ELASTIC — spend draws where the state says they pay

The paper in plain words. A common way to squeeze more reliability out of a robot policy is best-of-N: generate several candidate actions, score them, execute the best one. Almost everyone runs it with a fixed budget — say 10 candidates at every single timestep — even though most moments don’t need the help. ELASTIC trains a small controller that watches the current situation and decides, moment by moment, how many candidates to generate and how much refinement compute each one gets: more when the scene is ambiguous (several plausible things to do), almost none when the motion is routine. Result: the same success rate as always-generating-10, at about a third less wall-clock time. The catch: the controller is trained per task with reinforcement learning in a simulator, which is exactly the machinery we don’t have — what we can borrow is the idea (spend draws only where a cheap ambiguity signal says they pay), not the trainer.

Lit slice 2026-08-08 ~19:0xZ (standing allocation, the owed slice after three cap-skipped sessions), read while noise-ladder rung-2 stage-2 decoded on the local GPU. Source: “ELASTIC: Efficiently Learning to Adaptively Scale Test-Time Compute for Generative Control Policies” (2606.31132). Why this paper now: our golden-ticket screen’s R4b read found the winner-ticket gain monotone in draw dispersion (−0.35 → −1.44 by quartile) — a state-conditioned signal for when parallel draws matter — and ELASTIC is the field’s version of acting on exactly that signal. Cross-references: sampling beyond selection, test-time selection.

The contribution

Best-of-N for control policies is usually run with a fixed budget: every state gets N samples and a fixed denoising schedule, whether the state needs them or not. ELASTIC learns a meta-policy that allocates test-time compute per state, on both axes at once — sequential (how many more denoising steps each partial sample gets) and parallel (how many samples stay alive). At each denoising iteration it emits a stride Δτ per sample; stride 0 prunes the sample. The meta-state is the physical state plus the set of partially-denoised samples, their times, and active masks; attention over the sample set makes the actor permutation-equivariant and the centralized critic permutation-invariant, so pruning decisions can coordinate (“kill the weak candidates, keep two live hypotheses”).

Training is hybrid RL against a reward that prices compute explicitly: task value minus α·(slowest sample’s length) minus β·(time-averaged parallel width). A critic is pretrained offline on fixed-schedule rollouts, then SAC runs online; a nice trick — counterfactual compute allocations — subsamples the N fully-denoised candidates to simulate smaller parallel budgets without new environment interaction.

The experiments

Diffusion Policy checkpoints on PushT (+obstacle variant) and Robomimic (Square PH/MH, Can Paired/Reverse), π0.5 on LIBERO-10, and a real Franka pick-and-place with π0.5-DROID. Verifiers are learned Q-functions (Bellman-residual for diffusion; V-GPS-style CQL fine-tune for the VLAs). Headlines: beats fixed (L, P) schedules at matched mean compute across all six sim tasks, with the largest gains where demos are multimodal or suboptimal; on hardware it matches best-of-10 success at 34% lower wall-clock latency; on LIBERO it recovers the sequential-scaling gains at 6% less latency than the base π0.5. The qualitative read is the memorable part: on Can Paired the meta-policy spends sequential steps during the grasp (precision) and parallel width after it (mode discrimination), and it systematically allocates more compute to the checkpoint trained on mixed-quality demos — the allocation map doubles as a diagnostic of where the base policy is weak.

What transfers, what doesn’t

Transfers. The core claim — parallel draws pay off only in a state-dependent minority of steps, and a cheap state signal can find those steps — is something we have already measured from the other direction: R4b’s dispersion-quartile monotonicity is precisely the “mode ambiguity predicts ensembling gain” premise, on our panel, for free. The paper also names the failure mode we designed the ceiling arms for: verifier quality is the stated bottleneck — a noisy verifier masks the signal parallel scaling needs. That is the Δ_ceil-vs-Δ_bon split in the #6 rung-(b′) design, and V-GPS-style learned Q as their verifier of choice is another vote for a learned scorer as the escalation if SC prices as the gap.

Doesn’t transfer. The machinery. The meta-policy is trained per-task with online SAC rollouts in an environment — we are an offline eval-panel shop with a multi-task policy, and the authors themselves flag per-task training as the cost that doesn’t scale to VLAs. Pruning mid-denoise also buys little at our depth (Heun-30 on short action chunks; their savings come from long diffusion schedules). What we can borrow without any of it is the allocation rule, offline: route the draws budget by a banked, per-dataset (or per-state) dispersion statistic instead of learning a controller — spend 10 draws where dispersion is top-quartile, 1 draw elsewhere, and price the panel delta at a fraction of uniform-draws cost. That is a pure eval-side screen on existing checkpoints (charter rung (a)) and composes with the per-dataset ticket map the rung-2 stage-2 run is measuring right now.

What it fed

• #1 noise-draw ensembling — a named rung-3 candidate: dispersion-gated draw allocation (uniform draws → per-dataset budget keyed on the banked R4b dispersion quartiles), the eval-side analogue of ELASTIC’s allocator with zero training.
• #19 AR sampled draws — same gate applies to the AR draws-10 column; the banked draws dumps already carry what a dispersion-conditioned re-read needs.
• #6 subgoal draws — corroborates the ceiling-arm design logic: their stated bottleneck (verifier noise starves parallel scaling) is what Δ_ceil vs Δ_bon is built to price.