Researchers from the National University of Singapore, Princeton, and the University of Illinois Urbana-Champaign have identified three crucial factors that enhance AI intelligence: data quality, algorithmic design, and reasoning strategies. Their study demonstrates that a 4-billion-parameter model, DemyAgent-4B, can outperform larger models (up to 32 billion parameters) when trained on authentic datasets rather than synthetic ones. A model trained on real-world data achieved 29.79% accuracy on AIME math benchmarks, compared to less than 10% for its synthetic counterpart. The algorithm GRPO-TCR, with token-based scoring and efficient exploration methods, propelled another model to 70.93% accuracy. Additionally, deliberative reasoning strategies, characterized by deeper thinking and fewer tool calls, yielded superior outcomes. DemyAgent-4B secured impressive scores in various academic benchmarks, validating that well-structured training methods trump mere computational power. Researchers have made their training data and model weights public for further innovation in AI development.
Source link
