VISTA framework optimizes prompts via multi-agent reflection

VISTA solves the "black-box" problem of reflective prompt optimization by making the internal reasoning of the optimizer explicit and searchable.

• –Decoupling hypothesis generation from rewriting allows for semantically labeled optimization traces that humans can actually audit.
• –The two-layer explore-exploit mechanism (random restarts + epsilon-greedy) effectively prevents the optimizer from getting stuck in local optima that plague simpler methods like GEPA.
• –Parallel minibatch verification ensures that prompt improvements are statistically stable across different data subsets, reducing over-fitting to specific examples.
• –Empirical results on GSM8K show a massive recovery from 13.50% to 87.57% accuracy when starting with "defective" seed prompts, proving its robustness.
• –This represents a shift from "trial-and-error" prompting to a more principled, agentic engineering approach for complex reasoning workloads.