Structured data
Organize teacher-generated reasoning into cognitive steps using a shared vocabulary of 23 tags, enabling consistent step extraction.
ICLR 2026 · Zhejiang University
Structured
Reasoning for LLMs
A unified framework for efficiency and explainability.
Make reasoning steps explicit. Trace their dependencies.
Optimize the paths that lead to an answer.
College of Computer Science and Technology, Zhejiang University
Reasoning, one step at a time
<rephrase>understand
↓
<formalize>represent
↓
<inference>derive
↓
<verify>check
↓
<summarize>answer
An illustrative step sequence. Dependencies are analyzed separately using model attention.
The framework
From free-form text
to explicit reasoning steps.
Structured Reasoning combines step-annotated training data, layer-wise dependency analysis, and structure-aware reinforcement learning. It studies how a model can reach answers with more efficient, stable, and interpretable reasoning.
Step dependencies
Aggregate model attention into step-to-step relationships and inspect how information flows across layers.
MaxFlow & LCS
Use complementary rewards to assess step contributions and align reasoning with reliable reference paths.
Explore reasoning dependencies
Open the existing research demonstration to inspect step attention and compare MaxFlow, TopK, and TopP pruning on example reasoning graphs.
Publication
Read the research.
ICLR 2026 · Published
Official proceedings ↗Structured Reasoning for LLMs: A Unified Framework for Efficiency and Explainability
Yubo Dong, Hehe Fan, Linchao Zhu, and Yi Yang
The original preprint was titled “Enhancing Large Language Models through Structured Reasoning.” The conference version uses the title and author list shown above.
Public release
Structured Reasoning dataset.
One complete training split with problem statements, tagged reasoning, answer targets, and ordered step annotations. Available in Parquet and JSONL.
516Examples
23Step types
MITRelease license
What does each example contain?
problem_id — stable identifierproblem — questionreasoning — paired step tags and reasoninganswer — answer targetcontent — final responsesteps — step ID, type, and textLoad with Hugging Face Datasets
from datasets import load_dataset
dataset = load_dataset(
"FreeFrank/Structured-Reasoning",
split="train",
)
example = dataset[0]
print(example["reasoning"])For reasoning-supervised training, retain both the reasoning and final answer in the assistant target. Check context lengths with your model’s tokenizer; a 32k context accommodates the checked serialization.
This is a curated release with editorial curation and step annotations. It is not claimed to be the exact experimental corpus from the paper. The dataset card documents provenance and limitations; no independent corpus-wide accuracy estimate is reported. Attention weights and dependency graph edges are not included in this text dataset.
Open-source plan
What is available.
What comes next.
The paper, dataset, and research demo are available now. The remaining releases are planned in stages; dates will be announced in the main repository.
01Available
Paper & project documentation
Conference paper, project overview, citation, and resource links.
02Available
Dataset & interactive demonstration
516 annotated examples on Hugging Face and the existing step-dependency visualization.
03Planned
Data tools & training implementation
Planned: annotation and validation utilities, structured SFT, and MaxFlow/LCS reward implementations.
04Planned
Reproduction & evaluation
Planned: environment setup, experiment configurations, evaluation scripts, and instructions for reproducing reported comparisons.
05Planned
Model checkpoints
Planned: model weights, model cards, and inference examples. Release timing and licensing are to be announced.
The dataset’s MIT license applies to the dataset release. Licenses for forthcoming code and checkpoints will be specified when those artifacts are published. This roadmap describes intended scope rather than a committed release date.
Citation
Build on this work.
BibTeX · conference version
@inproceedings{dong2026structuredreasoning,
title = {Structured Reasoning for LLMs: A Unified Framework
for Efficiency and Explainability},
author = {Dong, Yubo and Fan, Hehe and Zhu, Linchao and Yang, Yi},
booktitle = {International Conference on Learning Representations},
year = {2026},
url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/ad5b3f324b24c17cdc2f3712298c76bd-Abstract-Conference.html}
}