论文库

原文摘要

Probing studies of latent reasoning read a model's state where the answer begins, and take a successful probe as evidence that an intermediate value was computed and a failed one as evidence that it was not. With prompts that differ in one token, we show that a digit copied from the question is present there in two looped and two standard language models, while a probe that sees one prompt at a time reads it at no more than 0.52 even after 45,000 training questions. Every token we varied leaves such a trace, so a successful probe may be reading the inputs. We give a protocol that tests a candidate value against copies of the inputs, against generic mixing of them (an operation swap with a bar set on synthetic mixtures), and for use (a clamp that controls for the copy), and we check each test where the answer is known. The clamp detects values a model is about to write or reads four tokens later, an unwritten bridge entity in two-hop recall (and not where it is readable but unused), and, in small transformers trained from scratch, an unwritten value used eight tokens later. Our first operation swap failed its positive control; its variance-based replacement passed two fresh controls but, in the small transformers, fires wherever the value is present. In the looped Ouro-1.4B the protocol finds a small code of the unwritten intermediate whose sign follows the operation, but it cannot be told apart from the answer with a copied number added back, and whether the model uses it is unresolved. For a published claim that sub-answers are computed only while a chain of thought is written, weaker tests suggest an early sub-answer; in correctly answered direct answers, whole-state patches find the sub-answer not carried on its own past its equation and, with different answers, the partner's answer in 8% of pairs (below our bar), mostly from the "=" before the answer (post hoc).