ICML 2026 Drift is a Sampling Error - Heungwoo/research GitHub Wiki

Drift is a Sampling Error: CAPS — SNR-gated power-distribution search for long-horizon VLA planning

Venue: ICML 2026 (Poster) Category: Efficiency Affiliations: Kewei Chen, Yayu Long, Mingsheng Shang Traction (2026-06): 0 citations (arXiv)

Overview of the Context-Aware Power Sampling (CAPS) framework with metacognitive SNR-based gating (Figure 1 from Chen et al., 2026)

Problem

Vision-Language-Action (VLA) models excel at short-horizon control but suffer instruction drift on long-horizon tasks: as a task progresses, irrelevant observations dilute attention to the original instruction, and the robot executes actions that are locally reasonable but globally wrong. The paper reconceptualizes drift as a systematic sampling error — local greedy sampling collapses into "Negative Pivotal Windows": irreversible local optima with high local probability that sever global success pathways. Prior fixes (CoT prompt engineering; generate-verify rerankers like TACO) either stay open-loop or rely on parallel independent sampling and reranking, lacking iterative refinement of a single trajectory.

Method

The authors propose Context-Aware Power Sampling (CAPS), a training-free, inference-time framework requiring no parameter updates.

Power distributions for global sharpening. CAPS raises the model's conditional trajectory distribution to a power (sharpening coefficient α) to sharpen global trajectory probabilities, enabling lookahead search over the generative trajectory distribution. The paper argues (and proves) this is not equivalent to local temperature sampling.
Metacognitive SNR control. A KL-divergence-based Signal-to-Noise Ratio signal estimates drift risk. When contextual SNR/entropy is high-certainty (H ≤ γ), the system stays in fast "System 1" greedy decoding; when SNR drops below a critical threshold γ_SNR, it triggers slow "System 2" search.
Block-based autoregressive MCMC. Once triggered, an intra-inference MCMC loop performs stochastic resampling of future trajectory chunks (Proposal) followed by an Acceptance step, iteratively refining a single trajectory rather than reranking independent samples. This is framed as adaptive computation / optimal control and provides an effective horizon extension.

Mechanism of the System 2 inference loop: SNR-triggered proposal (resampling) and acceptance (Figure 2 from Chen et al., 2026)

Results

RoboTwin 1.0 (π₀): CAPS lifts π₀ to 47.4% (+15.2%), vs TACO's +6.1%; on the drift-prone "Dual Bottles Pick Hard," success rises from 48.0% to 61.0%.
Simpler-WindowX (OOD): CAPS reaches 60.5% average with π₀ (+12.5% over π₀'s 48.0%), beating SpatialVLA (42.7%); "Carrot on Plate" +19.0%.
RoboTwin 2.0 (large-scale, 100 seeds/100 scenes, π₀.₅): average 66.2% vs π₀.₅ 59.3% and π₀.₅+TACO 64.0%.
Libero-long (π₀.₅): average 97.6% vs π₀.₅ 94.8%, TACO 96.6%, OpenVLA 49.8%, Robomonkey 56.5%.
Ablation (RoboTwin 1.0): Full adaptive CAPS = 47.4% at 2.15× latency, vs Always-On CAPS 48.1% at 8.50×, Rejection Sampling 40.2%, w/o Sharpening 38.5%, base π₀ 32.2% (1.00×) — the SNR gate recovers nearly all of the always-on accuracy at a fraction of the compute.

Significance

CAPS reframes long-horizon robustness as an inference-time sampling/search problem rather than a training problem, and its SNR gate makes "slow thinking" cost-adaptive — spending extra compute only when drift is detected. As a model-agnostic, training-free add-on that improves π₀, π₀.₅, and OpenVLA-class policies, it is a clean instance of the 2026 trend toward adaptive inference-time scaling for embodied control.