{"slug": "niodoo-hidden-state-steering", "title": "Niodoo hidden state steering", "summary": "Developer Jason Van Pham released Niodoo, a runtime that uses hidden state steering to improve small language models' performance without fine-tuning, enabling self-correction and memory systems. The tool, which runs on CUDA, aims to help models punch above their weight by injecting noise to break loops and using topological data analysis to monitor internal states. Pham seeks community collaboration to refine the code and push the research forward.", "body_md": "Hidden state steering, fork it , hate it, write papers about it. Here’s the code. The work goes deep, if u run the code you’ll see that yourself, last time I posted this it got called slop. So I spent months, red teaming my own claims, while building upon it. Can a model steer it self, while producing correction packets that assist it, yes. Can models punch above their weight, yes the potential is untapped. I’m trying to fill that gap, the messy middle. I saw the llama beat the strawberry prompt about 30% of the time, and this has been my goal to take that 30% to 0 and push the smaller model past its potential.\n\nAll this work came organically, over course of failures and my stubbornness to have models understand better, and create memory systems that aren’t cheap wrappers and prompt engineering. If the names sounds like marketing terms, its cause that’s how me or the AI described it. Lastly, hate the method that got me here(collaborating with AI), but I saw words as particles, I saw memories as multifaceted shapes, I had hardware that could only support smaller models, and I refused to think a trillion parameters helps a model understand you or your work better.\n\nThe overall goal is not to work against or compare with other models, as this is not a model, its my runtime. In a perfect world you can point Niodoo at your own runtime, and it can learn over time by observing whatever big model your using, whatever decisions the human makes, and whatever decisions it makes itself. A collaborator that doesn’t use cheap tricks like brittle hooks or markdowns. This is one of literal thousands of claims this runtime can produce, right now only works on cuda. Trying to clean it up responsibly and make it reproducible for anyone, the real challenge. No one has taught me anything, and I don’t go searching online for answers, that comes after my runtime hits it. That being said if messed up something please just email me [jasonvanpham@niodoo.com](mailto:jasonvanpham@niodoo.com). School me please, how I’m using git wrong, rust, or attributions I might missed, and I will kindly learn and correct. Whatever it takes to push this research further I’m about that.\n\nThese ones are a little bit messy internal monitor is something I’m still working on it uses tda analysis to measure when the model is looping.\n\nHere more towel drying, its messy, but as you see previously we can produce clean runtimes. Also showcases the models never been fine tuned, yet can accurately emit its own telemetry tags. Locking when the model acknowledges the right answer, spike to inject “physics forces” that perturbs its tokens during inference. In short injecting some type of noise to help break itself from loops, later on feed that back as learning signals. As I said it runs deeper, but I’m not gonna over claim something I can’t produce on a whim. And because I’m self taught and constantly doing things that makes the work harder for myself it seems, it’s going to take time before I surface all of it. Nothing here was produce or claimed without human oversight, my job is to make it surface that so it can be fed back to the community. In hopes that people can take the work and maybe solve a problem they are having in their own codebase.\n\nThis last photo cracks me up, “I’m back.”\n\nAnyways, I hope one day people can look past build with ai badge and see the actual months of work, failures, time, money, etc, you know all the human cost stuff and take the work for what it is. If you have a community that can look past that, and want to share ideas, frustrations, and breakthroughs I’d wish to be apart of that. This work is has always been to share, I’d love to write a paper that’s written just by me, but I’m not going to waist time writing papers I got runtime to fix. Bloated code that needs to be shrunk and some embarrassing legacy code that needs to be hidden.", "url": "https://wpnews.pro/news/niodoo-hidden-state-steering", "canonical_source": "https://discuss.huggingface.co/t/niodoo-hidden-state-steering/177135#post_1", "published_at": "2026-06-24 20:50:51+00:00", "updated_at": "2026-06-24 21:22:37.172267+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-research", "ai-tools", "machine-learning"], "entities": ["Jason Van Pham", "Niodoo", "CUDA", "Llama"], "alternates": {"html": "https://wpnews.pro/news/niodoo-hidden-state-steering", "markdown": "https://wpnews.pro/news/niodoo-hidden-state-steering.md", "text": "https://wpnews.pro/news/niodoo-hidden-state-steering.txt", "jsonld": "https://wpnews.pro/news/niodoo-hidden-state-steering.jsonld"}}