The field of large language models is moving towards developing more efficient and effective reasoning capabilities. Researchers are exploring various approaches to improve the reasoning abilities of these models, including optimizing chain-of-thought reasoners, speculative search, and confidence-guided compression. These advancements have the potential to unlock the full potential of large language models in real-world applications, such as scientific discovery and decision support systems. Noteworthy papers include Llama-Nemotron, which introduces a series of open-source reasoning models with superior inference throughput and memory efficiency. Another notable paper is Optimizing Chain-of-Thought Reasoners via Gradient Variance Minimization, which proposes a dynamic sample allocation strategy to minimize stochastic gradient variance. Additionally, ConCISE and Elastic Reasoning are introducing innovative approaches to compress and control the output lengths of large reasoning models, making them more suitable for real-world deployment.
Efficient Reasoning in Large Language Models
Sources
Llama-Nemotron: Efficient Reasoning Models
Optimizing Chain-of-Thought Reasoners via Gradient Variance Minimization in Rejection Sampling and RL
A Survey of Slow Thinking-based Reasoning LLMs using Reinforced Learning and Inference-time Scaling Law
Accelerating Large Language Model Reasoning via Speculative Search
Recall with Reasoning: Chain-of-Thought Distillation for Mamba's Long-Context Memory and Extrapolation
Long-Short Chain-of-Thought Mixture Supervised Fine-Tuning Eliciting Efficient Reasoning in Large Language Models
When Reasoning Beats Scale: A 1.5B Reasoning Model Outranks 13B LLMs as Discriminator
Reward-SQL: Boosting Text-to-SQL via Stepwise Reasoning and Process-Supervised Rewards
ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning
Scalable Chain of Thoughts via Elastic Reasoning