Also Worth Noting - 2026-08-05

Also Worth Noting

02 [Training] LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models MoE diffusion language models do not follow the same scaling rules as autoregressive MoE models , the optimal batch size grows faster and the optimal learning rate decays more sharply with compute. IsoFLOP analysis also reveals a different model-to-data allocation curve than AR models produce. These are not minor deviations; they mean AR-derived scaling recipes will systematically misconfigure compute budgets when applied to diffusion LM training runs. Teams planning MoE dLLM training at scale need their own empirical sweep, not a port from the AR playbook. link

03 [Inference] RestoreKV: Recovering Full-Cache Behavior Under Aggressive Query-Agnostic KV Cache Eviction Selection and restoration are orthogonal axes for KV cache compression, and combining them beats either alone at the same memory budget. RestoreKV keeps the existing selection step but adds a small set of learned restore tokens that attend to the compressed cache after prefill and reconstruct a compact complement of what was evicted. The key insight is that the eviction pattern is context-specific, but the mechanism for generating its complement is learnable and shared across contexts. Teams hitting quality collapse under tight KV budgets can treat RestoreKV as a drop-in addition to any selection-based eviction scheme. link

04 [Eval] PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents Retained task history does not reliably translate into better future behavior for current personal agents , a finding that undercuts the recursive self-improvement assumption built into most personal-agent product roadmaps. PAST-Bench isolates this question by running agents through ordered sequences of fresh-session tasks under matched conditions that toggle retained experience on and off, across 26 scenarios. The benchmark exposes that accumulated preferences, tool routines, and learned skills are not being exploited effectively by today's systems. Any team shipping a personal agent with a "gets better over time" promise should run this benchmark before making that claim publicly. link

05 [RAG] Search, Inspect, Fetch: Exploiting Boolean Retrieval for Deep-Research Agents Treating web sources as flat pages rather than structured objects with inspectable fields is the root cause of irrelevant context flooding deep-research agent pipelines. SIEVE introduces a search-inspect-fetch interface built on fielded Boolean retrieval: it filters candidates over document fields like titles, headings, and metadata, ranks the admitted set, presents structure-rich result cards for agent inspection, and fetches only selected documents. No model fine-tuning is required. Any search-visit agent loop can swap in this interface as a drop-in to cut irrelevant context ingestion without touching the underlying retrieval model. link

06 [Agent] When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills Persona skill pipelines concentrate fragmented personal signals into reusable artifacts that are far easier to extract than the original records, defeating differential-privacy guarantees designed for individual memory entries. AntiSkillBench evaluates this attack surface end-to-end across a dataset of 7,500 persona-grounded dialogue traces, covering leakage, impersonation risk, and defense coverage. The amplification happens because the compiled skill aggregates and reuses signals across many interactions, not because any single record is exposed. Teams shipping persona-skill features should not assume that per-record privacy budgets bound the privacy risk of the compiled artifact. link