Onset-Commitment in Large Language Models: First-Token Competition

Paper 4D · Pødenphant Lund (2026) · Read on Zenodo

A language model commits to a route at the very first generated token, and how committed it is can be read there. The first-token competing-routes count (CR) — the number of candidate continuations still in contention at the commit point — becomes a measurement instrument with two orthogonal levers on one dial: training depth sets the commitment, and semantic pressure moves it. The signal is washed out in whole-response mean CR and sharp at the onset.

DOI (concept)	10.5281/zenodo.20562088
Status	v1 live on Zenodo (2026-06-17)
Author	Tomas Pødenphant Lund [ORCID]

Reading the commit point

The raw material is standard: next-token uncertainty, the same distribution that underlies calibration and selective-prediction work and the logit-lens tradition. This paper adds no new quantity. First-token CR is a discretized live-routes count over the same next-token distribution; across the pressure substrates its rank correlation with first-token Shannon entropy is 0.61–0.85. What is new is where and what for: reading uncertainty at the first generated token as a commitment measurement, and showing that two practitioner levers map onto it cleanly.

A first-token CR near 1 means the model has already committed — one route, no contest. A value of 2 or more means it is still holding competing routes open. The reading lives at the onset and is washed out by a whole-response average: mean CR sits at 1.05–1.20 regardless of pressure, where first-token CR carries the whole effect.

The pressure lever moves behaviour and accuracy

Holding a disposition fixed — a premise-checking overlay that makes a model flag premises rather than answer — and varying only the pressure wording, the rate at which the model reflexively flags valid premises moves monotonically with pressure. Low pressure raises the doubt-driven flagging; high pressure suppresses it (Llama-3-8B: 24.4% → 13.3%; Qwen2.5-7B: 11.1% → 0.0% across low → high). The flag is a deliberative re-route; the direct answer is the automatic default. Pooled across susceptible substrates, McNemar p = 0.024, with false-premise detection preserved.

When the deliberative route is the correct one, as in multi-step reasoning, the same shift raises accuracy. On GPQA Diamond, varying only the pressure wording, low pressure raises accuracy substantially on substrates that can solve the task (Llama-3.3-70B: +13.3pp low minus high). The lever is difficulty-gated: on a substrate at the task's floor, more deliberation hurts, because the extra room only lets the model talk itself out of a lucky guess. The sign is also task-type-gated: on a one-step application task, high pressure / first-instinct beats low pressure (+7.6pp, replicated at +7.9pp), the opposite direction from multi-step reasoning. Decisiveness aids the commit-step where deliberation aids the derivation before it.

The mechanism is the onset committor

The behavioural and accuracy effects are invisible in mean CR and visible, sharply, at the first generated token. First-token CR under low versus high pressure: Llama-3-8B 2.39 / 1.00; Qwen2.5-7B 1.39 / 1.00 (permutation p < 0.001 each). The first token is, in both conditions, the start of "ANSWER:". Under high pressure that token carries essentially all the probability mass. Under low pressure the same opening competes with deliberative openings, and the competing-routes count rises. Low pressure literally holds open the choice between committing and deliberating, at the first token.

The position-wise profile makes the mechanism concrete. Under low pressure the count spikes at token 0 and decays within two tokens; under high pressure the first three tokens are fully committed. Beyond token 2 the conditions are indistinguishable. In the language of the quenched-versus-annealed picture of route-competition this is a Kramers committor spike: the barrier-crossing decision is taken at the onset transition state, and first-token CR is the committor read-out. The terms are used in their established technical sense and applied by analogy to token-routing, as a vocabulary for the readout rather than a claim that the forward pass implements Langevin dynamics. The spike is induced by lowering pressure, so the Kramers-like state is temperature-dependent, not a fixed regime. Pressure is the cooling schedule that raises or flattens the committor.

Pressure is temperature; capacity is susceptibility

Whether the temperature knob does anything depends on the substrate, and not monotonically. Measuring the onset-CR pressure-gap across a capacity range gives an inverted-U: Qwen2.5-0.5B +0.30, Qwen2.5-1.5B +0.95, Qwen2.5-7B +0.38, Llama-3.3-70B ≈0 (flat). Susceptibility peaks at mid-capacity. This is stated at its honest resolution: a preliminary within-family capacity gradient — three Qwen points plus a flat 70B from a different family — whose rise-then-fall is consistent with an inverted-U, not an established law. The mechanistic reading (too small to hold competing deliberative routes open; mid-capacity coherent and maximally susceptible; a large well-calibrated model commits regardless of framing) fits the points but is under-determined by four of them.

Two dissociable levels: the fine-tune quenches the onset, the body survives

The two-level structure shows cleanest when the same disposition is installed two ways on the same substrate. On Qwen2.5-1.5B, comparing an in-context overlay against a LoRA fine-tune of the same premise-checking data, three things happen at once. The fine-tune installs a much deeper disposition (over-flagging valid items at 35% versus the overlay's 2.5%). The fine-tune quenches the onset race: where the overlay leaves the onset openable by pressure (gap +0.95), the fine-tune commits at the first token regardless of framing (gap 0.0). And despite that, the fine-tuned model's behavioural modulation survives downstream, in the body of the response (35% → 20% with pressure).

This is the training-depth lever read directly: the same disposition, the same substrate, the same items, and the onset moves from openable to committed purely as a function of how deeply the behaviour is trained in. First-token CR reads how instruction-resistant a route has become. Converted into a continuous dose curve on a second domain (training a wrong route at increasing repetitions), the trained route's ignition rate climbs 21% → 100% and its overridability collapses 63% → 1% — data deepens a basin toward a canyon, made continuous and read on the onset instrument.

Pressure therefore operates at two dissociable levels: an onset committor (first-token route-competition, an inverted-U in capacity, set by training depth and moved by pressure) and a body deliberation (downstream re-routing and reasoning depth) which survives fine-tuning and carries the capable-model accuracy gain. The onset readout captures the first level; it does not capture the second.

What pressure does not do

Two negatives keep the claim honest. Onset CR is a system-state read-out, not an item-level cause: within a pressure condition, an individual item's onset CR does not predict whether that item is over-rejected. Lowering pressure raises onset competition and over-rejection globally, as parallel consequences of the temperature change, not an item-level chain. And the which-answer decision is not CR-coupled at all. This is consistent with the friction series' position that route-competition is a measurement of state, not a driver of the specific decision.

The boundary

The temperature/annealing/Kramers reading is a claim about the language model's own decision physics, a route-competition state read at the commit point. It is not a claim that the model measures, or reproduces the cognitive machinery of, human deliberation under stress; the bridge from CR to a latent evidence-competition variable in human decision models is outside this paper's scope. Where a human parallel is suggestive — pressure-induced collapse onto an automatic route resembling stress-driven reversion to habit — it is offered as a tentative substrate-analogy the model demonstrates, never as a measurement of people.

Connections to other papers in the series

• Paper 1 (Friction Theory) — the substrate-universal framework whose RACE-pressure this paper names from the measurement side; the friction-theoretic pressure knob is the same dial read at the commit point.
• Paper 3 (Friction-Guided Inference) — using competing-routes to decide when to extend computation; the onset readout sharpens where the commitment is made and read.
• In-Context vs Fine-Tuned Memory — the depth axis read behaviourally; this paper reads the same in-context-to-fine-tuned contrast as an onset quench.
• Paper 21 (Mount Stupid) — CR/logprobs as a measurement model for bounded decision; the onset committor is the commit-point instance.
• Paper 14 (Logic as Reactance) — the position-localised reactance cliff that the onset committor's tokens-0–2 localisation matches.

Read the paper

The full paper is on Zenodo (concept DOI 10.5281/zenodo.20562088):

Read on Zenodo → · Plain English version · Dansk version