Event
Denoising Counterfactual Protocol Measures WAM Optional Margin
1 reports · 1 sources
Haojia Gao and other authors propose a denoising counterfactual protocol that uses bitwise-reproducible branches, repeated continuations, held-out oracles, and public random numbers to measure how much optional space a world action model (WAM) can achieve when selecting candidate actions at test time, and on this basis construct a failure-gated selector. Across four WAMs, two benchmarks, more than 100,000 branch rollouts, and preregistered tests, the authors report that bounds obtained from 17 continuations per candidate show that the optional share of three WAMs ranges from 24% to 66%; the selector raises closed-loop task success rates from 76.9% to 81.6% on held-out episodes of training tasks and from 62.6% to 69.8% on unseen tasks. These improvement magnitudes are results reported by the authors.
Report timeline
- How Much Can Best-of-N Improve a World Action Model: Measuring the Selectable Space and Realizing It in Closed Loop
The paper proposes a denoised counterfactual protocol that uses bitwise-reproducible branching, repeated continuations, held-out oracles, and common random numbers to measure the selectable space a world action model (WAM) can achieve when selecting candidate actions at test time, and builds a fail-gated selector on that basis.
arXiv cs.RO ·