Value Forcing: enforcing Bellman consistency at every noise level for accurate value functions
_onionesque · x · 2026-11-04
Researcher chongyiz1 introduces Value Forcing. Flow matching has succeeded for robotics and RL policies, but prior attempts to apply it to value functions produced biased, collapsed predictions. The team found the key to accurate value estimation is enforcing Bellman consistency at every noise level. The result is an accurate value function estimator that works even with very small amounts of data.
More from Research
- TLAPS-Bench: 265 real-world TLA+ proof targets; top model GPT-6.1 Sol solves only 44.3% — tianyin_xu · 2026-11-04
- StreamDuo pairs a commentator with a wiki research agent to commentate unseen games — Zixian Gao · 2026-11-04
- Redwood researcher: AI research automation and runaway recursive self-improvement can come apart — deanwball · 2026-11-04
- UMAP 0.6 adds Adam and momentum layout optimizers, becoming defaults in 0.7 — leland_mcinnes · 2026-11-04
- UMAP 0.6 preview release candidate lands on PyPI with major changes — leland_mcinnes · 2026-11-04
- Training microscopic deformations in a sheet of glass to classify images using nothing but light — louisvarge · 2026-11-04