Driving · 2026-11-03
pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
Keivalya Pandya
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
pico-WAM is a compact world action model with a shared latent and explicit physical state readout to isolate whether control failures come from proposal generation or trajectory scoring. Using Push-T (206 demos, 15 seeds), counterfactual branch training improves cost ranking and increases policy-restarted vs warm-started coverage by 4.8 points. Using exact simulator costs for scoring boosts coverage by 12.2 points while keeping learned proposals; simulator-based proposal generation alone shows no detected gain.
Why it matters
The authors diagnose which component (ranking/scoring vs proposal generation) improves planning for compact world action models under controlled interventions. They show that better candidate ranking does not necessarily translate into better closed-loop control, and that exact scoring is an effective improvement target in their tested setting. The paper also provides a reproducible protocol for isolating policy–dynamics interactions and separating data-by-search and scoring effects.
Method
- Introduce pico-WAM: a shared latent with an explicit physical state readout feeding a policy and residual dynamics, enabling traceable observe→propose→imagine→score→execute steps.
- Train with demonstration-only windows vs windows augmented by counterfactual branches (Gaussian-noised action perturbations at selected states), then test planning with warm-started vs policy-restarted search.
- Run component factorial tests that separately replace learned trajectory scoring with exact simulator costs (while optionally retaining learned policy proposals).
Limitation
These interventions explain behavior in a controlled setting; they do not establish a new architecture, hardware transfer, or general superiority of planning. The shared-model tests use supervised state in Push-T, and the door experiment uses a different architecture.
Abstract (from arXiv)
When a World Action Model fails at control, is the bottleneck proposal generation or trajectory scoring? We introduce pico-WAM, a compact model with supplied physical state for tracing policy-dynamics interactions. Using 206 Push-T demonstrations and 15 training seeds, we compare data regimes and search configurations. Counterfactual branch training improves short-horizon cost ranking and enlarges the coverage advantage of policy-restarted over warm-started search by 4.8 percentage points; restart wins under both data regimes. Replacing learned trajectory scores with simulator costs improves coverage by 12.2 points while retaining learned proposals. Substituting simulator states in proposal generation alone shows no detected gain. Component diagnostics and a separate door-opening study show why better ranking need not yield better control or outperform imitation. These controlled interventions identify scoring as an effective improvement target in the tested setting and provide a reproducible protocol for diagnosing compact world models.
Related papers
- Rollback: Experience Backtracking in World Models for Policy Improvement
- How Much Planning Is Enough? Reducing Search and Computation in World-Model Planning
- When Low Prediction Error Misleads Planning: Diagnosing Representation, Dynamics, and Decision Failures in Latent World Models
- Computer-Use Agents Gain Little Benefit from a World Model When Conditioned on Tutorials
- AnyGoal: On World Model Generalization Beyond Expert Rollouts
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-03 · Indexed by Pub-AI: 2026-11-03 UTC
pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
First announced by arXiv: 2026-11-03 · Original record ↗
- arxiv
- 2611.01039
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-03 · First detected: 2026-11-03 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.01039v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
When a World Action Model fails at control, is the bottleneck proposal generation or trajectory scoring? We introduce pico-WAM, a compact model with supplied physical state for tracing policy-dynamics interactions. Using 206 Push-T demonstrations and 15 training seeds, we compare data regimes and search configurations. Counterfactual branch training improves short-horizon cost ranking and enlarges the coverage advantage of policy-restarted over warm-started search by 4.8 percentage points; restart wins under both data regimes. Replacing learned trajectory scores with simulator costs improves coverage by 12.2 points while retaining learned proposals. Substituting simulator states in proposal generation alone shows no detected gain. Component diagnostics and a separate door-opening study show why better ranking need not yield better control or outperform imitation. These controlled interventions identify scoring as an effective improvement target in the tested setting and provide a reproducible protocol for diagnosing compact world models.
