Optimization Variable Role Description

A short role string on each optimizable variable steers what textual-gradient descent is allowed to rewrite — answer-only vs reasoning-plus-answer, and more.

In TextGrad, each graph variable carries a role description — free-form text naming what the variable is for (e.g. “system prompt reused across queries”, “final numerical answer”, “reasoning + final prediction”). Empirically this is a primary steering channel: the same textual gradient descent step updates a number-only variable to a number, but keeps chain-of-thought text when the role says reasoning is in scope.

Treat role strings as harness config next to system prompt architecture and instance vs prompt optimization. Wrong roles silently change the optimization surface (stripping useful reasoning, or letting the optimizer rewrite frozen fields). Combine with requires_grad-style flags so predecessors that must not change never receive textual gradients.

---

Sources

• TextGrad (Yuksekgonul et al.), ch. 6