2026.09 — Present
Master's Student
University of Science and Technology of China · School of Cyber Science
Research interests: LLM agents, agent memory, and long-context language models.
01About
I'm Ruizhe Li (李睿哲), a master's student at USTC. My research focuses on LLM agents, agent memory, and long-context language models.
LLM Agents
Agent Memory
Long-Context LLMs
02News
- 2026.08Our agent memory system ReFind ranked #2 globally on the Academic Textual track of the 2026 Agent Memory Challenge (AML).
- 2026.08Our paper When Your Agent Opens the Chat App introduces ReFind, agent-controlled search over raw chat logs without semantic index construction.
- 2026.07Our paper Keep It InMind introduces a benchmark for the implicit-association blind spot in agent memory.
- 2026.01Our benchmark DeepResearch Bench II is out on arXiv; the public leaderboard is live at agentresearchlab.com.
- 2025.09Started research on agent memory and long-horizon retrieval for LLM agents.
- 2025.08Our paper Automated Creativity Evaluation for LLMs is accepted to Findings of EMNLP 2025.
03Education
2022.09 — 2026.06
B.S. in Data Science
University of Science and Technology of China · School of Artificial Intelligence and Data Science
04Publications
When Your Agent Opens the Chat App: Agent-Controlled Search over Raw Chat Logs Rivals Structured Memory
arXiv:2608.12888 · 2026
ReFind gives an agent iterative, controllable search over unmodified chat logs, combining lexical retrieval with session-aware ranking, local context, and temporal narrowing. It reaches the highest mean accuracy among compared systems on MemoryAgentBench without constructing summaries, trees, or knowledge graphs.
Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory
arXiv:2607.24368 · 2026
InMind is a 125-task benchmark showing that a memory can be crucial even when it is not textually similar to the current query. Its paired controls isolate whether failures come from storage, missing world knowledge, or retrieval and routing.
DeepResearch Bench II: Diagnosing Deep Research Agents via Rubrics from Expert Reports
arXiv:2601.08536 · 2026
A benchmark of 132 expert-grounded tasks across 22 domains and 9,430 fully-verifiable binary rubrics, scoring deep-research reports on information recall, analysis, and presentation. Built with a four-stage LLM+human pipeline and 400+ expert-hours — even the strongest systems satisfy fewer than 50% of rubrics.
Automated Creativity Evaluation for Large Language Models: A Reference-Based Approach
Findings of EMNLP 2025 · arXiv:2504.15784
A reference-based method that scores the creativity of large-language-model outputs against human references, offering a more reliable and automatic alternative to costly human creativity ratings.
Reformulating Conversational Recommender Systems as Tri-Phase Offline Policy Learning
CIKM 2024 · arXiv:2408.06809
Recasts conversational recommendation as a three-phase offline policy-learning problem, decoupling interaction-policy learning from costly online exploration.
05Projects
ReFind★ 9
Agent-controlled search over raw chat logs for long-term memory. ReFind combines iterative lexical retrieval with chat-native controls and rivals structured memory systems without building summaries, trees, or knowledge graphs.
Python · Agent MemoryRepo ↗
InMind★ 17
A 125-task benchmark for the implicit-association blind spot in long-term agent memory, with paired controls that isolate storage, world-knowledge, retrieval, and application failures.
Benchmark · DatasetRepo ↗
DeepResearch-Bench-II★ 83
Evaluation suite and leaderboard for deep-research agents, grounded in expert-written reports. Rubric extraction, batched LLM-as-judge scoring, and token accounting.
PythonRepo ↗
06Experience
2025.02 — 2025.08
LLM Algorithm Engineer Intern
Metastone Technology
Designed the memory system for a role-play product.
07Service · Skills
- President, Student Union School of AI & Data Science, USTC
- Teaching Assistant, DS4001 Data Science course · 2025 Spring
- Member, Tang Zhongying Charity Society Volunteering
PythonPyTorchLLM / AgentsRAG
LaTeXGitHTML / CSS / JSC