Shiwen Ni

Shenzhen University of Advanced Technology · China

Unclaimed · AI-structured profileOpenAlex

LLM-agent and software-engineering researcher: benchmarking coding-agent memory and multimodal document understanding

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Output is unscored: this OpenAlex identity reports h-index 0 and i10-index 0 with no consolidated venue record, so total-output metrics cannot yet be computed. This axis would measure cumulative works, venues and per-year volume.

Impact—

Impact is unscored: h-index and i10-index both read 0 in the snapshot, so citation-based influence cannot be estimated. This axis would measure citations, h-index trajectory and uptake of the benchmarks in follow-up work.

Focus100/100

Scores 100/100, the strongest axis. Topic assignments concentrate in one lane: Software Engineering Research (8) and Software Testing and Debugging Techniques (4) dominate, with AI-based Problem Solving and Planning (3) and Evolutionary Algorithms and Applications (3) adjacent; the fieldDist entry names LLM agent memory evaluation. Both reviewed 2026 papers — coding-agent memory benchmarking and document-understanding training — sit squarely in this software-engineering/LLM-agent line.

Collaboration53/100

Scores 53/100, a mid-range value. The raw record shows 13 distinct coauthors across 13 recent works, with 12 recurring partners including Liyang Fan, Yingcheng Shi, Yongbin Li, Min Yang and Jieping Ye — a stable, group-embedded pattern. Notably, both reviewed papers place Shiwen Ni in non-corresponding positions (middle on VibeMemBench, last on the training study), so leadership signal is limited.

Activity77/100

Scores 77/100. countsByYear shows 1 indexed work in 2025 jumping to 12 in 2026, and recent3yWorks totals 13 — a steep, very recent ramp. Both reviewed papers are 2026 arXiv preprints, so the cadence is high but young, and peer-reviewed validation has not yet appeared in this snapshot.

Reproducibility55/100

Scores 55/100 (reproMean 55 from 1 assessed paper). Both reviewed works carry medium repro-readiness: VibeMemBench is an executable benchmark but reports confidence intervals crossing zero and flags selection bias; the conditioned-training study is called rigorous, honest yet small-scale. Empirical practice looks careful, though the reproducibility evidence base is still thin.

Research summary

Shiwen Ni is a researcher at Shenzhen University of Advanced Technology (CN) whose profile, name-resolved from OpenAlex bibliographic data, sits at the intersection of software engineering and large language model agents. The topic profile is highly concentrated: Software Engineering Research (8 assignments) and Software Testing and Debugging Techniques (4) lead, with AI-based Problem Solving and Planning (3), Evolutionary Algorithms and Applications (3) and Topic Modeling (2) rounding out the record. Output is recent and fast-moving — 1 indexed work in 2025 followed by 12 in 2026, 13 works across the recent three-year window — while the h-index and i10-index both read 0, indicating a citation record that has not yet accumulated. Two 2026 arXiv papers reviewed by Evolve define the current agenda. VibeMemBench is an executable benchmark evaluating persistent memory systems for coding agents on SWE-bench-style repository tasks; it finds direct experience injection helps only modestly (confidence intervals cross zero) and 11 of 12 memory-system pairings fail, with selection bias limiting scope. A second paper shows mixed joint training gives no cross-granularity reinforcement across three document-understanding corpora, while conditioned training succeeds on two, framed via learning using privileged information. Evolve rates both as rigorous and honest, with medium repro-readiness. Focus (100/100) and activity (77/100) are the strongest radar dimensions; collaboration scores 53/100 across 13 distinct coauthors.

Collaboration pattern

Co-authorship is team-embedded: 13 distinct coauthors appear across the 13 recent works, and 12 recur frequently — including Liyang Fan, Yingcheng Shi, Yongbin Li, Chenghao Sun, Xin Chen, Xander Xu, Hu Wei, Min Yang, Jieping Ye, Chengguang Gan, Yunhao Liang and Hanjun Wei. In both reviewed papers Shiwen Ni holds non-corresponding positions (middle and last), consistent with contributing within a group at Shenzhen University of Advanced Technology rather than solo leadership.

Research keywords

coding agentspersistent memorySWE-bench style tasksexperience reuseretrieval systemsbenchmark evaluationmultimodal document understandingconditioned trainingmutual reinforcement effectmulti-task learningnegative transfersoftware testing and debugging

Highlights

VibeMemBench: an executable memory benchmark for coding agents
Built on real repository coding tasks in the SWE-bench style, the benchmark evaluates persistent memory systems for coding agents. Direct experience injection yields only modest gains whose confidence intervals cross zero, and Evolve flags selection bias as a limit on scope.
11 of 12 memory-system pairings fail to help
Across the 12 memory-system configurations tested on repository tasks, 11 pairings fail to improve coding-agent performance — a sharply negative result that tempers enthusiasm for current agent-memory solutions while leaving direct experience reuse as the only modestly positive signal.
Conditioned training, not naive joint training, builds cross-granularity synergy
On three multimodal document-understanding corpora, mixed joint training shows no cross-granularity mutual reinforcement, while conditioned cross-granularity training — framed via learning using privileged information — succeeds on two. Evolve calls the study rigorous, honest and small-scale.

Timeline

2025First visible bibliometric activity: countsByYear records 1 indexed work in 2025.
2026Output jumps to 12 indexed works, the most productive year in the snapshot (13 works across the recent three-year window).
2026VibeMemBench released on arXiv: an executable benchmark evaluating persistent memory systems for coding agents on real repository tasks.
2026Conditioned cross-granularity training study released on arXiv, showing where joint training fails for multimodal document understanding.

Papers in the Evolve catalog

Frequent co-authors

Liyang Fan×1 Yingcheng Shi×1 Yongbin Li×1 Chenghao Sun×1 Xin Chen×1 Xander Xu×1 Hu Wei×1 Min Yang×1 Jieping Ye×1 Chengguang Gan×1 Yunhao Liang×1 Hanjun Wei×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

All bibliometrics come from a single OpenAlex identity; the h-index and i10-index of 0 likely reflect incomplete citation indexing for a very new profile rather than assessed merit.
No ORCID is provided; the profile is name-resolved from public bibliographic data, and identity consolidation with other researchers named Shiwen Ni is unverified.
Only two papers have been reviewed by Evolve, both 2026 arXiv preprints; no peer-reviewed venue information exists in this snapshot.
Output and impact radar axes are unscored because citation-based metrics are unavailable in the underlying data.