SMDD-Bench - Lead Optimization: leaderboard

Success rate (%) on the 340 lead-optimization tasks, which ask for a molecule that improves target ADMET properties while holding constraints in SMDD-Bench (2026), where LLM agents with a minimal tool harness and a limited budget of Boltz-2 structure and ADMET-AI property oracles must submit molecules that pass hidden evaluators. Higher is better.

Metric: Success rate (%) on the 340 SMDD-Bench lead-optimization tasks (improve ADMET objectives while holding constraints and the binding interaction), minimalist ReAct agent harness with RDKit-style tools, 8 Boltz-2 and 15 ADMET-AI oracle calls per task, temperature 1.0, at most 100 turns; success means the submitted molecule passes the task's hidden evaluator; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4Medium57.6#135
2Gemini 3.1 Pro (Preview)Medium55.6#99
3Claude Sonnet 4.653.5#173
4Kimi K2.5Thinking43.5#266
5Qwen 3.5 397B A17B40#264
6DeepSeek V3.234.7#419
7MiniMax-M2.727.1#539

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=smdd-bench-lead-optimization · How It Works · Data refreshed daily, snapshot 2026-11-21.