Welcome to प्रragya(IPA – /praːgjaː/ 🔊) Lab @ BITS Goa, India

Deception is the intentional practice of twisting information. This research explores three forms: lies of omission, lies of commission, and lies of influence, focusing primarily on lies of omission. We introduce a novel NLP framework supported by an annotated dataset of 876,784 samples from merged fake-news sources and Times of India headlines. Each sample is labeled for omission type (e.g., speculation, bias), “color” of lie (black/white/gray/red), intent (influence, prestige), and topic (political, educational, religious, etc.). Our multi-task learning pipeline, which merges fine-tuned language models, attains an F1 score of 0.87 across all annotation layers. We also examine correlations between omission-based lies and propaganda techniques—e.g., loaded language often signals opinion—and will release both model and dataset under an open license.

We present DPO Kernels, combining kernel density estimation with divergence-based objectives for direct preference optimization in large language models. By incorporating semantics via custom kernel functions, DPO Kernels significantly improve the alignment between model outputs and human preferences. Experiments on public preference datasets show gains of 3-5% over existing DPO approaches, demonstrating improved generation quality and user satisfaction.

YINYANG-ALIGN is a novel benchmark and training framework for text-to-image alignment that concurrently optimizes for coherence and diversity, conceptually contradictory objectives. We propose multi-objective DPO training that balances these via Pareto optimality. Evaluated on standard T2I datasets, YINYANG-ALIGN achieves higher user preference and diversity metrics without sacrificing fidelity.

This work investigates LLMs’ capability on analogical reasoning (e.g., A:B :: C:? queries) with knowledge-enhanced prompts. We design knowledge retrieval modules that inject relevant facts into prompts, enhancing model performance by  12% on proportional analogy datasets. Analyses reveal improvements when symbolic knowledge complements contextual embeddings.

We introduce the Hindi AI Detectability Index (ADIhi), a benchmark suite and evaluation metric for detecting AI-generated Hindi text. We compile a corpus of 20K human-written and LLM-generated texts and evaluate detectors including token-level, syntactic, and stylometric features. Results indicate existing detectors struggle on code-mixed content and slang; using ADIhi, we rank LLMs and propose improvements tailored for Hindi.

We present FACTIFY3M, a multimodal factual verification benchmark that combines images, text, and explanations grounded in 5W QA. The dataset includes 50K multimodal claims with annotated 5WQA pairs and supporting sources. Baseline models using joint vision-language transformers achieve 74.5% accuracy, but leave room for improvement; the 5W explanations improve interpretability and diagnostic evaluation.

This paper introduces a formal definition and taxonomy for LLM hallucinations, proposes automatic metrics for hallucination quantification (e.g., hallucination density, factual error scores), and presents remediations such as knowledge grounding, self-critique modules, and multi-step prompting. In experiments across GPT-3, BERT, and open-source LLMs, these countermeasures reduce hallucination by up to 35%.

This paper demonstrates that standard token- and syntax-based detectors poorly generalize to AI-generated text in Hindi and code-mixed domains. We propose CT2 detection models that leverage stylometric and structural features, improving cross-domain robustness by 12%. We release CT2 dataset and models for future research.