Sankalp Pathak

I study what happens when you merge fine-tuned language models, and whether you can tell what will break just by looking at the weights.

Research at IIIT Allahabad; B.Tech in computer science from JUET (2025). One paper published, five under review. I write my predictions down before running the experiments that test them, and I report the ones that fail too.

pathaksankalp04@gmail.comGitHubLinkedIn

Papers

  1. Affordable Smartphone-Assisted Diagnostics: Computer Vision on Paper Microfluidics for Uric Acid Detection

    P. Mishra, S. Khanna, P. Gupta, S. PathakPublished, Chemical Engineering Science 2025

    Uric acid is a key biomarker: too high points to gout, hypertension and kidney disease, too low to neurodegenerative disorders.

  2. Looped Transformers Learn One of Two Algorithms, and Training Moves the Odds

    S. Pathak, S. GargUnder review, ICLR 2027

    Depth-recurrent (looped) transformers trained on prefix-product tasks do not learn a spectrum of more or less efficient solutions.

  3. Sign Conflicts in LoRA Adapter Merging Are Pervasive, Not Localized

    S. Pathak, S. GargUnder review, TMLR 2026

    Sparsification-based merging methods assume that sign conflicts between LoRA adapters are concentrated in a few weights that can be trimmed away.

  4. Weight-Space and Behavioral Interference in Multi-Objective LoRA Debiasing for Indian Social Bias: A Cross-Architecture Study

    S. Pathak, N. SainiUnder review, ARR for EACL 2027

    MOOLoRA treats debiasing Indic-language models across four Indian social axes as multi-objective optimisation over LoRA merge coefficients, solved with NSGA-II.

  5. Same Models, Different Verdicts: An Evaluation-Sensitivity Study of LoRA Adapter Merging for Multi-Task Code Analysis

    S. Pathak, S. GargUnder review, IEEE TSE 2026

    An audit of 1,020 configurations of LoRA adapter merging for code-analysis tasks.

  6. A Rate-Distortion Function for Model Merging

    S. Pathak, S. GargPreprint, Zenodo 2026

    Practitioners merge many LoRA fine-tunes into one deployable model, but no principled limit says how well this can be done at a fixed storage budget.

Abstracts, and a small interactive demo for each

Experience

IIIT Allahabad Since May 2026
Research intern with Dr. Naveen Saini. Built MOOLoRA, which merges debiasing adapters for Indic-language models as a Pareto problem.
With Prof. Sanjay Garg Since 2025
Independent research on adapter merging, looped transformers, and how evaluation choices change conclusions.
GreyB Aug to Dec 2025
Research analyst: prior-art searches and technical reports on AI, vision and networking.
AWS Certified AI Practitioner July 2026
Scored 942 out of 1000.

Teaching and awards

Taught game development to juniors at JUET on Saturdays: 45 students in my second year, 117 in my third, and more than two hundred in my last. Won RideHack 2024 and JUET Builds; joint secretary of the VR/AR/MR club.

Things I’ve made