• Linear Attention's Real Bottleneck Is State-Edit Granularity, Not Speed. Gated DeltaNet-2 splits the scalar gate into channel-wise erase and write gates. It tops Mamba-2, KDA, and Mamba-3 in head-to-head training, with the biggest gains on long-context retrieval.

• Tabular Agents Enter the RL Training Era. Spreadsheet-RL builds a multi-turn sandbox and lifts Qwen3-4B's SpreadsheetBench Pass@1 from 12% to 23.4%. The doubling is real, but the absolute number still sits short of production.

• Reasoning Doesn't Have to Be Text. LatentOmni interleaves audio-visual state inside a unified latent space instead of compressing to discrete tokens. It dodges the language-prior pull that bends CoT toward grammatical sentences.

• A 4B Orchestrator Beats GPT-5 and Gemini-2.5-Pro on Ten Benchmarks. Maestro uses outcome-based RL to schedule frozen experts. Training stability under sparse hierarchical reward, however, is something the abstract skips.

Also Notable

• Transit Planning via Continual Pretraining on 13M Transfer Records, No Routing Engine. TransitLM tests directly whether structured tasks can be served by pretraining alone instead of a specialized system. Not another RAG augmentation.
• MLLMs Score Big Five Traits From Person Videos, Grounded in Observed Behaviors. Separates "perception" from "stereotyping" in the evaluation. Methodology generalizes to other subjective-judgment tasks.
• CUSP Predicts Post-Cutoff Scientific Progress From Pre-Cutoff Knowledge. Cross-disciplinary event-level evaluation. Closer to the actual definition of forecasting than "can AI write a paper."
• Sensor2Sensor Converts Dashcam Video Into the AV Fleet's Sensor Configuration. Long-tail coverage becomes a sensor-conversion problem instead of a data-collection problem.
• SpaceDG Adds Motion Blur, Low Light, and Compression Artifacts to Spatial Reasoning. Almost all current benchmarks assume clean visual input. Adding degradation will likely cut current SOTA scores meaningfully.
• SceneAligner Extends "You Are Here" Localization to Real Raster Floorplans of Public Buildings. Past work assumed vector floorplans and small-scale environments. This one runs in real public buildings.

Read the full edition →