ReGDiff: regulated latent diffusion boosts metamaterial novelty 46% over baselines
机器之心 · wechat · 2026-10-22
A NeurIPS paper from Virginia Tech (Dawei Zhou's lab) and Meta tackles the trade-off between geometric validity and novelty in generative metamaterial design. ReGDiff first regularizes the autoencoder latent space with RAS (inter-class repulsion on perturbed negative samples, intra-class repulsion, and central sink), then adds Short-Range Repulsion (SRR) during DDPM sampling to push samples away from nearby known designs only at short log-decaying range. The team also releases MetaTruss, a 10K-sample 48³ truss voxel dataset, plus five metrics (symmetry, periodicity, connectivity, novelty, diversity). Against the strongest baseline, it improves validity 8.9%, novelty 46.4%, and diversity 128.6%, generating 1000 structures in 31 seconds vs 326s for DiT-3D. Ablations show both modules are necessary.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23