Latent Dynamics · 2026-10-26

When Does Dynamics Randomization Test Context Inference?

Oleg Smirnov, Sofiane Ennadir, John Pertoft, Bjartur Hjaltason, Sara Karimi

arXiv:2610.32414PDF

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 argue that dynamics randomization may not actually test context inference if the current state and action already contain most predictive evidence or if kernel differences vanish on visited support. They define “predictive headroom” (one-step log-loss difference) and compare kernels under fixed vs visited state–action distributions to separate context coverage from context reweighting effects, across discrete and continuous synthetic environments.

Why it matters

When evaluating world models with randomized dynamics, a small benefit from adding context could reflect limited predictive headroom rather than failure of context inference. The paper proposes measuring this missing information under the benchmark’s visitation/support and context-weighting, so conclusions about whether context is needed can be interpreted relative to what the inputs currently withhold.

Method

  • Define predictive headroom as the optimal one-step log-loss gap between using the predictive mixture over context and using the true context-specific kernel at a query (s,a).
  • Compare “struct” (fixed reference queries) vs “match” (same queries but visited state–action distribution) vs “visited headroom” (reweight contexts by d(θ|s,a)) to attribute reductions to query coverage vs context weights.
  • Compute headroom analytically/with known kernels in MiniGrid variants and with known kernels in a continuous Point-Mass-Attractor, using logged state–action episodes under uniformly random actions.

Limitation

The three environments are synthetic benchmarks with known transition kernels, and all data was collected under uniformly random actions; the diagnostic concerns one-step prediction on the collection distribution and reported values are predictor estimates rather than exact mutual information. For learned policies or physical robots, headroom would need measurements under their own prediction/control conditions.

Abstract (from arXiv)

Environments with randomized dynamics are widely used to ask whether world models need explicit context inference. However, randomization can leave little predictive information for a context encoder to add, because distinct latent parameters may produce similar transitions at visited states, or those states may already reveal the active context. We formalize the remaining one-step information as predictive headroom and compare kernels under fixed and visited state--action distributions to distinguish changes in query coverage from changes in context weights. We analyze discrete and continuous environments and find low headroom both when kernels differ little and when visitation reduces an otherwise large discrepancy. At the same time, resetting state components increases headroom, showing that context evidence accumulates in state. We propose reporting this headroom alongside the randomized parameters, so that gains from explicit context inference can be interpreted relative to the information missing from the current state and action.

Related papers

Publication record

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

First public release: 2026-10-23 · Indexed by Pub-AI: 2026-10-26 UTC

preprint

When Does Dynamics Randomization Test Context Inference?

Oleg Smirnov, Sofiane Ennadir, John Pertoft, Bjartur Hjaltason, Sara Karimi

Preprint posted: 2026-10-23 · Original record ↗

arxiv
2610.32414
Explicit source versions (1)
  • v2 · 2026-10-30
  1. Explicit source version detected · v2

    Source date: 2026-10-30 · First detected: 2026-11-07 UTC · Historical baseline

    Source evidence (1)
  2. Preprint available

    Source date: 2026-10-23 · First detected: 2026-11-07 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

Environments with randomized dynamics are widely used to ask whether world models need explicit context inference. However, randomization can leave little predictive information for a context encoder to add, because distinct latent parameters may produce similar transitions at visited states, or those states may already reveal the active context. We formalize the remaining one-step information as predictive headroom and compare kernels under fixed and visited state--action distributions to distinguish changes in query coverage from changes in context weights. We analyze discrete and continuous environments and find low headroom both when kernels differ little and when visitation reduces an otherwise large discrepancy. At the same time, resetting state components increases headroom, showing that context evidence accumulates in state. We propose reporting this headroom alongside the randomized parameters, so that gains from explicit context inference can be interpreted relative to the information missing from the current state and action.