DSeq-JEPA: Discriminative Sequential Joint-Embedding Predictive Architecture

Conference: ECCV 2026
Paper: ECCV 2026 official page / Project Page
Area: Self-Supervised Learning
Keywords: self-supervised learning / joint-embedding predictive architecture / discriminative region selection / sequential latent prediction / attention saliency

TL;DR

DSeq-JEPA adds two inductive biases on top of I-JEPA's latent prediction objective — an attention-derived saliency map that selects and ranks regions by discriminativeness (where to predict), and a predictor that then predicts the embedding of the next region along that order (in what order to predict) — improving ImageNet linear probing by 1.1/0.8/1.3 points over I-JEPA (ViT-B/L/H) with consistent gains on fine-grained recognition, COCO/ADE20K, and CLEVR, at zero inference overhead.

Background & Motivation

Self-supervised learning has come a long way along two routes: instance discrimination and pixel reconstruction. Contrastive methods (SimCLR, MoCo, SwAV) rely on large batches or momentum queues for instance-level separation, non-contrastive methods (BYOL, SimSiam) show that asymmetric prediction plus a momentum target avoids collapse without negatives, and masked image modeling (MAE, BEiT, SimMIM) simply asks the model to fill masked patches back in. Joint-Embedding Predictive Architectures (JEPAs) more recently moved the objective from "reconstruct pixels" or "contrast instances" to prediction in latent space: I-JEPA predicts the embeddings of masked target regions from visible context, avoiding both the capacity spent on low-level appearance recovery and the dependence on pairs or negatives. Follow-up work has mostly changed what is predicted — DMT-JEPA aggregates features of semantically similar neighbouring patches into the target, C-JEPA adds contrastive regularization for stability, and LeJEPA aims for a more principled, heuristic-free formulation.

All these JEPA variants, however, keep one default from I-JEPA: target regions are treated uniformly and predicted in parallel and independently. That default does not match how visual information is actually distributed — a few regions carry the primary semantic cues that define the object (a bird's beak and forehead, a car's grille), while large areas are merely background or secondary context; and human perception itself is selective and ordered, latching onto salient cues first and refining the interpretation with additional context afterwards. I-JEPA samples rectangular targets from a random distribution, which spreads the learning capacity evenly over all content and leaves the N targets with no dependency on each other. Autoregressive sequential prediction as in iGPT and RandSAC is a related direction, but these methods still reconstruct pixels or tokens and do not explicitly model which regions are more informative.

This paper therefore asks a very direct question: can predictive self-supervised learning benefit from an explicit notion of where to predict and in what order? DSeq-JEPA answers affirmatively by introducing a discriminative sequential inductive bias into I-JEPA: an attention-derived saliency map localizes and ranks the primary discriminative regions (where), and next-region embedding prediction then follows the trajectory from most to least discriminative (in what order), so that pre-training naturally forms a curriculum-like semantic progression from primary to secondary cues. Viewed through representation learning, this breaks the permutation symmetry over target regions implicit in independent region prediction and replaces it with a semantically grounded prediction order. Core idea: promote "where to predict" and "in what order to predict" from sampling details to explicit inductive biases — discriminative ranking supplies the ordering prior, and the sequential predictor turns that prior into structured latent-prediction supervision.

Method

Overall Architecture

The input is an unlabeled image and the output is a backbone encoder transferable to classification, fine-grained recognition, detection/segmentation, and low-level reasoning. The pipeline is a two-tower JEPA with a predictor, but adds two stages over I-JEPA. ① Discriminative region prioritization: the target encoder produces a CLS–patch similarity saliency map at a chosen transformer block; after normalization, Otsu-based adaptive binarization, 8-neighbourhood connected-component labeling and fragment filtering, several irregular regions are ranked by their average saliency, the top N−1 become primary discriminative regions and the N-th is the complement of the rest to guarantee full coverage. ② Probabilistic curriculum: a probability λ ramping linearly from 0 to 1 over pre-training epochs switches each sample between discriminative selection and random rectangular sampling, so that the unstable early saliency maps do not contaminate optimization. Given the ordered region sequence, the context encoder encodes regions one by one and the predictor performs sequential next-region embedding prediction: at step k it uses only the context representations of the first k regions plus a positional token to predict the embedding of region k+1 under the target encoder, aligned by a Huber loss. Sequential prediction exists only during pre-training; only the encoder is kept at inference.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Unlabeled image + target encoder<br/>CLS–patch similarity saliency map"] --> B["Discriminative region prioritization<br/>normalize → Otsu → components → top-N ranking"]
    B --> C["Probabilistic curriculum<br/>λ ramps 0→1, per-sample switch"]
    C --> D["Context encoder<br/>encodes the selected regions"]
    D --> E["Sequential next-region embedding prediction<br/>predict step by step from most to least discriminative"]
    E --> F["Huber alignment<br/>predicted embedding ↔ target-encoder embedding"]

Key Designs

1. Discriminative region prioritization: replacing uniform random sampling with regions ranked by semantic importance

I-JEPA samples random rectangular target patches — it treats all content equally, so a substantial share of the predictive supervision lands on background or texture, and the rectangles may overlap and straddle several semantic units. DSeq-JEPA replaces exactly this step. Given an input image, it first computes the similarity between the CLS token and all patch embeddings at a selected transformer block of the target encoder to obtain an \(h\times w\) saliency map \(A\). After normalizing \(A\) to \([0,1]\), it binarizes with an Otsu threshold chosen adaptively per image (a fixed threshold would select everything or nothing depending on each image's saliency dynamic range), labels connected components under 8-neighbourhood connectivity, discards fragments smaller than \(0.15hw\), and assigns each remaining region a discriminative score \(\rho_k\) equal to the average normalized saliency inside it. Sorting by \(\rho_k\) descending gives the top \(N-1\) primary discriminative regions, while the \(N\)-th region is defined as the complement of the image with respect to the first \(N-1\):

\[\mathcal{R}=\{R_k\}_{k=1}^{N},\qquad R_N=\Omega\setminus\bigcup_{k=1}^{N-1}R_k\]

This last term is not cosmetic housekeeping. Unlike I-JEPA's (semi-)random masks, the selected regions are irregular and non-overlapping, and the complement region guarantees gapless coverage of the whole image, which makes region-based representation learning more stable; it also keeps the remaining secondary content as a prediction target instead of discarding it. The whole selection process introduces no new parameters and needs no gradients — the cost is a thresholding plus connected-component pass on the data side. It works because attention similarity is already a free semantic prior: the model can itself say which patches contribute most to global semantics. Otsu makes that prior adaptive at the per-image scale, and the resulting ranking becomes the ordering prior required by the next design.

2. Probabilistic curriculum: keeping unreliable early saliency out of region selection

The selection mechanism above has an obvious risk window: early in pre-training the encoder has not learned meaningful structure yet, the CLS–patch similarity is close to noise, and forcing "primary discriminative regions" out of a noisy saliency map injects a wrong semantic prior into the pre-training objective. DSeq-JEPA handles this by turning region selection into a probabilistic curriculum: \(\lambda=\min(1,\max(0,t/T))\) ramps linearly from 0 to 1 over epochs, and each sample independently draws a Bernoulli(\(\lambda\)) bit — when it is 1 the discriminative regions are used, when it is 0 the pipeline falls back to I-JEPA-style random region sampling (random center and random size). Training thus moves smoothly from "almost all random regions" to "all discriminative regions" rather than hard-switching at some epoch.

Per-sample Bernoulli sampling instead of a per-epoch switch matters in practice: a hard switch creates an abrupt change in the target distribution, whereas per-sample sampling lets both modes coexist inside the same batch, keeps the transition continuous, and naturally produces a mid-training mixture of "some random, some discriminative" regions. This design also explains why a crude saliency proxy suffices: replacing CLS similarity with a label-free Grad-CAM-style proxy costs only 0.1 point of ImageNet linear probing (73.5 → 73.4), and the two proxies agree at a Top-20 patch IoU of 0.41. Conversely, merely adding an auxiliary CLS token of the same structure to I-JEPA (without changing any region-selection logic) yields nothing at all (72.4 → 72.4). The gains come from constructing an order out of a proxy, not from the extra token.

3. Sequential next-region embedding prediction: turning flat parallel prediction into a semantically directed causal chain

I-JEPA predicts the N target patches in parallel and independently, with a loss that is a sum of N terms and no dependencies among targets — equivalent to invariance to target permutation, which gives the model no incentive to understand which regions come first or which later predictions should depend on which earlier ones. DSeq-JEPA turns the sorted region sequence \(R_1,\dots,R_N\) into an autoregressive chain: at step \(k\) the predictor receives only the representations of the preceding regions plus the positional token of the next target region, and estimates the embedding of region \(k+1\) under the target encoder, i.e. \(\hat{\mathbf{s}}_{R_{k+1}}=g_\theta(\mathbf{p}_{R_{k+1}},\mathbf{h}_{R_1},\dots,\mathbf{h}_{R_k})\), where \(\mathbf{h}\) comes from the context encoder and \(\mathbf{p}\) is the region's positional token. Target embeddings still come from the target encoder, so alignment happens in latent space and never touches pixels.

This change turns "order" itself into part of the supervision: later predictions can condition on earlier, more discriminative regions, so pre-training forms a curriculum from primary to secondary cues. The ordering ablation gives hard evidence. Holding the same set of discriminative regions fixed, a random order (71.7) is worse than no order at all (72.0), showing that autoregression without a meaningful trajectory only forces the model to learn from mixed-difficulty targets; the inverse order (least to most discriminative, 71.3) is worse still, showing that the direction of the trajectory matters more than the presence of autoregression; only the discriminative order reaches 73.5. The cost is confined to pre-training: sequential region tokens and causal masking raise peak memory from 31.5 GB to 38.2 GB, wall-clock time from 24.2 h to 26.5 h, and total compute from 96.4 to 111.6 GFLOPs. At inference, parameters (86.6 M) and per-image compute (17.7 → 17.8 GFLOPs) are essentially unchanged, because the mechanism is only active during pre-training.

A Worked Example

Take a bird image from CUB (ViT-B/16, 224×224 input split into \(14\times14=196\) patches, \(N\) set to I-JEPA's default of 5). Computing CLS-to-patch similarity at the selected block yields a 14×14 saliency map; after normalization and Otsu binarization, 8-neighbourhood connected components produce several candidate regions, fragments smaller than \(0.15\times196\approx29\) patches are dropped, and the survivors may correspond exactly to the beak, the forehead, and wing feathers. Ranking them by average in-region saliency, the top 4 become the primary discriminative regions and the 5th is the complement covering everything else (sky and branch). Early in training \(\lambda\) is still small, so this image most likely goes through random rectangular sampling; by epoch 450 \(\lambda=1\) and everything follows the discriminative order — the predictor sees the encoding of the beak first, then predicts the forehead, the wing feathers, and finally the whole background complement, each step Huber-aligned to the ground-truth embedding from the target encoder. The visualization matches: I-JEPA's target masks are diffuse and often straddle multiple semantic units, whereas these regions are compact and numbered from most to least discriminative (regions ①–⑤ in Fig. 6).

Loss & Training

The pre-training objective is a smoothed \(\ell_1\) (Huber) loss between predicted and target embeddings, averaged over the \(N-1\) prediction steps and the \(D\) feature dimensions:

\[\mathcal{L}_{\text{pred}}=\frac{1}{N-1}\sum_{k=1}^{N-1}\frac{1}{D}\sum_{j=1}^{D}\psi\!\left(\left(\hat{\mathbf{s}}_{R_{k+1}}-\mathbf{s}_{R_{k+1}}\right)_j\right),\quad \psi(x)=\begin{cases}\tfrac{1}{2}x^{2}, & |x|<\delta\\ \delta\left(|x|-\tfrac{1}{2}\delta\right), & |x|\ge\delta\end{cases},\ \delta=1\]

The authors' rationale is that it makes the alignment between predicted and target embeddings both stable and discriminative while remaining more robust to outlier dimensions than \(\ell_2\) — which matters for a sequential process that accumulates error step by step. Pre-training follows the standard I-JEPA / C-JEPA protocol: ViT-B/16 and ViT-L/16 for 600 epochs, ViT-H/16 for 300 epochs at 448 resolution (the cached text prints this cell as ViT-H/16448 300, read here as 448 resolution plus 300 epochs, ⚠️ refer to the original paper). Training uses a single view with no view-based data augmentation, \(N=5\) follows the I-JEPA setting, and the only architectural addition is an auxiliary CLS token used solely to produce the saliency map — no extra layers and no extra parameters. Evaluation follows the same I-JEPA/C-JEPA downstream protocols: linear probing and full fine-tuning on ImageNet, linear probes on the fine-grained datasets, a linear probe on frozen ViT-L/16 features for CLEVR, and the same detection/segmentation fine-tuning configuration for COCO/ADE20K. DSeq-C-JEPA stacks C-JEPA's contrastive regularization on top, to test whether the two improvements are complementary.

Key Experimental Results

Main Results

Classification and fine-grained recognition (linear-probe / full fine-tuning Top-1; FGVC numbers are linear probes; the last column averages the four datasets):

Backbone	Method	ImageNet linear	ImageNet fine-tune	iNat21	CUB	Cars	Overall Avg.
ViT-B/16 (600ep)	I-JEPA	72.4	83.5	35.9	65.3	65.9	64.6
ViT-B/16 (600ep)	DSeq-JEPA	73.5 (+1.1)	84.0 (+0.5)	36.4 (+0.5)	66.2 (+0.9)	67.3 (+1.4)	65.5 (+0.9)
ViT-B/16 (600ep)	C-JEPA	73.5	84.2	36.2	64.9	66.1	65.0
ViT-B/16 (600ep)	DSeq-C-JEPA	73.8 (+0.3)	84.3 (+0.1)	36.6 (+0.4)	66.5 (+1.6)	67.4 (+1.3)	65.7 (+0.7)
ViT-L/16 (600ep)	I-JEPA	77.1	86.6	38.8	66.9	68.1	67.5
ViT-L/16 (600ep)	DSeq-JEPA	77.9 (+0.8)	86.8 (+0.2)	39.5 (+0.7)	68.1 (+1.2)	68.9 (+0.8)	68.2 (+0.8)
ViT-L/16 (600ep)	DSeq-C-JEPA	78.4 (+0.4)	87.2 (+0.3)	39.7 (+0.7)	68.3 (+2.5)	68.8 (+1.1)	68.5 (+1.0)
ViT-H/16 (300ep)	I-JEPA	81.1	87.1	38.9	67.2	67.7	68.4
ViT-H/16 (300ep)	DSeq-JEPA	82.4 (+1.3)	87.8 (+0.7)	39.3 (+0.4)	68.9 (+1.7)	70.1 (+2.4)	69.7 (+1.5)

Dense prediction and low-level reasoning (COCO/ADE20K with ViT-B/16 at 600 epochs; CLEVR with a linear probe on frozen ViT-L/16 features):

Task / Dataset	Metric	I-JEPA	DSeq-JEPA	C-JEPA	DSeq-C-JEPA
MS-COCO	AP^box	49.9	50.5 (+0.6)	50.7	50.9 (+0.2)
MS-COCO	AP^mask	44.5	45.0 (+0.5)	45.3	45.7 (+0.4)
ADE20K	mIoU	47.6	48.1 (+0.5)	48.7	48.9 (+0.2)
CLEVR (Count)	Top-1	85.6	86.4 (+0.8)	86.8	87.1 (+0.3)
CLEVR (Dist)	Top-1	71.2	71.5 (+0.3)	71.6	71.9 (+0.3)

Additional controls: with ViT-B/16 the authors reran DSeq-JEPA across three random seeds and obtained 73.8±0.4 linear-probe accuracy, so the gain is not single-run noise. NEPA's training configuration is not strictly matched (ViT-B/14 and ViT-L/14 pre-trained for 1600/800 epochs respectively), so it is only a reference point — on ImageNet fine-tuning DSeq-JEPA / DSeq-C-JEPA reach 84.0/84.3 (ViT-B) and 86.8/87.2 (ViT-L) versus NEPA's 83.8 and 85.3.

Ablation Study

Are the two key components synergistic? (ViT-B/16; ImageNet and iNat21 linear probing.)

Region generation	Prediction strategy	ImageNet	iNat21	Note
Uniform sampling	Flat (parallel)	72.4	35.9	the I-JEPA baseline
Uniform sampling	Sequential	72.3	34.9	order exists but is meaningless; drops below the baseline
Discriminative selection	Flat (parallel)	72.0	35.7	good regions but no dependency modelling; drops as well
Discriminative selection	Sequential	73.5	36.4	both components together give the largest gain

Prediction order and other design choices (ImageNet linear probing, ViT-B/16; ordering variants hold the same set of discriminative regions fixed):

Configuration	ImageNet	Note
Flat (I-JEPA style, no autoregression)	72.0	reference point of the ordering ablation
Random order	71.7	random order falls below no order at all
Spatial order (row-major by region center)	72.7	purely geometric structure helps, but only a little
Inverse order (least to most discriminative)	71.3	reversed trajectory direction loses the most
Truncating (predict only the Top-3)	73.0	dropping the tail costs 0.5
DSeq order (full Top-5)	73.5	this paper's order
\(N=3\) / \(N=5\) / \(N=7\)	72.9 / 73.5 / 73.4	insensitive to the number of regions
I-JEPA + auxiliary CLS token	72.4 (I-JEPA also 72.4)	the token alone brings no representational gain
Grad-CAM-style proxy instead of CLS similarity	73.4	a label-free proxy costs only 0.1; robust to the proxy

Pre-training overhead and inference efficiency (ViT-B/16): DSeq-JEPA pre-trains in 26.5 h with 38.2 GB peak memory and 111.6 GFLOPs total, versus 24.2 h / 31.5 GB / 96.4 GFLOPs for I-JEPA; at inference both use 86.6 M parameters and 17.7 vs 17.8 GFLOPs per image.

Key Findings

• Both components are required; enabling either one alone drops below the baseline. This is the most convincing result in the paper: uniform sampling + sequential (72.3/34.9) and discriminative selection + flat (72.0/35.7) both fall short of I-JEPA's 72.4/35.9, and only the combination reaches 73.5/36.4. The authors' reading is that a meaningless order amounts to supervision with mixed difficulty, while regions without dependencies collapse inter-region structure into independent targets — discriminative selection supplies a semantic trajectory and sequential prediction consumes it, so the two are complementary rather than individually effective.
• The direction of the order matters more than the presence of an order. Inverse (71.3) sits 0.7 below no order at all (72.0) and Random (71.7) is 0.3 below; spatial order gains only 0.7, showing that geometric ordering provides some structure but far less than semantic ordering (+1.5). Chain length matters too: predicting only the Top-3 gives 73.0, 0.5 below the full Top-5, so the later low-saliency regions still contribute complementary supervision.
• The discriminative order induces an implicit easy-to-hard curriculum. Using the epoch-450 ViT-B checkpoint over 10,000 ImageNet images, the per-step prediction loss is clearly lower for Top-2/Top-3 than for Top-4/Top-5 — the model first learns to predict stable, highly informative regions and then progressively integrates weaker, more context-dependent cues.
• The gain comes from constructing an order out of a proxy, not from the proxy itself. Adding an auxiliary CLS token to I-JEPA does nothing (72.4 → 72.4), while swapping CLS similarity for a Grad-CAM-style proxy still yields 73.4; the two proxies agree at a Top-20 patch IoU of only 0.41, so performance is not tied to a specific saliency estimator.
• Representations self-organize into semantic structure during pre-training. Visualizing patch-level 4-cluster assignments shows fragmented, noisy clusters early on that gradually align with object parts (Fig. 7), consistent with the claim that sequential prediction drives a structured semantic progression.
• The efficiency boundary is clear. All extra cost lives in pre-training (+2.0 h, +15.5 GFLOPs total, +6.7 GB peak memory) and inference is unchanged. The whole method runs under a single-view JEPA recipe for 600 epochs (300 for ViT-H) and still matches or beats 1600-epoch, multi-view-augmented iBOT on dense prediction (ADE20K mIoU 48.1 vs iBOT's 50.0 remains lower, but COCO AP^mask 45.0 vs 44.2 is higher), and at ViT-H scale it is the highest under both linear probing and fine-tuning among comparable settings.

Highlights & Insights

• Sampling details promoted to inductive biases. How I-JEPA samples target regions has always been treated as an implementation detail; this paper argues that "where to predict plus in what order" is itself a designable structural prior, and it offers the "breaking target permutation symmetry" framing on top — a viewpoint more transferable than the specific saliency implementation.
• The saliency proxy is chosen with restraint: CLS–patch similarity adds no parameters, no gradients, and no extra forward pass; Otsu handles per-image dynamic-range differences adaptively, and the λ curriculum absorbs early noise. The mechanism is highly reproducible, which is also why swapping the proxy (Grad-CAM) still works.
• The ablation design is honest. The authors deliberately include two controls — adding only the CLS token, and replacing the saliency proxy — pre-empting the objection that "a token or a different proxy is what actually helps." Negative results from enabling a single component are reported rather than hidden behind the best combination.
• Zero inference cost is a pattern worth reusing: sequential prediction introduces causal dependencies among targets only during pre-training, and the predictor is discarded afterwards, leaving just the encoder. Any idea of "shaping the pre-training signal with extra structure and paying nothing at inference" can land the same way. Transferring it to video self-supervision is natural — replace region order with temporal order, or saliency ranking with motion-magnitude ranking.

Limitations & Future Work

• Absolute gains are modest and trade places with C-JEPA. ViT-B/16 linear probing ties with C-JEPA at 73.5, and ViT-L/16's 77.9 is slightly below C-JEPA's 78.0; the advantages concentrate in fine-grained recognition (CUB/Cars) and dense prediction. The authors explain this as more "part-aware" representations, but that remains a plausible inference without quantitative part-localization evidence.
• Cross-method comparisons are not fully aligned. The authors themselves note that the NEPA comparison is not strictly matched (different backbone, resolution and epochs), and the comparison with DINO/iBOT pits single-view against multi-view augmentation — two recipes with different costs, so no direct verdict is warranted.
• The reliance on the saliency prior is not examined deeply. All conclusions rest on the assumption that attention similarity reflects semantic importance. When objects are small, occluded, or the scene is semantically scattered (cluttered multi-object images), it is unclear whether the Otsu + connected-components + top-N pipeline merges different objects into one region or shatters one object into fragments; the \(0.15hw\) fragment threshold and \(N=5\) are also fixed, with no cross-dataset sensitivity study.
• Validation is image-only. Vision-language pre-training is listed as future work; no multimodal or video experiments are reported, and whether sequential prediction corresponds to something like "name the main subject first, fill in details later" on the language side remains entirely open.
• Concrete improvement directions: replace the saliency proxy with a learnable scorer (multi-block attention fusion, or ranking supervised by a downstream task); make λ adaptive to per-region confidence instead of linear in time; introduce finer-grained sub-ordering within a region (part-to-whole) to probe the upper bound of the "order" degree of freedom.

Related Work & Insights

• vs I-JEPA: I-JEPA samples rectangular targets from a random distribution and predicts them in parallel, learning from a mean-reduced N-term loss; DSeq-JEPA keeps the same two-tower-plus-predictor skeleton and the same Huber objective, adding structure only in "which regions are selected" and "in what order they are predicted." The difference is whether a semantic order is imposed on target regions; the advantage here is consistent gains on fine-grained and dense tasks (+0.5 to +1.5) at zero inference overhead, the disadvantage roughly 15% more pre-training wall-clock time and 6.7 GB more memory.
• vs C-JEPA: C-JEPA adds contrastive regularization to I-JEPA, changing the loss term; this paper changes how predictions are organized. The two are orthogonal, so DSeq-C-JEPA stacks them for the best JEPA-family results — evidence that sequential structure is not a substitute for contrastive regularization.
• vs DMT-JEPA / LeJEPA: DMT-JEPA changes what latent target is predicted (aggregating features of semantically similar neighbours) and LeJEPA focuses on removing heuristics and providing a provable formulation; neither touches the order of, or dependencies among, target regions, so the angle taken here does not conflict with them.
• vs iGPT / RandSAC: they also use autoregressive sequential prediction as a self-supervised inductive bias, but the order comes from a fixed token order (or random segments) and reconstruction happens in pixel/token space; DSeq-JEPA makes the order a discriminativeness-derived, semantically grounded trajectory and aligns in latent space, avoiding capacity spent on low-level appearance.
• vs MAE / DINO / iBOT: the latter two need 1600 epochs (and multi-view augmentation for DINO/iBOT), whereas this method reaches the highest numbers among comparable settings at ViT-H scale with a single-view 600-epoch (300 for ViT-H) JEPA recipe, and stays competitive on dense prediction; the cost is that ImageNet linear probing at ViT-B scale remains below DINO/iBOT, so single-view JEPA is not uniformly better in pure semantic separability.

Rating

• Novelty: ⭐⭐⭐⭐ Pushes JEPA's target regions from "random sampling, parallel prediction" to "semantic ranking, sequential prediction," with a unifying permutation-symmetry framing; but neither component (saliency-based selection, autoregressive prediction) is new in itself.
• Experimental Thoroughness: ⭐⭐⭐⭐ Covers four task families (classification / fine-grained / detection & segmentation / low-level reasoning) across three backbone scales, with ablations on component synergy, ordering scheme, region count, CLS token, saliency proxy and overhead; the deductions are for unmatched cross-method configurations and the absence of saliency-failure and threshold-sensitivity analyses.
• Writing Quality: ⭐⭐⭐⭐ The motivation chain (where + in what order) is clear and the figures, tables and formulas are restrained, with honest controls in the ablations; some key implementation details (how region tokens are built, how the target encoder is updated) are not spelled out in the main text.
• Value: ⭐⭐⭐⭐ The gains are steady rather than dramatic, but "sequential latent prediction" is a modular design with low transfer cost and zero inference penalty, directly relevant to follow-up JEPA-style pre-training, especially for video and multimodal directions.