Daily ML / AI Representative Research Report · July 29, 2026

INSID3: Can one frozen self-supervised backbone
segment any concept from a single example?

CVPR 2026 Oral. The authors dismantle the entire multimodel in-context segmentation pipeline, leaving just one frozen DINOv3. The obstacle turns out to be not insufficient capability, but a previously neglected layer of positional bias.

CVPR 2026 Oral Training-Free +7.5 mIoU / one-third the parameters TU Darmstadt · PoliTo · TU Munich Open-source code available

01 Today's topic-selection process

To ensure broad long-term coverage rather than following the same popular topics daily, today's process builds a broad pool from top-conference award pages and recent arXiv work, then draws randomly instead of choosing the hottest paper by preference.

The candidate pool contains 45 papers, from ICLR 2026 Outstanding / Honorable Mention, CVPR 2026 awards and selected Orals, ICML 2026 Outstanding / Honorable Mention / Position, ACL 2026 Best Theme / Resource / Social Impact / Demo, and high-profile April–July 2026 arXiv preprints (full list in the appendix).

$ shuf -i 1-45 -n 3
18
39
14

First draw No. 18 → INSID3: Training-Free In-Context Segmentation with DINOv3(CVPR 2026). Materials were sufficient—CVPR Oral, arXiv HTML, open-source code, project page, and official explanation video were accessible—so it was adopted without redrawing.

Why this draw is welcome

Random sampling risks incremental SOTA papers with a module swap and a two-point gain but little to discuss. INSID3 instead subtracts complexity: its central conclusion is that previous additions were unnecessary, enabled by a concrete, independently testable technical discovery. This makes it substantially more informative than average.

02 Link list

The table lists all accessible entry points for the core paper and five key related works. “Not released” means actual searching found none, not omission.

Additional backward references include SAM (2304.02643), DINOv2 (2304.07193), Painter (2212.02499), PerSAM (2305.03048), SINE (2410.04842), DiffewS (2410.02369), together forming the evolution in Section 4.

03 Research background

What problem does this field address?

In-context segmentation (ICS) addresses a fundamental segmentation limitation: traditional semantic models recognize only dozens or hundreds of predefined training classes, while real users often want to segment this particular thing—a specific scratch on a car, a dog named Bella, an unnamed cellular structure in microscopy, or the third clip on a part. These concepts cannot be enumerated beforehand and have no labeled training set.

The ICS setting is consequently direct: Given one reference image and one annotated mask, segment the same concept in any new image. “Concept” is intentionally broad: a whole object, an object part, or a personalized instance. Borrowing NLP's in-context learning, task adaptation uses one demonstration without updating weights.

The difficulty is that two capabilities are simultaneously required, often with conflicting optimal representations:

• Semantic matching: Recognize that two regions depict the same kind of thing across views, lighting, and instance appearances, requiring abstract representations invariant to within-class variation.
• Fine-grained segmentation: Produce pixel-accurate masks with sharp boundaries, requiring localized representations that preserve spatial detail.

Invariance and locality are a classic tension in representation learning. Recent ICS history can be read as successive engineering compromises between them.

Widely accepted standard settings

The following conventions are broadly accepted in this field and followed by INSID3:

The column to watch most closely

It is parameter count, not mIoU. Training-free ICS is often marketed as an advantage, yet a DINOv2 + SAM pipeline approaches one billion parameters and needs two forward passes. Saved training cost shifts to every inference. Compare mIoU together with parameters/latency—the central ground on which INSID3 argues its case.

Major frontier directions and disagreements

Current ICS research has three broad routes, differing not just technically but fundamentally on where generalization originates:

Route A · Train a general-purpose in-context model

Examples: Painter, SegGPT, DiffewS. Treat reference image + reference mask + target image → target mask as image-to-image translation, supervised across many segmentation tasks. Belief: Sufficiently diverse training tasks will naturally produce in-context capability. Cost: Requires extensive mask-level supervision; granularity and style become tied to training distributions, degrading on unseen domains such as X-rays and remote sensing.

Route B · Frozen backbone + lightweight trained decoder

Examples: SEGIC, SINE. Accept that DINOv2 self-supervised features already contain emergent correspondence; train a small decoder to translate it into masks. Belief: Representations are sufficient; the missing component is readout. Cost: The decoder still needs supervised training and remains tied to its distribution. Papers repeatedly report in-domain gains but out-of-domain losses.

Route C · Combine frozen foundation models without training

Examples: Matcher, GF-SAM, PerSAM. Clear division of labor: DINOv2 handles semantic matching → generates point/box prompts → SAM produces masks. Belief: Generalization survives only without any task-specific training. Cost: Complex architectures, doubled parameters, high latency; worse, SAM fixes the granularity—you get the mask granularity SAM produces. If you want a wheel but receive the whole car, the pipeline offers little intervention.

The central disagreement is: “Must generalization be protected by avoiding training, or can better representations provide it?” Route C bets on the former and accepts complexity; Route B bets on the latter but cannot fully escape training. INSID3 occupies an interesting position: it simultaneously rejects both training and multimodel composition, and must establish that one self-supervised backbone can perform both semantic matching and segmentation. That is both its value and its decisive challenge.

04 Core paper: INSID3

Abstract (my summary)

INSID3 proposes that in-context segmentation needs neither fine-tuning nor multiple foundation models: one frozen DINOv3 is enough. The authors first identify previously untreated positional bias in DINOv3 dense features: patches at similar spatial positions across images are highly similar regardless of semantics, contaminating cross-image matching. A training-free trick removes this bias: pass a noise image through the encoder, perform SVD on its features, and take the subspace spanned by leading right singular vectors. With noise as input, this subspace contains no semantics, only the encoder's positional structure. Project real features onto its orthogonal complement to remove positional bias while retaining semantic information.

After debiasing, the method replaces pixel-level matching with a region-level approach: agglomerative hierarchical clustering forms semantically coherent target regions, reverse nearest-neighbor matching filters candidates, seed regions are selected, and cross-image semantic similarity × within-image self-similarity expands seeds into complete masks. The clustering threshold directly controls granularity, allowing object-, part-, and instance-level segmentation in one pipeline and avoiding SAM pipelines' fixed-granularity limitation.

Results: Average across 9 benchmarks 55.1 mIoU, exceeding previous training-free SOTA GF-SAM (47.6) by 7.5 points, with 304M vs 945M parameters (approximately one-third), speed approximately 3.4×, and no mask-level or class-level supervision throughout.

Motivation: The authors' observed pain points

The paper begins with something that seems wrong: ICS architecture complexity keeps rising, but the reason is historical baggage rather than present technical necessity.

Matcher/GF-SAM's division of DINOv2 matching + SAM masks rests on the premise that DINOv2 features are insufficiently localized for sharp masks and therefore require SAM. This held in the DINOv2 era. But DINOv3's Gram anchoring, addressing dense-feature degradation under long training, and high-resolution post-training already produce strong spatial structure. Is SAM still necessary?

Testing this, the authors find direct cross-image cosine matching with DINOv3 disappointing. The key diagnosis is: Failure comes from a nonsemantic positional component in features, not insufficient semantic capability. A target's top-left patch strongly matches a reference's top-left patch merely because both occupy that position—a systematic false match. The paper states that DINOv3's bias is much stronger than DINOv2's, a side effect of post-training for denser localization: solving one problem creates another.

The paper's most elegant step

turns a diagnosis directly into an approach. Once features are seen as semantic + positional components, a semantic-free probe can isolate position—and a noise image is ideal: it contains no semantics, leaving mainly the encoder's own positional priors. The entire operation requires no training or labels and only a few lines of code, the kind of construction that seems obvious in hindsight.

Narrative

The narrative is a tight causal chain, each step clearing the next obstacle:

1. Ask a subtractive question: Can one self-supervised backbone perform both semantic matching and segmentation without supervision or auxiliary models?
2. Verify the premise: DINOv3 dense features genuinely have strong spatial structure and semantic correspondence—segmentation capability exists.
3. Encounter and diagnose failure: Cross-image matching fails because of positional bias, more severe in DINOv3 than DINOv2.
4. Remove the obstacle: Noise + SVD identifies the positional subspace; orthogonal-complement projection removes it.
5. Change the reasoning level: Replace pixel matching with region reasoning—cluster regions, select by matching, complete through aggregation. The clustering threshold naturally controls granularity.
6. Key design insight: Position is noise across images but a useful signal within an image, where nearby positions tend to share objects. Therefore use two feature sets asymmetrically: debiased across images, original within images.
7. Deliver the promise: SOTA on 9 benchmarks, one-third the parameters, controllable granularity, cross-domain robustness, and no supervision.

Step 6 is more than an engineering trick; it completes the paper's logic. If position were entirely useless, it would simply be a bug. Because it becomes useful in another setting, asymmetric use is a substantive design principle rather than mere denoising.

Method: Four stages and key equations

Φ is the frozen DINOv3 encoder, Fr, Ft are reference and target dense features, respectively, and D is feature dimension. Denote the reference mask by Mr.

Stage 1 · Positional-bias correction (unlocking feature space)

Probe the encoder's positional prior with a pure-noise image:

Then project reference and target features onto the orthogonal complement of B :

Key design: Debiased features F̃ are used only for cross-image operations. Original features F remain intact in within-image operations because spatial structure is beneficial there.

Stage 2 · Fine-grained clustering

The target image's original features Ft undergo agglomerative hierarchical clustering, yielding mutually exclusive regions covering the target image {𝒢₁, …, 𝒢_K}.

Choosing agglomerative clustering over k-means is deliberate: it requires no preset cluster count, only a distance threshold τ. This τ is precisely the granularity control—small thresholds produce fine, part-like regions; large thresholds merge toward objects. The paper describes spatially coherent clusters representing consistent object- and part-level regions. This structurally enables one method to address semantic, part, and personalized tasks.

Stage 3 · Seed-region selection

Two substeps. First, reverse matching locates candidates: for each target patch i, find its reference-image nearest neighbor

Direction matters. Conventional forward matching starts from reference-mask patches and searches the target, ignoring information outside the reference mask. Reverse matching makes every target patch compete against the entire reference image—patches more like reference background than foreground are eliminated. Unannotated reference regions thus implicitly serve as negatives, providing free discriminative signal. Clusters overlapping candidate patches form candidate set C_cand.

Then perform seed selection: compute debiased prototypes of candidate clusters and reference foreground p̃tk, p̃r, selecting the highest cross-image similarity as seed

Stage 4 · Region aggregation

Seeds usually cover only the concept's most discriminative part, such as a dog's head. Other same-concept regions must be added without swallowing background. This uses the product:

A product, rather than weighted sum, ensures that either weak signal suppresses the result; one strong signal cannot compensate for the other. This helps avoid absorbing background.

Hyperparameter summary

The pipeline has no learnable parameters. Computation comprises one encoder pass, plus a cacheable noise-image pass, one SVD, one agglomerative clustering operation, and several dot products.

Experimental design and results

Benchmark coverage: 9 datasets across three tasks and four visual domains—semantic segmentation (LVIS-92ⁱ, COCO-20ⁱ, ISIC skin lesions, SUIM underwater, iSAID aerial imagery, X-Ray), part segmentation (PASCAL-Part, PACO), and personalized segmentation (PerMIS). This deliberate combination exposes any overfitting of self-supervised features to natural images through medical, underwater, and remote-sensing tests.

Baseline selection: Two clearly motivated groups: methods trained with mask supervision(Painter, SegGPT, SINE, DiffewS, SEGIC), and training-free methods (PerSAM, Matcher, GF-SAM). The authors additionally include a GF-SAM variant using DINOv3 features—a particularly fair control directly addressing whether gains merely come from a stronger backbone.

Table: mIoU (%) on 9 benchmarks. All INSID3 entries and its overall mean were verified. The three baselines'means and parameter counts were verified, but their per-dataset entries were not individually retrieved; consult Table 1 for exact comparisons. Fine-tuned baselines range from 310M–890M parameters.

efficiency: 304M (DINOv3 only) versus 945M (DINOv2 + SAM), approximately one-third; the project page reports 3.31 FPS, approximately 3.4× of the comparison method. Savings extend beyond memory: removing a large-model pass matters for batch processing and interactive annotation.

Ablations(paper §4.3):

• Positional-bias correction: Improves semantic correspondence by up to +6.6% PCK . This decisive ablation turns positional bias's existence and substantial harm into measurable evidence supporting the method's premise.
• Asymmetric feature use: Validates debiased cross-image features and original within-image features, testing the design principle itself rather than simply adding/removing a module.
• Clustering granularity τ: Shows how thresholds control output granularity.

Appendix extensions(Sections C–E): k-shot extensions, controls for dense representations' necessity, and supplementary comparisons with the newer SAM 3 (a meaningful addition given SAM 3's CVPR 2026 Best Paper Honorable Mention).

Value analysis: What does it actually contribute?

I see three layers of contribution, decreasing in value but all genuine:

Layer 1 · A reusable technical discovery (highest value)

“Dense visual-foundation-model features contain a separable positional component that noise probes + SVD can remove without training”—this applies far beyond in-context segmentation. Any cross-image dense-matching task may suffer similarly: semantic correspondence, visual localization/relocalization, point tracking, image retrieval, few-shot detection, or temporal video association. It is a standalone component that can enter other pipelines at almost no cost. If one thing survives from this paper, it will be this.

An interesting broader lesson follows: DINOv3's localization-oriented post-training created the stronger positional bias. Foundation-model improvements are not always monotonic downstream; optimizing an upstream objective can obstruct other uses. This matters across the foundation-model-plus-adaptation research paradigm.

Layer 2 · Convincing architectural simplification (high value)

Compressing a 945M dual-model pipeline into a 304M backbone while gaining 7.5 points is not merely lighter at parity, but lighter and stronger. In an ICS field accustomed to adding models for performance, this corrects the record: substantial complexity was historical baggage, not necessity. Checking whether old premises still hold before adding components is a broadly useful methodology.

Layer 3 · Practical granularity control (moderate but underestimated value)

SAM pipelines' fixed granularity is a longstanding pain point: requesting a part and receiving a whole object offers little recourse. INSID3 makes granularity a continuous threshold τ, obtained for free as a natural byproduct of agglomerative clustering. Interactive annotation tools directly benefit from the ability to adjust an incorrect granularity. The paper arguably underemphasizes this.

So what does it not contribute?

To be clear, it introduces neither a new problem—ICS is already well-defined—nor a new learning paradigm—there is no training. Its backbone capability comes entirely from others' DINOv3 work. INSID3 is accurately characterized as diagnosis + simplification: identify a specific mistaken premise in prior systems, repair it, then show that much complexity can be removed. Such methodological value often exceeds apparent novelty; its CVPR Oral selection suggests reviewers recognized this.

Limitations

The paper discusses limitations/future work in Section F, but I could not retrieve its complete text. Labels below distinguish: [Acknowledged by the paper] for confirmed mentions, and [My analysis] for independent judgments based on method structure.

[Acknowledged by the paper] Empty-mask cases

The method degrades when reference annotations correspond to no meaningful image region. Training-free matching methods share this weakness: no learned prior provides a fallback.

[Acknowledged by the paper] Extreme viewpoint changes and severe occlusion

Matching deteriorates under these conditions. Self-supervised semantic correspondence itself weakens under large viewpoint changes, and this method has no geometric or temporal prior to compensate.

[My analysis] Practical fragility from three hyperparameters

s, τ, α are all part of a training-free method, meaning they cannot be learned from data and require manual settings or validation search. If tuned separately on cross-domain benchmarks, zero-supervision claims need qualification because tuning consumes supervision. I could not verify shared versus domain-specific settings, but this is the first issue to confirm for reproduction and deployment. Real one-shot scenarios often lack even a validation set.

[My analysis] Patch resolution limits mask precision

Outputs are unions of patch-level clusters. ViT patches are typically 14 or 16 pixels; even high-resolution post-training may not match SAM's pixel-sharp boundaries. The paper wins on mIoU, which is relatively tolerant of boundary errors. Boundary-sensitive metrics such as Boundary IoU or F-measure could substantially narrow, or even reverse, the gap to SAM pipelines. Such metrics are unreported, a genuine evaluation gap. Applications requiring cutout-quality precision—film compositing, medical delineation, industrial metrology—should verify this independently.

[My analysis] Agglomerative-clustering complexity

Agglomerative hierarchical clustering typically costs O(n²) to O(n³ log n). At high resolutions, patch count n grows quickly, potentially making clustering the new bottleneck. Although 3.31 FPS beats the comparison, the scaling curve with resolution is unspecified. SAM pipelines mainly pay fixed model-forward costs, growing relatively gently with resolution, whereas clustering is superlinear. At sufficiently high resolution, efficiency gains may disappear.

[My analysis] Strong dependence on DINOv3

The method assumes DINOv3 dense features are sufficiently localized. Backbone transferability is unspecified: would future DINOv4, SigLIP, or AIMv2 work? Is positional bias DINOv3-specific or general to ViTs? The latter is more scientifically valuable. Only a limited DINOv2 comparison, noting much weaker bias, touches this without systematic analysis.

[My analysis] The boundary of zero supervision needs precision

The paper accurately claims no mask- or class-level supervision. However, DINOv3 was trained on Meta's LVD-1689M, approximately 1.7 billion images, an unreleased dataset. Unsupervised here means no task-specific supervision; system capability still rests on extremely expensive pretraining outsiders cannot reproduce or audit. This is not the paper's fault, but belongs in the accounting of training-free claims.

05 Related work and technical evolution

Five key related papers

① SegGPT: Segmenting Everything in Context (ICCV 2023)

What it solved: First fully brought NLP's in-context learning paradigm into segmentation. A unified in-context coloring representation lets one model use reference images/masks to perform instance, stuff, part, contour, and text segmentation on images and videos. It established the ICS setting itself.

What remained: Extensive mask supervision; training distributions fix granularity/style, causing degradation on unseen medical/remote-sensing domains. INSID3 explicitly includes it among supervised-training controls.

② Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching (ICLR 2024)

What it solved: Established a feasible training-free ICS route: DINOv2 supplies general feature matching to locate targets, and SAM generates precise masks. The core insight is that their capabilities are complementary: one understands semantics, the other boundaries.

What remained: Multistage architectures incur substantial computation; SAM fixes the granularity; moreover, INSID3 explicitly notes that this composition cannot fully exploit correspondence–segmentation synergy: modules operate separately, compressing interface information into a few point prompts.

③ Bridge the Points: Graph-based Few-shot Segment Anything Semantically (GF-SAM, NeurIPS 2024)

What it solved: Refines Matcher with graph relationships among candidate points and masks, converting DINOv2 matching into better SAM prompts and raising the training-free ceiling. It was the preceding training-free SOTA at 47.6 average mIoU and INSID3's main competitor.

What remained: Still coordinates multiple VFMs with all 945M parameters and dual-forward costs; SAM still constrains granularity. More importantly: INSID3's GF-SAM-with-DINOv3 control shows that upgrading the backbone alone cannot achieve its gains. The pipeline architecture itself is the bottleneck, not feature strength.

④ SEGIC: Unleashing the Emergent Correspondence for In-Context Segmentation (ECCV 2024)

What it solved: Shows that dense correspondence usable for ICS has already emerged in self-supervised DINOv2 features; it need not be learned from scratch. A lightweight decoder atop the frozen backbone reads correspondence into masks, representing the belief that only readout is missing.

What remained: Fine-tuning ties the model to training distributions, limiting unseen-domain and unseen-granularity flexibility—described as in-domain gains and out-of-domain harm. Historical significance: SEGIC established that correspondence is already in the features; INSID3 removes the final decoder. In this sense, INSID3 is the logical endpoint of SEGIC's insight.

⑤ DINOv3 (Meta AI, 2025-08)

What it solved: The source of all INSID3 capability. Through Gram anchoring it solves dense-map degradation during long training, a persistent self-supervised-learning problem where global representations improve as dense features blur. High-resolution post-training and post-hoc text alignment further yield spatial structure substantially stronger than DINOv2. This makes eliminating SAM possible.

What remained: It also creates INSID3's target problem: localization post-training strengthens positional bias, polluting cross-image matching. Unreleased LVD-1689M also prevents independent auditing of capability boundaries.

Evolution: Four paradigm shifts

Connecting these works logically rather than chronologically reveals a clear convergence:

The interesting feature of this progression is that complexity first increases, then decreases. Stage I→II adds models for generalization; II→III→IV subtracts twice, first removing a large model, then the decoder. Each subtraction is enabled by discovery: emergent correspondence for III, sufficiently localized dense features plus separable positional bias for IV.

In other words, Stage II's complexity was the right solution then; its justification disappeared as backbones improved, but no one promptly revisited the premise. This recurring foundation-model-era pattern is worth recording: Downstream complexity accumulates to bypass specific upstream defects. Once upstream models improve, much becomes removable historical residue, but removing it requires actively questioning old premises.

Unresolved questions

After Stage IV, these questions remain:

• The boundary-precision ceiling persists: Removing SAM gains simplicity/speed but relinquishes pixel-sharp masks. Across this entire lineage, no method simultaneously achieves training-free, single-model, pixel-level boundaries. This is the clearest gap.
• Where do training-free hyperparameters come from?: s, τ, α Such thresholds are a shared weakness. Real one-shot settings often lack validation data, yet unsupervised methods require manual tuning. Can reference images themselves determine thresholds adaptively?
• Is positional bias general or DINOv3-specific?: This is the most scientifically valuable unresolved question. If general to ViT dense representations, noise-probe debiasing should become standard preprocessing far beyond ICS. Current qualitative DINOv2/DINOv3 comparisons fall short of cross-architecture/objective studies.
• Granularity intent remains poorly expressed: INSID3 uses a continuous threshold τ , a major improvement over SAM's fixed granularity, but users intend “wheel, not whole car”—a semantic request, not a guessed number. Reference masks implicitly encode granularity: a small marked region suggests a part. Could the reference mask directly determine τ？
• Evaluation protocols miss boundaries and failure modes: Heavy reliance on mIoU hides boundary quality and distinctions between complete failures and roughly correct but coarse masks. Standard reporting lacks Boundary IoU, failure rates, and empty-mask/multiple-instance edge cases.
• Multiple instances and open-world combinations: What should happen when several same-class instances occur, or the concept is absent? Benchmarks cover these common deployment cases weakly.

Promising next steps

Four directions ranked by expected benefit / implementation difficulty:

1. Validate debiasing systematically as an independent cross-task component. The easiest and potentially most impactful step is inserting noise-probe + SVD debiasing into semantic correspondence, point tracking, visual localization, few-shot detection, and image retrieval, systematically across DINOv2/v3, SigLIP, AIMv2, and MAE backbones. If broadly effective, a stage in one CVPR paper becomes standard preprocessing. Low cost and worth doing immediately.
2. Improve boundaries without reintroducing SAM. Refine patch masks below patch scale using original/high-resolution DINOv3 features, or classical training-free guided filtering, CRFs, or low-rank matting propagation. Preserve one model and zero training while matching SAM-pipeline Boundary IoU; this would directly broaden applicability.
3. Make hyperparameters adaptive. Estimate from reference-mask area fraction, shape complexity, and within-reference feature clusters τ(a small marked area implies part level) and α. This resolves unsupervised-yet-tuned inconsistency and changes granularity from a numeric knob to intent expression. Fully training-free implementation is possible, making it an elegant, cost-effective improvement.
4. Revise evaluation protocols. Require mIoU + Boundary IoU + failure rate and add absent concepts, multiple instances, and ambiguous granularity. Such work introduces no method but redirects the field's optimization; mIoU currently shapes every method quietly.

One-sentence conclusion

INSID3 is worth reading not merely for its 7.5-point gain, but for demonstrating an increasingly important research move in the foundation-model era: After upstream models improve, revisit downstream complexity built around old defects; doing so can deliver simpler and stronger systems simultaneously. Its noise-probe debiasing component may travel farther than the paper itself.

06 Appendix: Today's candidate pool (45 papers)

Random-draw pool. ★ marks the selected paper.

ICLR 2026 award papers

1. Transformers are Inherently Succinct (Outstanding)
2. LLMs Get Lost In Multi-Turn Conversation (Outstanding)
3. The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm (Honorable Mention)

CVPR 2026 awards and selected Orals

4. Efficiently Reconstructing Dynamic Scenes One D4RT at a Time (Best Paper)
5. Native and Compact Structured Latents for 3D Generation (Best Student Paper)
6. NitroGen: An Open Foundation Model for Generalist Gaming Agents (HM)
7. SAM 3D: 3Dfy Anything in Images (HM)
8. ChordEdit: One-Step Low-Energy Transport for Image Editing (Best Student HM)
9. B³-Seg: Camera-Free, Training-Free 3DGS Segmentation
10. Depth Any Panoramas: A Foundation Model for Panoramic Depth Estimation
11. EgoX: Egocentric Video Generation from a Single Exocentric Video
12. A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens
13. Real-World Point Tracking with Verifier-Guided Pseudo-Labeling
14. MSPT: Efficient Large-Scale Physical Modeling via Parallelized Multi-Scale Attention
15. FMPose3D: Monocular 3D Pose Estimation via Flow Matching
16. VidEoMT: Your ViT is Secretly Also a Video Segmentation Model
17. MatAnyone 2: Scaling Video Matting via a Learned Quality Evaluator
18. ★ INSID3: Training-Free In-Context Segmentation with DINOv3
19. TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment

ICML 2026 award papers

20. The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models (Outstanding)
21. High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions (Outstanding)
22. Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes (HM)
23. Motion Attribution for Video Generation (HM)
24. How Much Can Language Models Memorize (HM)
25. A Random Matrix Perspective on the Consistency of Diffusion Models (HM)
26. To Grok Grokking: Provable Grokking in Ridge Regression (HM)
27. Position: The Alignment Community is Unintentionally Building a Censor's Toolkit

ACL 2026 award papers

28. CoSToM: Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in LLMs (Best Theme)
29. Mapping the Circumplex of Affect: Geometric Analysis of Emotion Representations (Best Theme)
30. HSCodeComp: Expert-level Agent Benchmark for Hierarchical Rule Application (Best Resource)
31. ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in LLMs (Best Resource)
32. Audio MultiChallenge: Multi-Turn Evaluation of Spoken Dialogue Systems (Best Resource)
33. VeriTaS: The First Dynamic Benchmark for Multimodal Automated Fact-Checking (Best Resource)
34. Your Students Don't Use LLMs Like You Wish They Did (Best Social Impact)
35. The olmOCR Project: Building Fully Open OCR using VLMs (Best Demo)

High-profile April–July 2026 arXiv / journal papers

36. Nemotron-3 Super: Open Efficient MoE Hybrid Mamba-Transformer (2604.12374)
37. Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention (2605.22791)
38. Experience Augmented Policy Optimization for LLM Reasoning (2606.30420)
39. Multi-scale Mixture of World Models for Embodied Agents (2607.00457)
40. Reward as An Agent for Embodied World Models (2606.19990)
41. World Action Models: The Next Frontier in Embodied AI (2605.12090)
42. ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration (2605.03042)
43. Towards End-to-End Automation of AI Research (Nature)
44. DFlash: Block Diffusion for Flash Speculative Decoding (2602.06036)
45. GenericAgent: Token-Efficient Self-Evolving LLM Agent (2604.17091)