Xinyu Guo

Xinyu Guo

Ph.D. student @UArizona, B.E. @SDU

I’m Xinyu Guo, a third-year Ph.D. student in Computer Science at the University of Arizona. I am fortunate to be advised by Prof. Mihai Surdeanu in the CLU Lab. Before that, I received my B.E. in Computer Science from Shandong University in 2023.

Large language models increasingly reason through knowledge-intensive and open-ended tasks. My research focuses on LLM-based reasoning in these settings, along two directions:

  • RL credit assignment for reasoning supervision: credit assignment is key to RL for reasoning, which needs fine-grained credit beyond RLVR’s outcome-only rewards. I tackle it from two angles: (1) efficient intermediate rewards, self-derived from the model’s own correct reasoning without annotation or an LLM judge (CogRE); and (2) from rewards to advantages, disentangling the varying query-specific rubrics in critic learning to estimate per-rubric values and advantages (DVE).
  • Reasoning mechanisms: looking inside LLMs to understand how concepts and their logical structure support reasoning, by aligning sparse autoencoder features with concepts (AlignSAE) and taxonomies (TaxoSAE).
🔥 I am actively looking for Summer 2027 research internships! My interests lie in reinforcement learning for LLM reasoning, though I am open to related directions. Please feel free to reach out if you think I'd be a good fit: xinyuguo@arizona.edu. CV

🔥 News

  • 2026.10: “TaxoSAE: Rethinking Sparse Autoencoders through Taxonomic Logic” is released on arXiv.
  • 2026.08:  🎉 “The Answer Lies Within: Self-Derived Rewards Enable Explainable Relation Extraction” (CogRE) is accepted to the EMNLP 2026 Main Conference!
  • 2026.08:  🎉 “AlignSAE: Concept-Aligned Sparse Autoencoders” is accepted to TMLR!
  • 2026.08: “Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL” is released on arXiv.
  • 2026.06:  🏆 I received the Galileo Circle Scholarship from the University of Arizona!
  • 2026.05: “Dr. DocBench: A Comprehensive Benchmark for Expert-Level and Difficult Document Parsing” is released on arXiv, and accepted to the LCFM Workshop @ NeurIPS 2026! [OpenReview]
  • 2025.11: “Deep Research: A Systematic Survey” is released on arXiv.
  • 2025.09:  🏆 I received the Science Graduate College Fellowship from the University of Arizona!
  • 2025.07:  🎉 “QUEST: Query Optimization in Unstructured Document Analysis” is accepted to VLDB 2025!

📝 Publications

RL Credit Assignment for Reasoning Supervision

EMNLP 2026 Main
CogRE

The Answer Lies Within: Self-Derived Rewards Enable Explainable Relation Extraction

Xinyu Guo, Zhengliang Shi, Minglai Yang, Mahdi Rahimi, Mihai Surdeanu

PDF | Code | arXiv

  • Identified two failure modes of LLMs on realistic one-shot relation extraction: attending to irrelevant tokens, and predicting relations at a granularity mismatched with human annotation schemas.
  • Proposed Hit@Dict, a self-derived intermediate RL reward that credits relation-bearing phrases in the reasoning trace, using a dictionary mined from the model’s own correct explanations; no LLM judge needed.
  • Improved F1 by +73.7% (relative) over accuracy-only RL on one-shot NYT29 and human-rated explanation quality by +54%.
Under Review
DVE

Disentangled Value Estimation: Rethinking Scalar Critics in Rubric-Based RL

Xinyu Guo, Minglai Yang, Chen Xu, Lingyong Yan

PDF

  • Identified a bottleneck in the critic model for rubric-based RL: the critic sees only the query and response prefix, yet is trained on a scalar reward that collapses many query-specific rubrics.
  • Proposed a shared critic conditioned on one rubric at a time that estimates per-rubric values and token-level advantages, aggregated with the rubric weights.
  • Outperformed PPO and GRPO in 13 of 15 model–benchmark settings (up to +8.0 points over PPO), with lower value error and advantages concentrated on rubric-relevant tokens.
Under Review
Rubric Dropout

Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL

Minglai Yang, Xinyu Guo, Utkarsh Tyagi, Mian Zhang, Razvan Dumitru, Sunjie Hou, Yunzhong He, Daniel Yue Zhang, Ying Liu

PDF | arXiv

  • Showed that rubric-as-reward RL reward-hacks out of distribution: the training judge’s score keeps rising while a stronger gold judge’s score peaks and falls by 22 points on ResearchQA.
  • Proposed Rubric Dropout, which randomly drops rubric criteria at every step, raising the out-of-distribution gold score by +6–7 points on ResearchQA with no in-domain cost.

Reasoning Mechanisms

Under Review
TaxoSAE

TaxoSAE: Rethinking Sparse Autoencoders through Taxonomic Logic

Xinyu Guo, Minglai Yang, Clayton T. Morrison, Eduardo Blanco, Mihai Surdeanu

PDF | Code | arXiv

  • Showed that existing SAEs learn a flat feature space, so concept readout lacks taxonomic consistency and control lacks the granularity to scope an edit.
  • Proposed TaxoSAE, which mirrors a taxonomy with stacked SAEs and a taxonomic alignment loss; only terminal concepts need supervision.
  • Improved hypernym readout of unseen abstract concepts by +31.3 points and swap success by +25.5 points; suppressing a concept removes only its own answers.
TMLR 2026
AlignSAE

AlignSAE: Concept-Aligned Sparse Autoencoders

Minglai Yang, Xinyu Guo, Zhengliang Shi, Jinhe Bi, Steven Bethard, Mihai Surdeanu, Liangming Pan

PDF | Code | OpenReview

  • Showed that unsupervised SAEs fragment human-defined concepts across many entangled features.
  • Proposed a “pre-train, then post-train” curriculum that binds each ontology concept to a dedicated latent slot.
  • Enabled reliable concept swaps (94.8%) and 9–∞× better disentanglement than standard SAEs; used the aligned slots to probe multi-hop reasoning and grokking-like dynamics.

Other Publications

📖 Education

  • 2024 - now: Ph.D. in Computer Science, University of Arizona, Tucson, AZ. Advisor: Prof. Mihai Surdeanu. GPA: 3.86/4.0.
  • 2019 - 2023: B.E. in Computer Science, Shandong University, China. Graduated with Honors. GPA: 3.8/4.0.

💻 Internships

  • 2026.05 - 2026.10, Research Scientist Intern, Abaka AI, Mountain View, CA.
    • Agentic RL Training: built an internal agentic RL training framework supporting MCP-based tool use and RL training; post-trained Qwen3-8B on 2K synthesized tasks, raising Pass@1 on τ²-bench Retail from 31.6% to 60.2% (RetailServe blog).
    • Document parsing: contributed to Dr. DocBench (NeurIPS 2026 LCFM Workshop), and helped organize the Dr. DocBench challenge at the DocInsights Workshop @ EMNLP 2026.

🔬 Research Experience

  • 2025.03 - now, Graduate Research Assistant, CLU Lab, University of Arizona, Tucson, AZ.
  • 2023.09 - 2024.09, Undergraduate Research Assistant, CAES Research Center, Shandong University, China. Advisor: Prof. Zhongxing Yu.
    • C++ type conversion bugs: built a dataset of 226 real-world type conversion bugs from 10 major C++ projects (e.g., V8, PyTorch, OpenCV, MySQL), summarized root-cause and fix patterns, and designed a tool to automatically detect and repair them.
    • Code clone detection & repair (undergraduate thesis): applied SemanticCodeBERT, which extends GraphCodeBERT with semantic flow graphs and five structure-aware pre-training tasks; outperformed GraphCodeBERT on BigCloneBench clone detection and Transformer baselines on Bug-Fixes code repair.
  • 2022.02, Oxford Winter Exchange Program, Oxford University, UK. Advisor: Prof. Jeremy Pitt. Topic: Human-centered AI.

🎖 Honors and Awards

  • 2026: Galileo Circle Scholarship, University of Arizona
  • 2024, 2025: Science Graduate College Fellowship, University of Arizona
  • 2023: Outstanding Undergraduate Thesis Award, Shandong University
  • 2023: Outstanding Graduate Student, Shandong University
  • 2022: Academic Scholarship, Shandong University
  • 2021: 18th Certified Software Professional Examination, Top 4.57%, CCF

🤝 Grants and Services

  • Co-PI, ABC SEED Grant, Arizona Astrobiology Center
  • Reviewer, ACL Rolling Review (ARR) 2026, NeurIPS Workshop (LCFM) 2026

📚 Teaching

  • 2024, 2025: Teaching Assistant, University of Arizona. Courses: CSC337 Web Programming, CSC460 Database, CSC477/577 Computer Vision.
  • 2023: Teaching Assistant, Shandong University. Course: Software Engineering.

profile views