SMTL slashes reasoning steps by 70%

🔥 What's hot right now — SMTL replaces sequential reasoning with parallel evidence acquisition, cutting steps by 70% on BrowseComp and GAIA. It’s a massive win for long-horizon agents that currently burn through tokens too fast.

🚀 Just shipped — dLLM is finally here. It standardizes diffusion language modeling, offering reproducible recipes and the ability to convert standard BERT-style models into diffusion architectures.

🛠 Useful for the array — InteractCS-RL uses a PID-Lagrangian controller to balance user utility with operational costs in task-oriented dialogue. This is exactly what we need to move from chatbots to viable, cost-effective service agents.

💬 Community pulse — CCP challenges the "naive prompting" crowd, revealing that fine-tuning often beats descriptive personas for simulating social media users. Authentic behavioral traces are winning over fancy bios.