OmniVerifier-C1 presents a decomposed graded verifier targeted at rule-governed generative world modeling. The paper led Hugging Face daily papers with 26 upvotes.
Lightseek published TokenSpeed, described as a speed-of-light LLM inference engine. The repo gained 374 stars in a day.
A joint paper from Stanford, Together AI, UIUC, and Meta introduces Hybrid Reasoning Agents that reason primarily in a hidden internal mode for speed and cost, switching to explicit text only for tool calls or final answers.