Research

A scalable autonomous training data generation pipeline is proposed, and it is shown that both supervised finetuning and reinforcement-learning on the generated data can significantly improve VLM reasoning performance both on spatiotemporal ordering tasks as well as downstream robotics and visual reasoning tasks.

VeruSteer, which learns to select error types from the outcomes of alternative repair trajectories, is presented, which constructs repair trees offline and compares paths from the same initial program by verification success, repair rounds, and final verification progress.

This work introduces JUICE (Joint Unified multimodal modeling In Continuous Encoder spaces), a single flow matching Transformer over frozen image and text representations, in which each modality has its own noise level.

This paper shows that JEPA-style latent world models (LWMs) can learn across diverse domains and data sources by adopting the language-conditioned interface of video generation models by introducing LaLa-World, a Language-conditioned Latent World model that learns from simulation and robot trajectories together with broader image and video data.

Consistent gains in physical-quantity probing and distracted-to-clean retrieval indicate that ConfigWM learns states that preserve task-relevant information while remaining robust to distractors, supporting reliable world modeling and planning in noisy worlds.

Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware, achieves state-of-the-art performance for its size class on agentic benchmarks, setting a new standard compared to the models of its size.

Local appearance purification (LAP) is introduced, which improves generation quality and makes encoders such as MAE more effective alignment targets by removing locally predictable content and aligning to the residual.

Symplectic Hamiltonian JEPA (LeHami), which replaces the standard predictor with an action-conditioned Hamiltonian flow over a structured latent state integrated using an exactly symplectic leapfrog step, shows how geometric structure can provide inference-time guarantees that are complementary to representation regularization in latent world models.

Through stage-by-stage analysis of agent trajectories, it is shown that the biggest bottleneck is identifying the buggy behavior: agents often reach the relevant feature but never trigger the failure or fail to recognize it.

This work provides a design principle for decoding for future AR models, diffusion models, and anything in between as paths on one corruption lattice, and defines the cost of a schedule as the dependence its parallel steps discard.

It is shown that LLM-generated specifications fail not merely because industrial code is harder, but because failures are structured, stage-dependent, and often repaired in the wrong direction, and future evaluation and tooling should move beyond final pass/fail outcomes and account for where specifications break.

It is proved that exact action reconstruction makes the encoder injective on the finite-horizon reachable subspace, so the encoder cannot discard any state direction reachable by an action sequence within $H$ steps, and it is shown that, as $H$ grows, the dominant eigenspace of the finite-horizon controllability Gramian converges to the controllable unstable subspace.

H-JEPA is introduced, an end-to-end recipe for training a hierarchy of action-conditioned JEPAs in which each level predicts farther ahead in its own learned latent space in which each level's predictions become subgoals for the planner below it.

SheetSage2, a unified music transcription framework that combines synthetic data, task-specific structured decoding, and autoregressive distillation, is presented, substantially improving over SheetSage1 and surpassing task-specific models on several benchmarks.

This work introduces Learnable Subspace Projections (LSP), which instead learns the subspaces to discard end-to-end in each weight matrix, and outperforms baselines, and its advantage widens as compression increases.

Across both image and video generation, the choice of training objective fundamentally determines whether a few-step model inherits its teacher's distribution coverage or trades it away for single-draw quality, and it is shown that this tradeoff extends to few-step causal video generation.

Mixed-pattern training on five completed pairs reduces large single-pattern transfer errors, although destination-trained references usually remain more accurate, and the benchmark and findings support systematic evaluation of observation-pattern sensitivity and provide a reusable workflow for developing methods under changing measurement conditions.

SCOPE (Sparse-Context Observability-aware Predictive Embeddings) is proposed to recover complete PDE fields from sparse observations by coupling full-field latent prediction with physical reconstruction, and outperforms mask-aware neural operators on all ten forward and inverse tasks.

This paper systematically evaluates existing action-conditioned joint-embedding predictive architecture (JEPA) world models, including LeWM, DINO-WM, and JEPA-WM for end-to-end autonomous driving (E2EAD), and proposes a SIGReg-regularized learnable projector applied to projected patch embeddings, which introduces a SIGReg-regularized learnable projector applied to projected patch embeddings.

The results show that representation need not remain a by-product of generation or merely a tool for improving it: it can be directly optimized as a first-class capability of diffusion pretraining alongside generation.

This work introduces Relative Generalization Invariance (RGI), the invariance of the validation-loss difference between any two tokens across models, and shows that RGI approximately holds across a wide range of optimizers and moderate architectural variations, suggesting that these choices induce an approximately uniform shift in token-wise losses.

This work introduces FuseReg, which replaces heuristic feature selection with training over random subsets of encoder layers, and theoretically analyze the underlying mechanism: subset sampling preserves the full-layer latent mean in expectation while explicitly penalizing sensitivity to cross-layer disagreement.

This technical report details the infrastructure behind Game Arena and describes the three pilot game environments: Chess, Poker, and Werewolf, enabling a systematic study of models's strategic planning, adaptation, and robustness under uncertainty.

A small concept generator is optimized with reinforcement learning so that its concepts maximize the downstream success of a larger, frozen answer generator, and can thus be trained into an effective, reusable search policy for a much larger one.

Unified Latent Diffusion Reasoner improves multimodal reasoning by weaving it into a single thread, where the model predicts successive thoughts in a common representation space.

The new TabPFN-3.5 sets a new state of the art on standard tabular prediction in TabArena, and extends it to the data practitioners encounter in practice: non-i.i.d. data with temporal or grouped splits, tables with strings, text and images, high-cardinality categorical features, and wide tables with many features.

DyAb is a pair-wise representation built on top of a pre-trained protein language model that achieves a Spearman rank correlation of up to 0.85 on binding affinity prediction across monoclonal antibodies targeting three different antigens.

LeVJEPA is introduced, the first video encoder trained under LeJEPA's collapse-free objective, which indicates that video becomes a viable and in several respects preferable substrate for general-purpose visual pretraining.

This work introduces LpWorldModel, a JEPA model regularized with Rectified Distribution Matching Regularization to match encoder features to a Rectified Generalized Gaussian distribution, yielding non-negative sparse codes, and finds that the learned sparse representations are mode-factored.

This first release of Prior Labs in relational learning opens-source three pieces of software that expect to accelerate research in the field towards meaningful real-world impact and releases an initial version of RPI, an open-source, model-agnostic interface that enables early adopters to easily define problems on new databases and apply any model implemented in RelArena, including TabPFN-Rel, to these problems.

A multi-agent collaborative framework, StarVerus, to automate the verification of industrial Rust code and introduces a planner-repairer-actor-rewriter multi-agent paradigm to further enhance the proof repair capabilities.

This study provides empirical clarity through a systematic exploration of multimodal pretraining and derives efficient pretraining recipes that achieve strong generative performance using only 5% of the compute budget.

It is shown that self-supervised learning (SSL) can convert raw DXA images into representations of systemic health, and prognostic information in DXA images that conventional readouts discard is revealed, learnable with relatively little data and modest compute.

Horsepower-JEPA organizes each graph into an ordered bank of coarse-to-fine partition resolutions and performs context-target latent prediction separately at each resolution using an online encoder, an exponential-moving-average target encoder, and a latent predictor.

A framework for analyzing openness across the AI stack is proposed, reviewing prior approaches and highlighting the diverse motivations for pursuing openness, showing how these map onto different forms of openness at both the model and system levels.

This work proposes to learn a world model of piano sound using JEPA by framing music as an action-conditioned system, and shows that the learned model captures the relationships between musical actions and their resulting sound.

This work proposes CrossBERT, a two-part architecture that separates the learning of high-quality encoded representations from the rigid grounding of token reconstruction, and demonstrates monotonic scaling and superior performance on MTEB(eng, v2) and frozen GLUE benchmarks.

This work proposes Anchored Self-Play (ASP), which anchors self-play with a small reference set by adding a code-embedding similarity reward for generation and mixing reference bugs into fixer training, and achieves the best fix rates across bug sources.

Off‐the‐shelf LLMs are not yet suitable for autonomous requirement quality evaluation; yet could be more suitable to assist human experts or serve specialized roles in agentic AI frameworks.