fangwu97/DeepSearch-1.5B
DeepSearch-1.5B by Fang Wu et al. is a 1.5 billion parameter reasoning model based on Nemotron-Research-Reasoning-Qwen-1.5B v2, trained with Reinforcement Learning with Verifiable Rewards (RLVR) and enhanced by Monte Carlo Tree Search (MCTS). This model integrates MCTS directly into training for systematic exploration and efficient replay buffering, achieving state-of-the-art accuracy among 1.5B reasoning models. It is optimized for complex mathematical and reasoning tasks, demonstrating improved performance on benchmarks like AIME and AMC while being 5.7 times more compute-efficient than extended RL training baselines.
Loading preview...
DeepSearch-1.5B: Enhanced Reasoning with MCTS
DeepSearch-1.5B is a 1.5 billion parameter reasoning model developed by Fang Wu et al. from institutions including Stanford University and Amazon AWS. It is built upon the Nemotron-Research-Reasoning-Qwen-1.5B v2 base model and introduces a novel training approach that integrates Monte Carlo Tree Search (MCTS) directly into Reinforcement Learning with Verifiable Rewards (RLVR).
Key Innovations & Capabilities
- Integrated MCTS Training: Unlike traditional methods, DeepSearch incorporates MCTS into the training loop, enabling systematic exploration, fine-grained credit assignment, and efficient replay buffering.
- State-of-the-Art Reasoning: Achieves leading accuracy among 1.5B reasoning models, demonstrating superior performance on various mathematical benchmarks.
- Compute Efficiency: Despite its enhanced capabilities, DeepSearch-1.5B is 5.7 times more compute-efficient than extended RL training baselines, reducing training costs and carbon footprint.
- Benchmark Performance: Shows an average accuracy improvement of +1.25 points over the best prior 1.5B model across benchmarks like AIME, AMC, MATH500, Minerva, and Olympiad.
Training Details
The model was trained on the rigorously decontaminated DeepMath-103K dataset using 16 NVIDIA H100 GPUs for approximately 330 GPU hours. The training employed a sophisticated search strategy including Global Frontier Selection, Entropy-based guidance, and a replay buffer with solution caching.
Ideal Use Cases
DeepSearch-1.5B is particularly well-suited for applications requiring strong mathematical reasoning, problem-solving, and complex logical deduction, especially where computational efficiency is a key concern.