FoE: Forest of Errors Makes the First Solution the Best in Large Reasoning Models

Researchers discovered that in Large Reasoning Models like DeepSeek-R1, the first solution is often the best, with alternative solutions being detrimental due to error accumulation. They propose RED, a new framework that achieves up to 19% performance gains while reducing token consumption by 37.7-70.4%.

• →Large Reasoning Models show a surprising 'First is Best' phenomenon where initial solutions outperform alternatives.
• →Errors accumulate in a 'Forest of Errors' structure that grows with test time, challenging accepted scaling laws.
• →The RED framework improves first solutions and discards problematic subsequent ones through dual-consistency.
• →RED demonstrates significant efficiency gains with 37.7-70.4% reduction in token consumption across benchmarks.
• →This research challenges conventional wisdom about test-time scaling in AI reasoning systems.