ByteDance (China)

Streaming video understanding must retain evidence before anyone knows which details later queries will need, yet existing methods flatten event transitions when compressing history and cannot recover details once a frame has passed. DeltaStreamer explicitly records visual changes between adjacent frames in a hierarchical textual memory, and after a query arrives a query-driven planner decides whether to retrieve, keep observing, or replay historical frames. It scores 66.07% overall on OVO-Bench and 59.53% on StreamingBench, above the representative streaming methods it is compared with.

Streaming video understanding must retain evidence before anyone knows which details later queries will need, yet existing methods flatten event transitions when compressing history and cannot recover details once a frame has passed. DeltaStreamer explicitly records visual changes between adjacent frames in a hierarchical textual memory, and after a query arrives a query-driven planner decides whether to retrieve, keep observing, or replay historical frames. It scores 66.07% overall on OVO-Bench and 59.53% on StreamingBench, above the representative streaming methods it is compared with.

PhysVista introduces a benchmark that links physical state perception, physical dynamics reasoning, and physical plausibility assessment into a closed loop, includes both real-world and AI-generated videos, and evaluates 16 VLMs, finding that models do relatively better on observable state perception but fall short on physical dynamics prediction, counterfactual reasoning, and fine-grained plausibility scoring and ranking, with performance generally better on real-world than generated videos.

PhysVista introduces a benchmark that links physical state perception, physical dynamics reasoning, and physical plausibility assessment into a closed loop, includes both real-world and AI-generated videos, and evaluates 16 VLMs, finding that models do relatively better on observable state perception but fall short on physical dynamics prediction, counterfactual reasoning, and fine-grained plausibility scoring and ranking, with performance generally better on real-world than generated videos.

The work identifies a "uniformity trap" in which language-style token-wise MoE routing, combined with uniform expert-usage regularization, scatters neighboring patches of the same object across different experts in video diffusion, causing routing fragmentation and structural distortion; it proposes SplitMoE, which bifurcates the expert pool into semantic and generic experts and uses prototype-guided routing with pull-push regularization so tokens cluster by semantic attributes, outperforming traditional load-balanced MoE on VBench-2.0 and T2V-CompBench under an equivalent activated-parameter budget, while revealing a coarse-to-fine denoising pattern in which semantic-expert weight peaks in the early high-noise stage around steps 10-15 and then shifts toward generic experts.

The work identifies a "uniformity trap" in which language-style token-wise MoE routing, combined with uniform expert-usage regularization, scatters neighboring patches of the same object across different experts in video diffusion, causing routing fragmentation and structural distortion; it proposes SplitMoE, which bifurcates the expert pool into semantic and generic experts and uses prototype-guided routing with pull-push regularization so tokens cluster by semantic attributes, outperforming traditional load-balanced MoE on VBench-2.0 and T2V-CompBench under an equivalent activated-parameter budget, while revealing a coarse-to-fine denoising pattern in which semantic-expert weight peaks in the early high-noise stage around steps 10-15 and then shifts toward generic experts.

This work studies the training stability of full-pipeline FP8 reinforcement learning for LLMs, identifies that compounded FP8 quantization noise distorts the importance ratio so that negative-advantage tokens are over-clipped at the lower bound and their gradients are erroneously zeroed, and proposes Calibrated Clipping, which dynamically realigns clipping bounds by matching the lower-bound clipping quantile of a BF16 reference and rebalancing the upper bound, eliminating entropy surges and restoring near-BF16 performance across GRPO and DAPO, 8B to 32B models, and multiple FP8 scaling granularities, while reaching up to 1.5x the BF16 training throughput.

This work studies the training stability of full-pipeline FP8 reinforcement learning for LLMs, identifies that compounded FP8 quantization noise distorts the importance ratio so that negative-advantage tokens are over-clipped at the lower bound and their gradients are erroneously zeroed, and proposes Calibrated Clipping, which dynamically realigns clipping bounds by matching the lower-bound clipping quantile of a BF16 reference and rebalancing the upper bound, eliminating entropy surges and restoring near-BF16 performance across GRPO and DAPO, 8B to 32B models, and multiple FP8 scaling granularities, while reaching up to 1.5x the BF16 training throughput.

Researchers from Shanghai Jiao Tong University and ByteDance Seed propose PISA, a block-sparse attention mechanism that uses coarse-to-fine pyramid Top-K selection with LogSumExp scoring and hardware-aware Triton kernels to reduce block-selection complexity from O(N²/C) to O(N log N); across 418M, 1.47B, and 2.67B scales it matches BSA, NSA, and HiLS on language modeling and commonsense reasoning, achieves the highest average accuracy among sparse methods on six containment tasks, and speeds up block selection over BSA by 2.86x, 5.31x, and 9.95x at 64K, 128K, and 256K.

Researchers from Shanghai Jiao Tong University and ByteDance Seed propose PISA, a block-sparse attention mechanism that uses coarse-to-fine pyramid Top-K selection with LogSumExp scoring and hardware-aware Triton kernels to reduce block-selection complexity from O(N²/C) to O(N log N); across 418M, 1.47B, and 2.67B scales it matches BSA, NSA, and HiLS on language modeling and commonsense reasoning, achieves the highest average accuracy among sparse methods on six containment tasks, and speeds up block selection over BSA by 2.86x, 5.31x, and 9.95x at 64K, 128K, and 256K.