This article reviews three ICML 2026 papers that introduce a user‑centric benchmark for mobile GUI agents, a path‑likelihood optimizer for diffusion language models, and a graph‑structured framework that prunes training datasets without sacrificing accuracy.
The paper introduces DLLM‑Searcher, which equips diffusion large language models with a two‑stage training pipeline and a P‑ReAct inference scheme, allowing the model to issue tool calls while simultaneously reasoning, yielding 14‑22% end‑to‑end speedup and matching or surpassing traditional autoregressive agents on multi‑hop QA benchmarks.
The article analyzes a recent Duke University paper that reveals a "scratchpad" mechanism in diffusion large language models, proposes the DPad method to prune redundant suffix tokens before decoding, and demonstrates up to 61.4× faster inference with unchanged or even improved accuracy across multiple benchmarks.
Researchers from UCLA and Meta AI introduce d1, a two‑stage post‑training framework that combines supervised fine‑tuning and a novel diffu‑GRPO reinforcement‑learning algorithm to enable efficient reasoning in masked diffusion large language models, achieving state‑of‑the‑art performance on multiple math and logic benchmarks.
