Reading Between the Dots: Decoding Hidden Computation Across Filler Tokens

(To quote one of the authors)

If you ask a frontier LLM a multi-hop reasoning question, e.g., "Who won the Nobel Prize for Chemistry in (1900 + Mozart's age when he died)?", it usually can't answer correctly immediately (no thinking)

BUT if you ask the same question & append 300 dots, suddenly it can answer? (Because the LLM starts thinking "during" the dots).