Publications - Abhishek Bhandari

Schematic of the pipeline: an OCR sentence is matched by CharBM25 to similar examples, which prompt an LLM to produce corrected text.

Evaluating In-Context Learning and Retrieval Strategies for Devanagari Post-OCR Correction

Abhishek Bhandari, Gaurav Harit

arXiv preprint arXiv:2609.21595 · 2026

The first systematic evaluation of 3B-32B LLMs for training-free post-OCR correction in Hindi and Marathi. A proposed character n-gram BM25 retrieval (CharBM25) for in-context examples outperforms random selection by up to 4 points of absolute WER.

Paper Dataset

A Framework and Dataset for Contextual Post-OCR Correction

Abhishek Bhandari, Gaurav Harit

ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP) · 2026

A correction framework uses the preceding sentence as context to resolve OCR errors in Hindi, Marathi, and Gujarati. The accompanying dataset supports evaluation with and without sentence context.

Paper Dataset

Figure 1 from the paper: the proposed multi-view architecture with character-wise gated fusion.

Post-ASR Correction for Low-Resource Rajasthani Language

Abhishek Bhandari, Gaurav Harit

ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP) · 2026

A character-level correction model combines transcripts from Whisper and MMS through gated fusion, using complementary recognition cues to correct errors in low-resource Rajasthani speech transcription.

Paper

Mind the (Language) Gap: Towards Probing Numerical and Cross-Lingual Limits of LVLMs

Somraj Gautam, AS Penamakuri, Abhishek Bhandari, Gaurav Harit

Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025) · 2025

MMCRICBENCH-3K tests numerical and cross-lingual reasoning on English and Hindi cricket scorecard images. It exposes difficulties in interpreting table structure and transferring visual understanding across scripts.

Paper Dataset

TabComp: A Dataset for Visual Table Reading Comprehension

Somraj Gautam, Abhishek Bhandari, Gaurav Harit

Findings of the Association for Computational Linguistics: NAACL 2025 · 2025

TabComp pairs table images with questions and generative answers to evaluate OCR-free reading comprehension. It reveals limitations in how current models interpret complex visual tables.

Paper Dataset Code

Reversible data hiding using multi-layer perceptron based pixel prediction

A Bhandari, S Sharma, R Uyyala, R Pal, M Verma

Proceedings of the 11th International Conference on Advances in Information Technology · 2020

A multi-layer perceptron predicts image pixels for reversible data hiding, allowing embedded information to be extracted while restoring the original image exactly.

Paper