Lalit Chandra Routhu

Author directory

2026

With 1.5 billion people speaking over 120 major languages, India exemplifies the challenges of multilingual AI evaluation. Current multilingual VLM benchmarks suffer from unverified auto-translations, narrow task coverage, small sample sizes, and lack of culturally grounded content. We present HinTel-AlignBench, a comprehensive evaluation framework and benchmark for Hindi and Telugu vision-language models with English-aligned samples. Our framework combines semi-automated translation with human verification to generate 4k QA pairs per language across five domains: adapted English datasets (VQAv2, RealWorldQA, CLEVR-Math) and native Indic sets (JEE for STEM, VAANI for cultural grounding). Evaluation of state-of-the-art open and closed-source VLMs reveals consistent performance regression from English to Indic languages, with average drops of 8.3 points for Hindi and 5.5 points for Telugu across four of five tasks. We identify key failure modes and establish reproducible baselines for multilingual multimodal evaluation.
Humor generation remains a challenging problem for language models due to its subjective, creative, and culturally sensitive nature. While large language models (LLMs) demonstrate strong generative capabilities, their deployment is often constrained by computational cost. We propose a lightweight, corrective feedback-driven framework in which a compact generator (GPT-2, 140M parameters) produces punchlines under the guidance of an LLM acting as a critic. Unlike traditional knowledge distillation, our method leverages the LLM as a feedback provider rather than a teacher to be mimicked, enabling the student model to iteratively refine its outputs through corrective supervision. We systematically evaluate multiple LLMs (LLaMA3-8B, Qwen3-8B and Mistral-7B) as guides under different strategies, including frozen and learnable configurations, to study their impact on humor quality and diversity. To support responsible humor generation, we curate a high-quality setup–punchline dataset from Reddit and explicitly filter out harmful content. Experimental results show that corrective feedback-guided training substantially improves humor quality, fluency, and diversity over standard fine-tuning and knowledge distillation baselines, while preserving the efficiency of compact models. We observe that, among the evaluated guides, Mistral-7B consistently provides particularly strong guidance on the curated joke dataset. Our code and resources are available at: https://github.com/Ramakrishna176420/Corrective-Feedback-Driven-Humor-Generation-for-Tiny-Language-Models.