Comment by pimeys
1 day ago
All of these flash models have this. You have to build your harness so that it deals with it. Infinite loops are solved by having an error message that says what to do differently on failure, invalid tool calls are solved by making the tool schema less strict and detect things in the runtime etc.
Hallucinations you can't fix. Gemini is a bit worse there than DeepSeek, but there's not much research on how to fix that. The only one is the CaMeL paper by Google, where you tag every prompt and result and then for every assistant response or tool call you first check where it got that data and error if you notice fabrication. This one is really annoying to implement.
With larger models the fabrication starts when the context grows or if you have too many tools, for flash models it's much earlier. We use the flash models for repetitive agentic tasks, where the prompt defines clearly what to do and how. The whole run is about 4-5 steps typically, and context size stays in the comfort zone.
Can you share what tools and processes you're using to do this?
I've been using Pi to build custom extensions and wrapping workflows in shell processes to make it more deterministic and enforce certain validations, all guided by Fable. This isn't production work, though, just playing llm factorio at home.
What you want is a bunch of sessions to replay. Something anonymized if it's not yours, and something that's not depending on state.
You replay all your sessions against your harness, and then store all logs all output, everything to a safe place.
Finally use a blind judge to check everything, and score the output.
Then fix your harness, iterate again until better until you are in a point where it's just the model's weakness. If you get to that, use a bigger model.