The Infinite Problem Generator (IPG) uses executable code to synthesize and verify 100% accurate physics reasoning data, overcoming LLM hallucination in data scaling.
arXiv · March 17, 2026 · 2603.14486
The Takeaway
It provides a 'Formula-as-Code' framework that creates high-complexity reasoning traces that are mathematically guaranteed to be correct. This solves the data bottleneck for fine-tuning LLMs on 'hard' sciences where synthetic text often fails.
From the abstract
Training large language models for complex reasoning is bottlenecked by the scarcity of verifiable, high-quality data. In domains like physics, standard text augmentation often introduces hallucinations, while static benchmarks lack the reasoning traces required for fine-tuning. We introduce the Infinite Problem Generator (IPG), an agentic framework that synthesizes physics problems with guaranteed solvability through a Formula-as-Code paradigm. Unlike probabilistic text generation, IPG construc