What Should a Finite World Model Preserve? Learning Consequences for Decision-Weighted Capacity Allocation
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Finite world models must decide which distinctions deserve limited representational resolution. We formulate this as context-conditioned allocation under observational support, decision relevance, and finite capacity. Our Learn–then–Allocate factorization predicts rate-conditioned representational consequences and solves the current budget explicitly. Uniform consequence error bounds allocation regret and yields exact recovery below the oracle margin; measured distortion curves reproduce of brute-force optima. A representation-specific ablation shows that the essential structure is rate-dependent marginal consequence: static importance and constant-marginal surrogates cannot reproduce scarcity-sensitive allocation, while a generic rate model captures only part of it. Under joint OOD, Learned-DP reaches exact match and regret versus and for Direct-MLP+Slack. In learned-latent control, one-step regret is halved relative to Uniform, but freezing one allocation over longer rollouts erodes and eventually reverses that advantage. Finite representation allocation is therefore context-conditioned: learn rate-dependent consequences, optimize current scarcity, and do not treat one allocation as permanent.
- Primary area:
- learning theory