# WAM Survey: How Prediction Serves Control

WAM Survey: How Prediction Serves Control

Full article: https://haiguangboy.com/posts/wam-survey

## Core Methods

↳ ① Modeling: joint prediction, generating future states and actions simultaneously; inverse dynamics, predicting future states first, then inferring actions from state changes. The survey does not determine which route is generally better.

↳ ② Architecture: shared backbone emphasizes unified representation and joint training; dual-system separates the world module and the action module, connecting them through an intermediate interface, which facilitates separate adaptation but makes interface design more complex. The modeling route and how modules are organized need to be compared separately.

↳ ③ Data: internet and human videos mainly provide visual dynamics, motion, and interaction priors; robot trajectories provide the correspondence between observations, actions, and outcomes. Larger-scale video data still needs action alignment before it can connect to control. Frozen parameters, alignment layers, or residual action heads are all adaptation methods that preserve priors.

## Key Results

World models pursue prediction accuracy, while policies pursue higher returns, and the two objectives are not naturally aligned. The authors caution that joint optimization may produce overly optimistic future predictions that are exploited by the policy. Training the world model and policy separately is one route, but model bias and module coordination issues still remain.

## Comparison of Related Routes

- Opposing route · [OpenWAM: The right answer changes with scale](https://haiguangboy.com/posts/openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra) `openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra_2026_09`: OpenWAM: The right answer changes with scale
- Opposing route · [PAVE: Failure data can also feed the policy](https://haiguangboy.com/posts/pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies) `pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies_2026_09`: PAVE: Failure data can also feed the policy
- Opposing route · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: World prediction and action selection can be optimized separately contradicts route bet: the behavior/action part must be a unified model, opposing modular assembly
- Opposing route · [World Models: Should we imagine at inference time](https://haiguangboy.com/posts/omega-eva) `omega_eva_envision_verify_and_act_with_latent_interactive_world_models_2026_08`: World prediction and action selection can be optimized separately contradicts core claim: what the world model should do is not assist during training, nor serve as an external simulator, but act as a verifier inside the action generation step
- Opposing route · [Lower prediction error, but the representation has already collapsed](https://haiguangboy.com/posts/slim-05b) `slim_05b_learning_action_grounded_predictive_latents_for_robot_manipulation_2026_08`: Neural simulators can support closed-loop policy improvement contradicts core claim: future prediction should be a training signal that shapes representations, not the model's output
- Same route · [D4RT_Efficiently_Reconstructing_Dynamic_Scenes_One](https://haiguangboy.com/posts/d4rt_efficiently_reconstructing_dynamic_scenes_one) `d4rt_efficiently_reconstructing_dynamic_scenes_one_2026_06`: 2D video prediction does not automatically guarantee 3D and multi-view consistency validates D4RT output is relative/up-to-scale, not metric—this is the crux when serving as an embodied 3D data source

## Boundaries

⚠️ In deployment, three kinds of "good" cannot be equated
· Realistic visuals do not prove actions are executable; low trajectory error does not guarantee closed-loop safety; a high success rate also cannot alone explain whether the gains come from prediction or from policy priors.
· 2D video coherence does not guarantee geometric, occlusion, and multi-view consistency, and errors will propagate to planning and action decoding.
· Multi-step sampling, imagined trajectories, and action decoding all delay feedback. Sampling steps, prediction horizon, number of candidate trajectories, and replanning frequency must all be factored into control latency.

## Author's Judgment (not paper content, cross-paper synthesis view)

💡 Viewed together with OpenWAM
OpenWAM found in its evaluation: without embodied pretraining, unidirectional information flow is slightly superior; after adding pretraining, the advantages of bidirectional interaction emerge. This adds a condition to the survey's classification: architectural pros and cons change with pretraining conditions, and one cannot choose sides based only on "unified" or "decoupled."

Looking back half a year later, what is worth testing is: with fixed data and compute budgets, can separately optimizing the world and the policy reduce real execution degradation caused by model bias. This result is what helps judge the value of separate optimization.