# Lower prediction error, but the representation has already collapsed

A world model can predict the future, but that doesn't mean it can control it—given actions it predicts outcomes, given outcomes it infers actions, with only 0.47B

Full article: https://haiguangboy.com/posts/slim-05b

## Core method

↳ ① Change the modeling unit. Past world models modeled "how the current state becomes the future state"; SLIM switches to "state-action-state" triples: actions are no longer a side condition when predicting the future, but the core variable for judging whether a representation is qualified for control. It's not the first to feed actions into a world model, but it makes "action identifiability" an explicit criterion for representations.
↳ ② The objective must therefore be bidirectional. Mask the action, and infer it from surrounding observations; mask the future latent, and predict it from current observation plus action. Forward and inverse are not two stitched-together losses—they are two directions of the same control relation.
↳ ③ Future prediction is only a training signal, not an output. Stage 2 removes future latents entirely from inputs and supervision, leaving only shaped predictive slots for the action stream to read; no future is generated at inference.

## Key results

· 0.47B: LIBERO 97.5%, zero-shot LIBERO-Plus 77.45%, CALVIN ABC→D 4.556.
· Single inference 60.6ms / 490 GFLOPs per action block, vs π0.5's 193.1ms / 4715. Note both use their native configs (SLIM 8-step blocks with 4 samples, π0.5 with 10 and 10), not a strictly matched comparison.
· Real-robot five tasks: average progress highest across nominal, distractor, and lighting tiers.

## Related route comparison

- Opposite route · [More accurate future prediction makes robots more sluggish instead](https://haiguangboy.com/posts/omega-0) `omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08`: Stage 2 removes future latents entirely, leaving only predictive slots shaped by Stage 1 contradicts coupling mechanism: motion query instead attends to video query, injecting predicted future visual dynamics into action representations
- Opposite route · [Tactile prediction should only be used for training, not shown to actions](https://haiguangboy.com/posts/tacwam) `tacwam_anchor_guided_world_action_model_with_mechanics_aware_tactile_prediction_2026_08`: MoT dual-stream: observation and action streams share joint attention, language only does per-stream cross-attention contradicts AGT attention: action tokens are forbidden from reading future visual/tactile prediction tokens—an information isolation mechanism, not action-conditioned dynamics
- Opposite route · [Predictive features cannot be fed directly to diffusion models](https://haiguangboy.com/posts/leapbot-wa) `leapbot_wa_world_anchor_action_models_via_predictive_latent_alignments_2026_07`: Defines the objective as action-grounded: infer actions from observation changes and predict latent transitions from actions contradicts key design: the world model only sees intent, not actions—deliberately depriving low-level motor commands to prevent collapsing into a control shortcut
- Opposite route · [First human-to-robot zero-shot transfer scaling law](https://haiguangboy.com/posts/dyna-2-scaling-law) `dyna_blog_dyna_2_20260811_2026_08`: Problem setup: VLAs only implicitly learn action-observation change relations, and pixel-level prediction wastes compute on appearance details contradicts route bet: the cause of emergent cross-embodiment transfer is world modeling (predicting future video) itself, not just data scale—WAM architecture outperforms pure VLA
- Same route · [JEPA-WAM: An innovative method to make world models better understand state changes](https://haiguangboy.com/posts/jepa-wam) `jepa_wam_learning_vision_language_action_policies_with_joint_embedding_world_modeling_2026_08`: Defines the objective as action-grounded: infer actions from observation changes and predict latent transitions from actions validates core claim: what should be predicted is the transition relation, not the absolute future state
- Same route · [LAWM: Why action labels become a burden](https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling) `latent_action_pretraining_through_world_modeling_2026_07`: LAWM: Why action labels become a burden

## Limitations

Stage 1 mixes in LIBERO-90, and only Stage 2 uses the four target suites—so Stage 1 gains can't be fully credited to the bidirectional masking; the paper lacks a data-matched ablation. Also, it loses on the background-change tier: 49 vs π0.5's 54, despite criticizing pixel prediction for wasting capacity on background texture. Real-robot runs are only 10 per cell, reporting partial scores rather than success rates.

## Author's take (not from the paper; cross-paper synthesis)

The most counterintuitive finding is the EMA ablation. Removing the EMA target drops the future latent prediction MSE from 0.245 to 0.166—lower. But effective rank collapses from 61.28 to 13.95, and LIBERO-Plus drops 10.6 points. Looking only at loss, you'd conclude "removing EMA is better."

This is already the third work showing "offline prediction metrics diverge from real-robot control performance." LeapBot-WA's SIGReg ablation shows the same dimensional collapse: rank 38.1→92.3, success rate 42.5%→71.3%. omega-0 is a different failure—swapping in the Wan encoder gives more accurate offline future reconstruction, yet real-robot execution becomes sluggish and hesitant, with no collapse there.

So the shared conclusion isn't "latent prediction always collapses," but: a single prediction loss is insufficient to judge whether a representation serves control. Effective rank and energy concentration only catch geometric degradation; control probes like action identifiability and closed-loop success rate must be added separately.