7 articles — 🟡 4 important , 🟢 3 interesting
TruDi is the first method that enables diffusion policies to train stably in a massively parallel on-policy RL setting, by applying trust-region optimization with a KL-divergence constraint enforced across the entire diffusion trajectory. Across 73 tasks spanning 4 benchmarks it achieves results competitive with or better than strong baselines, with clear gains on humanoid control.
DLWM addresses visual ambiguity in multimodal reasoning by constructing multiple diverse latent hypotheses simultaneously, preventing their collapse through orthogonal regularization, and adaptively allocating computation via an RL policy. Across multiple benchmarks it achieves a 2–5 percentage-point accuracy improvement while simultaneously reducing memory footprint by 24%.
Researchers from UNC Chapel Hill (Yixuan Wang et al.) identify 'causal heterogeneity' in LLM agent skill libraries: individual skills help on some task types while simultaneously harming others. The ASSAY framework separates skill generation from empirical correction and, through per-task masking, achieves 69.3% task goal completion on AppWorld with DeepSeek-V3 — a relative gain of 47.4% and a new state-of-the-art without weight modification.
Researchers from Renmin University of China introduce CODA-BENCH, the first benchmark that simultaneously evaluates code generation and data retrieval in a realistic Linux sandbox with Kaggle datasets. The benchmark covers 1,009 tasks across 31 communities, with an average of 980 files per environment, and leading agents reach only 61.1% success — accepted at ICML 2026.
Researchers introduce UCS-Bench, a benchmark with over 170 hours of egocentric video and 8,100 temporally annotated questions for testing the spatial intelligence of AI agents, alongside the DirectMe framework that incrementally builds structured spatial memory from streaming observations. Accepted at ICML 2026.
