%): 89.86 -> 88.74 | Hurt |
| loosen similarity threshold M from 10 to 30 (ablation) | RAG retrieval · ORD test | RAG-based Exact Match (%): 94.70 -> 87.33 | Hurt |
| loosen threshold M from 10 to 100 (coverage rises) (ablation) | RAG retrieval · ORD test | Proportion with similar cases (%): 81.70 -> 99.10 | Improved |
| scale backbone from 7B to 32B (ablation) | Reaction-Thinker · ORD test | Total Accuracy (%): 83.69 -> 89.86 | Improved |
| scale reasoning backbone from 7B to 32B (ablation) | reasoning-based LLM · ORD test | Exact Match (%): 56.12 -> 68.24 | Improved |
| apply RAG-based predictor to samples with no retrieved similar case | Reaction-Thinker (RAG-based) · ORD test | Exact Match (%): 94.70 -> 31.19 | Hurt |
| add GRPO to the 7B distillation model after SFT | DeepSeek-R1-Distill-Qwen-7B · USPTO-MIT test | Exact Match (%): 22.4 -> 35.9 | Improved |
Closest papers by meaning
- From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning (ICLR 2026 · similarity 0.88)
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.88)
- ReactGPT: Understanding of Chemical Reactions via In-Context Tuning (AAAI 2025 · similarity 0.87)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.87)
- Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets (NeurIPS 2025 · similarity 0.86)
- Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning (NeurIPS 2026 · 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.
