Research map / Large language models / Agents, code and math
When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs
Bogdan Zagribelnyy, Ivan Ilin, Maksim Kuznetsov, Nikita Bondarev, Roman Schutski, Thomas MacDougall, Rim Shayakhmetov, Zulfat Miftahutdinov, Mikolaj Mizera, Vladimir Aladinskiy, Alex Aliper, Alex Zhavoronkov
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 on CREED instead of USPTO-full (ablation) | C3LM-LFM2 (LFM2 2.6B SFT) · URSA-expert-2026 | Av. PT-Top-10 CC: 0.31 -> 0.38 | Improved |
| train on CREED instead of USPTO-full (ablation) | C3LM-LFM2 (LFM2 2.6B SFT) · USPTO-50K-test | Av. PT-Top-10 CC: 0.57 -> 0.45 | Hurt |
| add USPTO-full to CREED training (ablation) | C3LM-LFM2-CREED · URSA-expert-2026 | Av. PT-Top-10 CC: 0.38 -> 0.40 | Improved |
| add USPTO-full to CREED training (ablation) | C3LM-LFM2-CREED · URSA-expert-2026 | unique valid reactant sets: 58% -> 42% | Hurt |
| add USPTO-full to CREED training (ablation) | C3LM-LFM2-CREED · USPTO-50K-test | Av. PT-Top-10 CC: 0.45 -> 0.65 | Improved |
| add GRPO RFT with Molecular Transformer reward (ablation) | C3LM-LFM2-CREED+USPTO · URSA-expert-2026 | Av. PT-Top-10 CC: 0.40 -> 0.43 | Improved |
| add GRPO RFT with CREED multiple-answer exact-match reward (ablation) | C3LM-LFM2-CREED+USPTO · URSA-expert-2026 | Av. PT-Top-10 CC: 0.40 -> 0.40 | No clear effect |
| add GRPO RFT with Molecular Transformer reward (ablation) | C3LM-LFM2-CREED+USPTO · USPTO-50K-test | Av. PT-Top-10 CC: 0.65 -> 0.68 | Improved |
| evaluate on novel URSA-expert-2026 instead of USPTO-50K-test | LLMs overall · URSA vs USPTO-50K | Av. PT-Max CC: Claude 4.5 Sonnet 3.37 -> 1.45 | Hurt |
| fine-tune small LFM2 2.6B on CREED+USPTO | LFM2 2.6B · URSA-expert-2026 | Av. PT-Top-10 CC: 0.00 -> 0.40 | Improved |
| use open-weight general LLMs | Qwen3 8B / DeepSeek 3.2 · URSA-expert-2026 | Av. PT-Top-10 CC: 0.00 and 0.05 | Hurt |
| use LLMs rather than conventional SSRS models | LLMs · URSA-expert-2026 | ChemCensor plausibility | Hurt |
Closest papers by meaning
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.93)
- URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment (NeurIPS 2026 · similarity 0.91)
- Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers (NeurIPS 2026 · similarity 0.89)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.89)
- RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction (AAAI 2024 · similarity 0.89)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · 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.
