Three pointed papers do the day's heavy lifting. Agents' Last Exam asks why benchmark wins don't translate to real outcomes and proposes harder, broader tests. FlashMemory-DeepSeek-V4 makes ultra-long-context inference workable with lookahead sparse attention, attacking the KV-cache bottleneck the silicon market repriced last week. Why Muon Outperforms Adam delivers a curvature-perspective explanation for Muon's ~2x training-efficiency gain. SpatialWorld and Skill-3D push interactive spatial benchmarks toward real-world tasks, and Echo-Memory tests memory mechanisms inside action-world models. Press cycle quiet.
Agents' Last Exam names the gap between benchmark gains and real-world outcomes; SWE-Explore probes how coding agents actually navigate large repos; SlimSearcher cuts the cost of deep research agents with adaptive reward gating; DuMate-DeepResearch builds an auditable rubric-grounded multi-agent system.
Position-and-benchmark paper arguing strong scores on existing agent suites don't translate to real outcomes — proposes harder, broader evaluation.
Benchmarks the exploration behavior of coding agents in real repositories — past SWE-bench-style single-task scoring.
Trains deep-research web agents with adaptive reward gating — cuts the compute cost of strong agentic search.
Multi-agent deep-research system with recursive search and rubric-grounded reasoning — auditable by design.
FlashMemory-DeepSeek-V4 ships lookahead sparse attention for ultra-long context; End-to-End Context Compression at Scale attacks KV-cache growth structurally; Why Muon Outperforms Adam gives a curvature explanation for the optimizer that's been quietly displacing Adam in pretraining.
Lookahead sparse attention plus a lightning index let DeepSeek-V4 serve ultra-long contexts without holding the full KV cache.
End-to-end learned context compression for long-context inference — addresses KV-cache growth without per-token surgery.
Explains Muon's ~2x training-efficiency gain over Adam via a local-curvature analysis — clean theoretical backing for an optimizer already in widespread use.
SpatialWorld pushes interactive spatial reasoning into real-world tasks; Skill-3D evolves agentic 3D skills; Latent Spatial Memory and Echo-Memory study memory in action world models; AHA-WAM proposes asynchronous horizon-adaptive world-action modeling.
Spatial-reasoning benchmark grounded in real-world interactive tasks rather than synthetic 3D puzzles.
Controlled study of how memory mechanisms shape action-conditioned world models.
Latent spatial memory replaces explicit point clouds for 3D-consistent video world models.