TLDR: Researchers from the National University of Singapore, Princeton, and the University of Illinois Urbana-Champaign have identified three crucial elements for developing more intelligent AI agents: high-quality and diverse data, optimized algorithm design with token-based scoring, and a deliberative reasoning strategy. Their findings demonstrate that a smaller 4-billion-parameter model, DemyAgent-4B, can achieve performance comparable to or exceeding models with up to 32 billion parameters by focusing on these factors.
Researchers from the National University of Singapore, Princeton, and the University of Illinois Urbana-Champaign have identified three key factors that significantly enhance the intelligence of AI agents: data quality, algorithm design, and reasoning strategy. Their study, published on October 25, 2025, highlights that strategic training can enable smaller models to outperform much larger counterparts.
The first crucial factor is data quality. The type of data used during training is paramount, with models trained on authentic learning trajectories achieving significantly higher accuracy compared to those trained on artificial data. For instance, a 4-billion-parameter model trained on real data achieved 29.79% accuracy on AIME math benchmarks, whereas the same model using synthetic data scored under 10%. The researchers emphasize that real data captures the full reasoning workflow, including pre-tool analysis, guided execution, error correction, and self-reflection, which synthetic data cannot replicate. Data diversity is equally important; a mixed dataset of 30,000 examples from math, science, and programming dramatically accelerated learning, reaching 50% accuracy in just 150 training steps, compared to 220 steps for a math-only dataset.
The second factor is algorithm design, specifically focusing on token-based scoring. The team tested three algorithm variants to optimize performance, with a method called GRPO-TCR proving to be the most effective. This approach combines token-level scoring (grading each word chunk), broader clipping for more exploration, and a reward system designed to discourage overly long answers. This optimized method achieved 70.93% accuracy on one math benchmark and 68.13% on another, with token-based scoring outperforming sentence-based methods by approximately 4%. This allows agents to improve both exploration and precision simultaneously through tool interactions, a notable advancement over traditional reinforcement learning.
The third finding pertains to the reasoning strategy, summarized as ‘think more, act less‘. Researchers identified two primary reasoning styles: reactive (characterized by short thinking and frequent tool use) and deliberative (involving longer thinking and fewer tool calls). Models employing the deliberative strategy consistently achieved over 70% success rates in tool use, while reactive models performed poorly due to ineffective or incorrect rapid-fire tool calls. Interestingly, current long-chain-of-thought models, despite being optimized for extended thinking, often struggle with tool integration, tending to avoid tool calls entirely and relying solely on internal reasoning processes.
Also Read:
- AI Agent Systems Mimic Human Scientists to Tackle Complex Design Challenges
- OpenEvolve AI Agent Achieves Significant Algorithm Optimization
Applying these insights, the researchers developed DemyAgent-4B, an AI agent with only 4 billion parameters. The results are compelling: DemyAgent-4B achieved 72.6% on AIME2024, 70% on AIME2025, 58.5% on GPQA-Diamond science tests, and 26.8% on LiveCodeBench-v6 programming benchmarks. This performance rivals or surpasses much larger models, some with 14 to 32 billion parameters, demonstrating that intelligent training and strategic design can overcome brute-force scaling. The training data and model weights for DemyAgent-4B have been publicly released, encouraging further research and development in the field.
