Leyang Hu 胡乐阳

PhD Student in Computer Science, Brown University

I am a PhD student in Computer Science at Brown University, advised by Prof. Randall Balestriero. My work connects the theory and practice of deep learning: theory guides practice, and practice inspires new theory.

Concretely, that runs in two directions.

News

Selected Publications

The complete and most current list is on Google Scholar.

  1. Classification accuracy against reconstruction loss: level-set walks span a far wider accuracy range than conventional training reaches at the same loss
    TAG-DS 2026

    Learning by Reconstruction is an Ill-Defined Prior for Perception: A Level Set View

    Leyang Hu, Matteo Gamba, Akshay Ghandikota, Akash Nagaraj, Randall Balestriero

    Encoders with identical reconstruction loss span 4%–74% linear-probe accuracy, so reconstruction underdetermines representation quality — what separates them is optimization’s implicit bias.

  2. Accuracy relative to a supervised baseline for six self-supervised methods across 41 datasets, grouped by domain
    Under Review 2026

    Is SSL Ready for In-Domain Pretraining? A Cross-Dataset Benchmark and Analysis

    Sami BuGhanem, Leyang Hu, Haodong Zhang, … Randall Balestriero

    A benchmark of six modern SSL methods across 41 image datasets, showing that modern SSL can be competitive with supervised learning for limited-data in-domain pretraining.

    Code

  3. Lowering the curvature parameter beta from 1.0 (ReLU) smooths a pretrained network’s decision boundary and regression fit
    NeurIPS 2025

    Curvature Tuning: Provable Training-Free Model Steering From a Single Parameter

    Leyang Hu, Matteo Gamba, Randall Balestriero

    Fine-tuning that updates activation functions instead of weights, provably reshaping decision boundaries — 10.20% higher downstream accuracy than LoRA with under 60% of the parameters.

    arXiv Code Blog

  4. Sentiment words separate cleanly in perception space but collapse together in autoregressive space
    Preprint 2025

    Next Token Perception Score: Analytical Assessment of Your LLM Perception Skills

    Yu-Ang Cheng, Leyang Hu, Hai Huang, Randall Balestriero

    An analytical score for how well autoregressive pretraining features align with a downstream task, tracking linear-probe accuracy across 12 datasets and eight LLMs.

    arXiv Code Blog

  5. First-order prompt optimization stalls in a local optimum across iterations while REVOLVE keeps improving
    ICML 2025

    Revolve: Optimizing AI Systems by Tracking Response Evolution in Textual Optimization

    Peiyan Zhang, Haibo Jin, Leyang Hu, … Haohan Wang

    LLM self-refinement that tracks how responses evolve across iterations, using discrete second-order differences for more stable textual optimization.

    arXiv Code

  6. Harmful and harmless prompts separate in Llama-2-7b embedding space, with DROJ shifting harmful prompts across the boundary
    Preprint 2024

    DROJ: A Prompt-Driven Attack Against Large Language Models

    Leyang Hu, Boran Wang

    An embedding-level jailbreak that optimizes adversarial prompts directly, reaching 98.46% attack success on AdvBench with Llama 2 7B.

    arXiv Code