• Token-Level Routing Breaks Existing Batching Assumptions First: TokenRouter separates request-level routing logic from asynchronous per-model execution, raising decode throughput by 2.01–64.15x across routing methods, workloads, and model combinations.

• Practice Worlds May Set the Ceiling for Agent Capabilities: AgentGarten connects verifiable programmatic state with a shared neural renderer through one interface. Its “four rounds versus millions” result compares different learning methods, so it is not a direct sample-efficiency ratio.

• Mixed-Modality Retrieval Falls Into a V-Shaped Trough: Text representations may receive an inherent scoring advantage, making irrelevant text more damaging than the same amount of irrelevant imagery. Stress-test real modality ratios and noise patterns before launch.

Also Notable

• Test Cross-Task, Cross-Scene Navigation Without Fine-Tuning — SuperNav assigns understanding and decision-making to a general multimodal model, then delegates movement to specialized tools.
• Spatial Markers Resolve Which Object to Edit — VibeEdit uses instructions placed on a two-dimensional canvas to identify targets among several similar objects.
• Poster Generation Moves From Linear Prompts to Spatial Composition — Compo expresses two-dimensional composition through explicit semantic, identity, text, and pixel bindings.
• The Best of 58 Reasoning-Based Editors Reaches Only 56.6% Accuracy — RISEBench++ covers six reasoning dimensions and 1,000 human-annotated test cases.
• More Recurrent Transformer Steps Are Not Always Better — InfiLoop preserves reliable intermediate states with recurrence-aware residuals and keeps improving beyond 20,000 effective steps on Sudoku-Extreme.
• Natural-Text Pruning Calibration Misses Real Decoding Activations — SparseDecoding calibrates for generation and optimizes its SpMV kernel, delivering up to 1.48x end-to-end decoding speedup on an A100.
• Half the Visual Tokens Retain 99.5% of Performance — V-CoLA redesigns token selection and merging for linear attention, accelerating prefill by 1.86–6.15x.
• Visual Attention Sinks Change Across Network Depth — The first and last layers repeatedly attend to fixed regions, while middle layers respond to the question. This creates layer-specific intervention points for reducing hallucinations.
• Turning Off “Thinking Mode” Does Not Stop Explicit Reasoning — Open-ended questions most clearly expose persistent reasoning habits and the tradeoff between concise answers and accuracy.
• RAG Context Does Not Necessarily Reduce Training-Data Leakage — It lowers exposure for some examples near the memorization threshold, yet can make previously inaccessible samples extractable.
• Just 197 In-Domain Proteins Can Improve Generation — RefineMix uses out-of-domain data at selected discrete-diffusion timesteps to improve generalization without biasing the sampling distribution.
• Change Cameras at Deployment Without Explicit 3D Reconstruction — VersaCamVLA compresses any camera set into fixed-size scene tokens, then injects them into a pretrained VLA policy.

Read the full edition →