Researchers from Tsinghua University and NUS introduced ParallelWorld, integrating test-time computation into high-level planning for embodied agents. By simulating multiple candidate action paths and comparing future observations, the method allows robots to anticipate consequences before execution, boosting task success rates from 56% to 83% and offering a new paradigm for embodied decision optimization.
Test-Time Compute
测试时计算
Strategy increasing computational resources during inference to improve output quality.
Curated Intelligence
Anthropic introduced 'Dynamic Workflows' (Flows) at AI Engineer NYC, a new multi-agent orchestration feature for Claude Managed Agents. Designed for ambitious workloads, Flows optimize test-time compute allocation to enhance model capabilities and make complex scientific discoveries more accessible via API.
Anthropic officially introduced 'Dynamic Workflows' (Flows) at AI Engineer NYC, a new multi-agent orchestration feature designed to coordinate Claude Managed Agents for complex workloads. The system optimizes test-time compute allocation to accelerate capability emergence on hard problems, aiming to make scientific discovery broadly accessible via API.
OpenAI released the GPT-5.5 system card, detailing a model designed for complex real-world tasks with enhanced tool use, self-verification, and autonomous execution capabilities. The document discloses red-teaming results for cybersecurity and biology, emphasizing the deployment of the strictest safety safeguards to date. GPT-5.5 Pro, sharing the same underlying model but using parallel test-time compute, underwent separate evaluation for specific risk scenarios.
The Project Numina team released Kimina-Prover, applying test-time reinforcement learning (RL) search to large formal reasoning models. This approach aims to significantly enhance accuracy in mathematical theorem proving and complex logical deduction by dynamically optimizing inference paths, marking a new advancement in AI symbolic reasoning.
Only the title of the Kimina-Prover paper is available, concerning test-time RL search on large formal reasoning models. Without the body text, specific technical details, experimental results, or innovations cannot be assessed.