LightReasoner: training a big model on the 10 percent of tokens that matter

The premise: most of a reasoning trajectory teaches nothing

LightReasoner is the code release for the paper Can Small Language Models Teach Large Language Models Reasoning, accepted as an Oral Presentation at the ACL 2026 Main Conference. The argument starts from three weaknesses of supervised fine-tuning. It is data-intensive, because it relies on human-labelled or rejection-sampled datasets. It applies uniform learning, training all tokens equally even though only a small portion truly matter. And it has a ground-truth dependency, which hinders adaptability to new domains and reasoning formats. The paper's framing is that 90 percent of the compute goes into what the model already knows, while the critical 10 percent that drives breakthroughs is under-invested. The method that follows is to allocate effort the other way round, and the project's own summary of it is that strategic token selection, rather than exhaustive training, is what releases the latent reasoning ability of a model.

Amateur and expert, and why the smaller model helps

The central mechanism is a pair of models with a specific relationship. LightReasoner uses smaller amateur models to identify critical reasoning moments where stronger expert models should focus their learning, and the selection is done with what the project calls Expert-Amateur KLD detection, a divergence measure between the two models' predictions. The project states the assumption behind this as expertise over scale: domain expertise gaps, rather than model size, drive effective contrast, so even same-sized models with different knowledge can generate strong teaching signals. That is the claim that distinguishes this from ordinary distillation, where the small model is the student and the large one the teacher. Here the small model is the detector, and it works because it is wrong in the places the big model is also weak rather than uniformly worse. The learning rule is described as KL-guided learning, using the behavioural divergence between expert and amateur predictions to pinpoint reasoning bottlenecks, and the project says this works without requiring ground-truth labels at all.

The three stages, and the fine-tuning script

The framework is described as a three-stage lightweight pipeline. First, critical step selection via Expert-Amateur KLD detection, which is the stage described above. Second, contrastive supervision that captures the behavioural differentials between expert and amateur. Third, self-distillation for internalising the expert's strengths. The repository layout matches those stages, with LightR_sampling.py for the sampling and selection stage, LightR_finetuning.py for the training stage, data_prep.py for preparing data and merge.py for merging adapters. The models trained are Qwen2.5-Math and DeepSeek-R1, and the repository carries a gitmodules file, an analysis/ directory, an evaluation/ directory, and a LRsamples/ directory with pre-collected training samples. The dependency set is modest: transformers at 4.44.0 or later, accelerate, datasets, huggingface_hub, tqdm and peft for parameter-efficient fine-tuning.

What the reported numbers actually say

The results are specific and worth reading carefully, because the spread between model pairs is the interesting part. Qwen2.5-Math-1.5B gains 28.1 percent on GSM8K, 25.1 percent on MATH, 7.2 percent on SVAMP and 11.7 percent on ASDIV. DeepSeek-R1-Distill-Qwen-1.5B gains 4.3 percent on GSM8K, 6.0 percent on MATH and 17.4 percent on OlympiadBench. Qwen2.5-Math-7B gains 10.4 percent on GSM8K, 6.0 percent on MATH, 9.3 percent on SVAMP and 7.9 percent on ASDIV. The instruct variant of Qwen2.5-Math-1.5B gains much less, 1.9 percent on GSM8K and 2.6 percent on Minerva Math, and the project also reports strong generalisation, having trained only on GSM8K and improved across 7 benchmarks. The testing is described as spanning 7 benchmarks and 5 models, and the zero-shot pass@1 metric is the one used.

The efficiency claim, which is the more quotable half

The efficiency figures are the ones the project puts first, and they are larger than the accuracy gains. Taking Qwen2.5-Math-1.5B as the example, total time drops 90 percent, from 4 hours to 0.5 hours. The number of sampled problems drops 80 percent, from 3,952 to 1,000. And the number of tuned tokens drops 99 percent, from 1.77 million to 20 thousand. The reason those three numbers move together is that the selection stage filters the sampling stage, so fewer problems are sampled and each contributes far fewer tokens to the training set. The project frames this as smarter, not blindly harder, and the interesting consequence is for anyone who has a compute budget rather than a data budget, because the bottleneck in many fine-tuning setups is the cost of collecting the rejection-sampled data at all.

The LRsamples dataset, for people who want to skip the sampler

Two of the releases are about the data rather than the code. LRsamples is a set of pre-collected LightReasoner training samples, released so the pipeline can be reproduced without running the full sampling pipeline, which the project says shortens reproduction efforts and accelerates downstream research. And the enhanced models are on the Hugging Face Hub as a collection, described as ready-to-use models fine-tuned with the approach for immediate deployment and experimentation. The baselines are the Qwen2.5-Math models. Both releases are dated 2025/10/14, and the core implementation with Qwen2.5-Math and DeepSeek-R1 came a couple of days earlier on 2025/10/12. The paper itself is on arXiv as 2510.07962, with a Hugging Face papers page, and the code is MIT licensed with Python 3.10 or newer required.

Where this will not help, and what to check first

The mechanism has a precondition, and it is the pair of models. The gains are large when the expert and amateur diverge in a useful way, and small when they do not, which is exactly what the instruct variant's 1.9 percent on GSM8K suggests. If your candidate expert model is uniformly stronger on the kind of problems you care about, with the same reasoning trace shape as the amateur, there may be no critical moments to find, and the method becomes an expensive way to train on less data. Two practical checks come before anything else. Run the analysis/ directory on your own model pair and look at where the divergence lands, because the whole method depends on that distribution. And be careful with the benchmarks, because the headline gains are on mathematical reasoning sets, and the reported generalisation is across 7 benchmarks within that family rather than across task types. The project also asks for contributions to be discussed on the Feishu and WeChat groups linked from the overview rather than on the issue tracker, which is a hint about how a paper release is being run.

Editorial conclusion

LightReasoner fits a research team that has a strong model and a weak one, and wants to spend its fine-tuning budget on the steps where the strong model is actually uncertain rather than on every token of a full trajectory, and it fits anyone who wants the pre-collected LRsamples dataset to reproduce the pipeline without running the sampler themselves. It does not fit a team that needs a supported training stack, because the repository publishes no GitHub releases and the last push was on 2026-05-22, and it does not fit anyone looking for a plug-in fine-tuning recipe, because the method depends on the pair of models behaving differently on the same problems, which is an assumption about the model pair rather than a property of your data. Before you run it, read the analysis/ and evaluation/ directories rather than the headline numbers, because the reported gains vary sharply by model, from plus 28.1 percent on GSM8K for Qwen2.5-Math-1.5B to plus 1.9 percent for its instruct variant.

Frequently asked questions

Official sources

1. HKUDS/LightReasoner on GitHub
2. Issues
3. License: MIT
4. Project website
5. README