A new paper by researchers from Google Research and UC Berkeley shows that a simple sampling-based search approach can enhance the reasoning abilities of large language models (LLMs) without needing specialized training or complex architectures.
ByteDance Research has released DAPO (Dynamic Sampling Policy Optimization), an open-source reinforcement learning system for LLMs, aiming to improve reasoning abilities and address reproducibility issues. DAPO includes innovations like Clip-Higher, Dynamic Sampling, Token-level Policy Gradient Loss, and Overlong Reward Shaping, achieving a score of 50 on the AIME 2024 benchmark with the Qwen2.5-32B model.
Alibaba's Qwen team aims to find out with its latest release, QwQ. Despite having a fraction of DeepSeek R1's claimed 671 billion parameters, Alibaba touts its comparatively compact 32-billion 'reasoning' model as outperforming R1 in select math, coding, and function-calling benchmarks.
Improving the reasoning capabilities of large language models (LLMs) typically requires supervised fine-tuning with labeled data or computationally expensive sampling. We introduce Unsupervised Prefix Fine-Tuning (UPFT), which leverages the observation of Prefix Self-Consistency -- the shared initial reasoning steps across diverse solution trajectories -- to enhance LLM reasoning efficiency. By training exclusively on the initial prefix substrings (as few as 8 tokens), UPFT removes the need for labeled data or exhaustive sampling. Experiments on reasoning benchmarks show that UPFT matches the performance of supervised methods such as Rejection Sampling Fine-Tuning, while reducing training time by 75% and sampling cost by 99%.
The article introduces Huginn-3.5B, a novel AI reasoning model developed by researchers from multiple institutions. It utilizes a recurrent depth approach for efficient and scalable reasoning by refining its hidden state iteratively within a latent space, rather than relying on external token generation. This allows it to dynamically allocate computational resources and perform efficiently across various tasks without needing specialized training data.
A new test-time scaling method called budget forcing boosts LLM reasoning without increasing model size, outperforming OpenAI's o1-preview.
This method, developed by researchers at Stanford University, controls the computational effort an LLM expends during inference, allowing it to either stop reasoning early or think longer. The researchers created a curated dataset called s1K to test this method and found that their model, s1-32B, outperformed OpenAI’s o1-preview model on competitive math benchmarks by up to 27%.
Dolphin 3.0 R1 is an instruct-tuned model designed for general-purpose reasoning, coding, math, and function calling. It is designed to be a local model that businesses can control, including setting the system prompt and alignment.
The article discusses the implications of DeepSeek's R1 model launch, highlighting five key lessons: the shift from pattern recognition to reasoning in AI models, the changing economics of AI, the coexistence of proprietary and open-source models, innovation driven by silicon scarcity, and the ongoing advantages of proprietary models despite DeepSeek's impact.
AI researchers at Stanford and the University of Washington trained an AI 'reasoning' model named s1 for under $50 using cloud compute credits. The model, which performs similarly to OpenAI’s o1 and DeepSeek’s R1, is available on GitHub. It was developed using distillation from Google’s Gemini 2.0 Flash Thinking Experimental model and demonstrates strong performance on benchmarks.
Researchers at UC Berkeley have developed Sky-T1-32B, an open-source reasoning-focused language model trained for less than $450, which surpasses OpenAI's o1 in benchmarks like Math500, AIME, and Livebench. This model uses optimized training processes to balance computational efficiency with robust performance, making it accessible to a broader audience and fostering inclusivity in AI research.
