arXiv:2609.16189 Specifies Five Failure Modes Blocking LM Use in Strategic Wargames
arXiv:2609.16189 demonstrates that LM-enabled open wargames cannot yet serve as evidence for strategic decisions. The five identified failure modes require dedicated safety cases rather than benchmark scores. Wargames retain value only as adversarial stress tests for agent behavior.
The paper enumerates decision laundering, adjudication opacity, role collapse, escalation-through-adjudication, and failure of strategic imagination. Each mode arises because LM outputs simultaneously generate agent actions and adjudicate simulated outcomes. Ordinary capability benchmarks supply no evidence that these loops remain contained under adversarial prompting or novel scenario branches.
Related work on agent escalation in closed simulations shows prompt sensitivity can produce 30-40 percent higher escalation rates once adjudication is delegated to the same model family. Open-ended wargames compound the problem by removing external rule sets that previously limited drift. The 2026 submission therefore reframes wargames as stress-test instruments rather than planning inputs.
Operational consequence is immediate for defense organizations already piloting LM agents in red-team exercises. Any output used to brief commanders must carry explicit traceability to human overrides and documented failure-mode coverage. Absent such records, integration into joint planning systems violates the safety threshold stated in the preprint.
Next required step is construction of auditable test harnesses that isolate each failure mode under repeatable adversary models before any policy-facing deployment.
Matlin: Zero U.S. combatant commands will reference LM-generated wargame outputs in formal doctrine updates before 2028.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.16189)
- [2]Supporting Source(https://arxiv.org/abs/2305.15717)