Publications on Abhiraj Bibhar

Attention-Guided Federated Learning for Privacy-Preserving Clinical NLP

Mon, 01 Jun 2026 00:00:00 +0000

Abstract

Training clinical natural language processing models typically requires centralizing patient records, which is often legally or ethically infeasible across institutions. We introduce an attention-guided federated learning protocol that allows clinical language models to be trained collaboratively across multiple hospital systems without any patient data leaving its host institution.

Human-in-the-Loop Interfaces for Trustworthy AI Decision Support

Thu, 01 May 2025 00:00:00 +0000

Abstract

We report on a mixed-methods study of 48 clinicians interacting with three AI-assisted diagnostic interfaces that varied in how they presented model confidence and supporting evidence. Confidence calibration displays significantly improved appropriate reliance on model suggestions, and clinicians consistently preferred interfaces that surfaced underlying evidence rather than raw confidence scores. Time pressure substantially altered how much clinicians engaged with explanatory content — a factor largely absent from prior lab-based evaluations of AI decision support.

A Survey of Explainability Methods for Graph Neural Networks

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

Abstract

Graph neural networks (GNNs) are increasingly deployed in high-stakes settings, yet the tools available for explaining their predictions lag behind those for standard deep networks. This survey organizes the growing landscape of GNN explainability methods along three axes — instance-level versus model-level, perturbation-based versus gradient-based, and faithfulness versus plausibility — and evaluates each family against a consistent set of faithfulness benchmarks.

Low-Resource Adaptation of Large Language Models for Regional Dialects

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

Abstract

Large language models perform unevenly across regional dialects that are underrepresented in web-scale training corpora. We propose a parameter-efficient adaptation method that fine-tunes a small set of dialect-specific adapter layers using as few as 5,000 labeled examples, substantially closing the performance gap on downstream tasks including sentiment analysis, named entity recognition, and machine translation for four regional dialect groups.

Fairness-Aware Ranking in Multi-Stakeholder Recommender Systems

Sat, 01 Jun 2024 00:00:00 +0000

Abstract

Recommender systems increasingly serve multiple stakeholders — consumers, item providers, and platform operators — whose interests are not always aligned. We formalize a fairness-aware ranking objective that balances consumer relevance against provider-side exposure fairness, and propose an efficient re-ranking algorithm that achieves Pareto-improving trade-offs relative to relevance-only baselines on three large-scale recommendation datasets.

Toward Reproducible Benchmarks in Federated Learning Research

Fri, 01 Dec 2023 00:00:00 +0000

Abstract

Federated learning papers frequently report results on incompatible experimental setups — differing client counts, non-IID partitioning strategies, and communication budgets — making cross-paper comparison unreliable. This position paper proposes a minimal standardized benchmark protocol and reports a re-evaluation of six widely cited federated learning methods under matched conditions, finding that reported performance gaps shrink substantially once experimental setup is controlled for.

Interactive Visualization Techniques for Deep Model Debugging

Sun, 01 Oct 2023 00:00:00 +0000

Abstract

Debugging deep learning models remains largely a manual, intuition-driven process. We present GlassBox, an interactive visualization system that lets researchers inspect layer activations, attention weights, and gradient flow simultaneously across a model’s forward and backward passes, with linked views that highlight where behavior diverges from expectations.

Cognitive Load in Multimodal Tutoring Systems

Thu, 01 Sep 2022 00:00:00 +0000

Abstract

Multimodal tutoring systems that combine text, audio, and visual explanation can improve learning outcomes, but they can also increase extraneous cognitive load if poorly designed. Across three classroom studies with a combined 240 students, we measure cognitive load using dual-task response times and find that adaptive modality selection — choosing which modality to emphasize based on a learner’s real-time performance — reduces extraneous load without harming learning gains.