Does Verbose Chain-of-Thought Really Help? A Dual-Validator Replication and Mechanistic Analysis

Chain-of-thought (CoT) prompting improves large language model reasoning, but the source of this improvement remains contested: does CoT help because the intermediate steps carry useful semantic content, or because conditioning on more tokens provides additional computational affordance before the model must commit to an answer? To disentangle these accounts, we ask whether two reasoning traces that express the same semantic content—the same facts, operations, and intermediate values, verified through directed acyclic graph (DAG) equivalence—produce different outcomes when one trace is more verbose than the other. We evaluate across four model targets and eight reasoning benchmarks using a dual-validator design pairing an algorithmic DAG-informed equivalence validator with a more flexible LLM judge, together with multi-sample rewrites, number-redacted completion, and stratified bootstrap confidence intervals. Verbose traces do improve accuracy—25 of 32 benchmark–target cells are positive under at least one validator—but the effects are modest (typically 1–4 percentage points) and depend on the quality of the verbose prose, not merely its length. Under maximum numerical redaction, where every number is replaced by a placeholder, the verbose-helps effect is amplified (median 3.24× across four arithmetic benchmarks), suggesting that the benefit is more consistent with prose-level reasoning structure than with numerical scaffolding visibility. A complementary causal probe—using quiz-bowl questions whose progressively-revealed clues let us vary input information independently of trace verbosity—localises the effect to additive in-context entity priming: naming an entity in the trace raises that entity's probability by 12–17 nats without correspondingly suppressing competing answers, so verbose elaboration helps when it surfaces the right entity and hurts when it does not. This is difficult to reconcile with a purely length-based forward-pass account, yet the persistent small advantage of well-written verbose traces means the semantic-content-only account is not the full story either. The picture that emerges is that the mechanism sits between the two hypotheses: reasoning content is necessary, but verbose elaboration by a competent author provides a further boost that neither pure compute nor pure semantics alone explains.