At a Glance
Representing which agent produced each action (novice vs expert) makes models far better at recovering from mistakes they never see experts make, even when experts never enter those bad states.
ON THIS PAGE
What They Found
Training on interaction traces (novice mistakes corrected by experts) exposes parts of the problem space that experts alone never visit, letting models learn how to recover. When datasets explicitly label which agent generated each action, models gain an additional boost and more reliably infer the expert’s intent. However, interaction traces without agent labels can fail to teach expert-like avoidance behavior in some hazardous situations. explicit agent indicators amplify the benefit by helping models infer who acted when.
Key Data
1Each evaluation used three held-out test sets (safe, hazardous, recovery) with 320 examples each to compare models’ behavior.
2Training/data setup used 10 different environments on 20×20 grids with rewards +100 for goal, −20 for high-cost cells, and −1 otherwise; interventions triggered after 2 consecutive high-cost steps.
3Control and interaction dataset sizes cited include 500 expert-only traces paired with ~99,500 novice/interaction traces (variants also matched token counts at ~93,500 interaction traces).
What This Means
Engineers building multi-agent systems or AI tutors should care because marking who acted in a transcript can improve an agent’s ability to correct others and handle edge cases. Technical leaders curating training data can use these insights to decide when to include interaction logs with clear agent labels to increase robustness. Researchers studying agent trust or agent-to-agent evaluation can adopt the representation tricks to probe learning from corrective dialogues. multi-agent systems agent-to-agent evaluation
Need expert guidance?We can help implement this
Key Figures
Fig 1: Figure 1: Example traces under expert-only and interaction policies. Grids show the agent’s trajectory at 4 different time steps. Sequences are the representations of the full trajectory seen by the model, with the underlines indicating the part corresponding to the sub-trajectories at each timestep.
Fig 2: Figure A1: Overview of datasets with source indicators. A: example of a no-source trace and the corresponding with-source trace. B: depiction of an example dataset containing expert-only data and interaction data.
Fig 3: Figure A2: Performance on hazardous trials with varying frequencies of agent indicator tokens, under both prediction elicitation strategies.
Fig 4: Figure A3: Performance on hazardous trials for models trained on 3 types of datasets: control datasets (500 expert-only traces, 99,500 novice-only traces), token-matched interaction datasets (500 expert-only traces, 93,500 interaction traces; total number of tokens is matched with control datasets), and the original interaction datasets (500 expert-only traces, 99,500 interaction traces).
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Findings come from a synthetic grid planning task, so effects may differ on natural language or more complex environments. The intervention rule and novice error model were simplified; real-world mistakes and corrections can be more varied. Interaction data without agent identity can degrade learning of expert avoidance behavior in some scenarios, so adding labels is necessary but not sufficient for all settings. Inter-Agent Miscommunication
Methodology & More
A controlled study used a simple spatial planning task (20×20 grid) to test how learning from observed interactions compares to learning from expert-only traces. Two data-generating regimes were used: an expert-only policy that always avoids high-cost cells, and an interaction policy where a novice follows an incorrect objective (ignoring high-cost cells) and an expert intervenes once the novice spends two consecutive steps in costly cells. Trajectories were encoded as sequences (start, goal, states and actions) and transformers were trained on datasets that either contained only expert traces, interaction traces, or interaction traces augmented with tokens marking which agent produced each step. Models trained on interaction traces were better at recovery tasks—producing valid or optimal paths when starting from high-cost states that experts never visit—where expert-only models failed. Interaction-trained models without agent labels struggled to learn expert-like avoidance of hazardous paths, showing that interaction alone can be ambiguous. Adding explicit agent indicators (labels that show who acted when) amplified the benefit: models used the correction structure to infer expert intent and generalize to unseen recovery situations. Practical takeaway: when collecting or curating interaction data for training, include clear signals about agent expertise to capture corrective information that improves robustness and recovery behavior. Agent Service Mesh Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
No notable affiliations, very low h-indices, arXiv only and no citations — limited identifiable credibility.