WTF: simulation-free RL fine-tuning for flow maps cuts training compute by up to 280x
yeewhye · x · 2026-10-15
A NeurIPS 2026 Spotlight paper introduces Wasserstein-Tilted Flow Maps (WTF), enabling simulation-free RL fine-tuning for diffusion and flow models. Instead of expensive rollouts, WTF replaces KL regularization with a transport-based alternative that moves each sample toward higher reward. One step beats 50 steps of the base model, with up to 280x less training compute.
Related event: WTF Flow Maps Enable Simulation-Free RL Fine-Tuning, 280x Cheaper(3 posts)→
More from Research
- Bittensor Subnet 55 completes 10M synthetic CRISPR experiments with 248 competing miners — markjeffrey · 2026-10-15
- Logit lens shows Qwen3-1.7B's refusal to be shut down depends on how many layers you train — Sauers_ · 2026-10-15
- DeepSeek's first LLM paper was literally titled "Scaling Open-Source Language Models with Longtermism" — teortaxesTex · 2026-10-15
- Vals AI Puts Frontier Models in Real-Time MuJoCo Robot Simulations — scaling01 · 2026-10-15
- AI Post-Processing Tool Extends Extreme Weather Warnings From 10 Days to 6 Weeks — LesterMackey · 2026-10-15
- DuplexJev Cuts Autoregressive Decoding for Voice Agents, Answering 80 Speech Decisions in 0.1s — adventists-ai · 2026-10-15