What's this about?
Training reliable long-horizon tool-calling agents using critique-aware supervision — training on critiques of trajectories rather than only on the trajectories themselves.
What is critique-aware supervision?
Standard training shows the model successful runs and says "do this." Critique-aware training adds the commentary: this step was wrong and here is why, this recovery worked, this tool call was unnecessary. It is the difference between watching someone work and having them explain their reasoning as they go, including the parts they got wrong. For long-horizon work it matters more than for short tasks, because a long trajectory has many decisions and only one final outcome — the outcome alone cannot tell the model which decision was the good one.
Why does this matter to you?
- Long-horizon tool-calling is exactly the shape of an automated scanning or remediation pipeline.
- It joins DART-SD and "what makes good agentic data" in this month's run of papers arguing the supervision signal, not the volume, is the bottleneck.
- Critiques are expensive to produce, which is the practical objection — the paper's contribution is making that scalable.