Language Models Need Sleep: Periodic Context Consolidation via Fast Weights and SSM Blocks
This paper proposes a sleep-like consolidation mechanism for transformer-based LLMs to address the quadratic scaling of attention with context length. During 'sleep' phases, the model performs N offline recurrent passes over accumulated context, updating fast weights in state-space model (SSM) blocks via a learned local rule, then clears the KV cache. The approach is evaluated on synthetic tasks (cellular automata, multi-hop graph retrieval) and math reasoning, where standard transformers and SSM-attention hybrids fail, with performance scaling with sleep duration N.
Related guides (4)
Related events (8)
A Sleep-Like Consolidation Mechanism for LLMs
A preprint on arXiv proposes a sleep-like memory consolidation mechanism for large language models, drawing an analogy to biological sleep-based memory consolidation in neural systems. The work appears to address how LLMs might better retain and integrate new information over time, a key challenge in continual learning and knowledge updating. The paper attracted notable community attention on Hacker News with 164 points and 122 comments, suggesting broad interest in the approach.
Sleep paradigm for LLMs enables continual learning and memory consolidation via distillation and RL
A new arXiv preprint proposes a 'Sleep' paradigm for language models that enables continual learning by consolidating short-term in-context memories into long-term parameters. The framework has two stages: Knowledge Seeding (distilling a smaller model's memories into a larger network via on-policy distillation combined with RL-based imitation learning) and Dreaming (self-improvement via RL-generated synthetic curricula without human supervision). Experiments cover long-horizon tasks, continual learning, knowledge incorporation, and few-shot generalization, addressing a known weakness of current LLMs in retaining temporal knowledge across contexts.
Test-Time Training End-to-End (TTT-E2E) Retrains Model Weights to Handle Long Inputs
Researchers from Astera Institute, Nvidia, Stanford, UC Berkeley, and UC San Diego introduced TTT-E2E, a method that compresses long context into transformer weights by training the model during inference via meta-learning. The approach uses sliding-window attention restricted to 8,000 tokens and updates only the fully connected layers of the last quarter of the network on each 1,000-token chunk at inference time, keeping per-token generation latency roughly constant as context scales to 128,000 tokens. TTT-E2E slightly outperforms vanilla transformers on next-token prediction loss across long contexts and matches efficient architectures like Mamba 2 and Gated DeltaNet on inference speed, but fails dramatically on Needle-in-a-Haystack retrieval beyond 8,000 tokens and incurs substantially higher training latency. The work reframes long-context handling as a training-inference trade-off rather than an architectural design problem.
QK-Restore: Fixing long-context recall degradation caused by CoT fine-tuning in hybrid LLMs
Researchers find that chain-of-thought supervised fine-tuning systematically degrades long-context recall in hybrid linear-attention models (HypeNet, Jet-Nemotron), with Needle-In-A-Haystack performance collapsing dramatically—e.g., HypeNet-9B dropping from 67.2% to 9.4% at 256K context. The root cause is identified as CoT-SFT biasing attention gradients toward short-range patterns, corrupting the query-key projections responsible for long-range routing. The paper proposes QK-Restore, a training-free fix that restores only W_Q and W_K from the pre-SFT checkpoint, recovering long-context capability while preserving reasoning gains.
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.
Local linear structures in LLM weights and activations are dynamic, not fixed global directions
A new arXiv paper investigates the nature of linear structures in transformer weights and activations, finding strong local low-rank task-gradient structure but rejecting the hypothesis that fixed task planes exist. The authors show that useful bases drift substantially within 100 optimization steps, yet early recovery updates form a trajectory-prefix basis capturing 77% of LoRA recovery displacement. They also establish a formal connection between parameter perturbations and activation steering, finding a 0.58 cosine similarity between gradient-step-induced activation shifts and CAA steering vectors, suggesting linear structures are evolving local geometries rather than stable global task directions.
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.
ConSA: Learned FA/SWA allocation for efficient hybrid attention in LLMs
ConSA is a framework that learns optimal assignments between full attention and sliding-window attention layers under a user-specified sparsity target, using L0 regularization and augmented Lagrangian constraints. Evaluated on 0.6B and 1.7B parameter models, learned allocations consistently outperform hand-crafted rule-based baselines, with KV-head-wise granularity outperforming layer-wise. A consistent structural pattern emerges: SWA concentrates in bottom layers while FA clusters in contiguous middle-layer blocks, diverging from the evenly interleaved patterns used in existing hybrid architectures.