lux-liang.top

Undergraduate researcher · Xinjiang University

Jialiang Liang 梁嘉亮

Finding useful context
for coding agents.

I study how coding agents explore repositories and select useful context. My current interests extend to user intent and memory in long-running agents.

B.Eng. in Information Security · Expected graduation: June 2028

Last updated in September 2026

02 / RESEARCH

Research

How can coding agents find, retain, and reuse the information that matters? My work so far focuses on repository exploration, evaluation, and context efficiency.

Coding AgentsAgent EvaluationContext Engineering

Looking ahead

How can an agent recover user intent from an evolving conversation?Which experiences should become useful long-term memory?

Selected papers & preprints

2026

Co-first author · NeurIPS 2026 · under review

SWE-Explore

Benchmarking How Coding Agents Explore Repositories

Evaluating whether coding agents find the repository context they need, using line-level context selection and downstream issue-repair validation.

SWE-Explore benchmark construction and evaluation framework
Benchmark framework · Figure 2

My contributionCo-led benchmark construction, evaluation protocol design, large-scale baseline experiments, and manuscript preparation.

848 instances · 203 repositories · 10 languages

Comparison of end-to-end issue resolution and direct evaluation of repository exploration
Repository exploration · Figure 1
2026

Co-author · Submitted to ACL Rolling Review · Aug. 2026

SWE-Pruner Pro

The Coder LLM Already Knows What to Prune

Selecting useful context from relevance signals in a coding model’s hidden representations, without a separate pruning LLM.

SWE-Pruner Pro agent trajectory and pruning-head framework
Method overview · Figure 3

My contributionContributed to experimental evaluation, result analysis, and manuscript preparation.

39.4% fewer total tokens on SWE-QA-Pro

Qwen3-Coder-Next vs. no pruning: 607K → 368K total tokens; mean judge score 7.60 → 7.84.

Comparison of a separate goal-conditioned pruner and pruning integrated into the coding agent
In-agent context pruning · Figure 1

03 / EXPERIENCE

Education

  1. 2024.09 — 2028.06 (Expected)

    Xinjiang University

    B.Eng. in Information Security

    GPA 4.16 / 5.00 · Rank 5 / 78

03 / EXPERIENCE

Research Experience

  1. Aug. 2026 — Present

    AI for Multimodality and Science Lab (AIMS)

    Research student · advised by Prof. Jun Xia (夏俊)

    HKUST(GZ) · Lab page ↗
  2. Dec. 2025 — Present

    Visual Perception and Security Group (VPSG)

    Research student · advised by Prof. Zhiqing Guo (郭治卿)

    Xinjiang University · Lab page ↗
  3. Jan. 2026 — Jun. 2026

    LLM for Software Engineering Lab (LLMSE)

    Research student · mentored by Prof. Xiaodong Gu (顾小东)

    Shanghai Jiao Tong University · Lab page ↗

04 / AWARDS

Awards & Honors

National competition results and selected academic distinctions.

Competitions

05 / LIFE

Beyond the lab.

Careful observation, honest experiments, and a willingness to be wrong shape how I work. Outside research, I enjoy mountain trails, photography, live music, and quieter moments.

A winter mountain ascent
Above the snowline
A pine tree covered in fresh snow
First snow
Sunset filtering through trees
Last light
A golden desert poplar beneath a clear blue sky
Desert gold
Rocks and white walls in a classical garden
Quiet geometry
A bright flag above Sayram Lake and snowy mountains
Lake wind
Shanghai skyline at night
City lights
A live band playing in a warm music venue
Live notes
A monument against a snowy mountain sky
At the summit
A snowy mountain behind a quiet town
Cloud line
A campus bonfire glowing at night
A little wild
A cat resting in warm sunlight
A slower afternoon
A cat sitting quietly on a bench at night
Night watcher
A winter mountain ascent
Above the snowline
Sunset filtering through trees
Last light