The paper, titled Beyond Corrected Actions: Training LLM Agents on Textual Critiques, addresses the growing use of large language models in multi-turn agentic tasks. These tasks involve extended sequences of interactions with users and environments, where agents must make decisions over time. The authors note that a common approach to improving such agents is to use more—though the abstract cuts off before specifying what that common approach entails.
The core proposal, as implied by the title and the opening lines, is to train agents on textual critiques rather than relying solely on corrected actions. This shift suggests that feedback in the form of natural-language critiques may provide richer or more generalizable guidance for agent behavior than simple action corrections.
Because the abstract is truncated, the full methodology and experimental results are not available from the source. However, the central idea is clear: moving beyond action-level corrections to incorporate textual critiques as a training signal for LLM agents. The paper appears to target a key limitation in current agent training, though specific claims and evidence remain outside the provided text.