Records found: 6
SwiReasoning alternates between silent latent reasoning and explicit chain-of-thought based on next-token entropy, improving accuracy and dramatically cutting token usage on math and STEM benchmarks.
'R-Zero is a co-evolutionary framework where a Challenger creates hard problems and a Solver learns from them, enabling zero-data training that boosts reasoning accuracy across benchmarks.'
ProRLv2 scales RL training to 3,000 steps and combines regularization and exploration techniques to expand reasoning capabilities in compact LLMs, showing strong benchmark gains across math, coding, logic and STEM tasks.
GURU introduces a multi-domain reinforcement learning dataset and models that significantly improve reasoning abilities of large language models across six diverse domains, outperforming previous open models.
'Researchers introduce Soft Thinking, a training-free method that allows large language models to reason with continuous concept embeddings, enhancing accuracy and efficiency in math and coding tasks.'
Researchers at UC Berkeley and UCSF have developed Adaptive Parallel Reasoning, a novel method that allows large language models to dynamically distribute inference tasks across parallel threads, enhancing reasoning performance without exceeding context window limits.