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

arXiv 2026 38.5 method

TLDR

Introduces operator-on-F, a diagnostic for latent world models that measures planning-relevant errors, outperforming reward-prediction error in correlating with return loss.

Reasoning

Strengths: Novel diagnostic with strong empirical correlation (-0.90) to return loss, demonstrated on TD-MPC2 size sweep and cross-architecture comparison. Weaknesses: Limited to a single environment (cheetah-run) and small number of model sizes (n=5), lacking broader validation.

Read-first score

Read-first score 38.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

62.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 255.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Operator-on-F: a planning-time diagnostic that compares a model's k-step latent pushforward to the environment's on an observable subset F, using the model's own predictor, complementing value-equivalence.

Methodology

The diagnostic is evaluated on a TD-MPC2 model size sweep over the cheetah-run task, measuring operator error, reward-prediction error, and Bellman residual. Rank correlation between operator error and return loss is computed with anchor-bootstrap 95% confidence intervals, and a cross-architecture comparison between TD-MPC2 and a pure-SSL latent world model is performed.

Key Results

Reward-prediction error varies only ~3x across model sizes (0.028–0.091) and correlates weakly with return (Spearman -0.30), while operator error spans 0.28–2.62 and correlates with return loss at -0.90 (95% CI [-0.90, -0.70]). At 317M parameters, operator error spikes to 2.62 and planning return collapses to 0.9, despite reward error remaining in the same narrow range.

Resources

Paper