• DoSA Safely Skips Candidates That Cannot Reach Top-K. Provable bounds avoid full scoring for rejected candidates, delivering up to 1.94x and 1.50x time-to-first-token speedups over SGLang and LiteTopK, respectively.

• Real Behavior Traces Can Train General Abilities. Vela generated 100,000 tasks from anonymized behavior records of 100,000 users and retained roughly 75,000 after filtering. Fine-tuning gained up to 11.5 points on user-related tasks and averaged four points on general benchmarks. Reuse still requires consent, representativeness, and bias controls.

• Deep Research Agents Still Struggle With Everyday and Professional Work. LifeSearch evaluates 203 tasks derived from real-user query seeds against more than 4,000 detailed rules covering numbers, dates, and URLs. The best evaluated agent scores only 63.4%.

Also Notable

• Put Tool Noise, Recovery, and User Corrections in One Reflection Pipeline — Agent-Reflex scores 58.6 across five long-horizon tool-use benchmarks. It outperforms strong environment-scaling and reflection baselines, showing better recovery from noisy interactions.
• Repair Environments, Tasks, and Verifiers Together — WorkRover uses dependency-closure replay to find the smallest compatible fix, then converts workspace progress into per-turn credit signals.
• A GUI Agent's Current Step Should Choose Its Memory Format — SwitchMem routes between textual and visual memory at each step, beating fixed memory representations on three GUI benchmarks.
• Even Self-Distillation Teachers With Answers Lock Onto Shortcuts Too Early — Constraining privileged information through the system prompt alone improves Pass@1 in 11 of 12 settings, averaging 3.9 percentage points.
• Online Harness Improvement Needs No Ground-Truth Verifier — Harness² extracts lessons from changes made during one task, creating a continual improvement mechanism for later tasks.
• Generate Multiscale Concept Codes, Then Refine Text in Parallel — Next-Concept Modeling delivers 10x to 31x speedups across tasks, offering a non-autoregressive path beyond diffusion language models.
• Multimodal Data Ratios Are Asymmetric — Across 268 experiments, audio benefits more from larger models, while vision data significantly disrupts audio learning. Repeating vision or audio data beyond two passes adds little.
• Enterprise Retrieval Tests Need Coverage, Not Just Natural Queries — CoverSynth generates queries for coverage. On RepLiQA, it cuts low-discrimination samples answerable by all 12 systems from 91% with human queries to 12%.
• Next-Token Prediction Can Form Recoverable Taxonomic Trees — Experiments inject synthetic facts from a latent taxonomy into Pythia and OLMo 2. At fixed data budgets, induction weakens with model scale, suggesting larger models may favor memorizing individual facts.
• Polite Stereotypes Can Break Unrelated Tasks — Experiments find lower safety and honesty. Existing defenses improve safety results but largely preserve the underlying bias.
• Later Handoffs Do Not Beat the Initial Router — Starting from Qwen3.5 35B-A3B, later handoffs show no advantage on BFCL or MultiWOZ. A revised supervision objective raises trajectory pass scores by 7.8 percentage points over stepwise labels on TwinRouterBench.
• Route Length Can Directly Train an Optimizer — ResizeCO turns k-opt into GPU tree search. This lets a diffusion-based combinatorial optimizer extrapolate to 10,000 nodes without large collections of optimal labels.

Read the full edition →