DataPrep-Bench sets first unified standard for LLM training data preparation; Skill Self-Play from Qwen co-evolves diverse agent skills via RL; NVIDIA's Molt offers a lean PyTorch-native framework for agentic RL at scale

The data-centric AI movement gets its first unified benchmark. DataPrep-Bench from Peking University (41 upvotes) is quietly significant — it treats data construction and data quality evaluation as co-equal targets rather than studying them in isolation. Their Distributional Alignment Score (DAS) using MMD is the only metric achieving r > 0.70 correlation with downstream performance simultaneously across Math, Science, and Medical domains. This could reshape how teams evaluate synthetic data pipelines.

Skill Self-Play bridges a fundamental tension in self-improving LLMs. The Qwen team's insight that agent skills serve as a natural middle ground between environment-bound verification (reliable but narrow) and open-ended self-generation (broad but unreliable) is elegant. The co-evolutionary loop between proposer, solver, and skill controller is reminiscent of population-based training but operates at the skill abstraction level. The claim of 'striking turnarounds for initially misaligned models' deserves close attention.

NVIDIA's Molt is the anti-complexity framework. In an era of increasingly baroque training infrastructure, Molt's explicit design goal — being 'compact and clean enough for a researcher to hold in their head, and for an AI coding assistant to read and reason about' — feels deliberately provocative. That it matches state-of-the-art Megatron performance while maintaining this simplicity is the real contribution. The paper also notably emphasizes never training on tokens the agent didn't generate, a principled stance on data provenance in RL.

Context management for agents is getting formalized. The Agentic Context Management paper from Maximem AI (17 upvotes) names and decomposes what many production agent teams are struggling with ad hoc. Their economic analysis — showing that naive context accumulation grows token cost quadratically while validated compaction achieves linear cost with preserved fidelity — crystallizes the problem well. The 92% LongMemEval score suggests practical viability.

Sleeper hits to watch: Interactive Training 2 from Waterloo proposes an auditable control plane for steering live training runs through a shared protocol — relevant for the growing interest in human-in-the-loop and agent-guided training. Three-Body Scattering Modeling achieves competitive one-step image generation (FID 1.63 on ImageNet-256) using a physics-inspired energy distance framework, offering a genuinely novel alternative to diffusion and GAN paradigms.