UltraBERT: Beyond the Limit of BERT Pre-Training UltraBERT revisits encoder pretraining, finding that extra compute saturates after a few trillion tokens and that masking ratios should differ by task.

StreamEval: Evaluating Interactive Streaming Video Generation Models StreamEval benchmarks interactive streaming video across five tasks, measuring execution, state consistency, visual fidelity, and audiovisual coordination.

Deep Research Agents is the fastest-rising task on Papers with Code. 43 papers in the last 30 days, 4.2x its share of new papers in the 180 days before.

UniAVC: Unified Audio-Video Creation via Progressive Training and Asymmetric Scheduling UniAVC separates unimodal competence from cross-modal learning, first training audio and video branches independently before using scarce paired data for correspondence.

InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views InfiniSplat v2 separates geometry from Gaussian attribute prediction, freezes the geometry branch, aligns Gaussians using training-time camera poses, and supervises large-angle target views.

Congrats @free001style on achieving SOTA on HumanEval! DMax-Coder-MMD scores 85.9 Pass@1. That puts it ahead of NeoHorse-1-9B at 0.98.

JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs Decision models can replace many LLM tool-choice calls, but isolated accuracy poorly predicts agent success.

Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling Intern-Decision evaluates structured choices in one causal pass by placing decision placeholders inside a complete assistant JSON skeleton.

OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory OpenVLHarness combines heterogeneous visual tools with persistent multimodal memory, rendering outputs as model-readable evidence and enabling code over stored artifacts.

Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds Scaling continued pre-training improves in-world gameplay, but diversity and model size govern transfer differently.