Recent impressive results from large reasoning models like DeepSeek’s R1 have been interpreted as triumphs of Chain of Thought (CoT) reasoning. These models appear to “think through” problems step-by-step before arriving at answers. But are these intermediate tokens actually functioning as meaningful reasoning steps, or are we simply anthropomorphizing model behavior?
A critically examines this interpretation by investigating how the semantics of intermediate tokens—often described as “thoughts” or reasoning traces—actually influence model performance. Using formally verifiable reasoning traces and solutions, the study reveals a surprising finding: models trained on noisy, meaningless traces can perform as well or better than those trained on correct traces. This challenges fundamental assumptions about how reasoning works in language models and suggests we may be .
Understanding the History of Reasoning Traces
Previous research has explored training transformer models on algorithmic execution traces, including Searchformer (transformers trained on A* search algorithm execution) and Stream-of-Search (transformers trained to internalize search processes). These studies demonstrated improved performance but didn’t fully evaluate whether the traces themselves were semantically meaningful.
