A study of early-stopping methods for large reasoning models finds that cutting off generated reasoning can reduce token use, but the trade-off with accuracy and reliability differs across models and tasks. The researchers evaluated methods on mathematical and scientific reasoning, code generation, and factuality using five models and a shared replay protocol.
The paper introduces Target-Integrated Token Efficiency (Tite), a measure that compares a method’s token use with an oracle across different accuracy targets. In the reported experiments, the oracle—which selects among discrete reasoning budgets—achieved higher accuracy than the models’ natural, full-length reasoning in several settings. For example, on mathematical and scientific reasoning, gpt-oss-20b scored 0.894 with the oracle, compared with 0.824 using full reasoning. On code generation, R1-Qwen3-8B reached 0.514 with the oracle versus 0.382 with full reasoning; on factuality benchmarks, the same model scored 0.730 versus 0.540.
That potential does not mean existing stopping techniques reliably find the best point to end reasoning. The study reports that more aggressive methods, including DTSR-1 and CoDE-Stop, often used fewer tokens but experienced substantial accuracy losses. DEER, EntroCut, and PUMA showed more favorable accuracy-efficiency trade-offs in some cases, but no method performed best consistently across all settings.
The results also show why accuracy alone can be misleading. On mathematical reasoning with gpt-oss-20b, Statistical Early Exit matched full-reasoning accuracy at 0.824 but recorded a TITE score of 0.191. DEER and EntroCut had lower accuracy, at 0.773, but higher TITE scores of 0.381 and 0.380. The authors say current methods still have difficulty locating dependable stopping points while keeping the cost of checking them low.
The researchers additionally tested whether answers remain correct as reasoning continues. They found cases where correctness fell at later points by more than their model of decoding noise predicted. Their reliability checks also found continued reasoning after forced termination and weaker supporting reasoning in some outputs that reached a correct answer early. The findings suggest that evaluating early-stopping requires looking beyond token savings and final-answer accuracy to whether the retained reasoning supports the answer.
Sources: arXiv cs.AI.