• Natural-Language Specifications Can Compile Into Local Neural Functions. Compile by Training spends about one minute training reusable capabilities and reaches 83.6% semantic accuracy on a FuzzyBench-Hard subset.

• Random KV Eviction Can Match Complex Selectors. Random Attention protects the prompt, allocates budgets per attention head, and raises vLLM throughput by 32% to 43%.

• LLaDA-Image Publishes a Staged Image-Generation Recipe. Its 6B model scores 53.53 and 53.38 on Qwen-Image-Bench’s English and Chinese tracks. Distilled variants need only two to four inference steps.

• Translation Evaluation Turns Failures Into Actionable Rules. LTB combines peer-reviewed multimodal cases with manually verified rules, making results useful for diagnosis and regression testing.

Also Notable

• On-Policy Distillation Keeps Improving on a Single Query. Training coverage may come from students visiting new states, not just from the original sample count.
• DRACO Uses Adaptive Rubrics for Long-Horizon Credit Assignment. It adds within-trajectory reward signals when agent training lacks programmatic verifiers.
• FlashRender First Corrects Camera-Control Bias From Sampling Steps. Few-step distillation then turns slow viewpoint changes into rendering that takes seconds.
• Frame Selectors Don't Explain Every Long-Video Gain. Controlled experiments isolate the contributions of frame selection, compression, and budget reinvestment.
• One-Off Coding-Agent Traces Can Become Reusable Terminal Environments. Terminal-Universe turns recorded runs into environments for repeated queries, execution checks, and post-training.
• Recurrent Layers in Hybrid Models May Tolerate 4-Bit Quantization. Full-layer NVFP4 W4A4 experiments retest whether errors accumulate through recurrent state.
• Test Contamination Inflates Scores but Rarely Reorders Models. Deployment decisions should measure absolute score distortion separately from ranking distortion.

Read the full edition →