Research

Research and projects

Papers

Joint first author, submitted to AAAI and under review (2026)

Memory Response Paths and Path-Structured Context Compilation

With Jerrold Huang, I study how individual memories change LLM-agent decisions. I proposed Memory Response Paths to estimate a record's incremental contribution across four horizons, then designed paired-availability tests across 180 environments, model API workflows, statistical analysis, and fixed-evidence context ablations. Equal final responses concealed earlier path differences in 56.1% of environments. In a separate 144-item MemGym-DR evaluation, the paper's PSCC method improved LLM-judged quality by 6.9 percentage points using 32.8% fewer API tokens than the benchmark's A-Mem baseline. A-Mem is a prior SOTA conversational-memory framework, reporting results on LoCoMo and DialSim in its original study.

The manuscript and code remain private.

Data and further reading

A note on the evaluations
MRPBench and the separate PSCC evaluation are reported as different experiments.

Joint first author, accepted at ICAIF 2026 and selected for oral presentation

Sequential Tone Residualization and Abnormal Returns in Earnings Calls

With Jerrold Huang, I evaluate earnings-call language as a candidate input to investment research. I proposed the application, co-designed financial and dialogue information sets for tone residualization, built the time-aligned Capital IQ, IBES, and CRSP panel, and proposed dictionary and LLM validation. Across 1,901 held-out 2025 calls, a one-SD answer-tone residual was associated with +61.5 bps in one-day market-adjusted return, with standardization fixed from the development sample. Tone prediction achieved positive OOS R², and prespecified joint pre-event placebo tests found no significant association.

The manuscript and code remain private.

Data and further reading

Capital IQ transcripts, IBES, CRSP, historical S&P 500 constituents
Accessed through institutional subscriptions and WRDS.

Research assistant projects

Research assistant at UBC Sauder (Apr 2026–Present)

Age-Biased Technological Change

As the sole RA on this faculty-led study, I measure technological exposure by linking CPS, O*NET, and USPTO data, including 60+ harmonized historical O*NET releases. I validate PatentsBERTa measures with Qwen 3 and sampled manual reading, then construct variables, select fixed effects, and fit weighted FE OLS/GMM models of wages and occupational sorting. R and Python workflows on HPC/Slurm support large-scale estimation, bootstrap inference, and robustness checks. I maintain the replication package and draft the empirical section.

The project is led by J. Frank Li and Cheng Chen at UBC Sauder.

Data and further reading

Research assistant at UBC Sauder (May–Aug 2026)

Financial Media Multimodal Analysis

The project studies multimodal financial information across media. I assessed RavenPack news coverage against a ProQuest corpus of over 3.2 million articles. I independently built reproducible yt-dlp ETL that generates clean panels from metadata, archives and deduplicates video and audio, and validates subtitle availability, preparing financial alternative data for subsequent analysis.

I worked with Allen Hu at UBC Sauder.

News corpus pipeline

Data and further reading

Teaching and open materials

Research assistant at VSE and Praxis UBC (Jan 2025–May 2026)

NLI tutorial in first revision at Programming Historian

Data Science Notebooks, Courseware, and Workshops

Through COMET and Praxis UBC, I developed and deployed notebooks and HTML courseware for economics and social science students, applying computer vision, NLI, and sentiment analysis to historical and economic research. I led proposals, built image-embedding and classification workflows, and co-developed a zero-shot NLI tutorial with human-label validation and robustness checks. I also designed and taught a computer vision workshop.

Supervised by Jonathan Graves and Laura Nelson, in collaboration with the Praxis UBC team. Linked materials retain their individual coauthor credits.

Computer vision workshop Kouroi image-embedding notebook

Data and further reading