Budget-Aware Reasoning in Large Language Models: A Survey of Token-Budget Control for Chain-of-Thought
Note: Preprint; drafted with substantial AI assistance under the author’s direction; independent human verification is planned for a later version. Reasoning models raise accuracy by generating long chains of thought (CoT) and pay for it in tokens. The literature treats three different problems as one: efficiency (shorter reasoning on average), adaptiveness (length that tracks difficulty) and controllability (length that respects a budget someone specified). This survey takes controllability as its organizing problem. It covers 503 works through October 2026, organizes mechanisms by the pathway through which a budget signal enters the model (prompt text, decoding-time enforcement, in-sequence control tokens, training objectives, representation space) and by who sets the budget (user or system, model, hybrid), and extracts 19 core method families and three further control interfaces in detail. Four results organize the synthesis. First, the word budget carries seven senses across papers and vendor APIs (hard cap, soft cap, exact target, minimum, effort level, ratio, self-estimate) that decide whether stopping early counts as success or as deviation. Second, learned adherence to a budget, adherence enforced by truncation, and correctness within a budget are three different measurements; a metric family with at least five names mixes them, and we specify the counting unit, aggregation and identification conditions under which each can be recovered. Third, within-study comparisons support setting-specific conclusions on reward design (stateful budget signals help; RL after SFT improves accuracy at a budget; tightening curricula beat fixed budgets; one-sided gates can be satisfied by vacuously short answers), while, beyond that one gate-versus-band run, no controlled single-chain study isolates the penalty shape, and cross-study differences in counting, enforcement and protocol preclude a pooled ranking. Fourth, observed accuracy declines under forced continuation describe how current policies spend tokens, not a property of larger budgets: the constrained optimum cannot fall when a larger cap still permits earlier stopping. We derive the resulting protocol pitfalls, propose a reporting checklist with a per-response record schema, and give analytic checks that qualify three published theoretical claims. Retrieval was restricted to documented queries, and extraction depth varies across the corpus; both limits are recorded in the study registry released with the survey.
Authors
- Salam M. Ghazi Al-Mohanna (ORCID: https://orcid.org/0000-0002-6409-8874)
Institutions
- Xi'an Jiaotong University (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24195009
- Primary Topic
- Topic Modeling
- Type
- preprint