Institution profile

Mahindra University

Academic institutionasia · in
Official website
Research library9linked papers
Opportunities0open roles
Selected work

Representative Papers

Lost in Visual Translation: A VLM-Assisted Perceptual-Semantic Coherence Framework for EEG-to-Image Reconstruction

Jul 14, 2026

Current evaluation methods for EEG-to-image reconstruction struggle to disentangle visual fidelity from semantic recoverability, often overestimating reconstructions that are semantically accurate yet visually blurry. To address this limitation, this work proposes the BCI-Coherence Score (BCS), which introduces, for the first time, a perception–semantic consistency framework leveraging vision-language models (VLMs). The approach employs structured semantic probes to guide multiple VLMs in generating dual-dimensional tolerance-aware scores, which are then fused into a unified metric. Evaluated on T-PAS (MAE=0.079, r=0.700) and T-SAS (MAE=0.082, r=0.850), BCS significantly outperforms conventional metrics and demonstrates strong alignment with human judgments (Cohen’s κ=0.882), thereby establishing a new evaluation paradigm tailored for brain–computer interface–based image reconstruction.

0 citationsRead paper

Phantom transitions in language model fine-tuning

May 25, 2026

This work addresses the "silent failure" phenomenon in language model fine-tuning—where correct tokens fail to outcompete semantically similar alternatives despite a steadily decreasing cross-entropy loss—by introducing a novel analytical framework based on density matrices. The authors construct an order parameter that integrates predictive distributions with geometric overlap in token embedding space, decomposing prediction dynamics into signal and background drag components. This approach reveals, for the first time in non-orthogonal embedding spaces, two distinct failure mechanisms: kinematic and structural failures, and clarifies the nature of pseudo-phase transitions. Combining geometric embedding analysis, LoRA comparisons, and gradient-step-level dynamics monitoring, the study identifies a universal dimensionless quantity under full-parameter fine-tuning that accurately predicts the critical learning rate for leave-one-out architectures, achieving a prediction error of only 2.1%.

0 citationsRead paper

How effective are VLMs in assisting humans in inferring the quality of mental models from Multimodal short answers?

Feb 10, 2026Conference of the European Chapter of the Association for Computational Linguistics

This study addresses the challenge of assessing the quality of students’ mental models in multimodal short-answer responses, which requires deep reasoning beyond the capabilities of traditional scoring methods that often fail to capture conceptual understanding. To this end, the authors propose MMGrader, a novel approach that integrates vision-language models (VLMs) with concept graphs to jointly model the semantic content and structural organization of multimodal answers, enabling interpretable and structured evaluation of mental model quality. Evaluated across nine open-source models, MMGrader achieves state-of-the-art performance with an accuracy of 40% and a prediction error of 1.1 points, while its score distributions align closely with human grading trends. This work offers educators an effective AI-driven paradigm for diagnosing collective classroom understanding at scale.

0 citationsRead paper
Recent publications

Latest Papers

Lost in Visual Translation: A VLM-Assisted Perceptual-Semantic Coherence Framework for EEG-to-Image Reconstruction

Jul 14, 2026

Current evaluation methods for EEG-to-image reconstruction struggle to disentangle visual fidelity from semantic recoverability, often overestimating reconstructions that are semantically accurate yet visually blurry. To address this limitation, this work proposes the BCI-Coherence Score (BCS), which introduces, for the first time, a perception–semantic consistency framework leveraging vision-language models (VLMs). The approach employs structured semantic probes to guide multiple VLMs in generating dual-dimensional tolerance-aware scores, which are then fused into a unified metric. Evaluated on T-PAS (MAE=0.079, r=0.700) and T-SAS (MAE=0.082, r=0.850), BCS significantly outperforms conventional metrics and demonstrates strong alignment with human judgments (Cohen’s κ=0.882), thereby establishing a new evaluation paradigm tailored for brain–computer interface–based image reconstruction.

0 citationsRead paper

Phantom transitions in language model fine-tuning

May 25, 2026

This work addresses the "silent failure" phenomenon in language model fine-tuning—where correct tokens fail to outcompete semantically similar alternatives despite a steadily decreasing cross-entropy loss—by introducing a novel analytical framework based on density matrices. The authors construct an order parameter that integrates predictive distributions with geometric overlap in token embedding space, decomposing prediction dynamics into signal and background drag components. This approach reveals, for the first time in non-orthogonal embedding spaces, two distinct failure mechanisms: kinematic and structural failures, and clarifies the nature of pseudo-phase transitions. Combining geometric embedding analysis, LoRA comparisons, and gradient-step-level dynamics monitoring, the study identifies a universal dimensionless quantity under full-parameter fine-tuning that accurately predicts the critical learning rate for leave-one-out architectures, achieving a prediction error of only 2.1%.

0 citationsRead paper

How effective are VLMs in assisting humans in inferring the quality of mental models from Multimodal short answers?

Feb 10, 2026Conference of the European Chapter of the Association for Computational Linguistics

This study addresses the challenge of assessing the quality of students’ mental models in multimodal short-answer responses, which requires deep reasoning beyond the capabilities of traditional scoring methods that often fail to capture conceptual understanding. To this end, the authors propose MMGrader, a novel approach that integrates vision-language models (VLMs) with concept graphs to jointly model the semantic content and structural organization of multimodal answers, enabling interpretable and structured evaluation of mental model quality. Evaluated across nine open-source models, MMGrader achieves state-of-the-art performance with an accuracy of 40% and a prediction error of 1.1 points, while its score distributions align closely with human grading trends. This work offers educators an effective AI-driven paradigm for diagnosing collective classroom understanding at scale.

0 citationsRead paper