• Local Quality Did Not Translate Into Deployment Success. In OR-License, several models beat the expert reference on conditional objective quality. None matched its DAQ score when infeasible deliveries counted as zero.

• SpecMRL Puts High-Information Directions First. It beat standard MRL at every tested compression level and surpassed full-size MRL with half the embedding dimensions.

• Familiar Factors Do Not Guarantee Compositional Generalization. AnyGoal found failures across several model families during goal transfer, suggesting planners need compositional holdout tests.

Also Notable

• MemT-1.0 Builds Memory Updates Into Model Computation — At a 128K history length, Nano delivers about 1.86× end-to-end speedup and cuts persistent storage 66× versus a full-context KV cache.
• HarnessPro Recursively Improves Execution With Model-Generated Tests — It offers three deployment-budget tiers. Heavy mode raises average pass rate by 2.0 percentage points over RHO across three benchmarks with GPT-5.5 and Codex, while cutting per-task cost by 26.1%.
• BlenderWorld Tests Deliverables on Realistic 3D Tasks — The benchmark contains 223 tasks and 2,651 task-specific binary criteria. Current models still struggle with fine-grained spatial consistency and accurate structural reproduction.
• OpenVLHarness Turns Tool Outputs Into Addressable Evidence — With GPT-6 Luna at medium reasoning effort, it beats the maximum-effort base model, Codex, and Visual Sketchpad by 7.7 to 8.5 points while using at most 37% of their model tokens.
• AdaExploreBench Tests Active Evidence Gathering and Rule Drift — Results suggest the main bottlenecks are acquiring evidence and detecting when it expires, not reasoning after enough evidence is available.
• TokenWorth Selects Memories With Execution Value — It uses sparse feedback to guide retrieval and bounded storage management under joint resource constraints.
• Mirror Learns Video-Action Pairs Through Interaction — A visual planner and inverse dynamics model form a self-improving loop without pre-collected action data or hand-built rewards.
• IGPO Estimates Each Reasoning Step's Causal Contribution — It replaces attention or entropy proxies with parameter interventions, cutting average token use by up to 62% across two model sizes and nine benchmarks.
• Maestro Makes Provers Report Why They Failed — A global coordinator can add guidance, repair a faulty lemma, or retract a subtree and replan instead of spending more budget on a bad decomposition.
• LatComm Separates Context Encoding From Question-Specific Reading — Another frozen model can reuse the fixed compressed representation. On QuALITY, it reduces receiver input 57× and triples document question-answering speed.
• TRACE Organizes Multimodal Clinical Timelines by Hospital Day — It evaluates mortality, length-of-stay, and readmission prediction across cohorts comprising nearly 180,000 total hospital stays.
• MexRegQA Audits Whether Retrieved Evidence Is Sufficient — Of 135 questions marked as covered by retrieval, 35 lacked some or all required facts. Removing evidence-defective items changed nominal significance in 14 of 55 model comparisons.

Read the full edition →