Symbio: A Self-Fine-Tuning AI Loop That Learns from Corrections
The core loop is deceptively simple: you run a task, the LLM produces a response, and then you either accept the output or supply a corrected version. Symbio captures that pair — the wrong reasoning and the right answer — and uses it as a training example in a lightweight fine-tuning pass. Over a session, the model starts nudging its behavior toward the kinds of corrections you give most often. For me, this has been the missing piece in a lot of agent workflows: the agent fails on something, I fix it, and the same failure happens again on the next run. Symbio closes that gap.
Here's roughly how I'm using it:
1. Set up a task with a clear evaluation. For example, ask the agent to write a Python script that parses a messy CSV. The model tries, fails on edge cases.
2. When it fails, paste the corrected approach (or mark the correct answer) into the feedback field.
3. Symbio bundles the original context, the failed response, and your correction into a training sample, then runs a short LoRA fine-tune (or full fine-tune depending on size).
4. Re-run the same task or a similar one. The model now has a concrete example of the mistake it made and the fix, which is far stickier than generic "follow instructions" prompting.
This is essentially online learning for LLMs. For a while I was skeptical — fine-tuning still feels heavy for something you'd do after every mistake. But with small base models and low-rank adapters, a single correction can be incorporated in a few minutes. You don't need a cluster; a single GPU with 24GB