One question through a reasoning model, and common misconceptions

One question through a reasoning model, and common misconceptions A workflow diagram generated by Archify. EX / Misconceptions at answer time 02 / One question at answer time (main) 03 / Training time and memory EX / Misconceptions about training and cost Uses the exact budget · budget is a soft cap · Misconceptions at answer time Uses the exact budget budget is a soft cap A new architecture · still a Transformer · Misconceptions at answer time A new architecture still a Transformer Aha from nothing · base models self-check · Misconceptions at answer time Aha from nothing base models self-check Longer is more right · longer, not surer · Misconceptions at answer time Longer is more right longer, not surer CoT shows real thought · not always faithful · Misconceptions at answer time CoT shows real thought not always faithful Prompt + budget · effort, budget, mode · One question at answer time (main) Prompt + budget effort, budget, mode Prefill · one forward pass · One question at answer time (main) Prefill one forward pass Decode thinking · thousands of tokens · One question at answer time (main) Decode thinking thousands of tokens End of thinking · self-stop or cut off · One question at answer time (main) End of thinking self-stop or cut off Decode answer · much shorter · One question at answer time (main) Decode answer much shorter Return · hidden, summary, or full · One question at answer time (main) Return hidden, summary, or full Verifier · checks right/wrong · Training time and memory Verifier checks right/wrong Weights after RLVR · same Transformer · Training time and memory Weights after RLVR same Transformer KV cache · grows every token · Training time and memory KV cache grows every token N runs + select · optional: vote, verifier · Training time and memory N runs + select optional: vote, verifier Cannot be hacked · verifiers have holes · Misconceptions about training and cost Cannot be hacked verifiers have holes RL adds new skills · still debated · Misconceptions about training and cost RL adds new skills still debated Thinking is cheap · billed as output · Misconceptions about training and cost Thinking is cheap billed as output Voting fits all tasks · needs comparable answers · Misconceptions about training and cost Voting fits all tasks needs comparable answers reward weights write K, V sample N times Legend compute step training time: RL and verifier budget control and answer selection memory: KV cache common misconception