CoT Traces: Pretty Lies, Right Answers

🔥 What happened Researchers at Arizona State University used the synthetic iGSM benchmark to test whether correct answers actually come with valid reasoning traces. On the hardest problems, 31.6% of correct answers had invalid traces — over half failed semantic dependency checks, not syntax or arithmetic. 💡 Why it matters If you rely on CoT monitoring to audit agents, you're trusting an artifact that doesn't causally drive the answer. Shuffling tokens in just 10% of training trace sentences barely dents accuracy — even though zero traces pass verification. ⚡ Our take CoT traces are marketing for the model, not a debugger. Selling them as an audit log is security theater.