SPHERE Project Page

Real Fig1 (seed=2)

This page can load real Fig1 geometry data exported from: code/data/visual_sphere/outputs_fig1_real_series_seed2.

Generate website JSON:

python scripts/build_real_fig1_dataset.py \
  --src-dir /home/leadtek/Downloads/projects/SPHERE/code/data/visual_sphere/outputs_fig1_real_series_seed2 \
  --out-dir static/real_fig1/seed2 \
  --seed 2

Run locally (required for fetch()):

python -m http.server 8000 --directory .

Local preview

This repository is the static project page for the ICML 2026 accepted paper SPHERE: Mitigating the Loss of Spectral Plasticity in Mixture-of-Experts for Deep Reinforcement Learning.

Public site: https://sphere-rl.github.io/

The page is currently optimized for desktop / laptop reading, matching the expected paper-project-page audience. Mobile support should not drive layout decisions unless that requirement changes.

Publication boundary

The project page now exposes the public paper surfaces directly: arXiv https://arxiv.org/abs/2605.04712, code https://github.com/sphere-rl/sphere, Google Slides, and the tracked ICML 2026 poster PDF at static/sphere-icml2026-poster.pdf.

If arXiv replacement processing is in progress, the public arXiv source/PDF may temporarily lag behind the submitted replacement even though the project page link is already stable.

Do not commit unfinished camera-ready PDFs or local preprint drafts under this static site root; otherwise static hosting can expose them even without a visible link. Keep unpublished drafts outside the repo (for example ../local-unpublished/) until they are ready to link.