NeurIPS paper: steering discrete diffusion cuts molecular optimization oracle calls 58x
yisongyue · x · 2026-11-05
A NeurIPS 2026 paper from Caltech and Genesis Molecular AI studies how to spend test-time compute to align pretrained molecular generators with specific protein targets.
- Pretrained models know medicinal chemistry's grammar, but random sampling rarely yields target-binding molecules; reward-tilted distributions p(x) ∝ ppre(x)·exp(r(x)/β) concentrate mass on high-reward valid molecules
- Test-time budget must balance oracle evaluation calls against model-side sampling and update costs
- Their steering approach matches Beam Search's 5,000-evaluation performance with only 86 oracle calls — 58x fewer
Takeaway: which test-time algorithm wins depends on the relative cost of oracle calls vs model inference.
More from Research
- STOC 2027: ~75% of 1,665 submissions report AI involvement in proofs and results — fortnow · 2026-11-05
- MIT shows GPT-6 Astra editing Qwen weights directly, boosting MATH-500 from 7.9% to 48.1% — lateinteraction · 2026-11-05
- Michael Levin preprint proposes topoiesis mechanism for group morphogenesis rescue — drmichaellevin · 2026-11-05
- WebGPU demo fits your voice recording and visualizes vocal tract motion in-browser — vincesitzmann · 2026-11-05
- MoE beats dense in matched OOD finetune: DeepSeek V3.1 hits 0.804 CE vs Llama 405B — kalomaze · 2026-11-05
- diffTLM: a differentiable vocal tract simulator wins NeurIPS 2026 Spotlight — vincesitzmann · 2026-11-05