JEPA Deep Dive: From I-JEPA to V-JEPA 2-AC — How Predictive Representation Learning Leads to World Models

In my previous world model roundup, I covered Cosmos, Genie 3, and Marble in detail but only sketched JEPA.

Not because JEPA isn’t important. Quite the opposite — JEPA may be the most theoretically deep path in the current world model landscape. It’s backed by Yann LeCun, has a complete technical evolution from I-JEPA to V-JEPA 2-AC, has been validated by AMI Labs’ $1.03 billion funding round, and represents a fundamentally different technical approach from world models whose primary training objective is pixel generation.

This article provides a complete technical breakdown of the JEPA series from the first paper to the latest.

1. JEPA’s Core Idea: The One-Sentence Version

Don’t predict pixels — predict representations.

This sounds simple, but it forms a sharp contrast with a class of world models that use pixel or observation generation as their primary training objective. JEPA does not claim that generating pixels has no value; rather, it argues: if the goal is to learn abstract world representations useful for prediction, understanding, and decision-making, then requiring the model to reconstruct all observation details may not be the optimal learning objective.

Generative world models work like this: given historical observations, predict future pixels. JEPA works like this: given historical observations, predict abstract representations of future observations — make predictions in representation space, not pixel space.

Why does this distinction matter? Because pixel space contains a large amount of task-irrelevant details: lighting changes, texture details, random noise. Requiring the model to precisely predict all these details wastes model capacity and introduces useless gradient signals that slow down learning.

LeCun made this point clearly in his 2022 position paper: for learning high-level semantics and world state representations, requiring the model to precisely predict all pixels is not an ideal learning objective.

This isn’t to say pixel-level prediction has no value — video generation is certainly impressive. JEPA’s argument is: if your goal is to learn world representations useful for downstream tasks, then predicting in representation space is a more efficient learning strategy.

But this strategy leaves a core question that runs through this entire article: the “unpredictable pixel details” that JEPA actively filters out — could they be precisely the information that robot control needs? I’ll return to this in section 9’s open questions.

2. I-JEPA: Proving It Works on Images (2023)

Paper Info

Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture, Meta AI, CVPR 2023. Paper, Code.

Architecture

I-JEPA has three core components:

Context Encoder: A Vision Transformer that receives visible image patches and outputs their representations.

Target Encoder: Same architecture as the Context Encoder, but parameters are updated via exponential moving average (EMA) — it doesn’t directly participate in gradient computation but slowly tracks the Context Encoder’s parameter changes. It encodes masked regions into target representations.

Predictor: A much lighter Transformer (embedding dimension 384 vs the encoder’s larger dimension) that receives the Context Encoder’s visible representations and predicts the masked regions’ representations in the Target Encoder’s space.

Masking Strategy

This is where I-JEPA differs fundamentally from MAE (Masked Autoencoder).

MAE masks at the pixel level — randomly masking 75% of patches and asking the model to reconstruct those pixels.

I-JEPA’s masking strategy shares a similar “mask-and-predict” structure with MAE, but the key difference lies not in “where to mask” but in what the prediction target is. Specifically: I-JEPA uses a multi-block strategy, selecting 4 target regions (each 15%-20% of the image) and 1 context region (85%-100%), removing overlaps. The Context Encoder only receives patches from the context region, the Target Encoder encodes patches from the target region, and the Predictor maps context representations to target representations. The final optimization uses representation-space prediction loss, not pixel reconstruction loss.

This means: the model doesn’t need to reconstruct masked content from the pixel level. Instead, it needs to infer what the masked regions should look like in abstract representation space from the visible regions’ representations.

Key Results

I-JEPA with ViT-H/14 on ImageNet:

• Linear probing: 79.3% top-1 accuracy (vs MAE’s 77.2%)
• 448 resolution: 81.1% top-1
• 1% low-shot: 73.3% (vs MAE’s 59.8%) — this gap is enormous
• Full fine-tuning: 87.1% (300 epochs), compared to MAE’s 1600 epochs at 87.8%

The last point is particularly noteworthy: at comparable final performance, I-JEPA requires approximately 5.3× fewer fine-tuning epochs. This suggests strong training efficiency potential for the representation prediction objective, though it cannot be simply equated to 5.3× end-to-end training time savings — different methods may have different batch sizes, data augmentation, GPU utilization, and other factors.

Why It Matters

I-JEPA’s significance isn’t about SOTA numbers. It cleanly demonstrates one thing: self-supervised learning that doesn’t generate pixels and only predicts in representation space not only works but shows strong training efficiency in visual representation learning tasks.

This is the first cornerstone of the JEPA path.

3. V-JEPA: From Images to Video (2024)

Paper Info

V-JEPA is Meta’s representative work on the video feature-prediction path. The core paper Revisiting Feature Prediction for Learning Visual Representations from Video systematically demonstrates this paradigm.

Core Extension

If I-JEPA proved “predicting image regions in representation space works,” V-JEPA goes further: can we predict masked spatiotemporal regions in video within representation space, and thereby learn dynamics-relevant visual representations?

This is much harder than images. Masked regions in images are spatially fixed — the model just needs to understand spatial structure. But spatiotemporal prediction in video involves motion patterns, temporal dependencies, and scene evolution, requiring modeling of far more complex temporal structure than spatial prediction on images.

V-JEPA extends the I-JEPA framework to the spatiotemporal dimension. The Context Encoder receives spatiotemporal patches from historical video frames, and the Predictor predicts abstract representations of the masked spatiotemporal regions in representation space.

Technical Highlights

V-JEPA maintains JEPA’s core design philosophy:

• Target Encoder still updated via EMA
• Prediction still happens in representation space, never returning to pixel space
• Masking strategy extends from spatial to spatiotemporal — masking parts of spatiotemporal regions

Key Contribution

V-JEPA’s core contribution is proving the JEPA framework can naturally extend to video, and the learned representations perform well on downstream tasks like action recognition and video understanding. Under frozen representation evaluation, V-JEPA’s ViT-H/16 model achieved 81.9% on Kinetics-400, 72.2% on Something-Something-v2, and 77.9% on ImageNet-1K.

From the perspective of the JEPA path’s evolution, V-JEPA serves as a key transition from image representation learning to video temporal modeling: it demonstrated that latent video prediction can learn powerful spatiotemporal representations. What pushed this path further into “action-conditioned prediction + planning” was V-JEPA 2.

4. V-JEPA 2: From Representation Learning to World Model (2025)

Paper Info

V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction, and Planning, Meta AI, first released June 2025, updated March 2026. Paper.

This Is the Most Important One

If I-JEPA and V-JEPA are basic research, V-JEPA 2 is where the JEPA path first fairly completely demonstrates that action-conditioned prediction can serve real robot planning.

Architecture Upgrade

V-JEPA 2’s architecture has major upgrades:

Encoder: Vision Transformer, available in multiple sizes — ViT-L (~300M parameters), ViT-H (600M parameters), ViT-g (1B parameters). Video input is split into 2x16x16 tubelets (2 frames × 16 × 16 pixels). Positional encoding uses 3D-RoPE (3D Rotary Position Embedding), a key design for processing spatiotemporal sequences.

Action-conditioned variant (V-JEPA 2-AC): This is V-JEPA 2’s most important addition. A ~300M parameter predictor Transformer using block-causal attention, receiving action sequences as conditional input to predict future representations. Importantly, the predictor’s inputs are not just video representations and actions — in the robot experiments, they also include end-effector state (robot end-effector status). The paper explicitly states “conditioned on past video frames, actions, and end-effector states.” Therefore, V-JEPA 2-AC’s actual prediction form is closer to z_{t+1} = f_θ(z_{≤t}, a_{≤t}, s_{≤t}), where s_t is the robot end-effector state, rather than a simple Markov p(z_{t+1}|z_t, a_t). At an abstract level, it can be understood as an action-conditioned latent transition model; but the concrete implementation is an autoregressive predictor with historical context, actions, and end-effector state.

This means: the model can not only “understand” video but also “imagine” what the world would look like if a certain action were executed. This is a key step toward world model capability.

Pretraining

The pretraining scale is substantial:

• Pretrained on over 1 million hours of internet video and approximately 1 million images
• Progressive resolution strategy (starting from low resolution, gradually increasing)
• Mask-denoising feature prediction objective

Key Results

Video understanding:

V-JEPA 2 reported results across multiple video understanding benchmarks, organized by category:

Representation / Action Understanding: Something-Something-v2 77.3% top-1, Diving48 90.2%

Anticipation: EPIC-KITCHENS-100 39.7 R@5

Video QA / Physical World Understanding (after alignment with 8B language model): PerceptionTest 84.0, TempCompass 76.9, TemporalBench 36.7, TOMATO 40.3

It surpasses InternVideo2 and DINOv2. Note that these benchmarks use different evaluation metrics and cannot be simply averaged across categories.

Robot manipulation (this is the most exciting part):

V-JEPA 2-AC was post-trained on approximately 62 hours of robot trajectory data with action-conditioned post-training on top of the pretrained V-JEPA 2, then tested for robot manipulation without task-specific training or environment adaptation. The “zero-shot” here should more precisely be understood as robot-environment zero-shot / task zero-shot — generalization to new environments, new objects, and new tasks — rather than interaction-data zero-shot (learning robot control without any robot interaction data).

The specific task results reported in the paper:

Target reaching accuracy is less than 4 cm. Under the same RTX 4090, CEM planning setup reported in the paper (horizon=1, image-goal), Cosmos baseline requires approximately 4 minutes per action (80 samples), while V-JEPA 2-AC, even using 10× more candidate samples (800 vs 80), requires only about 16 seconds — approximately 15× faster. Note that horizon=1 means this result reflects single-step planning latency difference, not a long-horizon rollout advantage; the paper itself acknowledges that long-horizon rollout encounters prediction degradation and search-space explosion.

This result shows that under the robot, planner, and task setup given in the paper, V-JEPA 2-AC’s latent-space planning demonstrates clear advantages for object manipulation — V-JEPA 2-AC performs comparably to Cosmos on the reach task (100% vs 80%), and shows even larger advantages on object interaction tasks (Cosmos achieves 0% success on cup/box pick-and-place, while V-JEPA 2-AC reaches 80% and 50% respectively). However, this comparison is not yet sufficient to conclude that JEPA generally outperforms generative world models across all robot control tasks.

Why V-JEPA 2-AC Is the JEPA Path’s Turning Point

V-JEPA 2-AC demonstrates for the first time:

1. The JEPA framework can scale to the billion-parameter level
2. Action-conditioned prediction can serve real robot control
3. Predicting in representation space is not just theoretically elegant but also more efficient in practice than pixel-level prediction
4. After post-training on a small amount of robot data, the model can zero-shot generalize across multiple manipulation tasks — no need to retrain for each specific task

This is exactly the blueprint LeCun drew in his 2022 paper: using predictive representation learning to build AI systems that understand the physical world. Three years later, V-JEPA 2-AC has begun to demonstrate that this blueprint can be partially realized on limited-scale robot tasks — but the distance to a “complete world model” remains significant.

What Does V-JEPA 2-AC Actually Solve?

It’s worth breaking down V-JEPA 2-AC’s capabilities into three layers:

Layer 1: representation. Compress visual observations into task-relevant latent representations — filtering out unpredictable pixel details while retaining abstract information useful for downstream tasks.

Layer 2: prediction. Given action sequences, predict future latent states in representation space — not predicting pixels, but predicting “how the task-relevant state of the world will change.”

Layer 3: planning. Perform latent rollout over multiple candidate actions and select the action with highest expected return — this is what CEM planning does in the paper.

Connecting these three layers:

observation
    ↓
representation (encode current state)
    ↓
action-conditioned dynamics (predict action consequences)
    ↓
latent rollout (imagine multiple future trajectories)
    ↓
planning (select best action)
    ↓
action

This is actually very close to Dreamer’s core loop. But the starting points differ: Dreamer’s latent dynamics was designed from the beginning as the core loop of model-based RL; V-JEPA 2-AC started from a large-scale visual representation model and acquired action-conditioned prediction capability through robot trajectory data. One path goes “from control, learn good representations”; the other goes “from representations, acquire control capability.”

Sidebar: V-JEPA 2.1 (2026)

In March 2026, Meta released V-JEPA 2.1. Unlike V-JEPA 2-AC’s focus, V-JEPA 2.1 primarily advances dense visual representation: through dense predictive loss, deep self-supervision, and multimodal tokenizers, it makes representations more fine-grained and consistent in both spatial and temporal dimensions. The paper also reports further results on Ego4D short-term object interaction anticipation (7.71 mAP), EPIC-KITCHENS anticipation (40.8 R@5), Something-Something-v2 (77.7), navigation, depth estimation, and real robot grasping — with grasping improvement of approximately 20 percentage points over V-JEPA 2-AC.

Notably, V-JEPA 2.1 and V-JEPA 2-AC are not simply “the next version of action-conditioned world model” — rather, V-JEPA 2.1 is more like the continued evolution of the JEPA representation backbone, improving representation quality and spatial granularity rather than directly advancing action-conditioned dynamics. V-JEPA 2.1’s deeper significance may lie in providing empirical supporting evidence from another direction that “representation quality is a bottleneck for world-model capability”: even without directly expanding the action-conditioned dynamics model, improving dense spatial-temporal representation itself can significantly improve robot task performance. This is at least a strong empirical signal, but these results alone cannot yet prove that representation quality is the primary bottleneck for world-model scaling — this judgment is this article’s inference based on V-JEPA 2.1’s results.

Therefore, this article continues to use V-JEPA 2-AC as the key node for “JEPA moving toward action-conditioned world model” and does not mix V-JEPA 2.1 into this control path. The JEPA path continues to evolve, and V-JEPA 2.1 demonstrates the vitality of this framework.

5. AMI Labs: An Industrialization Observation (2026)

AMI Labs (Advanced Machine Intelligence Labs) officially launched in March 2026, announcing approximately $1.03 billion** (approximately €890 million) in seed funding at a pre-money valuation of approximately **$3.5 billion (according to company announcement and reports from TechCrunch, Reuters, and others) — one of the largest seed rounds in European AI foundation model history. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and others, with participation from NVIDIA, Temasek, Samsung and other institutional investors, as well as individual investors including Jeff Bezos and Eric Schmidt.

The core team includes Yann LeCun (Executive Chairman), Alex LeBrun (CEO), Saining Xie (Chief Science Officer), Michael Rabbat (VP of World Models), and Pascale Fung (Chief Research and Innovation Officer).

AMI Labs’ stated technical direction is: building world models that understand physical environments, retain long-term information, and execute logical planning — focusing on learning abstract representations from multimodal sensor inputs, filtering out unpredictable details, and predicting outcomes in conceptual space. Notably, AMI Labs explicitly emphasizes action-conditioned world models.

Based on public personnel backgrounds and technical vision, it is reasonable to speculate that JEPA-style predictive representation learning will be one of AMI Labs’ important research directions; however, this is inference based on publicly available information, not a confirmed architectural conclusion officially announced by AMI.

6. JEPA vs Dreamer/RSSM: Both Predict in Latent Space — What’s the Difference?

If you’ve read my RSSM deep dive, you’ll notice an obvious similarity between JEPA and RSSM/Dreamer: both predict in latent space rather than directly predicting pixels.

But they’re not the same thing.

Similarities

Both recognize pixel space isn’t a good prediction target. RSSM compresses observations into latent states via an encoder, then uses a dynamics model to predict the next state in latent space; JEPA encodes observations into representation space via an encoder, then uses a predictor to predict future representations.

Core Differences

Training objectives differ. RSSM/Dreamer’s latent dynamics model ultimately serves RL — it needs an explicit state → action → next state structure to support planning and control. JEPA’s training objective is self-supervised representation learning — it doesn’t explicitly model the separation of state and action, instead learning useful representations through masking and prediction.

Action’s role differs. In Dreamer, action is an explicit input to the dynamics model — state transitions directly depend on action. In the original I-JEPA and V-JEPA, there’s no concept of action. Only V-JEPA 2 introduced an action-conditioned variant, but its action conditioning differs from Dreamer’s RSSM — V-JEPA 2 uses block-causal attention to process action sequences and visual representations together.

Application scenarios differ. Dreamer’s core scenario is RL and control — imagining futures, evaluating actions, selecting optimal policies. JEPA’s core scenario is representation learning and understanding — learning world representations useful for downstream tasks. V-JEPA 2’s robot experiments show JEPA can also do control, but differently from Dreamer.

One-Sentence Summary

Dreamer is “using latent dynamics models for planning”; JEPA is “using predictive representation learning for understanding.” Both are clever but solve different problems.

In more formal terms: Dreamer/RSSM’s core is a latent state-space model designed for control — it explicitly maintains a deterministic recurrent state h_t and a stochastic latent state z_t, with action-conditioned transitions supporting imagined rollouts, value estimation, and policy learning. In summary, its dynamics can be expressed as h_t = f(h_{t-1}, z_{t-1}, a_{t-1}), z_t ~ p(z_t|h_t), rather than a simple p(z_{t+1}|z_t, a_t). For JEPA, V-JEPA 2-AC uses autoregressive feature prediction, predicting future video representations conditioned on past video representations, actions, and end-effector states. The paper also discusses error accumulation in multi-step autoregressive rollout and identifies long-horizon prediction stability as a direction for future work. sg denotes stop-gradient, and the prediction target is the representation output by the target encoder rather than the raw observation. The former is designed from the start to serve model-based RL rollout and planning; the latter’s core contribution lies in “what to predict” rather than “how to rollout.”

It’s also worth noting a structural difference: Dreamer/RSSM’s latent state is by design s_t = (h_t, z_t) — a deterministic recurrent state plus a stochastic latent state, together forming the state variable for the dynamics model and imagination. V-JEPA’s representation, by contrast, is first and foremost a high-dimensional patch-level visual representation learned for visual understanding and prediction — it is not a pre-defined Markov latent state. V-JEPA 2-AC demonstrates on top of this that these representations can further support action-conditioned dynamics, rather than assuming from the start that they are complete Markov states. This is also why V-JEPA 2-AC’s predictor needs to make future representation predictions at the feature map / token level.

7. JEPA’s Position in the World Model Landscape

Going back to the four technology paths from my roundup article:

• Latent Dynamics (Dreamer/RSSM): state → action → next state
• Generative Video (Cosmos): condition → future video frames
• Interactive (Genie 3): state + action → interactive future
• Spatial/3D (Marble): persistent spatial representation

JEPA doesn’t cleanly fit into any of these. It’s closest to Latent Dynamics since it also predicts in latent space. But unlike Dreamer, it doesn’t have an explicit state transition structure, nor is RL and planning its primary goal.

If I had to categorize it, I’d say JEPA represents a fifth path: Predictive Representation Learning (this “fifth path” classification is proposed in this article for analytical convenience, not a taxonomy uniformly adopted in existing literature). Its core contribution isn’t “how to model world dynamics” but “what objective function to use for learning world representations.” That said, more precisely, these five dimensions are not strictly mutually exclusive categories — future systems could well combine JEPA representation + RSSM dynamics + language conditioning + 3D spatial memory. JEPA is therefore more like a representation/prediction paradigm orthogonal to world-model architecture, not merely a fifth kind of world model.

More precisely, JEPA is a class of predictive representation learning architecture / objective family, not just a specific world-model architecture. From a higher-level perspective, it embodies a modeling philosophy of “first learn predictable, decision-useful abstract states, rather than reconstructing all observation details.”

When Does JEPA Become a World Model?

This question deserves separate discussion because it directly affects how we understand the entire technology path.

I-JEPA itself is better described as predictive representation learning, not a complete world model. Its goal is to learn image representations useful for downstream tasks — no temporal dynamics modeling, no concept of action.

V-JEPA introduces temporal prediction, but its primary goal remains learning video representations — it predicts future frame representations but doesn’t support action-conditioned planning.

V-JEPA 2-AC goes further by adding action-conditioned prediction and using model-based planning to predict action consequences on robots. The paper itself refers to it as a latent action-conditioned world model — latent dynamics that predict action consequences in representation space. Under this definition, V-JEPA 2-AC can reasonably be viewed as a latent action-conditioned world model; however, whether it has learned a sufficiently complete, long-term stable, control-sufficient world state remains an open question.

Therefore, to be precise, I would call I-JEPA / V-JEPA “JEPA representation learning” and V-JEPA 2-AC a “JEPA-based latent world model” under the looser definition. This distinction matters — it lets us understand more precisely which parts of the JEPA path are representation learning and which parts are world modeling.

JEPA Path Technical Overview

Pulling together the previous sections, the JEPA path’s evolution is not a simple linear chain “LeCun → JEPA → I-JEPA → V-JEPA → V-JEPA 2 → robots → AMI Labs.” A more accurate structure is a progressively branching tree:

JEPA (2022 theoretical blueprint)
│
├── I-JEPA (2023)
│     └── Image representation prediction: proving "predicting representations ≠ pixel reconstruction" works
│
├── V-JEPA (2024)
│     └── Video representation prediction: from spatial to spatiotemporal
│
└── V-JEPA 2 (2025.06 initial, 2026.03 updated)
      ├── Video understanding (avg 88.2 across 6 benchmarks*)
      │     *Note: 88.2 is the paper's average across six benchmarks, not a single dataset accuracy
      ├── Latent video prediction
      ├── Action-conditioned prediction (V-JEPA 2-AC)
      └── Robot planning demonstration (robot-environment / task zero-shot)
              │
              └── Still open:
                   ├── Long-horizon rollout stability
                   ├── Representation sufficiency (state sufficiency)
                   ├── Robust control
                   └── Causal / counterfactual validity

AMI Labs (officially launched March 2026)
└── Industrial research direction
      └── Highly related to JEPA philosophy, but specific architecture not yet public

Two things are worth noting about this structure. First, V-JEPA 2 is not a single achievement — it simultaneously encompasses four layers of contribution: video understanding, latent prediction, action-conditioned prediction, and robot planning. The first three are at the video/representation level; only the fourth directly involves control. Second, the relationship between AMI Labs and V-JEPA 2 is one of “technical philosophy continuity” rather than “architectural inheritance” — equating AMI Labs with simply “the company version of V-JEPA 2” is technically imprecise.

JEPA’s Core Problem Is Not Prediction, But State Sufficiency

One of the most worth-discussing questions in the JEPA path is the relationship between “predictable representations” and “complete world states.”

For a world model, what truly matters is not that the latent prediction error is low enough, but whether the latent state has sufficient statistical adequacy for future decision-making. More formally, an ideal latent representation z_t = E(o_{≤t}) should at least satisfy: for the future that downstream control cares about, conditioning on z_t should eliminate the need to access the compressed-away raw observation information. In other words, the question JEPA needs to answer is not “can it predict representations” but “does this representation retain the information needed for action-conditioned planning?”

This is precisely the watershed between representation learning and world modeling.

A representation can be excellent for video understanding or action recognition without necessarily retaining all the information needed for planning. An ideal world state should at least satisfy: from current state z_t and action a_t, being able to predict future relevant states accurately enough and supporting stable multi-step rollout for planning. But JEPA’s representations actively discard details that are unpredictable with respect to the prediction target — the question is: what information does the model judge as “unpredictable/irrelevant”? For control, some information that is unpredictable for observation prediction (micro-contact states, friction, deformable object state, occluded object state, tiny geometric differences) may be precisely the key variables for action outcome. If this information is discarded as nuisance information, then the representation is great for classification but potentially insufficient for control.

This is actually a deeper layer of the “representation sufficiency” question above: good representation ≠ complete state. A representation that is excellent for video prediction is not necessarily a state representation sufficiently complete for downstream decision-making. How to ensure that while filtering unpredictable details, we don’t simultaneously discard information that future decisions truly need — this is the core design problem for predictive world models.

8. What This Means for Practitioners

If You’re Doing Robot Control

V-JEPA 2-AC’s manipulation results deserve serious attention. Under the same RTX 4090 + CEM planning experimental setup reported in the paper, V-JEPA 2-AC’s planning time is approximately 16 seconds, while Cosmos baseline requires approximately 4 minutes; even with V-JEPA 2-AC using 10× more candidate trajectories (800 vs 80), it still maintains approximately 15× planning latency advantage. This result shows that latent-space planning has significant computational advantages under this experimental setup, but it cannot be directly extrapolated to “all JEPA systems are 15× faster than generative world models” — different planning horizons, candidate counts, and hardware configurations all affect actual latency. Additionally, 16 seconds/action itself is still far from real-time robot control. If your system can tolerate some planning delay rather than generating beautiful video, the JEPA path may be more suitable.

If You’re Doing Self-Supervised Learning

I-JEPA’s training efficiency performance (5.3× fewer epochs achieving comparable performance to MAE) is already attention-worthy. The JEPA framework provides a self-supervised learning paradigm that doesn’t rely on pixel reconstruction, which is particularly valuable as compute costs receive increasing scrutiny.

If You’re Doing World Model Research

The JEPA path provides an important alternative: not all world models need to generate pixels. If your downstream tasks require understanding rather than generation, predicting in representation space may be the better choice.

If You’re Following AMI Labs

AMI Labs’ team and funding scale tell us one thing: investors have strong interest in the “world models + physical intelligence” direction. But stay clear-eyed — this funding should not be directly interpreted as validation of JEPA’s specific technical path. AMI Labs’ public statements focus on world models, persistent memory, reasoning, planning, and action-conditioned intelligence, not on any confirmed JEPA product architecture. The distance from papers to products remains vast — as of now, AMI Labs has not released benchmark results that can independently validate its specific world-model architecture or product capabilities.

9. Open Questions for the JEPA Path

Finally, a few questions the JEPA path hasn’t fully answered yet:

Representation sufficiency. This may be one of the most critical questions for JEPA’s transition from representation learning to a true world model. JEPA’s greatest potential strength is also its greatest risk: the model actively filters out unpredictable pixel details. But for robot control, some seemingly low-level details (precise geometry, contact states, friction, fine object states, affordances) may be precisely what determines whether an action succeeds. A representation that works very well for video classification is not necessarily a sufficiently complete state representation for control. Therefore, the real question to answer is not “are representations better than pixels” but: does this representation retain all the information needed for downstream planning?

Representation degeneration and objective design. JEPA avoids representation degeneration through target encoder, EMA updates, and prediction objective design, but “why this mechanism can stably learn informative representations” remains an open research question. Especially as the predictor, target encoder, and masking strategy extend further into long temporal sequences and action-conditioned prediction, whether representations remain sufficient and stable is a more important question than simply avoiding collapse.

Long-range prediction stability. V-JEPA 2-AC demonstrated action-conditioned prediction and planning capability. Notably, the authors have already explicitly trained multi-step prediction through rollout loss to mitigate autoregressive error accumulation — the paper describes the predictor’s outputs being fed back for multi-step prediction. However, the planning horizon and task complexity demonstrated in the current robot experiments are not yet sufficient to prove stable world-model rollout capability for long-term, complex tasks. In comparison, Dreamer and other model-based RL methods place multi-step latent rollout and policy/value learning at the core of their training loop from the beginning, so the two approaches differ fundamentally in their technical paths for long-horizon planning.

Language alignment. A core capability of current VLA (Vision-Language-Action) models is language grounding. The JEPA path currently works primarily in visual and action space — how to integrate with language capabilities is an important open question.

Scalability. V-JEPA 2 has scaled to the billion-parameter level, but whether JEPA-style models have a stable, predictable scaling law still lacks sufficient evidence. The more critical question is not simply pursuing parameter scale, but whether representation quality, action-conditioned prediction accuracy, and downstream planning performance continue to improve as model scale, video data volume, and prediction horizon increase.

2026 New Development: WA-JEPA. The state sufficiency question has begun to be directly challenged in subsequent work. WA-JEPA (August 2026, Rethinking the Video JEPA Paradigm for World-Action Modeling in Autonomous Driving) re-examines V-JEPA’s applicability to autonomous-driving planning, arguing that random spatiotemporal masking and deterministic latent regression are not naturally suited for future-directed action modeling, and introduces future-masked prediction, latent flow matching, and joint future-action prediction. This shows JEPA is further differentiating from “predicting representations” toward “predicting actionably future states” — the gap between V-JEPA’s original representation-learning objective and true world-action modeling objective is being directly confronted by researchers.

The JEPA path’s core insight — that in tasks targeting visual representation learning, prediction, and planning, latent representation prediction can avoid the computational burden of pixel-level generation while achieving strong downstream performance — has been progressively validated through experiments from I-JEPA to V-JEPA 2-AC. From academic hypothesis to AMI Labs’ industrialization attempt, this path is moving toward broader applications.

But it’s not the only answer for world models. As I concluded in my roundup article, “world model” is losing its singular meaning. Different questions require different tools.

What truly makes JEPA worth attention is not that it proposes yet another “world model architecture,” but that it redefines what world models should predict. The pixel generation path typically learns world dynamics by predicting future observations; JEPA places its prediction target directly in the learned representation space. The former needs to explain and generate a large amount of observation details; the latter actively focuses prediction on factors that are more predictable and more useful for tasks in the representation. And V-JEPA 2-AC takes a further step: “after executing this action, how will the task-relevant state of the world change?”

But the deeper question is: how much observation information does an internal state for prediction and decision-making actually need to contain? This is the core question that the JEPA path truly challenges — not “pixels bad, latent good,” but “what information is sufficient for prediction and control.” I-JEPA/V-JEPA demonstrate the effectiveness of predictive representation learning; V-JEPA 2-AC is only beginning to validate whether representations can support action-conditioned dynamics. But there remains a gap not yet fully bridged between “representation learning works” and “representation is a control-sufficient state.” This is the true turning point of JEPA’s journey from representation learning to world modeling, and the aspect most worth continued attention.

Next up, I’ll break down VLA (Vision-Language-Action) — from RT-2 to OpenVLA to π₀, how this path end-to-end connects language understanding, visual perception, and robot actions, and where it fundamentally differs from the world model approach.