Kolibri turns reasoning length into a dial: low effort keeps 96% of the high score at 35% of the length
Synopsis
Aleph Alpha's Kolibri is a hybrid reasoning model whose system prompt carries one sentence selecting no reasoning, low, medium or high effort. Because its supervised fine-tuning traces carried no effort labels, the team annotated each trace post hoc by its reasoning length relative to traces from the same dataset at the same judged difficulty, with a 2B student model reproducing a larger model's difficulty grades. Median reasoning length orders low below medium below high on 23 of 24 benchmarks; low keeps 96% of the high score at 35% of the length, and the released model keeps 98% at 76% of the completion tokens.
Interpretation
Kolibri makes reasoning length a user-selected setting: one model answers directly or reasons at low, medium or high effort, with high as the default. Earlier approaches either trained a model to decide whether to reason at all, put a token budget into the prompt, enforced a budget at inference, or rewarded a length target during reinforcement learning. Median reasoning length orders low below medium below high on 23 of 24 benchmarks, and the mean score over 14 benchmarks is 38.4 without reasoning, then 64.9, 67.1 and 67.8 at low, medium and high.
The effort labels are inferred after the fact from each trace's reasoning length relative to comparable requests rather than written by annotators. The supervised fine-tuning data records no effort level for any trace, so no direct effort label was available. Grouping by dataset and difficulty grade leaves 176 groups in 69 datasets meeting the 200-exchange threshold, each with its own 30th and 70th percentiles; in the math dataset, low effort on a hard prompt has a median of 4.1K tokens, above the 2.3K of high effort on an easy prompt.
A 2B student model carries the difficulty grading, distilling a larger model's easy, medium and hard verdicts into a rule that can score every conversation. With datasets alone as the domain stand-in, the middle 95% of traces in one math dataset still spans 430 to 79K tokens, so difficulty accounts for much of the spread. The student grades each conversation four times, averages the grades and rounds toward the harder one, agreeing with GLM-5.2 84.8% of the time; split by grade, the math dataset's medians are 592, 2.6K and 18K tokens.
Perspective
The demonstrated scope is the reasoning length and scores of Kolibri's supervised fine-tuning checkpoint and released model on 24, 14 and six benchmarks, together with the labeling procedure that groups training data by dataset and difficulty grade.
The effort labels are assigned post hoc, so a model could predict length from the prompt and ignore the effort sentence; at a fixed difficulty the effect of the setting is smaller than the labels ask for. The dataset is only a proxy for the domain, the 2B judge will sometimes grade difficulty wrong, and reasoning efficiency is not yet addressed, so even at low effort the model can spend tokens that change nothing about the answer.