Eleven World Models Miss the Odds; ES Covers More Paths

Featured

01 Pretty Samples Can Hide Bad Probabilities

Given the same initial frame and action, do repeated generations produce each outcome at the right frequency? PAWBench asks this because real physical processes often allow several valid outcomes. A world model should reproduce their conditional distribution, not just one plausible trajectory.

Its companion evaluator, PAWEval, converts repeated video generations into empirical probabilities over physical outcomes. This shifts evaluation from “Does this video look real?” to “Does this prediction set have the right probability structure?”

Across 50 scenarios and 11 systems, no model consistently combined behavioral coverage with reference probability matching. One impressive sample may simply represent a lucky guess. Datasets need multiple outcomes per condition, and tests need repeated sampling. Probability alignment matters, but it does not certify general physical understanding or causal reasoning.

Key takeaways:

• Evaluate world models through repeated generation and outcome frequencies, not isolated showcase samples.
• Build datasets and tests around multiple valid futures from the same initial conditions.
• Use PAWBench for probability alignment, not as a complete test of physical understanding.

Source: PAWBench: How Far Are We from Probabilistically Aligned World Modeling?

02 Seeing the Street Isn't Enough

Strong street-scene recognition does not mean an MLLM can act reliably across a city. Visual clues expire as the agent moves, routes become unavailable, and pedestrians change position.

UrbanGround uses territory-wide 3D geographic data from Hong Kong to build a real-scale sandbox. It replaces static questions with continuous, first-person movement under physical constraints. Current models often recognize scenes and handle short-range spatial reasoning. Direction judgment, pedestrian avoidance, and long-range exploration remain unreliable.

Local errors accumulate without effective correction. The sandbox therefore tests how long perceptual evidence stays useful after the environment changes. Teams building embodied or urban agents should test evidence expiry and active recovery before deployment. A real-scale simulation still cannot replace validation in a real city.

Key takeaways:

• Static street-scene performance does not predict continuous urban movement.
• Add movement, route changes, and pedestrians to test when local evidence expires.
• Use real-scale sandboxes for predeployment stress tests, then validate in real environments.

Source: UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City

03 Successful Trajectories Need Causal Coherence

ACE represents agent data through four interdependent elements: environment, task, interaction, and success signal. Misalignment anywhere can produce experience that contradicts the task goal or the reported outcome.

The framework separates data generation into accuracy, complexity, and diversity. First, verify that an experience is genuine and internally consistent. Next, check whether its difficulty matches the current learner. Finally, ask whether added samples cover new behavior or merely rephrase existing experience.

This separation helps teams distinguish candidate construction, execution checks, and sample selection. Audit successful trajectories one link at a time. Every component should support the same outcome and teach something the current model lacks. Benefits still need validation across agent tasks and training setups.

Key takeaways:

• Audit the full causal chain behind successful trajectories, not only the final reward.
• Adjust data difficulty as model capability and execution settings change.
• Separate new behavioral coverage from superficial rewrites before expanding a dataset.

Source: What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents

04 Post-Training Competes on Reasoning Coverage

Evolution strategies improved Pass@1 while achieving higher Pass@K than GRPO. They also produced a wider range of reasoning paths. GRPO suffered entropy collapse during training.

The authors therefore propose sequential training: GRPO first, then ES. GRPO supplies stronger single-attempt performance, while ES broadens multi-path coverage.

Although ES caused clear parameter drift across the model, only a small number of large updates drove the gains. Held-out evaluations did not show inevitable catastrophic forgetting. Algorithm choice is therefore not only about memory use. Teams must compare reasoning coverage, effective updates, and sampling cost across their own models, tasks, and budgets.

Key takeaways:

• Track Pass@K alongside Pass@1 to expose differences in reasoning-path coverage.
• GRPO and ES may work better in sequence than as an either-or choice.
• Test ES against your model scale, task type, and compute budget before drawing broad conclusions.

Source: Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO