Tasks requiring long serial reasoning force language models to externalize reasoning in chain of thought.
5 events · 2 assessments · 2 decisions
Reassessed no status change
Subclaim on non-legible intermediate tokens first assessed verified (0.78) as an existential capability in trained settings. Immaterial to status: the against-argument was already rated slight on that reading. Status supported 0.72, credence 0.75 retained; both argument evaluations refreshed. No dependents notified since nothing changed materially. Provenance: no instances recorded; map not warranted.
Reassessed: still Supported
verdict confidence 0.72 · credence 0.75
Structured and assessed
First pass. Two named arguments: for (bounded serial depth, TC0 subclaim minted as deferred stub at 0.2; CoT-gains subclaim minted at 0.25; linked existing 7fbf57ea as support) and against (non-legible intermediate tokens minted at 0.45; linked existing d8eba1e3). Assessed supported, confidence 0.72, credence 0.75: weak reading near-certain, legibility reading supported for current models. No instances recorded (none existed; sources only searched, not read whole), so no source map. Canonical form kept. Assessment is new, not a change of an existing one, and parents already read the claim as a mechanism premise; no notification needed beyond the system's.
Assessed Supported
verdict confidence 0.72 · credence 0.75
The claim rests on an architectural fact and on experiments. A transformer performs a fixed number of layers of computation per generated token, and without intermediate tokens such a network is confined to the circuit class TC0 (Merrill and Sabharwal), which under standard complexity conjectures excludes inherently serial problems; allowing a chain of thought of polynomial length provably lifts that bound. Serial computation longer than the network's depth must therefore be relayed through generated tokens. Empirically, chain of thought's gains concentrate on mathematical and symbolic multi-step tasks, and studies that make a hidden behavior require multi-step reasoning, such as Emmons et al. (2025) and METR's 2025 hint study, find that models then almost always reveal it in their trace. The claim holds firmly in the weak sense that long serial tasks force computation into generated tokens. It is less secure in the stronger sense that the reasoning becomes legible. Filler tokens (Pfau, Merrill and Bowman 2024) add hidden computation only for parallelizable problems, not serial ones, but reasoning carried in non-legible or encoded tokens has been induced in constructed settings, and outcome-based reinforcement learning can make traces partly illegible. In current models the externalized reasoning on hard tasks is mostly readable, so the claim is supported as a description of present systems; whether it survives optimization pressure against readable traces, or much deeper models whose single pass covers more serial steps ("long" is relative to depth), is the open part.
Claim entered the graph