Sumit Yadav

Publications 9

Peer-reviewed papers, preprints and reports — mechanistic interpretability, AI safety and low-resource NLP. Much of the interpretability work is sole-authored.

1. 

   ACL 2026 Main · CORE A*

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

   Utsav Maskey, Sumit Yadav, Mark Dras, Usman Naseem

   Task-specific trajectories in embedding space separate refusal from non-refusal. Steering only over-refusal-prone tasks at inference time cuts over-refusals up to 73% — no retraining, minimal utility loss.

   Paper arXiv DOI Write-up

2. 

   NeurIPS 2026 Interp4Discovery workshop · Atlanta

   Your Intervention Was Too Narrow: Mechanism Width as a Precondition for Interpretability-Driven Discovery

   Sumit Yadav, Basanta Joshi

   A component is credited with a function when perturbing it changes behaviour — but the standard one-unit intervention recovers only ~10% of the achievable effect (8.6% one head on GPT-2 IOI, 9.6% one layer on Llama-3.1-8B, 11.1% rank-1 steering in DINOv2). Measure the mechanism's width k* first, widen to match, and you recover 95–100%. A null result from a too-narrow intervention is not evidence of absence.

   Paper OpenReview Preprint Project

3. 

   Preprint 2026 sole-authored

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

   Sumit Yadav

   Effective dimension — unsupervised and label-free — predicts generalization across vision and language (partial r = 0.75 over 52 ImageNet models, 13 architecture families). Causality runs both ways: degrade the geometry with noise and accuracy follows (r = −0.94).

   arXiv DOI Write-up

4. 

   GLOW @ IJCAI-ECAI 2026 sole-authored · best-paper candidate

   Geometric Phases of Mechanism Formation in Neural Networks

   Sumit Yadav

   Linear probes and CKA over dense checkpoints: classification mechanisms form output-layer-first, inside the first ~5% of training (Cohen's d = 3.68). The same deep-first pattern holds in the first ~200M tokens of LLM pretraining and reproduces on Pythia and OLMo-2.

   Preprint DOI

5. 

   LoResLM @ EACL 2026 Rabat, Morocco

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

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

   The first monolingual BERT for Maithili (~50M speakers), pre-trained on a newly built corpus. 87.02% on news classification — 5–7% over NepBERTa and HindiBERT — and open-sourced for downstream NER and sentiment work.

   Paper arXiv DOI Model

6. 

   J. Bus. Econ. Stud. 2024 1(6): 1–8

   Revolutionizing Currency Security: A YOLOv8-Based Approach for Automated Detection of Counterfeit Nepali Banknotes

   Santosh Giri, Loknath Regmi, Sumit Yadav, Binaya Basnet, Binit KC, Mohan Bhandari

   A YOLOv8 detector for counterfeit 1000-rupee notes: true-positive recall of 0.82 on the front face and 0.986 on the back, portable across hardware platforms.

   Paper

7. 

   B.E. Thesis 2024 Tribhuvan University (IOE, Pulchowk)

   Evaluating Auto-Encoding Transformer Language Models for Maithili Text Classification

   Sumit Yadav, Raju Kumar Yadav

   A Maithili masked language model built by transfer learning and fine-tuned on a curated news-classification set — the precursor work to maiBERT.

   Thesis

8. 

   Technical Report 2023 Tribhuvan University

   Machine Learning Analysis of Tirhuta Lipi

   Sumit Yadav, Raju Kumar Yadav

   Character recognition for the endangered Tirhuta script: MobileNet embeddings with logistic regression reach 0.97 accuracy, opening OCR and translation for Maithili.

   Report DOI

9. 

   2023

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

   Sumit Yadav et al.

   A TF-IDF + n-gram support-vector pipeline over 100+ imbalanced classes that beats neural baselines and retrains incrementally. Consolidates two 1st-runner-up competition submissions (LOCUS 2021, 2023).

   Paper