Deepak Pandita | Tharindu Cyril Weerasooriya

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

Sat, 01 Nov 2025 00:00:00 +0000

ARTICLE: Annotator Reliability Through In-Context Learning

Sat, 01 Mar 2025 00:00:00 +0000

When humans label data for AI (like deciding if a tweet is offensive), how do we know who’s doing good work versus who’s careless or biased? Traditional methods struggle because on subjective tasks, sincere people genuinely disagree. We built ARTICLE, which uses large language models to check if an annotator’s labels are internally consistent and align with what the AI predicts. This helps identify reliable annotators without silencing minority perspectives—crucial for building fair AI systems that don’t just reflect the majority view.

ProRefine: Inference-Time Prompt Refinement with Textual Feedback

Wed, 01 Jan 2025 00:00:00 +0000

When multiple AI agents work together (like one agent planning and another executing), they rely heavily on good prompts. But writing perfect prompts is hard, and even small mistakes can cascade through the system, causing failures. ProRefine solves this by creating a feedback loop: one AI agent evaluates how well the prompt worked and suggests improvements, then another agent refines the prompt based on that feedback—all happening automatically at inference time without any training.

We tested ProRefine on mathematical reasoning tasks and it improved accuracy by 3-37% compared to standard methods. Even better, it lets smaller, cheaper AI models perform nearly as well as much larger, expensive ones. This makes powerful AI more accessible and cost-effective, especially for complex multi-agent systems that are becoming critical in commercial applications.

Rater Cohesion and Quality from a Vicarious Perspective

Fri, 01 Nov 2024 00:00:00 +0000

When labeling controversial content (like tweets about politics), people’s own beliefs strongly influence what they find offensive. We used “vicarious annotation”—asking people not just “Is this offensive to you?” but also “Would a Democrat/Republican/Independent find this offensive?” This revealed fascinating patterns: Republicans were the worst at predicting how others would react, and disagreement spiked on hot-button issues like gun control and abortion. Understanding these patterns helps us build content moderation AI that doesn’t just reflect one group’s values while silencing others.