Also Worth Noting - 2026-08-14

Also Worth Noting

02 [Eval] PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives Frame-quality scores miss the test that matters: can an agent actually complete multi-step objectives inside a world model? PlayWorld replaces clip-level metrics with agent players that pursue long-horizon goals, such as rotating 360 degrees to check environmental consistency or walking into water to verify physics. That design catches failures invisible to per-frame evaluation. Teams building game or robotics world models should run this before trusting short-clip quality numbers. link

03 [Inference] Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation Standard video distillation pipelines let bidirectional teachers score frames that depend on future context the student never had, breaking online controllability at inference. Context-Matched Distillation enforces causal ordering during training so the teacher only scores what the student could have seen. The fix recovers online control without giving up step-count reduction. Any team distilling autoregressive video models for real-time applications should audit whether their teacher is peeking forward. link

04 [Agent] LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation Eager consolidation, calling an LLM after every interaction to update memory, makes costs compound as conversation length grows. LycheeMemory V2 defers consolidation to retrieval time instead, processing segments lazily only when a query arrives. That shift cuts LLM calls by roughly half on long sessions without measurable recall loss. Teams running long-horizon agents in production will recognize the cost curve this addresses. link

05 [Training] Maglev: Sliding Recurrent Memory Constant KV cache size at inference is normally incompatible with full-attention quality, but Maglev couples two models to close that gap. A full-attention prefiller produces memory targets; a recurrent decoder with sliding-window attention injects those targets as compressed KV state, keeping cache growth at zero. Training stays parallelizable because the prefiller runs over the full sequence offline. The paper claims parity with dense-attention baselines on long-context tasks, making this a practical efficiency tradeoff worth tracking for serving infrastructure. link

06 [Theory] Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing Injecting a free-form passage into a model is not the same as the model being able to use it: existing unstructured knowledge editors leave the edited model able to recite the passage while failing atomic questions about its facts, let alone multi-hop reasoning over them. The authors name this missing property composability and trace it to editors passively relying on the fixed passage as the only learning source, proposing proactive self-distillation instead. Teams patching production models with knowledge edits should test edited facts compositionally, not just by recall, before trusting the edit. link