The Illusion of Thinking: Models Don't Actually Reason (arxiv.org)

Recent research reveals significant limitations in Large Reasoning Models (LRMs), which have been designed to emulate human-like reasoning processes. Although LRMs have shown improved performance on reasoning benchmarks, this study highlights that their fundamental capabilities are not well understood. By employing controllable puzzle environments that isolate problem complexity, researchers discovered that LRMs experience a complete collapse in accuracy when faced with tasks of high complexity. Interestingly, they displayed a counterintuitive scaling behavior, initially improving their reasoning effort with increasing complexity before it sharply dropped, suggesting significant constraints in their ability to handle advanced reasoning tasks. The findings challenge the prevailing evaluation methods that focus solely on final answer accuracy, which can mask the underlying reasoning processes of these models. By comparing LRMs against standard language models, the study categorizes performance into three regimes: low, medium, and high complexity tasks, each presenting different strengths and weaknesses. This research underscores the intricate dynamics of model reasoning, emphasizing that while LRMs may excel in some scenarios, they ultimately struggle with exact computations and demonstrate inconsistency across scales, raising important questions about the authenticity of their reasoning capabilities within the AI/ML community.