Jiale Zhao

jialeuuz@gmail.com jialeuuz CV

Hi! I’m Jiale Zhao, a computer science graduate (B.Eng., 2025) from Chongqing University of Posts and Telecommunications. I currently work on rubric‑based RLVR for large language models during my internship at Li Auto.

I plan to begin a PhD in Fall 2026. My interests center on large language models and applied NLP, including:

• Human-centered human–AI interaction (HCI)
• Agent-based LLMs and multi-step reasoning
• Rubric-based RLVR
• Self-evolving systems
• Interpretability and controllability
• End-to-end multimodal interactive systems (e.g., GPT-4o)

Background

• Fall 2026 (planned): PhD studies (applications in progress)
• Sep 2021 – Jun 2025: B.Eng., Computer Science, Chongqing University of Posts and Telecommunications

Publications

• When and What to Ask: AskBench and Rubric-Guided RLVR for LLM Clarification — first author. Under review (ACL). arXiv: https://arxiv.org/abs/2602.11199.
• RubricHub: A Comprehensive and Highly Discriminative Rubric Dataset via Automated Coarse-to-Fine Generation — second author. Under review (ACL). arXiv: https://arxiv.org/abs/2601.08430.
• ThinkPilot: Steering Reasoning Models via Automated Think-prefixes Optimization — co-first author. Findings of EACL 2026. arXiv: https://arxiv.org/abs/2510.12063.

Ongoing Work

• ProcessRubrics — first-author work at Li Auto on process-level rubric learning for improving structured reasoning quality.
• Learning Persona as Behavior — second-author collaboration with Prof. Lu Cheng (UIC) on behavior-oriented persona learning with stronger cross-domain stability.

Selected Projects

All three were production business deliverables I shipped during my Li Auto internship.

• Data Flywheel for Code LLM — evaluation-centric loop (SFT → evaluate → data build → filtering → back to SFT) to continually raise coding capabilities.
  • Evaluation-first: Code-eval today is noisy—difficulty too low, specs ambiguous—so I standardized harnesses and rubrics to push harder tasks and capture real capability.
  • Linked loops: Evaluation feedback drives harder data construction; the same tooling filters low-quality samples; filtered data re-enters the generation stack for repair and resurfacing.
• Multi-step Reasoning + Tool Invocation Agent — code-LLM agent that plans, writes code, and executes tool calls for precise answers.
  • Multi-step reasoning: Breaks complex or code-debugging tasks into structured plans so context can be stitched into a single executable query.
  • Tool grounding: Integrates function calls/code execution for real-time data, external APIs, and environment actions when model priors or knowledge bases fall short.
• MindGPTo (GPT‑4o-style multimodal app) — end-to-end audio + vision application with paralinguistic control, built from scratch with a modular FE/BE split.
  • Mode coverage: Ships traditional audio→ASR→LLM→TTS, production audio2text→TTS pipelines, end-to-end audio2audio, and multimodal audio+image+video→text→TTS workflows.
  • Paralinguistic SFT: Large-scale audio data pipelines boost colloquial speech and nuanced cues (beyond laughter/pauses) such as age, gender, compound emotions, emotional actions, and ambient sounds.

Manuscripts for Resubmission

• Decoding the Ear: A Framework for Objectifying Expressiveness from Human Preference Through Efficient Alignment — third author; awaiting resubmission.
• Breaking the Exploration Bottleneck: Rubric-Scaffolded Reinforcement Learning for General LLM Reasoning — sixth author; awaiting resubmission.

Research Experience

• ⚡ Before college: I fell for AI and taught myself to code the summer before matriculation, so I arrived on campus already building small projects and craving research.
• 🎯 Freshman (first semester): Two weeks in, I earned a spot in an NLP lab after a selection process. I worked on traditional ML, but my advisor was rarely available, so I mostly pushed myself and learned by doing.
• 🚀 Freshman summer (2nd lab selection): I spent the summer in another competitive selection and joined a CV lab with weekly group meetings and structured training. That was my first real mentorship track and it gave me a rigorous deep‑learning foundation.
• 🔧 Sophomore fall: After reading transformer papers I was desperate to dive in, but compute limits meant I simply couldn’t run the transformer experiments I imagined. My advisor even sighed and told me, “Let’s drop the transformer idea—there’s no compute for you, just do classic CV.” I really didn’t want to pivot back, so for a while my “research” was just reading more papers and thinking through designs on paper, getting more obsessed but still unable to touch real experiments.
• 🤖 Sophomore spring: When ChatGPT‑3.5 launched, I used it daily to speed up coding and research. I switched labs to build “digital humans,” shipped a 24/7 virtual‑avatar livestream on Bilibili, and became one of the earliest to deploy that idea.
• 🧭 Junior → now: That digital‑human project led to a Li Auto algorithm/LLM internship. Junior year, with ample compute, I delivered an SFT data flywheel, a multi‑step tool‑use agent, and MindGPTo (a GPT‑4o‑style multimodal app). Senior year and onward, I moved into research mode, joined multiple papers, and focused on rubric‑driven RL/RLVR, agentic reasoning, and data‑centric LLM improvement.