VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
VideoMLA applies Multi-Head Latent Attention (MLA) to causal video diffusion, replacing per-head keys and values with a shared low-rank content latent and decoupled 3D-RoPE positional key, achieving 92.7% reduction in per-token KV memory. The paper investigates why MLA works despite pretrained video attention not being low-rank (unlike the spectral assumption motivating MLA in LLMs), finding that the MLA bottleneck itself determines effective rank rather than the pretrained spectrum. On VBench, VideoMLA matches short-horizon baselines, achieves best overall score at long horizons, and delivers 1.23x throughput improvement on a single NVIDIA B200 GPU.
Related guides (4)
Related events (8)
AdaCodec: Predictive Visual Coding for Efficient Video MLLMs
AdaCodec introduces a predictive visual code interface for video multimodal large language models that exploits temporal redundancy in video. Instead of encoding every sampled frame as an independent RGB image, it sends full visual tokens only for reference frames with high conditional predictive cost, and encodes inter-frame changes as compact P-tokens. Evaluated against a Qwen3-VL-8B per-frame baseline across eleven benchmarks, AdaCodec at 1/7 the token budget (32k vs 224k tokens) surpasses the baseline on all long-video benchmarks while reducing time-to-first-token from 9.26s to 1.62s.
Moment-Video: Benchmark Diagnosing Temporal Fidelity of Video MLLMs on Momentary Visual Events
Moment-Video is a new benchmark of 1,000 human-verified video-QA pairs designed to evaluate how well video multimodal large language models (MLLMs) handle brief, localized visual events that may span only a few frames. The benchmark covers 7 domains and 25 subcategories across four task types: Temporal Occurrence, Temporal Counting, Action Description, and Temporal Reasoning. Evaluation of 33 proprietary and open-source models reveals severe deficiencies: the best model (Seed-2.0-Pro) achieves only 39.6% accuracy, while most open-source models score below 25%. Diagnostic analyses show that denser frame sampling helps but does not resolve the bottleneck, pointing to fundamental limitations in how current video MLLMs represent and preserve transient visual evidence.
CLSA: Cross-Layer Sparse Attention with Shared Routing for Efficient Long-Context Inference
Researchers propose Cross-Layer Sparse Attention (CLSA), a method that builds on KV-sharing architectures (like YOCO) to share both the KV cache and the routing index across decoder layers. A single indexer computes token-level top-k selection once and reuses it across layers, reducing routing overhead while preserving fine-grained selectivity. Experiments on short- and long-context benchmarks show up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context, addressing pre-filling, KV-cache storage, and decoding bottlenecks simultaneously.
Squeezing Capacity from MLLMs for Subject-driven Image Generation via Dual Layer Aggregation
This paper proposes conditioning diffusion models on Multimodal Large Language Models (MLLMs) that jointly encode text and reference images, augmented with VAE-based identity conditioning to address copy-paste artifacts and identity preservation failures in subject-driven image generation. A Dual Layer Aggregation (DLA) module aggregates multi-level MLLM features, and a multi-stage denoising strategy progressively balances semantic and fine-detail identity signals during inference. Experiments show improved human preference scores on subject-driven generation benchmarks compared to prior approaches that encode text and reference images separately.
Latent Context Language Models (LCLMs) achieve competitive encoder-decoder KV cache compression at scale
Researchers introduce Latent Context Language Models (LCLMs), a family of encoder-decoder compressors that map long token sequences to shorter latent embeddings consumed by a decoder, targeting the KV cache memory bottleneck in long-context inference. The authors conduct architecture search and continually pre-train 0.6B-encoder/4B-decoder models on over 350B tokens at compression ratios of 1:4, 1:8, and 1:16. LCLMs improve the Pareto frontier across general-task performance, compression speed, and peak memory, and are demonstrated as efficient backbones for long-horizon agents that can skim compressed context and expand relevant segments on demand. The work closes a previously noted gap between encoder-decoder approaches and KV cache compression methods on the accuracy-efficiency frontier.
Survey: Human-View Video Understanding with MLLMs — Watch, Remember, Reason Framework
A new arXiv survey paper proposes a unified 'human-view' framework for analyzing multimodal LLM-based video understanding, organized around three functional abilities: watching (perception), remembering (memory), and reasoning. The authors introduce a formulation characterizing video understanding systems by perceptual representations, memory states, reasoning traces, and predictions, then survey methods, datasets, and benchmarks across these dimensions. The work covers challenges including spatio-temporal perception, long-video processing, streaming understanding, and faithful reasoning, with application domains spanning egocentric, sports, medical, and narrative video.
DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention for Long-Context LLMs
DashAttention introduces a two-stage hierarchical sparse attention mechanism that replaces the fixed top-k block selection used in methods like NSA and InfLLMv2 with an adaptive α-entmax transformation, allowing a variable number of KV blocks to be selected per query. The approach keeps the full hierarchy differentiable by using the first-stage selection as a prior for second-stage softmax attention. Experiments show comparable accuracy to full attention at 75% sparsity with a better Pareto frontier than competing methods, and a Triton GPU implementation achieves meaningful speedup over FlashAttention-3 at inference time.
Using LoRA for Efficient Stable Diffusion Fine-Tuning
This Hugging Face blog post explains how Low-Rank Adaptation (LoRA) can be applied to fine-tune Stable Diffusion models efficiently. LoRA reduces the number of trainable parameters by decomposing weight updates into low-rank matrices, enabling fine-tuning on consumer hardware with significantly less memory. The post covers practical implementation details using the diffusers library.