Sumit Yadav

SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering

Utsav Maskey, Sumit Yadav, Mark Dras, Usman Naseem

ACL 2026 (Main Conference)

LLM representations follow distinct per-task "constellation" trajectories in embedding space; SafeConstellations is an inference-time method that selectively steers only over-refusal-prone tasks toward non-refusal pathways — cutting over-refusals by up to 73% with minimal utility loss and no retraining.

paperdoi

On the Relationship Between Representation Geometry and Generalization in Deep Neural Networks

Sumit Yadav (sole author)

Preprint, 2026

Effective dimension — an unsupervised, label-free geometric metric — predicts generalization across domains: partial r = 0.75 over 52 ImageNet classifiers (13 architecture families), replicating on CIFAR-10 and generalizing to NLP (8 encoders on SST-2/MNLI, 15 decoder-only LLMs on AG News). Establishes bidirectional causality — degrading geometry with noise drops accuracy (r = −0.94), while PCA that improves geometry preserves it.

paperdoi

Geometric Phases of Mechanism Formation in Neural Networks

Sumit Yadav

GLOW 2026 — Workshop on Generalizing from Limited Resources in the Open World @ IJCAI 2026 (Poster)

Using linear probes and centered kernel alignment (CKA) across dense training checkpoints, finds that classification mechanisms form output-layer-first and within the first ~5% of training: output layers reach >70% probe accuracy by epoch 5 while input layers stay below 50% (Cohen's d = 3.68). The same deep-first pattern holds in the first ~200M tokens of from-scratch LLM pretraining (GPT-2 Small, SmolLM2-135M) and reproduces on public Pythia / OLMo-2 checkpoints — and isn't explained by gradient magnitude. arXiv & code coming soon.

MaiBERT: A Pre-training Corpus and Language Model for Low-Resourced Maithili Language

Sumit Yadav, Raju Kumar Yadav, Utsav Maskey, Gautam Siddharth Kashyap, Ganesh Gautam, Usman Naseem

LoResLM @ ACL 2026 (pp. 444–452), Rabat, Morocco

The first monolingual BERT for Maithili (~50M speakers), pre-trained with masked language modeling on a newly constructed Maithili corpus. Reaches 87.02% accuracy on news classification — outperforming regional models NepBERTa and HindiBERT with 5–7% gains across classes — and is open-sourced for downstream tasks like sentiment analysis and NER. (Earlier preprint: "Can maiBERT Speak for Maithili?")

paperarXivdoimodel

Revolutionizing Currency Security: A YOLOv8-Based Approach for Detecting Counterfeit Nepali Banknotes

Sumit Yadav et al.

J. Bus. Econ. Stud., 2024

Classifies Nepalese banknotes as genuine or counterfeit to help curb financial fraud. On 180 samples of the 1000-rupee note, YOLOv8 reaches a true-positive recall of 0.82 on the front face and 0.9863 on the back, with strong mAP — and the approach is adaptable across hardware platforms.

paper

Evaluating Auto-Encoding Transformer Language Models for Maithili Text Classification

Sumit Yadav, Raju Kumar Yadav

B.E. Thesis (Electronics & Computer Engineering), Tribhuvan University, 2024

Builds a Maithili masked language model via transfer learning and fine-tunes it on a curated Maithili news classification dataset — the precursor work to maiBERT, addressing the absence of dedicated language models and task-specific data for Maithili.

paper

Machine Learning Analysis of Tirhuta Lipi

Sumit Yadav, Raju Kumar Yadav

Technical Report, Tribhuvan University, 2023

Character recognition for the Tirhuta script (Maithili) — MobileNet embeddings with logistic regression reach 0.97 accuracy, enabling OCR and translation for a low-resource script.

paperdoi

Support Vectors Are a Better Way of Text Classification for Imbalanced Data

Sumit Yadav et al.

2023

A TF-IDF + n-gram support-vector pipeline for text classification across 100+ classes on highly imbalanced data. It outperforms neural baselines on the test set and supports incremental retraining as users supply fresh data. Consolidates the team's 1st Runner-Up submissions at Docsumo DataRush (LOCUS 2021) and DataVerse (LOCUS 2023).

paper