What they did

The authors created TRACE, a controlled diagnostic environment covering 12 root causes and fine-grained affected-segment attribution. Each episode begins with a hidden intervention whose simulated consequences appear as noisy, confounded and distributed evidence; agents use Python and SQL to investigate before naming the cause and, where applicable, the affected segment.

They trained Qwen3.5-35B-A3B first with supervised fine-tuning and then reinforcement learning. Because the simulator records the injected intervention, it can provide an objective reward without requiring expert adjudication of each diagnosis.

Key findings

• The unmodified Qwen3.5-35B-A3B scored 0.159 on FullAttr@1, which measures whether the top answer fully identifies the required cause and attribution.
• Supervised fine-tuning increased its score to 0.637; subsequent reinforcement learning with synthesized rewards raised it further to 0.757.
• The best prompted baseline, Claude Opus 5, scored 0.686 on the 235-episode held-out test set. The RL-trained 35B model also surpassed a prompted Qwen3.5-122B-A10B model.
• The trained policy used substantially fewer tool calls than the prompted 35B base model, although the abstract does not report the exact reduction.

Why it matters

The results suggest that diagnostic tasks without cheap real-world verification can still support reinforcement learning when a simulator supplies known interventions and objective labels. Within TRACE, access to this training signal mattered more than simply prompting a larger model, pointing to simulation as a practical route for training causal-investigation agents.

Caveats

TRACE is a controlled simulation in one domain, with 12 predefined causes and a 235-episode test set. The reported results therefore do not establish that the learned policy will transfer to real advertising systems, unfamiliar causes or other diagnostic domains, where simulators may omit important dynamics and the true cause may remain genuinely uncertain.