This repository hosts the code, W&B logs, and finetuned models for our paper. The proposed algorithm, PALU (Performance-Aware Length Update), frames concise reasoning as a performance-constrained length reduction problem, optimizing for the shortest reasoning path that still meets a target accuracy.
-
65% shorter generations while achieving +15% accuracy improvements across five benchmark reasoning tasks (AIME24, Math 500, AMC 23, OlympiadBench and Minerva-Math).
-
Cross-domain robustness across math, logic, and STEM tasks.
-
Model-scale generalization: consistent performance from 1.5B, 7B, to 14B parameter models.
-
Built on Lagrangian optimization principles, balancing reasoning efficiency and accuracy.
-
Evaluated on the DeepSeek-R1-Distill-Qwen model family.
| Resource | Description | Availability |
|---|---|---|
| 🧠 W&B Logs | Full experiment tracking and visualization | Coming soon |
| 🤗 Hugging Face Checkpoints | Fine-tuned PALU models | Huggingface page |
| 📄 Research Paper | Detailed methodology, experiments, and analysis | Arixv |
For a concise visual summary of PALU and its core contributions, please refer to our project poster below:
