#chain-of-thought

'NVIDIA's Reinforcement Learning Pretraining (RLP) injects dense, position-wise reinforcement into pretraining by rewarding chains-of-thought for information gain, improving reasoning and data efficiency across architectures.'

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.

'PDDL-INSTRUCT trains LLMs to produce logical chain-of-thought traces and uses VAL for stepwise verification, yielding up to 94% valid plans on Blocksworld and major gains across planning benchmarks.'

'Meta AI compresses recurring CoT reasoning into short named behaviors that models can retrieve or internalize, cutting reasoning tokens by up to 46% and improving accuracy in some settings.'

K2 Think is a 32B open-weight reasoning system from MBZUAI that matches or outperforms much larger open models on math by combining long CoT fine-tuning, RL with verifiable rewards, planning, and wafer-scale speculative decoding.

ASTRO, a novel post-training method, significantly enhances Llama 3's reasoning abilities by teaching search-guided chain-of-thought and self-correction, achieving up to 20% benchmark gains.

'ReasonFlux-PRM is a new trajectory-aware reward model that evaluates both reasoning steps and final answers in large language models, significantly improving their reasoning capabilities and training outcomes.'

ByteDance researchers introduce ProtoReasoning, a new framework leveraging logic-based prototypes to significantly improve reasoning and planning abilities in large language models across various domains.

OThink-R1 introduces an innovative framework that enables large language models to switch between fast and slow reasoning modes, cutting redundant computation by 23% without losing accuracy.

Mistral AI introduces the Magistral series, a new generation of large language models optimized for reasoning and multilingual support, available in both open-source and enterprise versions.

The Adaptive Reasoning Model (ARM) and Ada-GRPO introduce a dynamic approach to AI reasoning, significantly improving efficiency and accuracy by tailoring reasoning strategies to task complexity.

'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.'

Anthropic’s research exposes critical gaps in how AI models explain their reasoning via chain-of-thought prompts, showing frequent omissions of key influences behind decisions.

Microsoft launched the Phi-4-Reasoning family, a set of 14B parameter open-weight models optimized for complex reasoning tasks. These models demonstrate competitive performance on math, planning, and coding challenges with transparent training and open access.

New research shows that specialized reasoning models combined with efficient inference-time scaling methods like majority voting outperform non-reasoning models in complex tasks, offering insights into optimizing computational resources.