Sprint Paper Upload

_data/people.yml

@@ -14,6 +14,11 @@ jonsaadfalcon:
   url: https://jonsaadfalcon.com/
   title: PhD Student
 
+shayantalaei:
+  name: Shayan Talaei
+  url: https://www.linkedin.com/in/shayan-talaei-6b65a0229/
+  title: PhD Student
+
 # Visiting
 
 bradleybrown:

_pubs/sprint.md

@@ -0,0 +1,37 @@
+---
+title: 'SPRINT: Enabling Interleaved Planning and Parallelized Execution in Reasoning Models'
+authors:
+  - name: Emil Biju
+    equal: true
+    affiliation: Stanford University
+  - key: shayantalaei
+    equal: true
+    affiliation: Stanford University
+  - name: Zhemin Huang
+    equal: true
+    affiliation: Stanford University
+  - name: Mohammadreza Pourreza
+    affiliation: Google
+  - name: Amin Saberi
+    affiliation: Stanford University
+  - key: azaliamirhoseini
+    affiliation: Stanford University
+venue: preprint
+year: 2025
+date: 2025-06-06
+has_pdf: true
+doi: 10.48550/arXiv.2506.05745
+tags:
+  - machine learning
+  - artificial intelligence
+  - reasoning
+teaser: SPRINT enables LRMs to dynamically identify and exploit parallelization opportunities during reasoning, reducing sequential tokens by up to 39% while maintaining performance.
+materials:
+  - name: Paper
+    url: https://arxiv.org/abs/2506.05745
+    type: file-pdf
+  - name: Codebase
+    url: https://github.com/ShayanTalaei/SPRINT
+    type: code
+---
+Large reasoning models (LRMs) excel at complex reasoning tasks but typically generate lengthy sequential chains-of-thought, resulting in long inference times before arriving at the final answer. To address this challenge, we introduce SPRINT, a novel post-training and inference-time framework designed to enable LRMs to dynamically identify and exploit opportunities for parallelization during their reasoning process. SPRINT incorporates an innovative data curation pipeline that reorganizes natural language reasoning trajectories into structured rounds of long-horizon planning and parallel execution. By fine-tuning LRMs on a small amount of such curated data, the models learn to dynamically identify independent subtasks within extended reasoning processes and effectively execute them in parallel. Through extensive evaluations, we show that the models fine-tuned with the SPRINT framework match the performance of reasoning models on complex domains such as mathematics while generating up to ~39% fewer sequential tokens on problems requiring more than 8000 output tokens. Finally, we observe consistent results transferred to two out-of-distribution tasks of GPQA and Countdown with up to 45% and 65% reduction in average sequential tokens for longer reasoning trajectories, while achieving the performance of the fine-tuned reasoning model.