• AesCode Separates Verifiable Correctness From Aesthetics. Its 8B model improved visual quality by 31.1 percentage points over a same-size Qwen3-VL baseline using identical reference images. Results across products still need testing.

• ReTerm Turns Failures Into Recovery Starting Points. It raised Qwen3.5-9B's Pass@1 on Terminal-Bench 2.0 from about 20.2% to 34.8%. Nearly two-thirds of tasks remain unsolved.

• Jeva Replaces Text Generation With Dedicated Scoring. It treats null answers, permutation invariance, changing candidate sets, and cross-request caching as architectural constraints for low-latency classification and routing.

Also Notable

• Cache Structured Scores for Video Questions and Temporal Localization — Follow-up queries over the same visual history can reuse cached visual computation. Performance across four video-question-answering benchmarks matches the backbone's Thinking mode, with measured acceleration above 100×.
• Align 100 Hours of Wearable Data for Humanoid Learning — The dataset connects first-person video, full-body motion, finger joints, and touch during mobile manipulation in workplace settings. Against tracker-fitted references, body-point error falls 71.9% relative to the strongest baseline, alongside better hand stability and contact estimates.
• Turn Task Progress Into Dense, Verifiable Agent Rewards — The method first builds task worlds with complete hidden states, then derives stage-based progress rewards. Tev-Gym contains 1,170 verified task environments across seven domains.
• Separate Motion Shape From Physical Magnitude — One motion-residual prior transfers across cloth, animals, and soft robots without retraining. It also handles unseen geometry families without skeletons, registration, or cross-mesh correspondence.
• Scale Generative Reconstruction to Large Scenes — A global point cloud provides the scaffold, while chunked generation passes context between neighboring regions. Synchronized structure, geometry, and material context preserve scene-level consistency across overlapping chunks.
• Infer the Right Similarity Criterion at Test Time — A small amount of feedback identifies which provided similarity standard fits the current task. BASIS requires no training and handles criteria that vary between instances, but it cannot discover new standards.
• Condition Schrödinger Bridges Without Retraining — Time-adjusted twisted sequential Monte Carlo handles conditional generation and image inverse problems. A changing exponent downweights unreliable early guidance, while telescoping importance weights yield an asymptotically exact sampler for the posterior.
• Estimate Cyclic-Permutation MinHash Variance More Clearly — A more comparable approximate variance expression helps quantify estimator error for circulant-permutation MinHash. Numerical experiments align the expression with the estimator's empirical mean squared error.
• Guarantee Exact Persistent Homology in Generated Graphs — Provable local constructions and global constraint solving replace topology consistency as a soft penalty. Both approaches exactly reproduce target persistence diagrams across four graph-generation benchmarks.

Read the full edition →