CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

What this is

Personal academic website for Hamidreza Yaghoubi (Ph.D. student, UMD GAMMA Lab). Built on the Academic Pages Jekyll theme, which is itself a fork of Minimal Mistakes.

The site is deployed to GitHub Pages at https://hamidrezayaghoubi.github.io.

Local development

bundle install          # install Ruby dependencies (first time or after Gemfile changes)
jekyll serve -l -H localhost   # serve at localhost:4000 with live reload

Or via Docker:

docker build -t jekyll-site .
docker run -p 4000:4000 --rm -v $(pwd):/usr/src/app jekyll-site

Architecture

Two rendering paths coexist:

index.html — The homepage is a hand-crafted, self-contained HTML/CSS/JS file with a dark GitHub-style design. It does not go through Jekyll’s layout system or use Liquid templates. All styles are inline in <style> tags. This is intentional — it bypasses the Academic Pages theme entirely for the landing page.
Jekyll-rendered pages — Everything under _pages/ (about, CV, publications, talks, etc.) uses the Academic Pages theme. Layouts live in _layouts/, partials in _includes/, and styles in _sass/. The main content page is _pages/about.md, which also contains the publications list and work experience.

Key config files:

_config.yml — Site-wide settings: author info, social links, plugin list, collection definitions, permalink structure.
_data/navigation.yml — Controls which links appear in the masthead navigation. Currently only “CV” is active; Publications/Talks/Teaching are commented out.

Content collections (defined in _config.yml):

_publications/ — Individual publication markdown files
_talks/ — Talk files
_teaching/ — Teaching files
_portfolio/ — Portfolio items
_posts/ — Blog posts

Note: the _pages/about.md page maintains its own publications list as HTML, separate from the _publications/ collection. The collections exist from the template but are not currently linked in navigation.

Custom styling:

_sass/_custom.scss — All site-specific CSS overrides (layout widths, typography, component classes like .publications, .work-experience, .academic-service). Edit here for theme customization.
_pages/about.md includes inline <style> for page-specific overrides.

Markdown generators (markdown_generator/): Python scripts and Jupyter notebooks to batch-generate publication/talk markdown files from TSV data. Run publications.py or talks.py to regenerate collection files from publications.tsv / talks.tsv.

Content editing

Homepage (index.html): Edit directly as plain HTML. No Jekyll processing.
About/bio (_pages/about.md): Main content page with bio, publications list, work experience, and academic service. Uses HTML divs with CSS classes for structured sections.
Author sidebar: Controlled by author: block in _config.yml (name, bio, social links, avatar).
Navigation links: Edit _data/navigation.yml to add/remove/reorder header links.
CV link: Currently points to a Google Drive PDF in _data/navigation.yml.
Files (PDFs, etc.): Place in files/ directory; accessible at /files/filename.pdf.
Images: Place in images/; profile photo is images/profile.png.

Research papers (ground truth for website copy)

Papers are in papers/ (not committed to git). Do not claim things about these papers that are not stated here.

Recti-Q (IROS 2026) — Recti-Q.pdf

Full title: “Feature-Space Rectification for Out-of-Distribution-Robust Quantized Perception in Edge Robotics”
Identifies the “Quantization-Induced Robustness Gap”: PTQ models preserve clean-input accuracy but degrade significantly under distribution shift (sensor noise, weather, novel environments)
Solution: freeze the quantized backbone, train a small LoRA adapter on the classifier head using source data only. Less than 1% parameter overhead (as small as 6 KB)
Architecture-agnostic (CNNs and Transformers); evaluated on ImageNet-C and PACS
Authors: Hamidreza Yaghoubi, Parastoo Pilevar, Ming C. Lin

PolySona (IROS 2026, also under review at ICLR 2026) — PolySona.pdf

Full title: “Parameter-Efficient and Modular Latent Behavior Modeling for Traffic Simulation”
Mixture-of-Experts (LoRA) framework to extract latent driving styles in trajectory prediction models
Addresses rare/safety-critical scenarios where driving style diversity produces multi-modal trajectory outcomes
Enables counterfactual simulation and style-consistent predictions
NOT about temporal understanding in the VLM/video sense; NOT about real-time on-vehicle inference specifically
Authors: Laura Zheng, Hamidreza Yaghoubi, Tony Wu, Tianyi Zhou, Ming C. Lin

IROS 2025 — 2503.04994.pdf

Full title: “Quantifying and Modeling Driving Styles in Trajectory Forecasting”
Analyzes driving style quantification in real-world trajectory datasets; proposes style-based resampling to improve forecasting of rare/edge-case driving behaviors
NOT about temporal understanding or on-vehicle real-time operation
Authors: Laura Zheng, Hamidreza Yaghoubi Araghi, Tony Wu, Sandeep Thalapanane, Tianyi Zhou, Ming C. Lin

Decompose-and-Compose (CVPR 2024) — Decompose and Compose CVPR 2024 Paper.pdf

Addresses spurious correlation via compositional image manipulation
Uses class activation maps to separate causal from spurious components; creates counterfactual training data by combining image elements
Achieves better worst-group accuracy without group labels
Authors: Fahimeh Hosseini Noohdani, Parsa Hosseini, Aryan Yazdan Parast, Hamidreza Yaghoubi Araghi, Mahdieh Soleymani Baghshah

Annotation-Free Group Robustness (OOD-CV @ ICCV 2023) — 2312.04893.pdf

Proposes LFR (Loss-Based Feature Re-Weighting): infers data groupings from ERM model losses, resamples to create a balanced set, retrains only the last layer
No group annotations required; outperforms prior annotation-free methods, especially under high spuriosity
Authors: Mahdi Ghaznavi, Hesam Asadollahzadeh, Hamidreza Yaghoubi Araghi, Fahimeh Hosseini Noohdani, Mohammad Hossein Rohban, Mahdieh Soleymani Baghshah