• A 35B Agent Reaches Trillion-Scale Performance by Scaling Sideways. Agents-A1 doesn't grow parameters. It stacks 45K-token trajectories and heterogeneous skills to match 1T models like Kimi-K2.6 and DeepSeek-V4-pro — but only on some benchmarks.

• 194 Upvotes for Orca's Ambition. It wants a single world latent space and Next-State-Prediction to unify understanding, prediction, and action. The authors call it a preview. Big vision, early landing.

• The Cost of Async Pipelines May Be Overstated. Whether gradient staleness hurts depends on the optimizer. AdamW degrades; the newer Muon shrugs off one-step delay, closing the async-sync gap at 10B scale.

• Tabular Foundation Models Are "General" Only in Their Comfort Zone. Across 11 models and 142 datasets, TFMs lead on small-to-medium IID data. Scale up, add dimensions, or break IID, and tree models take over again.

• What Stalls Mobile 3D Rendering Is Spherical Harmonics, Not Image Quality. Flux-GS uses Monte Carlo energy aggregation to cut the cost of high-order SH, opening a cheaper path for AR and on-device 3D.

Also Notable

• Credit Assignment for Every Tool Call — Outcome rewards can't tell whether a tool call was useful, redundant, or misleading. TACO assigns credit and blame to single calls without an external judge.
• Deliberately the Opposite of Frontier World Simulators — A low-compute, real-time, keyboard-controllable world model for consumer GPUs. Being playable beats being big.
• Self-Correction for Masked Discrete Diffusion — Once a discrete token is unmasked it can't be changed, and that's the key weakness for high-resolution text-to-image.
• Video Models Learn Occlusion and Hand-Object Interaction Just to Predict the Next Frame — Which is exactly the prior that 4D hand-motion reconstruction lacks.
• Molecule Generation Benchmarks Have Long Been Hostage to Drug-Like Proxies — NMO moves the target to real scientific settings like quantum materials, exposing how poorly current models transfer.
• Benign Fine-Tuning on Harmless Data Quietly Reverts Prior Alignment and Unlearning — This paper offers a unified explanation for that post-training fragility.
• Everyone Hopes LLM Rerankers Fix Cold-Start — But a five-domain benchmark that separates rerank quality from recall coverage finds it: if recall misses it, semantic understanding is wasted.
• Fixed Top-N Retrieval Per Query Is Both Wasteful and Harmful — This paper turns "how many to retrieve" into a per-query budget, pulling fewer docs, or none, when the reader already knows the answer.
• Merging Multiple Skills Into One Model Always Robs Peter to Pay Paul — MOPD uses multi-teacher on-policy distillation for post-training skill integration, dodging the inefficiency and quality loss of Off-Policy Finetune and Mix-RL.

Read the full edition →