# JEPA-WAM: An Innovative Approach That Makes World Models Better at Understanding State Changes

Without Relying on Data Accumulation, This Innovative World Model Method Boosts π0.5 Real-Robot Success Rates by Up to 12.8 Percentage Points

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

## Core Method

Instead of attaching a prediction module alongside the policy, a single predictor—Qwen2.5-0.5B—handles two tasks at once: in one forward pass, outputs at visual token positions predict transitions, while action conditioning is read from dedicated tokens. Both paths share the same weights, and transition supervision updates the very parameters that generate actions

The prediction target: a frozen V-JEPA 2.1 encodes the current frame and the frame δ steps ahead into a joint target, which the predictor approximates. The loss is per-patch cosine distance, with no image generation involved throughout

Four design choices, each backed by ablation studies:
1️⃣ The target uses a joint representation of current and future states, not an isolated future—predicting only the future drops 1.9%
2️⃣ Preserving spatial correspondence between dense patches—mixing adjacent patches drops 4.5%
3️⃣ Supervision is applied to the final output of the shared backbone—applying it only to intermediate layers drops 2.7%
4️⃣ Actions read only dedicated action tokens, isolating world prediction from control interference—reading all predicted hidden states drops 6.1%
At deployment, the prediction head is simply removed, adding no inference burden

## Key Results

📚 Simply switching to V-JEPA 2.1 lifts LIBERO-Plus from 73.2% to 77.0%; adding the full transition learning mechanism reaches 79.2%

🧲 The comparison with LeapBot-WA is even more telling: both use V-JEPA 2.1, but the latter relies on a separate Anchor DiT to learn dynamics knowledge and distill it into an Action DiT, whereas JEPA-WAM lets transition prediction directly supervise the shared action backbone. On the same zero-shot benchmark, it's 79.2% versus 73.1%, a 6.1 percentage point gap—the gain can't be attributed solely to the visual encoder; how world knowledge enters the policy is equally critical!

When integrated into π0.5, LIBERO-Plus rises from 84.5% to 86.3%, proving transferability to mature VLAs. Across five real-robot bimanual tasks: ID improves from 77.5% to 90.3% (+12.8 percentage points), and OOD from 72.5% to 84.7% (+12.2 percentage points)—transferred knowledge ultimately translates into real success rates

## Comparison of Related Approaches

- Opposite route · [The more accurate future prediction, the more sluggish the robot becomes](https://haiguangboy.com/posts/omega-0) `omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08`: Core claim: transition supervision must directly shape the backbone that generates actions; external modules or using it as context don't work contradicts coupling mechanism: motion queries attend to video queries in reverse, injecting predicted future visual dynamics into action representations
- Opposite route · [Tactile as prediction target rather than observation input](https://haiguangboy.com/posts/n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens) `n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens_2026_07`: Tactile as prediction target rather than observation input
- 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`: Representations maintain patch-level dense structure, avoiding global pooling or compression into few tokens contradicts ISAE's three losses: semantically aware reconstruction + KL + SIGReg slice isotropy regularization
- 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
- Same route · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: Core claim: transition supervision must directly shape the backbone that generates actions; external modules or using it as context don't work validates route bet: behavior/action components must be a unified model, opposing modular assembly
- Same route · [π0.5 shudders in place when failing to grab a spoon, but Orca goes further with physics intuition from watching videos](https://haiguangboy.com/posts/orca) `orca_2026_07`: The key to a world model is readable states

## Limitations



## Author's Assessment (Not from the paper; cross-paper synthesis)

The more practical significance lies in this: well-funded companies can push scaling through data and compute, but resource-constrained teams can still make headway via training objectives, supervision signals, and information structure. Just as DeepSeek-R1 improved performance through training method innovations under compute constraints

🧲 JEPA-WAM also demonstrates: without competing head-on with giants on scale, method innovation can still enhance the capability conversion efficiency of mature foundations