[BACK]

[PAPERS]

[RESEARCH]

PEER_REVIEWED / SELECTED_PUBLICATIONS

[01]PREPRINT / RESEARCH SQUARE / 2025

Merging LoRA Adapters for Multi-Task Code Analysis: An Empirical Study of Linear Combination and Task Interference

SANKALP PATHAK

We investigate whether task-specific LoRA adapters — each fine-tuned independently on Meta-Llama-3.1-8B-Instruct — can be merged via weighted linear combination into a single adapter that preserves performance on both tasks. We evaluate 19 configurations (a 4×4 lambda grid plus three baselines) on synthetic static-code-analysis data (3,463 samples) and PrimeVul vulnerability data (9,858 expert-verified C/C++ samples). The best merged configuration retains 98% of solo vulnerability-detection performance while gaining code-analysis capability, and 91% of solo code-analysis performance while gaining vulnerability detection. Interference is asymmetric: vulnerability detection is more sensitive to the code-analysis adapter weight than vice versa.

LORALLMSCODE_ANALYSISVULN_DETECTION
[02]CHEMICAL ENGINEERING SCIENCE / 2025

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

P. MISHRA, S. KHANNA, P. GUPTA, S. PATHAK, ET AL.

Uric acid is a critical biomarker of purine metabolism — too high points to gout, hypertension, and renal disease; too low correlates with neurodegenerative disorders. This work introduces a μPAD + smartphone platform that quantifies uric acid between 1.5–25 mg/dL without external light sources, enzymes, or nanoparticles. Microfluidic channels are drawn directly onto Whatman G1 filter paper by a low-cost DIY XY plotter using technical pens filled with a PDMS/heptane hydrophobic solution. Detection runs on a chromogenic reaction between potassium ferricyanide and ferric chloride — uric acid intensifies the resulting bluish-green complex, and a custom Android app reads its luminance against eight on-chip reference zones.

MICROFLUIDICSCOMPUTER_VISIONBIOSENSINGANDROID
[03]PREPRINT / ZENODO / 2026

A Rate-Distortion Function for Model Merging

SANKALP PATHAK, SANJAY GARG

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. We cast merging as multi-source lossy source coding under worst-task distortion and prove the first rate–distortion theorem for it. For T tasks with rank-r updates, a per-task error radius B, and a budget of R bits, the worst-task distortion under a Stiefel-random worst case has a closed-form floor B²(1 − d_eff/(Tr)) plus a compression term Θ(2^(−2R/d_eff)). A Gaussian-QR rotation with uniform scalar quantization matches the lower bound up to a constant. Across 16 real LoRA adapters the task subspaces are linearly independent, so the floor is zero — yet current methods still leave 0.10–0.22 nats/token of worst-task error, showing the gap is algorithmic, not informational.

INFO_THEORYMODEL_MERGINGLORAQUANTIZATION

[PAPERS] / [3_PUBLICATIONS] / [CLICK_TO_READ_MORE]