• Artisan Brings 3D Assets Into Physical Validation. It recorded the lowest collision rate and highest stability rate among more than 10 representative models. The comparison is limited to the evaluated methods and specified benchmark protocols.

• Policy Residuals May Be Distillation’s Useful Signal. A reference policy filters shared behavior from the expert’s signal. Task-trained residuals may transfer better than differences caused by model scale.

• Evaluation Targets Can Shrink 271× to 611×, but the Substitute Is Not Lossless. Keeping all positives and sampling weighted negative strata cuts costs, while precision and F1 remain ratio estimates.

Also Notable

• World-Model Adaptation Raised Code-Repair Success From 26.3% to 31.2% With a Fixed Coder. This measures predictive selection and adaptation under specific conditions, not general self-evolution.
• ProgramGym Builds 1,648 Verifiable Long-Horizon Program-Reproduction Environments From Code Repositories. Environment repair, black-box test quality, failure recovery, and model fine-tuning all contribute to program-reproduction performance.
• Embodied Agents Benefit From Basic Task Experience but Stop Extracting More Useful Information as Samples Grow. More experience does not naturally become composable skills.
• Representations That Excel at Understanding May Still Underperform at Generation. On ImageNet 256×256, the 459-million-parameter RG-L/16 significantly beats larger pixel diffusion Transformers by directly involving internal representations in pixel prediction.
• Misleading Charts Cut Average Task Success by More Than 34 Percentage Points Across 13 Vision-Language Web Agents. The benchmark contains 280 task instances, arranged as matched comparisons using normal and misleading charts. Failures are also steered toward deceptive branches.
• OGPO Trains VLAs to Reach Semantically Important States Instead of Replaying Demonstrations Step by Step. Offline training can therefore supervise recovery after execution drifts.
• ChemDiffAgent Recasts Reasoning and Tool-Call Traces as a Masked Diffusion Process. Across six chemistry tasks with matched budgets, it improves overall performance by about 20% over autoregressive agents.
• Dynamic Vocabulary Compression Quality Depends on More Than the Compression Ratio. Across four models on general text, three changes close roughly half the gap between zip2zip and the uncompressed baseline at a similar compression ratio. They separate input and output embeddings, correct RoPE positions, and selectively block risky merges.
• X-AVDD Cuts End-to-End Data-Synthesis Cost by About 8× in GPU-Hours. Under VGGSound’s SPC=10 setting, top-1 test accuracy rises from 8.2% to 24.0%, demonstrating the value of explicitly modeling cross-modal correspondence.
• Secret Triggers Can Unlock Encrypted Neural Networks Embedded in Model Weights. Under cryptographic assumptions, weight access and common elicitation methods struggle to detect these arbitrarily hidden capabilities.

Read the full edition →