Sumi is an openly released 7B parameter uniform diffusion language model pretrained from scratch on 1.5T tokens that matches autoregressive models on several benchmarks.
super hub Canonical reference
Efficient Streaming Language Models with Attention Sinks
Canonical reference. 74% of citing Pith papers cite this work as background.
210
Pith papers citing it
32
external citations · Pith
Background
74% of classified citations
abstract
Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a "sink" even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.
hub tools
citation-role summary
background 36
method 7
baseline 2
other 2
citation-polarity summary
claims ledger
- abstract Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attent
authors
co-cited works
representative citing papers
Creates LoCoMo benchmark dataset for very long-term LLM conversational memory and shows current models struggle with lengthy dialogues and long-range temporal dynamics.
A 0.6B LM with length-aware attention adjustments performs competitive in-context retrieval at million-token scale on MS MARCO, NQ, and LIMIT benchmarks.
ISPA reduces KV cache size by up to 50% in AR video models by transitioning layers to local attention and applying instance-specific least-squares weight modulation to compensate for lost history.
Self-GC governs agent context as indexed objects with planner-proposed actions, achieving 84.85% no-impact on future continuations on a hard set versus 54-70% for baselines.
In 160M and 290M parameter models, a new residual-stream split into scratch and protected channels causes massive activations to re-emerge in the protected decode channel, more concentrated on the start token.
Distance-Adaptive Representation (DAR) keeps full KV dimensionality inside a local window and reduces it to 1/4 outside, matching full-dimensional baselines on pretraining (70M-410M) and 1B-scale fine-tuning while uniform reduction performs worse.
Reroute turns irreversible visual-token pruning into recoverable routing that reuses existing attention scores, improving grounding performance under aggressive reduction on LLaVA-1.5 and Qwen while preserving TFLOPs and KV-cache budgets.
AVLLMs route audio-visual information sequentially in video tasks and via parallel streams for interleaved items, allowing early token discard with little performance loss across models and scales.
STAR-KV applies differentiable soft thresholding for per-head and per-block adaptive low-rank KV cache compression, combined with hybrid decomposition and low-rank-aware quantization, achieving up to 75% compression and 3.1x throughput gains.
HACK++ is a head-aware KV cache compression framework for VAR models that decouples current-scale attention from historical cache under adaptive per-head budgets to achieve near-lossless generation at 30% attention and 10% cache budgets.
The same composed tasks are realized by different attention-head patterns in different models when the same selectivity-plus-ablation protocol is applied.
Characterizes high-dimensional phase structure of momentum under sparse updates via closed-form second-moment dynamics, with regimes matching SGD, unstable, or heavy-ball depending on retention-to-learning timescale ratio.
AVMP separates KV and SSM cache pools behind unified virtual addressing with failure-triggered migration, cutting OOM events 7.6% and raising throughput 1.83-13.3x on synthetic loads and 2.36x on ShareGPT traces.
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
HeadKV compresses KV cache for autoregressive image generation via head-aware budget allocation, early head-type identification from consistent patterns, and stratified token eviction.
EventPrune prunes 80% of visual tokens in Video-LLMs using event camera motion cues, yielding 1.89x speedup, 52% fewer GFLOPs, and slightly higher accuracy than full-token baselines on first-person dynamic spatial reasoning.
LongLive-2.0 delivers an NVFP4 parallel infrastructure that enables direct training of long multi-shot autoregressive diffusion video models and achieves up to 2.15x training and 1.84x inference speedups on Blackwell and other GPUs.
AdaFocus achieves better accuracy on long-video benchmarks with roughly 33 times fewer visual tokens by combining query-aware adaptive sampling and zero-cache disk-based refinement.
CausalCine enables real-time causal autoregressive multi-shot video generation via multi-shot training, content-aware memory routing for coherence, and distillation to few-step inference.
MISA routes to a small subset of indexer heads via block statistics, matching full DSA performance on LongBench with 4-8x fewer heads and 3.82x speedup while recovering over 92% of selected tokens.
FLUID is a continuous-time transformer using Liquid Attention Networks to model attention as stable ODE solutions that interpolate between discrete SDPA and CT-RNNs, with an explicit sink gate and liquid hyper-connections for better information flow.
TSCG compiles JSON tool schemas into token-efficient structured text, raising tool-use accuracy for small LLMs from 0% to 84.4% on benchmarks while cutting tokens by 52-57%.
OCR-Memory encodes agent trajectories as images with visual anchors and retrieves verbatim text via locate-and-transcribe, yielding gains on long-horizon benchmarks under strict context limits.
citing papers explorer
-
Sumi: Open Uniform Diffusion Language Model from Scratch
Sumi is an openly released 7B parameter uniform diffusion language model pretrained from scratch on 1.5T tokens that matches autoregressive models on several benchmarks.
-
Evaluating Very Long-Term Conversational Memory of LLM Agents
Creates LoCoMo benchmark dataset for very long-term LLM conversational memory and shows current models struggle with lengthy dialogues and long-range temporal dynamics.
-
Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale
A 0.6B LM with length-aware attention adjustments performs competitive in-context retrieval at million-token scale on MS MARCO, NQ, and LIMIT benchmarks.
-
Towards Memory-Efficient Autoregressive Video Generation via Instance-Specific Parametric Absorption
ISPA reduces KV cache size by up to 50% in AR video models by transitioning layers to local attention and applying instance-specific least-squares weight modulation to compensate for lost history.
-
Self-GC: Self-Governing Context for Long-Horizon LLM Agents
Self-GC governs agent context as indexed objects with planner-proposed actions, achieving 84.85% no-impact on future continuations on a hard set versus 54-70% for baselines.
-
Massive Activations Are Architecturally Robust: A Controlled Scratch/Commitment Residual Stream Test
In 160M and 290M parameter models, a new residual-stream split into scratch and protected channels causes massive activations to re-emerge in the protected decode channel, more concentrated on the start token.
-
Dual Dimensionality for Local and Global Attention
Distance-Adaptive Representation (DAR) keeps full KV dimensionality inside a local window and reduces it to 1/4 outside, matching full-dimensional baselines on pretraining (70M-410M) and 1B-scale fine-tuning while uniform reduction performs worse.
-
Reroute, Don't Remove: Recoverable Visual Token Routing for Vision-Language Models
Reroute turns irreversible visual-token pruning into recoverable routing that reuses existing attention scores, improving grounding performance under aggressive reduction on LLaVA-1.5 and Qwen while preserving TFLOPs and KV-cache budgets.
-
From Senses to Decisions: The Information Flow of Auditory and Visual Perception in Multimodal LLMs
AVLLMs route audio-visual information sequentially in video tasks and via parallel streams for interleaved items, allowing early token discard with little performance loss across models and scales.
-
STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control
STAR-KV applies differentiable soft thresholding for per-head and per-block adaptive low-rank KV cache compression, combined with hybrid decomposition and low-rank-aware quantization, achieving up to 75% compression and 3.1x throughput gains.
-
HACK++: Towards More Effective Head-Aware Key-Value Compression for Efficient Visual Autoregressive Modeling
HACK++ is a head-aware KV cache compression framework for VAR models that decouples current-scale attention from historical cache under adaptive per-head budgets to achieve near-lossless generation at 30% attention and 10% cache budgets.
-
Pattern Selectivity is Not Task-Causal Structure: A Cross-Architecture Mechanistic Study of Composed-Task Circuits in 1B-Class Language Models
The same composed tasks are realized by different attention-head patterns in different models when the same selectivity-plus-ablation protocol is applied.
-
Dynamics of Stochastic Momentum with Sparse Updates in High Dimensions
Characterizes high-dimensional phase structure of momentum under sparse updates via closed-form second-moment dynamics, with regimes matching SGD, unstable, or heavy-ball depending on retention-to-learning timescale ratio.
-
Asymmetric Virtual Memory Paging for Hybrid Mamba-Transformer Inference
AVMP separates KV and SSM cache pools behind unified virtual addressing with failure-triggered migration, cutting OOM events 7.6% and raising throughput 1.83-13.3x on synthetic loads and 2.36x on ShareGPT traces.
-
Tensor Cache: Eviction-conditioned Associative Memory for Transformers
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
-
Head-Aware Key-Value Compression for Efficient Autoregressive Image Generation
HeadKV compresses KV cache for autoregressive image generation via head-aware budget allocation, early head-type identification from consistent patterns, and stratified token eviction.
-
EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning
EventPrune prunes 80% of visual tokens in Video-LLMs using event camera motion cues, yielding 1.89x speedup, 52% fewer GFLOPs, and slightly higher accuracy than full-token baselines on first-person dynamic spatial reasoning.
-
LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation
LongLive-2.0 delivers an NVFP4 parallel infrastructure that enables direct training of long multi-shot autoregressive diffusion video models and achieves up to 2.15x training and 1.84x inference speedups on Blackwell and other GPUs.
-
AdaFocus: Adaptive Relevance-Diversity Sampling with Zero-Cache Look-back for Efficient Long Video Understanding
AdaFocus achieves better accuracy on long-video benchmarks with roughly 33 times fewer visual tokens by combining query-aware adaptive sampling and zero-cache disk-based refinement.
-
CausalCine: Real-Time Autoregressive Generation for Multi-Shot Video Narratives
CausalCine enables real-time causal autoregressive multi-shot video generation via multi-shot training, content-aware memory routing for coherence, and distillation to few-step inference.
-
MISA: Mixture of Indexer Sparse Attention for Long-Context LLM Inference
MISA routes to a small subset of indexer heads via block statistics, matching full DSA performance on LongBench with 4-8x fewer heads and 3.82x speedup while recovering over 92% of selected tokens.
-
FLUID: Continuous-Time Hyperconnected Sparse Transformer for Sink-Free Learning
FLUID is a continuous-time transformer using Liquid Attention Networks to model attention as stable ODE solutions that interpolate between discrete SDPA and CT-RNNs, with an explicit sink gate and liquid hyper-connections for better information flow.
-
TSCG: Deterministic Tool-Schema Compilation for Agentic LLM Deployments
TSCG compiles JSON tool schemas into token-efficient structured text, raising tool-use accuracy for small LLMs from 0% to 84.4% on benchmarks while cutting tokens by 52-57%.
-
OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory
OCR-Memory encodes agent trajectories as images with visual anchors and retrieves verbatim text via locate-and-transcribe, yielding gains on long-horizon benchmarks under strict context limits.
-
PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training
Stealth Pretraining Seeding plants persistent unsafe behaviors in LLMs via diffuse poisoned web content that activates on precise triggers and evades standard evaluation.
-
Stream-CQSA: Avoiding Out-of-Memory in Attention Computation via Flexible Workload Scheduling
Stream-CQSA uses CQS-based decomposition to stream exact attention computations for billion-token sequences on limited-memory hardware.
-
HybridGen: Efficient LLM Generative Inference via CPU-GPU Hybrid Computing
HybridGen achieves 1.41x-3.2x average speedups over six prior KV cache methods for LLM inference by using attention logit parallelism, a feedback-driven scheduler, and semantic-aware KV cache mapping.
-
On the Emergence of Syntax by Means of Local Interaction
A 2D neural cellular automaton spontaneously self-organizes into a Proto-CKY representation that exhibits syntactic processing capabilities for context-free grammars when trained on membership problems.
-
AsyncSparse: Accelerating Sparse Matrix-Matrix Multiplication on Asynchronous GPU Architectures
AsyncSparse presents BCSR and WCSR kernels that use TMA and warp specialization to accelerate SpMM, outperforming prior libraries by 1.47-6.24x on SuiteSparse and achieving 2.66x end-to-end speedup on Qwen2.5-7B at 90% block sparsity.
-
HeadRank: Decoding-Free Passage Reranking via Preference-Aligned Attention Heads
HeadRank lifts preference optimization into attention space via entropy-regularized head selection and distribution regularizers to sharpen discriminability for efficient listwise reranking.
-
Improving Sparse Autoencoder with Dynamic Attention
A cross-attention SAE with sparsemax attention achieves lower reconstruction loss and higher-quality concepts than fixed-sparsity baselines by making activation counts data-dependent.
-
A Mechanistic Analysis of Looped Reasoning Language Models
Looped LLMs converge to distinct cyclic fixed points per layer, repeating feedforward-style inference stages across recurrences.
-
Transactional Attention: Semantic Sponsorship for KV-Cache Retention
Transactional Attention uses semantic sponsorship from anchor patterns to retain dormant critical tokens in KV caches, achieving 100% credential retrieval at 16 tokens where all prior methods fail.
-
CodeComp: Structural KV Cache Compression for Agentic Coding
CodeComp uses Joern-extracted Code Property Graph priors for training-free structural KV cache compression, outperforming attention-only baselines on bug localization and code generation while matching full-context patch quality.
-
When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models
A V-sink/L-sink taxonomy plus a frozen-backbone, NTP-trained layer-wise key gate (LSG) improves LLaVA-1.5-7B by up to +1.55pp on MMStar and +3.08pp on CVBench.
-
MemDLM: Memory-Enhanced DLM Training
MemDLM embeds a simulated denoising trajectory into DLM training via bi-level optimization, creating a parametric memory that improves convergence and long-context performance even when the memory is dropped at test time.
-
STAC: Plug-and-Play Spatio-Temporal Aware Cache Compression for Streaming 3D Reconstruction
STAC compresses KV caches in streaming 3D reconstruction transformers via temporal token preservation with decayed attention, spatial voxel compression, and chunked multi-frame optimization, delivering 10x memory reduction and 4x faster inference at SOTA quality.
-
PrefixWall: Mitigating Prefix Caching Side Channels in Shared LLM Systems
PrefixWall mitigates APC side channels in multi-tenant LLM systems via selective prefix isolation, delivering up to 70% higher cache reuse and 30% lower latency than full-isolation baselines.
-
Visual Para-Thinker: Divide-and-Conquer Reasoning for Visual Comprehension
Visual Para-Thinker is the first parallel reasoning framework for MLLMs that uses visual partitioning strategies, Pa-Attention, and LPRoPE to extend test-time scaling benefits to visual comprehension tasks.
-
Exact Flow Linear Attention: Exact Solution from Continuous-Time Dynamics
Exact Flow Linear Attention derives a closed-form exact update for delta-rule linear attention from continuous-time dynamics, removing Euler discretization error while preserving linear complexity and structure.
-
Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling
Four Over Six adaptively scales blocks in NVFP4 quantization to smaller FP4 values, making representable value distributions more uniform and reducing quantization error especially for near-maximal values.
-
DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning
DELTA partitions layers into full, delta, and sparse groups to select salient tokens via aggregated attention scores, matching full-attention accuracy on AIME and GPQA while cutting attended tokens up to 4.25x and achieving 1.54x speedup.
-
Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention
Low-precision Flash Attention fails due to similar low-rank attention representations combined with biased rounding errors that accumulate and corrupt weight updates; a minimal fix to reduce rounding bias stabilizes training.
-
FastKV: Decoupling of Context Reduction and KV Cache Compression for Prefill-Decoding Acceleration
FastKV decouples prefill context reduction via Token-Selective Propagation from independent KV cache selection, delivering up to 1.82x prefill and 2.87x decoding speedups while matching decoding-only accuracy.
-
PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction
PyramidDrop accelerates LVLMs by staged, similarity-based dropping of visual tokens that become redundant in deeper layers, delivering 40% faster training and 55% lower inference cost with comparable accuracy.
-
Scaling and evaluating sparse autoencoders
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
-
Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention
Infini-attention combines compressive memory with masked local attention and long-term linear attention inside each Transformer block to support infinite context length with bounded resources.
-
Massive Activations in Large Language Models
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
-
Lost at the End: Primacy Bias in Multimodal Retrieval-Augmented Question Answering
In multimodal KB-VQA, gold evidence at the first prompt slot beats gold at the last by 16–26 points, flipping the classic U-shaped lost-in-the-middle pattern into primacy bias.
-
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.