publications

Public manuscripts and preprints. Each entry links to the paper, released code, datasets or artifact bundles, and the project page with the problem framing and supporting artifacts.

manuscripts and preprints

2026

  1. arXiv
    When Top-1 Fails: Calibrating LoRA Monitors for Masked Diffusion LMs
    Lucky Verma and Pratik Yadav
    arXiv preprint arXiv:2606.24119, 2026
    paper pdf code project page doi
    abstract

    This paper audits a transferred top-1-frequency collapse warning for LoRA fine-tuning of masked diffusion language models. In the reported cohorts, the top-1 warning fires broadly while observed collapse is zero at the audited horizons, motivating family-local monitoring with max LoRA gradient norm instead.

  2. arXiv
    Weight Decay Regimes in Grokking Transformers: Cheap Online Diagnostics
    Lucky Verma
    arXiv preprint arXiv:2605.20441, 2026
    paper pdf code dataset project page doi
    abstract

    This solo-author preprint studies cheap online diagnostics for grokking phase regimes in transformers. The public artifact includes a Python package, aggregate JSONs, provenance manifests, figure artifacts, Lean 4 checks, and a Hugging Face dataset.

  3. arXiv
    When Does Removing LayerNorm Help? Activation Bounding as a Regime-Dependent Implicit Regularizer
    Lucky Verma
    arXiv preprint arXiv:2604.23434, 2026
    paper pdf code artifacts project page doi
    abstract

    This empirical study examines when replacing LayerNorm with activation bounding helps or hurts transformer training. The released artifact bundle includes configs, result manifests, validation scripts, aggregate JSONs, and a public Hugging Face artifact dataset.