{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreif77luttzara7nfxdsuqtophc5mhxlpwjjg54nsmlxdoz3dkg3b2y",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mizmzvjgj4t2"
  },
  "path": "/t/seeking-arxiv-endorsement-cs-lg-for-new-rl-optimizer-hopper/175091#post_1",
  "publishedAt": "2026-04-09T00:21:40.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "Hopper — partial orthogonalization changes early reasoning behavior in RL",
    "numerical divergence in hybrid models."
  ],
  "textContent": "Hi everyone,\n\nI am preparing to submit an independent research paper to arXiv under the `cs.LG` (Machine Learning) category, and I am looking for an endorser to help clear the system’s submission threshold.\n\nThe paper formalizes “Hopper,” a variant of the Muon optimizer that I adapted specifically for RL fine-tuning pipelines like GRPO. I recently shared some of my empirical findings on the forum—specifically how reducing to `ns_steps=1` creates a “lazy orthogonality” that accelerates early reasoning discovery—which you can see here: Hopper — partial orthogonalization changes early reasoning behavior in RL.\n\nYou might also recognize me from my technical deep-dive earlier this month on numerical divergence in hybrid models.\n\nIf anyone here is an active arXiv author with `cs.LG` endorsement privileges and would be willing to take a quick look at my draft to endorse it, please let me know! I am more than happy to share the full PDF and the open-source training scripts via DM.\n\nThanks so much, Jen",
  "title": "Seeking arXiv endorsement (cs.LG) for new RL optimizer (Hopper)"
}