August 8, 2026

Researchers from Tsinghua, Zhejiang University and Meituan propose AgentOPSD, a method that redistributes sparse outcome rewards into turn-level credit signals for long-horizon multi-turn agent RL, using recursive Bayesian updates over teacher-student likelihood gaps instead of uniform trajectory-level advantage broadcasting (as in GRPO). It reaches 89.1% success on ALFWorld with Qwen2.5-7B without extra rollouts or learned critics.

A cross-university team built OSReward, a benchmark of 1,019 human-annotated computer-use trajectories across web, Windows, Ubuntu and mobile, to test VLM-as-judge reliability. Frontier judges score ~90% overall but drop to 70% on hard cases, mostly by wrongly accepting incomplete tasks as successes; the team also released open OS-Shepherd judge models (9B/35B) matching commercial judges at 30-60x lower cost.

Tencent Hunyuan researchers present WorldClaw, an agentic coarse-to-fine pipeline that turns text prompts into large, freely explorable 3D worlds via three stages: intent planning, global terrain generation, and regional object population with editable, textured meshes, while preserving global spatial coherence.

A joint academic-Tencent team propose EnvACE, where a single policy both acts and simulates its own environment responses ('world rehearsal') during RL training, removing the need for external simulators. This absorbs environment dynamics into the model's parameters as an internalized world model, improving both training efficiency and test-time performance on agent benchmarks.

ByteDance Seed and academic collaborators introduce GST-Bench, a 2,762-question video VQA benchmark testing whether vision-language models can build global spatial awareness from long egocentric videos, including inference from novel viewpoints and mapping to top-down layouts. The best zero-shot VLM scores 42.68 versus a 79.08 human baseline; proprietary models fail mainly at cross-frame integration, open models fail at both local and global reasoning.