arXiv:2608. 01130v1 Announce Type: new Abstract: A broad range of models face the mismatch where they are updated through trajectory losses but are evaluated by downstream task reward.
Scout-Refine-Correct (SRC) is a training recipe that equips flow-matching image generation models with an inference-time refinement capability. During sampling, the model performs a scout rollout to predict a full sample, uses this prediction as a self-conditioning signal, and then takes the actual step; this process can be repeated for multiple refinement rounds. On ImageNet-256, SRC improves LightningDiT-B’s FID from ~14.5 to ~7.26 with 550 normalised flow evaluations, matching a best-of‑N baseline that relies on an oracle verifier, and further reduces FID to 3.81 when combined with best-of‑N.
The paper introduces FHEAT, a fractional diffusion operator derived from the discrete cosine transform, to learn how much spectral mixing each stage of a 3D medical segmentation network should perform. By reparameterizing the operator with a semigroup time, the authors enable the optimizer to decide whether global mixing is needed, resulting in a lightweight U‑shaped architecture (Light‑UNETR) paired with a Kolmogorov‑Arnold mixer (KAN3D). In semi‑supervised experiments, FHEAT‑Seg achieves state‑of‑the‑art Dice scores while dramatically reducing FLOPs through learned spectral sparsification.
The paper investigates how the usefulness of training samples, as determined by coreset selection, depends on the learner rather than just the data. Experiments on ImageNet-100 and ImageNet-1k show that changing model width, input grid, stride, and architecture (e.g., ResNet vs. ViT) shifts the crossover point where different selection criteria (easy-first vs. geometric coverage) become optimal. These findings demonstrate that the relative value of a fixed subset of samples varies with the target learner’s capacity and structure, and that selection strategies must be tuned to the specific model they will train.
arXiv:2609.35767v2 Announce Type: replace-cross Abstract: Unified multimodal models can both look at and render images, so in principle they can repair their own generations: diagnose what an image g...
arXiv:2609.35910v1 Announce Type: new Abstract: Adaptive perception spends extra computation on inputs where perception is expected to improve. When perception feeds a downstream decision system, a b...
The paper presents a convergence framework for deep $V$‑learning over a finite horizon $H$, deriving explicit bounds on policy loss by decomposing the Bellman update error into six residuals. It shows how $L^s$ concentrability controls expected $L^1$ loss, quantifies the impact of shared sampling across horizon levels, and provides optimal and near‑optimal sample allocations for statistical error rates. The work also establishes sharp action‑gap bounds under a margin condition, transfers optimal‑gap results to frozen‑iterate gaps, and offers consistency guarantees for generative‑reset approximate‑ERM procedures with exact action scores.
HARTS is a training system that accelerates agentic reinforcement learning with hybrid‑attention models by sharing prefixes of irregular rollout trees. It uses a two‑level planner to create compression‑oriented microbatches, refines their membership and execution order, and accumulates gradients before a single backward pass through shared prefixes. The system supports distributed parallelism, activation recomputation, and preserves RL objectives while achieving 4.77‑10.45× higher update throughput on large hybrid‑attention MoE models.
arXiv:2608.21098v1 Announce Type: new Abstract: Fusing prior knowledge with data-driven learning is attractive where data is scarce, yet no controlled account says when it helps, is redundant, or har...
EPIG-Tree proposes a compute‑optimal branching strategy for gradient‑efficient reinforcement learning, arguing that branches should be placed where they most reduce policy‑gradient uncertainty per unit of compute. By deriving allocation laws from a law‑of‑total‑variance decomposition, the method introduces an EPIG‑Tree score that guides branch placement using already computed rollouts, estimating occupancy‑ and score‑weighted value uncertainty. Empirical results show EPIG‑Tree reduces gradient MSE in cloned‑state control, improves frozen‑LLM gradient calibration, and outperforms flat GRPO and entropy branching in both single‑turn math and multi‑turn Wordle tasks.
arXiv:2606. 08779v1 Announce Type: new Abstract: Reinforcement Learning (RL) has emerged as a pivotal post-training paradigm, yet it frequently suffers from unpredictable sub-optimum performance or even training collapses.
The paper introduces TRACK, a training‑free trajectory routing method that accelerates video diffusion by selectively switching between large and small models during denoising steps. A calibration process generates a disagreement score map, guiding the selection of the appropriate model at each step to maintain quality while reducing computational cost. Experiments on Wan 2.1, Cosmos 3, TurboDiffusion, and FastVideo show speedups ranging from 1.95× to 2.73× with comparable quality and diversity.