Textual Gradients

Natural-language criticisms of a variable’s role in a computation graph, produced by an LLM as the analogue of a numerical gradient.

In TextGrad, a compound AI system is a computation graph whose nodes are variables (prompts, predictions, code, plans) and whose edges are arbitrary transforms — LLM calls, simulators, interpreters — that need not be differentiable. A textual gradient on a variable is interpretable natural-language feedback describing how to change that variable to improve downstream evaluation, aggregated from every successor context where the variable was used (chain-rule analogue through black-box ops). This is the primitive behind compound AI system optimization.

Unlike white-box numerical prompt gradients that need model parameters (unusable for closed APIs), textual gradients work through black-box LLM calls and simulators. Prompt-only precursors (e.g. ProTeGi-style NL feedback on task mistakes) become full autodiff-style propagation across stacks — including instance vs prompt optimization, not only instruction rewrite. DSPy-style program optimizers are complementary: they structure layers and often bootstrap demos; TextGrad-style frameworks center backward NL feedback as the update rule.

The operator is domain-independent once implemented: given conversation context {x|y} and criticisms on {y}, an LLM explains how to improve {x}. Use textual gradients when you want automated search over unstructured artifacts (system prompts, solutions, tool schemas) without hand-tuning each component — then apply textual gradient descent to produce the update. Pair with prompt optimization tooling and offline prompt evaluation so feedback quality is measured, not assumed.

---

Sources

• TextGrad (Yuksekgonul et al.), ch. 2; §4