HKU and Tsinghua propose VAST: RL that chooses its own value-backup horizon per state
jiqizhixin · x · 2026-11-04
HKU and Tsinghua researchers introduce VAST (reported as a NeurIPS 2026 paper), arguing long-horizon control benefits from choosing the length of a value backup per situation rather than fixing one temporal stride for all states.
Three takeaways
- Flexible segment scoring: a G-function estimates discounted return between a current state and a future state over a chosen horizon, with compositional consistency encouraging segment estimates to agree across time scales
- Choose destination and duration: a stitching policy selects a future subgoal and its horizon, combining segment return with downstream value; a separate goal-conditioned execution policy turns the choice into actions without taking the predicted horizon as input
- Evaluation: strong or competitive performance across 50 state-based OGBench tasks, especially on difficult ones
The attached figure illustrates the G/V decomposition, cross-horizon stitching, and optimal target structure.
More from Research
- Jev model improves ICL demonstration retrieval via distillation into a lightweight MLP — kalyan_kpl · 2026-11-04
- Indie researcher open-sources 10 years of math work: 146k files, AI-accelerated over 3 years — HamOnWhy · 2026-11-04
- Decision models match Gemini for RAG conflict typing at 1/39 the cost and 6x speed — parvkhatri · 2026-11-04
- RLHF's scalar reward hides a 2,450-year-old problem: helpfulness and harmlessness are opposing forces — vasilisvj · 2026-11-04
- EvoTeam evolves multi-agent team organization and skills from execution experience — JLU-Deep · 2026-11-04
- FlowTranslator: training-free in-image translation via diffusion flow editing — Fudan-University · 2026-11-04