$ cat wiki/papers/2026/2608.18746-decision-metric-alignment.md
Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned Objectives for MPC Planning (arXiv:2608.18746)
TL;DR
JEPA-style latent world models plan by taking Euclidean distance to a goal latent as the cost for model-predictive control. This paper names the property that justifies doing so — decision-metric alignment — and shows it does not follow from a model probing well: strong decoding of task variables does not guarantee the cost ranks candidate action sequences by real task progress. Two rank-agreement diagnostics are proposed, and DA-LeWM closes the observed gap with action-conditioned heads while probe scores stay similar (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED; the paper was not read. Listed
on HuggingFace Daily Papers, 2026-08-23, 15 upvotes
(source).
Method
The named property. Decision-metric alignment is whether latent distance to a goal ranks candidate plans in the same order as real task progress. The paper's point is that this is a separate property from representation quality — a probe can decode task variables well from a latent space whose geometry still misranks plans.
Two diagnostics:
- Plan-Real Spearman — latent–real rank agreement on random plans.
- CEM-stage Spearman — the same agreement measured as cross-entropy-method search concentrates its proposal distribution, i.e. where the ranking actually has to hold.
Analysis. Sufficient conditions under which latent distance preserves real-cost rankings, with three controlling quantities identified: encoder distortion, terminal rollout error, and candidate margins.
Method. DA-LeWM augments LeWM with inverse-dynamics and demonstration-conditioned goal-action heads.
Results
- DA-LeWM accelerates convergence and achieves higher online success than LeWM across all reported experiments.
- Probe scores remain similar — the improvement is in the geometry the planner uses, not in what the representation encodes.
What the abstract does not give: absolute success rates, the environments, the size of the empirical alignment gap, and Spearman values before and after.
Significance
"Probe scores remain similar" is the whole result. Representation quality has been the standard evidence that a latent world model is working; this separates it from the property planning actually needs and gives two measurements for the second one. A model can look right by the usual instrument and still rank plans wrongly — which is the same failure shape as SoftVTBench: A Deformation-Aware Visuo-Tactile Dataset and Benchmark for Deformable-Object Manipulation (arXiv:2608.18701) (task success passes while physical interaction fails) and SemaPLC: A Project-Grounded, Verification-Gated Agent Harness for PLC Code Generation (arXiv:2608.18565) (static scores put every method within 10 points while dynamic behaviour spreads them), arriving on the same day from a third direction.
It is also a concrete instance of the caution Mechanistic Interpretability carries about probes. A probe answers "is the information there", never "is it arranged so the downstream user of it succeeds". Here the downstream user is CEM, and the gap is measurable.
Scope kept: JEPA-style latent models with Euclidean goal-distance costs and CEM search. Nothing here says the diagnostics transfer to learned critics or to planners that do not rank by latent distance — and the abstract's "across all our experiments" is a claim without a count attached.
Open Questions
- How large is the empirical alignment gap the method was designed against?
- Which environments, and how many?
- Do the two Spearman diagnostics disagree — i.e. is a model ever aligned on random plans and misaligned once CEM concentrates? That case is the one the second diagnostic exists for, and the abstract does not say whether it was found.
- Author list, affiliation, licence — unknown; the paper was not read.
Cite
Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned
Objectives for MPC Planning (2026). arXiv:2608.18746.