Driving · 2026-11-10

Planners Exploit Deterministic Error: Dithered Low-Bit World Models

Yong Hwi Song, Keon Myung Lee

arXiv:2611.11780PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors show that round-to-nearest low-bit activation quantization can cause sampling-based planners (MPPI/CEM) to collapse even when latent one-step prediction error stays comparable to a well-planning stochastic rounding scheme. They attribute the failure to persistent, candidate-shared deterministic quantization error exploited across planner iterations. A training-free fix, shared dither redrawn every forward call, restores performance and adds small latency.

Why it matters

Model-based control queries world models thousands of times per control step, so lowering precision can reduce compute cost. The paper identifies a failure mode specific to sampling-based planning: deterministic rounding error persists across iterations and is shared across candidates, creating selection bias that drives drift toward favored (incorrect) candidates. The authors propose a simple prescription—shared dither redrawn per forward call—that reduces planning damage without calibration or training.

Method

  • Quantize only hidden activations (per-layer uniform grids) and compare deterministic round-to-nearest (det) vs stochastic rounding variants that differ in how dither is shared and redrawn (ind, sh-iter, sh-step, sh-ep, det+shift).
  • Evaluate quantized TD-MPC2 (MPPI; 512 samples, 6 iterations, horizon 3) on Meta-World (success) and DMC (return) and DINO-WM (CEM open-loop on PushT/Wall/PointMaze; 300 samples, 30 iterations, horizon 5 on PushT).
  • Deploy an integer TD-MPC2 world model with fused low-bit kernels and CUDA graphs on an Ampere GPU to measure rollout/step latency; test block-scaled FP4 formats (NVFP4, MXFP4).

Limitation

The authors state the prescription assumes an iterating planner and that on DMC it helps in fewer than half of the affected cells; they also did not combine dither with calibration objectives.

Abstract (from arXiv)

Sampling-based planners such as MPPI and CEM query a learned world model thousands of times per control step, so low-bit world models are an obvious way to make model-based control cheaper on a robot. With round-to-nearest activation quantization and no calibration, we find that planning can collapse, by up to 100 points of success, while the world model's latent one-step prediction error stays no larger than that of a stochastic rounding scheme that plans well. At the same bit width, what decides the outcome is not the size of the rounding error but whether it persists across planner iterations and is shared across candidates. Round-to-nearest leaves an error that persists, and an iterating planner drifts toward the candidates that this error favors; in TD-MPC2, holding the same dither fixed within a step or an episode brings the drift back. Dither shared by all candidates and redrawn at every iteration removes most of the damage at the same bit width and without training: at least half of it in all 19 affected Meta-World cells (task $\times$ bit width) of TD-MPC2 (median 95%), in 4 of 5 DINO-WM cells, and in all 8 affected tasks of a 317M-parameter multi-task TD-MPC2. Captured as CUDA graphs on an Ampere GPU, a W8A4 integer TD-MPC2 world model runs its planning rollouts 1.9-3.3$\times$ and the whole planning step 1.3-1.5$\times$ faster than FP32, though not faster than BF16 with pre-converted weights; shared dither adds at most 2.3% latency. In TD-MPC2, block-scaled NVFP4 removes most of the damage and MXFP4 part of it, which ties the failure to coarse grids with a shared scale. Hypotheses, thresholds and analysis code were fixed before the data that test them, and the hypotheses that failed are reported.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-11-09 · Indexed by Pub-AI: 2026-11-10 UTC

preprint

Planners Exploit Deterministic Error: Dithered Low-Bit World Models

Yong Hwi Song, Keon Myung Lee

Preprint posted: 2026-11-09 · Original record ↗

arxiv
2611.11780
Explicit source versions (1)
  • v1 · 2026-11-09
  1. Preprint available

    Source date: 2026-11-09 · First detected: 2026-11-10 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

Sampling-based planners such as MPPI and CEM query a learned world model thousands of times per control step, so low-bit world models are an obvious way to make model-based control cheaper on a robot. With round-to-nearest activation quantization and no calibration, we find that planning can collapse, by up to 100 points of success, while the world model's latent one-step prediction error stays no larger than that of a stochastic rounding scheme that plans well. At the same bit width, what decides the outcome is not the size of the rounding error but whether it persists across planner iterations and is shared across candidates. Round-to-nearest leaves an error that persists, and an iterating planner drifts toward the candidates that this error favors; in TD-MPC2, holding the same dither fixed within a step or an episode brings the drift back. Dither shared by all candidates and redrawn at every iteration removes most of the damage at the same bit width and without training: at least half of it in all 19 affected Meta-World cells (task $\times$ bit width) of TD-MPC2 (median 95%), in 4 of 5 DINO-WM cells, and in all 8 affected tasks of a 317M-parameter multi-task TD-MPC2. Captured as CUDA graphs on an Ampere GPU, a W8A4 integer TD-MPC2 world model runs its planning rollouts 1.9-3.3$\times$ and the whole planning step 1.3-1.5$\times$ faster than FP32, though not faster than BF16 with pre-converted weights; shared dither adds at most 2.3% latency. In TD-MPC2, block-scaled NVFP4 removes most of the damage and MXFP4 part of it, which ties the failure to coarse grids with a shared scale. Hypotheses, thresholds and analysis code were fixed before the data that test them, and the hypotheses that failed are reported.