The Rank-One Corner: How Much Value Equivalence Does a Task Need from a World Model?

A learned world model is usually judged by how faithfully it reconstructs its observations or predicts reward, as though quality were something the model simply has or lacks. But what a task actually needs from a model is narrower: the few predictive coordinates its queries depend on, which we call the closure. We show that how much of that closure a latent comes to represent is set not by the model's capacity or its observations but by the dimensionality of the objective it is trained against, and we measure this directly on a DreamerV3 stack in a controlled environment with known ground-truth closure. An aligned scalar value signal -- the objective at the heart of value equivalence -- installs only a one-dimensional projection of a closure that needs several dimensions: read through a single linear probe, the recoverable structure rises from R^2=0.10 to 0.76 as the scalar is replaced by the full objective. Sweeping the objective's dimensionality from one to four installs exactly that many predictive directions through an auxiliary head, and the same staircase appears -- at attenuated magnitude but the same rank -- through the model's own value head, so the dissociation is dimensional rather than an artifact of head form. Capacity-matched comparisons and in-situ pressure checks rule out the obvious alternatives. The law governs a regime, and we measure its boundary: on a companion closed-loop task whose structure is observable frame by frame, reconstruction installs that structure and the scalar objective suffices -- the objective decides what a latent represents exactly where cheaper training signals cannot already recover it. Value equivalence is thus not all-or-nothing but dimensional: the familiar single-reward objective is its rank-one corner, and a model installs as much of a task's structure as the objective it is asked to predict.

Operator-on-F complements value-equivalence: a planning-time diagnostic for latent world models

World-model evaluation for model-based reinforcement learning typically asks whether the learned model predicts reward and value well, which can leave planning-relevant errors in the model's latent rollouts unmeasured. We introduce a complementary diagnostic, operator-on-F, that compares a model's k-step latent pushforward to the environment's on an observable subset F, using the model's own predictor. On a TD-MPC2 size sweep over cheetah-run, reward-prediction error stays within [0.028, 0.091] for every model size - only about 3x variation - so an unnormalized reward-fit check has narrow resolution to distinguish them; the (unnormalized) Bellman residual and reward error themselves have weak relationships with return (Spearman -0.10 and -0.30). Operator error spans 0.28 to 2.62 over the same sizes. At 317M the operator error is 2.62 - an order of magnitude above the 0.28-0.36 cluster - and the planning return collapses to 0.9, while reward-prediction error (0.091) is the highest of the five but stays within the same small [0.028, 0.091] range as the rest of the sweep. The rank correlation between operator error and return loss is -0.90 (anchor-bootstrap 95% CI [-0.90, -0.70] at n=5 sizes; leave-one-out removal of any single size leaves it at -0.80 or stronger). The operator also returns informative, architecture-discriminating estimates in a cross-architecture comparison between TD-MPC2 and a pure-SSL latent world model. The operator diagnostic complements value-equivalence rather than replacing it.

In-Context Reinforcement Learning through Bayesian Fusion of Context and Value Prior

In-context reinforcement learning (ICRL) promises fast adaptation to unseen environments without parameter updates, but current methods either cannot improve beyond the training distribution or require near-optimal data, limiting practical adoption. We introduce SPICE, a Bayesian ICRL method that learns a prior over Q-values via deep ensemble and updates this prior at test-time using in-context information through Bayesian updates. To recover from poor priors resulting from training on sub-optimal data, our online inference follows an Upper-Confidence Bound rule that favours exploration and adaptation. We prove that SPICE achieves regret-optimal behaviour in both stochastic bandits and finite-horizon MDPs, even when pretrained only on suboptimal trajectories. We validate these findings empirically across bandit and control benchmarks. SPICE achieves near-optimal decisions on unseen tasks, substantially reduces regret compared to prior ICRL and meta-RL approaches while rapidly adapting to unseen tasks and remaining robust under distribution shift.