• KGD Separates Knowledge Refresh From Task Adaptation. It beats strong baselines by 4%–12% across eight public benchmarks and sustains gains over a 90-day data stream and a Shopee production test.

• Persistent Memory Does Not Guarantee Continuous Improvement. PAST-Bench uses 26 scenarios and 204 task episodes to test whether saving, retrieving, and updating experience actually improves future behavior.

• Hunyuan3D-Buffalo Unifies 3D Understanding, Generation, and Editing. Its 87-million-sample multimodal corpus includes 12 million editing examples, targeting geometrically consistent local changes and integrated workflows.

Also Notable

• Queryable World Models for Cheaper, Safer Agent Feedback. Quo Vadis, World Modeling? extends world models beyond forecasting physical futures into controllable interaction proxies agents can query before acting.
• Confident Teacher Outputs Can Still Be Spurious Signals. The proposed on-policy distillation method identifies language priors and formatting templates unrelated to the task input.
• Video Research Benchmarks Expose Two Evaluation Illusions. Agents may bypass visual tools or present memorized parameters as genuine retrieval and execution. Video-DeepResearch tests exploration across continuous video streams and the open web.
• Diffusion MoE Models Need Their Own Scaling Laws. Optimal batch size, learning rate, and compute allocation cannot be copied from autoregressive models. LLaDA MoE v2 measures these behaviors systematically.
• Any-OPD Bridges Incompatible Teacher and Student Representations. It connects heterogeneous teachers to flow-matching distillation across different VAE latent spaces, architectures, and timesteps.
• Larger Skill Libraries May Not Produce Better Agents. ContinualSkillBench tests whether new skills can be reused, combined, and applied to improve later tasks.
• The LM Output Head Can Hide Quantization Costs. ARCHead compresses the high-precision output head after the Transformer body is quantized, using activation-aware residual correction.
• ChronoLens Compares How Languages Change Across Multiple Levels. Its cross-temporal, multilingual measurements test whether morphology, syntax, semantics, and pragmatics change in step.

Read the full edition →