SourcesSpargeAttention2 — 95% Sparsity 16.2x Attention SpeeduparXiv·high signalXBlueskyLinkedInCopy linkTsinghua hybrid Top-k plus Top-p masking with distillation fine-tuning. 95% attention sparsity, 16.2x speedup on video diffusion. arXiv 2602.13515.SourceSource pagearXiv↳ Follow the threadPolicy dependency / Stack layerTyped Provenance Guardrails Block All 19 Unsafe Releases From a Persistent Agent's Autobiographical MemoryarXiv 2609.02127Stack layerThe best code-embedding retriever returns a functionally broken near-clone as rank 1 in two thirds of queriesarXiv 2609.01865Stack layerSolarWM open-sources a 1.43M-clip world-model data engine with frame-aligned camera geometry, captions and provenancearXivStack layerTyped intention stores let an on-device model beat the best published prospective-memory scaffoldarXivStack layerCoGR trains LLMs to generate keywords on both the query and item side so generative retrieval drops into existing inverted-index infrastructurearXivStack layerSafin-1 builds safety into the architecture through memory routing rather than post-hoc alignmentarXivStack layer / Threat patternThree of Four Major Agent Frameworks Provide No Built-In Confinement for Delegated AuthorityarXiv 2609.00267Policy dependency / Stack layerDistilling 1,000 ML repos into 5,000 verified skills lifted an agent 134% on MLE-bench with the model and budget held fixedarXiv 2609.02749