Last login: on ttys000
guest@qsun:~$ ./init_profile.sh
        (
       ) )
      ( (
    ._____.
    |     |]
    \     /       |\__/,|   (`\
     `---'      _.|o o  |_   ) )
   -------------(((---(((--------

Qumeng Sun

> M.Sc. CS @ Göttingen
> Thesis Researcher @ NEC Labs
CONTACT:qsun [at] qumengsun.com|GITHUB|TWITTER/X|LINKEDIN
# ~/home/BIO.txt
I am finishing my Master's thesis at NEC Laboratories Europe. My work is simple: I try to give machines a memory, so they don't forget what you said five minutes ago.
Before this, I worked in industry. I built ML pipelines (from SQL fetching to visualization), and set up clusters so models could be trained without crashing. I care about System Stability and Practical AI.
I share ideas on X and my blog. If something resonates, feel free to reach out - I'd love to discuss.

>>./RESEARCH// What I study
>>./SYSTEMS// What I built
>>./WRITINGS// What I think

Research Engineer. I build systems that remember and pipelines that don't break.

guest@qsun:~$ tail -n 2 updates.txt

> [2025-12] Personal website launched.
> [2024-12] [PLACEHOLDER] Update details pending.
# ~/research/STATEMENT.md
LLMs have no memory. They talk to you, and then they reset. It's like talking to someone who wakes up every morning with a blank slate. I don't think that works for real applications. Models are forgetful and Data are chaotic. My job is to fix that.
Research Interests (RI) & Questions (RQ)
[RI-1] MEMORY & CONTINUAL LEARNING
Goal: Enable AI Agents to retain long-term context and learn continuously without catastrophic forgetting.
└─ RQ 1.1
Is mimicking human brain memory mechanisms effective for enhancing long-term context in LLMs?
└─ RQ 1.2
Is non-parametric memory more efficient than parametric updates for achieving continual learning in frozen LLMs?
[RI-2] ALIGNMENT & EVALUATION
Goal: Guide model behavior (Alignment) and rigorously measure performance beyond static benchmarks.
└─ RQ 2.1
How do we design the ultimate benchmark that goes beyond static datasets to evaluate true generalizability?
└─ RQ 2.2
What are the effective metrics for evaluating "plasticity" and adherence to instructions in RLHF?
[RI-3] AGENTIC SYSTEMS & INFRASTRUCTURE
Goal: Build robust Multi-Agent Systems for task automation and Software Engineering (AI4SE).
└─ RQ 3.1
How do we orchestrate multi-agent collaboration to solve complex software engineering tasks?
└─ RQ 3.2
What is the trade-off between agent autonomy and human-defined constraints in real-world automation?
[RI-4] INTERPRETABILITY OF ATTENTION
Goal: Understand why attention-based Transformers dominate, and leverage that insight to design better alternatives.
└─ RQ 4.1
What theoretical framework can unify existing hypotheses about how attention mechanisms work?
└─ RQ 4.2
Can a deeper understanding of attention lead to more efficient or generalizable architectures beyond Transformers?

    # SELECTED WORKS
    • [2026] Long-Term Memory Systems for AI AgentsM.Sc. Thesis
      Q. Sun, NEC Laboratories Europe
      Hierarchical L0-L2 memory architecture combining semantic stores and experience graphs.
      [WIP]
    • [2025] Graph-ReAct: Agentic RAG with Unified StateProject Report
      Q. Sun, Uni Göttingen
      Dynamic DAG execution for RAG, achieving 0.754 FactF1 on HotpotQA.
      [CODE]
    # ~/systems/ENGINEERING
    guest@qsun:~/home$