Xun Xu - Resume

Publications

Who Teaches Which Token? Verifier-Gated Multi-Expert On-Policy Distillation first author

Submitted to ICLR 2027

Proposed VG-OPD, a verifier-gated multi-expert on-policy distillation framework that selects who should teach which token based on criterion-level expert gains.
Introduced token-level supervision that localized expert guidance through teacher–student disagreement and weighted it by criterion importance within GRPO.
Achieved the best overall performance across seven scientific reasoning benchmarks for both 4B and 8B models, ranking first on five benchmarks at both scales; ablations showed that localizing verified supervision was critical to the gains.

OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation co-first author

Accepted to EMNLP 2026

Introduced OPD-Evolver, a slow-fast co-evolution framework that trains agents to evolve through experience rather than merely storing or retrieving it.
Built a four-level memory hierarchy for fast test-time evolution, with outcome-calibrated attribution and privileged-hindsight distillation to learn experience selection, execution, writing, and repository maintenance.
Evaluated on multi-domain benchmarks, surpassing ReasoningBank and Complementary RL under the reported setting; OPD-Evolver-9B further shows competitive performance against much larger counterparts such as Qwen3.5-397B-A17B and Step-3.5-Flash.

Experience

Research Assistant @ HKUST Jul. 2026 - Sept. 2026

Advised by Prof. Zaixi Zhang · Post-Training, RL & On Policy Distillation

Developed a verifier-gated multi-expert on-policy distillation framework for LLM scientific reasoning, enabling criterion-level expert routing and token-level supervision.
Conducted experiments to evaluate the effectiveness of the proposed framework in improving the reasoning capabilities of LLMs.

Research Intern @ NUS LV Lab Mar. 2026 - Jul. 2026

Advised by Prof. Shuicheng Yan · Self-Evolving Agents, Agent Memory & Agentic RL

Studied self-evolving agents, agent memory, and agentic RL, focusing on experience selection, reusable knowledge updates, and tool-mediated task improvement.
Built baselines and distilled agent-RL / memory-system literature into two coupled axes: action-policy learning and memory-operation design.
Designed reproducible evaluation loops linking policy behavior, memory updates, and downstream task success for controlled ablations over learning signals and memory operations.

Research Intern @ FDU FVL Lab JUN. 2025 - SEPT. 2025

Advised by Prof. Henghui Ding · Latent Memory for Long-Context LLM Reasoning

Proposed G-MemLLM, a gated latent-memory augmentation framework coupling a frozen LLM backbone with trainable persistent memory slots through GRU-style memory updates for long-context reasoning.
Designed memory encoding, cross-attention retrieval, and latent-state consolidation pipelines, and conducted controlled ablations over memory retention, overwrite dynamics, and multi-hop reasoning consistency.

SWE Intern @ Xiaohongshu (RedNote) Nov. 2025 - Feb. 2026

Designed LLM AIOps agent workflows for cluster diagnosis, applying agentic task decomposition, tool selection, observation parsing, and evidence-grounded reasoning to production infrastructure scenarios.
Refined Function Calling schemas, diagnosis prompts, API orchestration, and fallback policies to make the agent’s tool use more reliable and its cluster-diagnosis process easier to inspect.