SemanticKV compresses KV cache for agentic video generation
Authors report a 3.35x KV-memory cut and 1.22x speedup on Causal Forcing at compression ratio 0.7.
Stories per day · 30 days
2 RSS
Relateddiffusion5video-generation4distillation3multimodal3text-to-image3world-models3halcyon2training2vbench2audio-video1benchmark1compression1denoising1digital-humans1editing-framework1flow-models1glance-forcing1gpic1
34 stories, newest first · one card per story
Stories tagged #autoregressive
History Forcing improves long-horizon video world models
History Forcing is a distillation framework that strengthens historical supervision for few-step autoregressive video generators, achieving state-of-the-art results.
MoviEdit enables long-form multi-shot video editing
MoviEdit uses cascaded bidirectional-autoregressive design and 87K training videos to outperform open-source models on video editing.
TaoMate-HD streams high-resolution audio-video with progressive denoising
Authors report a 1.6x speedup versus direct high-resolution generation and an 11.18-point drop in character error rate.
Paper splits visual generation into autoregressive knowledge and diffusion rendering
A 3B text-to-image model trained in 6,600 H100 hours scored 86 on GenEval and 87 on DPG-Bench.
Future Chunk Prediction trains autoregressive video models on later chunks
The auxiliary objective predicts separated future chunks without intervening frames and does not change autoregressive inference.
Lafa scales continuous flow language models to 8B
Trained on fewer than 30B tokens, its self-speculative decoding reaches 2.54 times autoregressive throughput at comparable quality.
Halcyon trains text-to-image models without language pretraining
The authors report Halcyon-1.6B beating a larger JiT baseline on gFD with 275 times less pretraining compute.
Halcyon trains text-to-image models without language pretraining
Authors say Halcyon-1.6B beats a JiT baseline on gFD with 275 times less pretraining compute.
MT Forcing trains multi-turn image editing to limit drift
Single-Turn DMD and Teacher-Forcing Guidance target artifact buildup and unintended changes across edit turns.
Language Models Learn to Improve Language Models
LM2LM uses a research world model as an RL reward so Qwen3.8-27B proposes stronger model improvements than GPT-6-sol.
Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
Glance Forcing distills causal video diffusion with one PF-ODE trajectory and dual slow-fast experts.
RepForcing: Representation-Space Initialization for Autoregressive Video Distillation
RepForcing improves four-step autoregressive video distillation using representation-space initialization before DMD.
MARC: Multi-Bit Watermarking for Autoregressive Audio Generation against Codec Attacks
MARC embeds multi-bit watermarks in autoregressive audio that survive codec and overwriting attacks.
ForcingBooster: Adaptive Patch-wise Denoising for Fast Autoregressive Video Diffusion
ForcingBooster cuts autoregressive video diffusion steps about 50% by stopping easy patches early.
Long-WAM: Scaling the Context of World-Action Models
Long-WAM scales robot world-action context, lifting RoboCasa success from 63.3% to 78.7%.
SGF+: Decoupling Gradient Flows for Autoregressive Video Generation
SGF+ decouples video context writing from denoising, enabling 24-hour generation from 5-second training rollouts.
Disentangling Paradigm, Identifier, and Decoding in Generative Retrieval
Decoding choice, not just the paradigm, shifts diffusion generative retrieval Hit@1 by up to 13.7 points.
Empirical Variational Autoencoder
Empirical VAE learns autoregressive latent priors with one linear layer, matching diffusion quality at faster inference.
SheetSage2: Coherent Lead-Sheet Transcription with Synthetic Supervision
SheetSage2 transcribes coherent lead sheets using synthetic MIDI supervision and beats prior systems on most reported metrics.
Learning Latent Protein Languages for Autoregressive Generation
Learned latent protein languages improve autoregressive sequence and structure generation versus amino-acid tokens.
In-Distribution Forcing for Long Video Generation at Test Time
ID-Forcing keeps KV caches in-distribution so short-horizon video models generate minute-scale clips without drift.
SemanTok: Predictable Semantic Tokens for Efficient Autoregressive Video Generation
SemanTok aligns video tokens with frozen DINO features so short prefixes carry clip semantics.
Rollout-Marginal Distillation for Long-Horizon Autoregressive Video Generation
Rollout-Marginal Distillation scores video chunks independently so long autoregressive rollouts keep visual quality.