Interesting Finds — 2026-08-25 (4/4)
Sleep for hybrids, aggressive diffusion decoding, AI-found test-time controllers, a virtual cell, a DeepSeek J-Space report, and PagedAttention explained.
Final group — six to close the set and avoid a single-item remainder.
11. LLM Sleep — offline recurrence for hybrid attention-SSM — 2605.26099
arXiv:2605.26099 — “The authors introduce LLM Sleep, a training and inference framework for hybrid attention-SSM architectures that periodically runs N offline recurrent passes over active context to consolidate information into its fast-weight state-space model blocks before clearing the attention key-value cache.”
Why it matters: “Vanilla hybrid architectures fail to execute deep reasoning over context that has been evicted from active attention, regardless of their nominal storage capacity. By shifting iterative reasoning compute to an offline “sleep” phase, LLM Sleep decouples the computational depth required for memory consolidation from the strict latency constraints of real-time prediction, unlocking superior multi-hop reasoning and long-context comprehension without increasing inference latency.”
Tested on cellular automata, multi-hop graph retrieval, and math reasoning where both vanilla transformers and hybrids failed. More N helps most on examples needing deeper reasoning. The idea is old in a good way: consolidation need not be on the critical path.
12. DMax — aggressive parallel decoding for diffusion LMs
czg1225/DMax — “The authors introduce DMax, a novel training and inference framework designed to unlock aggressive decoding parallelism in Diffusion Language Models (dLLMs). It mitigates the cascading error accumulation that typical parallel decoding methods suffer from by reformulating the standard binary mask-to-token transition into a self-revising continuous trajectory in the embedding space. This is achieved via two core techniques: On-Policy Uniform Training (OPUT), which trains the model on its own predictive distribution to learn self-correction, and Soft Parallel Decoding (SPD), which represents intermediate decoding states as a probability-weighted interpolation between predicted tokens and mask embeddings.”
Why it matters: “While non-autoregressive parallel decoding has promised massive throughput improvements, existing masked diffusion models suffer severe generation quality collapse when forced to decode aggressively due to irreversible early errors. DMax successfully bridges this speed-accuracy gap. Applying DMax to the state-of-the-art LLaDA-2.0-mini baseline yields more than a 2.5x improvement in speed (tokens-per-forward) on complex reasoning (GSM8K) and coding (MBPP) benchmarks with negligible accuracy degradation, achieving over 1,300 tokens per second (TPS) on dual H200 GPUs.”
Parallel decoding’s failure mode is early error that cannot be undone. DMax’s continuous trajectory with self-correction is the counter — keep the speed, allow revision.
13. AutoTTS — discovering test-time scaling
zhengkid/AutoTTS — “Test-time scaling is currently dominated by manually engineered heuristics (e.g., self-consistency, early stopping). AutoTTS demonstrates a paradigm shift where researchers construct discovery environments rather than hand-crafting algorithms. The resulting AI-discovered strategy, the Confidence Momentum Controller, achieves superior accuracy-cost Pareto frontiers, reducing inference token usage by nearly 70% compared to standard self-consistency, all for a total discovery compute cost of just under $40.”
Manual heuristics for test-time scaling give way to discovered controllers. 70% token reduction at held accuracy for $40 of search is a strong efficiency claim; the method — build the environment, let the agent find the controller — is more reusable than the specific controller it found.
14. GenBio — Virtual Cell AI model
lifespan.io — GenBio Launches a Virtual Cell AI Model — GenBio’s virtual cell effort, flagged from TLDR. Details were light in the fetch, but the direction is clear: a cell-scale model for biology in the same way we have built atmosphere-scale models for weather. Track for what data it was trained on and what it can predict versus simulate; the name covers both.
15. DeepSeek-V4 J-Space Capability Realization Report
Tiger3807861189/DeepSeek-V4-J-Space-Capability-Realization-Report — community report analyzing DeepSeek-V4’s J-Space capabilities. Useful as a third-party reconstruction of what the model can do in that subspace, separate from DeepSeek’s own framing. Read for the experimental setup and failure cases more than the headline claims.
16. PagedAttention — Virtual Memory for the KV Cache
Nick Gustafson — PagedAttention — a clear walk through Kwon et al.’s PagedAttention as virtual memory for the KV cache. The core is paging 16-token blocks to eliminate 60-80% fragmentation waste, plus copy-on-write for shared prefixes.
“Each 16-token chunk becomes one physical block in the pool.” Multiple requests sharing a system prompt or parallel samples share physical blocks with refcounts; a write clones one block instead of duplicating the whole cache. For a 2k-token system prompt across 100 concurrent users, that saves roughly 200k tokens of cache — tens of gigabytes on a 70B model in fp16. SGLang’s RadixAttention generalizes the idea further.
If you serve LLMs, this is the page that makes the cost of your system prompt concrete.