$ ls research/ projects/

# research & projects

Publications, manuscripts, and things I've built.

── publications · 03

[1]

“A Toulmin-Based Discourse Annotation Method for Identifying Sycophancy in LLM Argument Feedback”

Ding, H., Uhde, A., Gohourou, D., & Cooper, E. W. (2026). Human Interface Symposium 2026, Okayama, Japan · Poster 3T-P08.

poster

An annotation method that identifies sycophancy by locating each argument's defect (Toulmin model) and checking whether the LLM's feedback addresses it. Across 5,920 responses (296 arguments × 5 stance cues × 4 LLMs), high sycophancy occurred in 1.8% of responses — almost entirely under pro-user cues — with a tenfold gap between models.

[2]

“WebMCP Tool-Use: A 14B Open-Weight Model Beats a Frontier API, and 32B Regresses”

Ding, H. (2026). Manuscript.

unpublished

Evaluation of 15 models (1.7B–284B) on 248 intents across four synthetic web apps. See the project write-up below.

[3]

“Using Random Forest for Future Sea Level Prediction”

Ding, H. (2023). SHS Web of Conferences, Vol. 174, art. 03008.

peer-reviewed · published · doi:10.1051/shsconf/202317403008

abstract · methods · bibtex

abstract

A machine-learning approach to forecasting global sea-level rise using Random Forest regression. Historical tide-gauge and satellite-altimetry data were used to train a predictive model; hyperparameter tuning was applied to improve accuracy. The results suggest Random Forest can serve as a viable, interpretable alternative to classical statistical sea-level models for medium-horizon forecasting.

methods

› Tide-gauge & altimetry preprocessing

› Random Forest regression (scikit-learn)

› Grid-search hyperparameter tuning

› Cross-validated accuracy evaluation

#random-forest#sea-level-rise#hyperparameter-tuning#climate-ml#regression

bibtex

@article{ding2023sealevel,
  title   = {Using Random Forest for Future Sea Level Prediction},
  author  = {Ding, Haolun},
  journal = {SHS Web of Conferences},
  volume  = {174},
  pages   = {03008},
  year    = {2023},
  doi     = {10.1051/shsconf/202317403008}
}

── projects · 03

WebMCP Tool-Use Evaluation2026-05

Independent research · manuscript

›Built a harness with schema validation and LLM-based semantic checks to test 15 models (1.7B to 284B parameters) on 248 intents across four synthetic web apps, serving local models via SGLang on two RTX A6000 GPUs.

›Open-weight Qwen3-14B beat the frontier API DeepSeek V4 Flash (0.835 vs. 0.822); Qwen3-32B called tools more often when it should refuse; three of five sub-metrics were near-saturated, masking real differences.

#python#sglang

Shogi assistant for Rokid glasses2026-09

First place · Rokid Glasses Hackathon, Osaka

›Built a shogi (Japanese chess) assistant for Rokid smart glasses.

›Now developing AI agents for the platform together with Rokid.

Spark Ontology2026-08 —

Internal AI agent platform · HI+ Camp, Hangzhou

›Sole developer of HI+ Camp's internal AI agent platform.

›Used company-wide to track 130+ startups from sourcing to funding, auto-summarize weekly reports, and turn meeting minutes into tracked to-dos.