What they did

The authors built DoCtOR, short for Diagnose-then-Correct PPO-enhanced Reflection. The framework automatically identifies the decisive error step and responsible agent, uses counterfactual reasoning to propose a corrected step, and limits reflection to that agent rather than updating every participant from the failed trajectory.

They evaluated the approach on HotPotQA for multi-hop question answering, ChartQAPro for chart reasoning, and Mind2Web for web interaction. Comparisons included Reflexion, Retroformer, and COPPER.

Key findings

• DoCtOR improved on initial success rates by 22% on HotPotQA, 26% on ChartQAPro, and 27% on Mind2Web.
• It outperformed Reflexion, Retroformer, and COPPER across the reported experiments.
• The authors report that the broader diagnose-then-correct approach generalized beyond a single experimental configuration.
• In low-resource settings, reflecting only on steps after the decisive error produced results comparable to reflecting on the complete failed trajectory.

Why it matters

The work challenges the assumption that every agent should learn from every collective failure. Assigning feedback only after locating the consequential mistake could make multi-agent improvement more precise and reduce the risk that correctly behaving agents store misleading lessons.

Caveats

The reported evidence covers three benchmarks, so it remains unclear how reliably automated attribution works in longer, more open-ended collaborations where responsibility is distributed across agents rather than traceable to one decisive step. The supplied results also do not quantify attribution accuracy, computational overhead, or margins over individual baselines.