The Big Picture
Selective, per-agent withdrawal of centralized guidance during training produces decentralized policies that perform much better at execution, especially when agents have limited or noisy views or teams change.
ON THIS PAGE
Core Insights
TANDEM reads a single centralized critic two ways — a full global view and a per-agent masked view — and uses a simple, parameter-free gate to pick which read to trust for each agent at each training step. Where the two reads agree on whether an action helped or hurt, the per-agent agreement signal the agent trains on the view it will actually have at execution; where they disagree, the centralized read centralized critic corrects the agent. Evaluated across three domains (partial observability, observation noise, and changing teams/tasks), TANDEM produced the best execution performance on most setups and was especially strong where the gap between global and local information was largest.
Test your agentsValidate against real scenarios
By the Numbers
1Top performer on 8 of 9 test setups and never significantly worse than any of six standard centralized-training baselines.
2Significant wins on 5 of 9 setups — e.g., discovery tasks: 6.91 vs next-best 5.48 (larger map) and 3.70 vs 2.63 (smaller map).
3Explained-variance gap between the critic's global and local reads reached 0.39–0.41 on discovery; sampling with observation noise 0.2 had execution return 37.79 vs MAPPO's 31.68.
Why It Matters
Engineers building multi-agent systems that must run without a central overseer — for example, fleets of robots, distributed sensors, or firefighting agents — because TANDEM yields policies that are robust when central information disappears. ML researchers and team leads evaluating multi-agent trust and reliability can use TANDEM's per-agent agreement signal as a practical diagnostic of when agents are ready to operate independently. central overseer
Key Figures

Fig 2: Figure 2. The three domains, left to right: discovery, sampling, and wildfire suppression.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreLimitations
TANDEM focuses on differences between the global state and each agent's observable state and does not directly address changes in an agent's own capabilities or goals. The gate only reacts when the two reads disagree in sign, so it cannot correct situations where both reads agree but are jointly wrong. Experiments isolated single forms of the information gap (partial observability, noise, openness) rather than combining all of them or testing large-scale benchmarks, so further validation is needed for more complex production settings. The idea of coordinating such patterns relies on orchestration concepts like the Orchestration.
Full Analysis
TANDEM narrows the mismatch between training and deployment when agents are trained with a centralized critic but must act on local observations at execution. Instead of two separate critics, TANDEM uses one centralized critic and reads it two ways each training step: a global read that sees full state and a masked read that only sees what an agent will observe at execution. A parameter-free gate compares the two per-agent advantage signs and mixes them: train on the masked (execution) advantage when both reads agree, otherwise fall back to the centralized read to correct misleading local views. Both reads are regressed toward the same centralized return target so the masked read steadily learns to match the global read where possible.
The method was implemented on top of a standard policy-gradient trainer and tested across three domains selected to isolate different causes of the information gap: discovery (partial observability), sampling (observation noise), and wildfire suppression (agents/tasks entering and leaving). With identical training budgets and tuned baselines, TANDEM achieved the best execution return on 8 of 9 setups and large gains where the critic's global and local readings differed most. Ablations show the gate is particularly important under team and task changes, while the shared regression target helps the local read converge toward useful global information. The result is a practical, low-complexity training tweak that gives teams of agents a clearer path to reliable autonomous behavior, plus a per-agent diagnostic (agreement between reads) to track readiness for deployment.
Not sure where to start?Get personalized recommendations
Credibility Assessment:
Authors have low reported h-index (one author h-index=3) and no listed institutional affiliations; arXiv preprint with no citations — limited evidence of established reputation.