Tharindu Cyril Weerasooriya

LPI-RIT at LeWiDi-2025: Improving Distributional Predictions via Metadata and Loss Reweighting with DisCo

Subasa - Adapting Language Models for Low-resourced Offensive Language Detection in Sinhala

ARTICLE: Annotator Reliability Through In-Context Learning

Using LLMs to identify high-quality human annotators by checking if their labels are consistent with AI predictions—helping build better training data while preserving diverse …

ProRefine: Inference-Time Prompt Refinement with Textual Feedback

ProRefine automatically improves AI prompts during inference by having one AI agent give feedback to refine another agent's prompts—boosting accuracy by 3-37% and helping smaller …

MultiGA: Leveraging Multi-Source Seeding in Genetic Algorithms

Mixing outputs from multiple LLMs (GPT-4, DeepSeek, etc.) using genetic algorithms to evolve better solutions for complex reasoning tasks—like breeding the best answer from diverse …

Blind Spot Navigation in Large Language Model Reasoning with Thought Space Explorer

Note Presenting at NeurIPS (Math-AI Workshop) Date: December 6, 2025 Time: Sat 3:30 p.m. - 4:15 p.m. Location: NeurIPS 2025 - Workshop Upper Level Ballroom 6A Session Type: Poster …

Rater Cohesion and Quality from a Vicarious Perspective

Asking people to predict how others with different political views would label content reveals hidden biases and improves data quality for content moderation AI.

Harnessing Business and Media Insights with Large Language Models

Vicarious Offense and Noise Audit of Offensive Speech Classifiers: Unifying Human and Machine Disagreement on What is Offensive

We ran a massive experiment: 9 different AI content moderation systems analyzed 92 million YouTube comments about US politics. The results were shocking—different AI systems …

Subjective Crowd Disagreements for Subjective Data: Uncovering Meaningful CrowdOpinion with Population-level Learning

CrowdOpinion uses unsupervised learning to group similar content and predict the full range of human opinions about it, rather than forcing everyone into a single 'correct' …