LLM Agents Interview Questions #6 - The AST Explosion Trap

Without an intermediate abstraction layer, your evolutionary loop is sampling an intractable syntax tree universe with no structural bias.

You’re in a Senior AI Engineer interview at Google DeepMind. The interviewer sets a trap:

You’re feeding 4K astronomical scatter plots into a Vision-Language Model (VLM) to run symbolic regression, but the loss curve is completely flat. The VLM fundamentally cannot map raw pixels directly to algebraic operators. What is the missing layer in your pipeline?

90% of candidates walk right into it.

---

Most candidates say, “We need to fine-tune the VLM on paired image-to-equation datasets.”

Or worse: “Extract the raw pixel coordinates using an OpenCV script to create a CSV, then pass that to the evolutionary algorithm.”

They think this is a simple data translation problem.

---

But they aren’t optimizing for coordinate translation, they’re optimizing for search space pruning.