#test-time scaling

'TUMIX ensembles diverse tool-using agents that share notes and use an LLM judge for early stopping, improving accuracy on hard reasoning benchmarks while cutting inference cost.'

'Google Research's ReasoningBank converts agent traces into high-level strategies and, with MaTTS, yields large gains in effectiveness and efficiency at test time.'

MetaStone-S1 introduces a unified reflective generative approach that achieves OpenAI o3-mini-level reasoning performance with significantly reduced computational resources, pioneering efficient AI reasoning architectures.

Salesforce AI releases GTA1, a powerful GUI agent that outperforms OpenAI's CUA by leveraging innovative test-time scaling and reinforcement learning techniques to improve task success and action grounding.

Research shows that increasing reasoning tokens at test time boosts multilingual reasoning accuracy in English-focused RLMs, especially for high-resource languages and STEM tasks, but struggles with low-resource languages and domain adaptation.