• FP8 RL Can Lose Its Negative Feedback Signal. Quantization noise can push negative-advantage tokens beyond clipping boundaries and erase their gradients. Calibrated Clipping removed the reported entropy spikes across GRPO, DAPO, and models from 8B to 32B.

• OmniEdu Organizes Nearly 70,000 Samples Around Four Educational Capabilities. Training improved curriculum understanding, problem-solving, and tutoring benchmarks across 4B, 9B, and 27B models. These results do not yet represent classroom performance.

• UltraTex Separates Two Sources of Waste in 2K Texture Generation. It removes background tokens, sparsifies foreground interactions, and adapts VAE decoding. On common in-dataset samples, training ran 20.6–91.1 times faster and end-to-end inference ran 22.3–74.6 times faster than the selected baseline.

Also Notable

• Solver-Verified Experience Can Stop Optimization Agents From Repeating Modeling Mistakes. OptiSkill stores problem-level modeling frameworks and local error-prevention patterns in a layered skill library. Candidate skills enter the library only after verification.
• Recommendation Models Can Grow Without Retraining From Scratch. Inherit4Rec supports both dense-to-dense scaling and dense-to-sparse expert conversion. It outperformed the evaluated inheritance baselines on two recommendation datasets.
• Higher Prediction Error Does Not Always Mean Model Degradation. A mechanism-controlled stress test separates increasing environmental unpredictability from predictor drift away from the conditional mean. Preset controls then test that attribution.
• Python Is Better Aligned With English, Not a Universal Transfer Hub. Results from three code models found no single center in cross-language geometry. Different transfer directions also preferred different intermediate languages.
• Train Molecular Models On Local Changes Caused By Substituents. MolSC contains 181,000 substituent-level training samples. Its 1,541 evaluation samples do not overlap with training at the scaffold, substituent, or molecule level.

Read the full edition →