arXiv cs.CL
·
18 hours ago
Researchers scaled zero reinforcement learning to a 1 trillion parameter model to study emergent reasoning capabilities without human-annotated data. The training pipeline incorporated algorithmic optimizations and demonstrated that scaling to 1 trillion parameters enhanced sample efficiency and performance while the model spontaneously developed advanced cognitive behaviors like self-verification and parallel reasoning. The resulting Ring-2.5-1T-Zero model achieved competitive performance on mathematical benchmarks and produced more structured and concise reasoning traces compared to existing approaches.
arXiv cs.CL
·
18 hours ago
Researchers identified how language models separate belief from reality in their computations, showing two mechanisms operate at different positions: a generic value slot that stores attributed values, and a router at the query position that selects which frame (belief or reality) to read from. The behavior emerges consistently across three architectures and five model families at 3B to 7B parameter scales. This separation mechanism could help explain how models handle multiple incompatible interpretations and may apply to other non-actual contexts like counterfactuals and fiction.
arXiv cs.AI
·
18 hours ago
Researchers created a benchmark to test how Large Language Models handle information retrieval and reasoning across massive context windows, finding that model performance varies significantly based on how facts are distributed in the text. The benchmark evaluated four models (Gemini-2.5-flash, ChatGPT-5-mini, Claude-4.5-haiku, and Deepseek-v3.2-chat) on literal extraction, logical inference, and hallucination, revealing two failure modes: performance drops when evidence is scattered across the document, and anti-hallucination prompts can cause models to refuse facts that are actually present. The findings indicate that improving context utilization and handling fact distribution are necessary for deploying large language models reliably in autonomous tasks that require extended reasoning over long documents.
arXiv cs.AI
·
18 hours ago
Researchers developed a theoretical framework to analyze how multi-agent systems with communication perform on reasoning tasks, examining state tracking, recall, and multi-hop reasoning. The framework derives bounds on the number of agents needed, communication bandwidth required, and achievable speedups as problem size increases. The analysis identifies when communication is beneficial and reveals tradeoffs between agent count and bandwidth, with experiments on pretrained language models confirming the theoretical predictions.
