Research archive / 01
I study how text becomes representation, how unwanted signals enter that representation, and how an intervention changes the decisions built on top.
Semantic modeling · Representation analysis · LLM reliabilityFirst author · IPM · 2026
Online reviews rarely express one clean preference. This work extracts aspect–opinion–sentiment structures with a QLoRA-adapted Qwen model, maps sentiment into a continuous space, and models compensatory consumer trade-offs for decision support.
First author · AEI · 2026
Platform bias is often handled after text has already been compressed into scalar scores. This work probes representation-level platform signals, removes their linear directions with LEACE, and aligns cross-platform semantics before ranking.
National undergraduate innovation project
I lead a project connecting the two research lines above: extracting decision-relevant meaning from real reviews, then making representations more comparable across platforms.
Current research interests
These topics extend my existing work in semantic modeling, representation intervention, and open-source LLM tooling.
Probe representations for platform, sentiment, and task signals; then measure what remains after a controlled intervention.
Evaluate knowledge edits for efficacy, locality, robustness, and consistency across devices and model implementations.
Test whether exchanged information improves receiver-conditioned reasoning under matched budgets and explicit controls.
Public links
Status labels describe the current process; they do not imply acceptance. Code and preprints will be linked only when they are ready for public use.
GitHub↗