• Better Interface Design Improves Desktop Agent Accuracy and Cost: Across eight applications and 368 tasks, atomic interfaces raised two models’ success rates by 3.8 and 3.5 percentage points over native SDK use. They also cut runtime token use by about 73%.

• Evidence Processing Is a Transferable Agent Skill: ViewCraft beat direct execution by 6.1 points on average across eight benchmarks. The advantage persisted under equal token budgets.

• Synthetic Supervision Must Keep Pace With the Policy: RSI-Nav reached 68.6% and 72.4% success on R2R-CE and RxR-CE, though the evidence remains limited to navigation benchmarks.

Also Notable

• SARATHI Separates Pruning Selection From Weight Reconstruction — Data-free probing reduces calibration-domain bias, while shift compensation addresses activation drift at high sparsity. Six models posted competitive results at 15% to 40% sparsity.
• Apertus 1.5 Expands Open Models Across Modalities — It adds image, audio, and 262,144-token context support. Specialized vision and audio models still lead in perception and high-resource-language transcription.
• Complementary Features May Live at Different Depths — SALC replaces layer-aligned fusion with sparse cross-layer connections. It generalizes better across four model pairings and three out-of-domain semantic segmentation settings.
• Tool-Using Labeling Agents Automate Short-Video Governance — After earlier supervised classifiers and domain-pretrained models, this framework cuts manual costs by up to 85% in production.
• Coding Agents Can Improve Behavioral Specifications From Failures — SpecForge uses multi-frontier search to manage cascading regressions from local revisions. It achieved the best results across five base models.
• DNA Models Must Beat Simple K-Mer Features — Only NTv3 showed a consistent advantage among five model families. It improved 46 of 57 datasets, with an average ΔMCC of +0.072.
• Beam Search Leads on Semantic-ID Recommendation Quality — In an industry-scale semantic-ID recommendation system, standard autoregressive beam search beat multi-token prediction on most product-critical ranking and content-QA metrics. At Qwen3-8B scale, its training cost was about 2.6 times lower, while multi-token prediction led on raw inference throughput.
• Video Segmentation Needs Identity-Drift Recovery — Dual-Agent verifies and corrects targets after errors enter the memory loop. It improved SAM3 by 1.2 J&F on LVOS-v2 and 2.7 on SecVOS.
• Sketches and Text Can Map Directly to CAD Code — HiCAD aligns both inputs with executable parametric CAD through collaborative multitask alignment. Spatial rewards enforce geometric and topological consistency.
• Learned Continuous Filters Transfer Across Resolutions — Mesh-Free Convolution learns scale-dependent filters and directional transport terms to improve the long-term stability of autoregressive PDE forecasts.

Read the full edition →