ToolGrad: Efficient tool-use dataset generation with textual "gradients"

Details

Training LLMs to use tools effectively requires large annotated datasets of (query, tool-call chain) pairs. Prior approaches like ToolBench and ToolACE start with a query and use a depth-first search agent to find a valid solution — an inherently expensive process with low pass rates, since most search paths fail. ToolGrad, presented at ACL 2026 by Google XR researchers Zhongyi Zhou and Ruofei Du, inverts this order: it first constructs a valid ground-truth tool-use chain and then generates the corresponding user prompt. Because a concrete tool sequence is more unambiguous than a natural-language query, this annotation step requires only a single LLM call rather than a full agent exploration loop — directly addressing the inefficiency. The practical results are a higher pass rate, lower generation cost, and more complex long-horizon chains than DFS-based methods. Models fine-tuned on ToolGrad data outperform those trained on baseline-generated datasets and, notably, match state-of-the-art proprietary LLMs on out-of-distribution evaluation sets featuring tools the model never encountered during training — a strong generalization signal. The framework incorporates 'textual gradients' (borrowed from TextGrad) where an LLM critic provides plain-text feedback to iteratively refine generated chains, analogous to how numerical gradients update weights in standard ML. The source content is somewhat truncated, so full benchmark details and code availability are not confirmed here.