Choosing LLMs for Strategic Agent Interactions

arXiv · Arastun Mammadli, Herbert Woisetschläger, Jawad Fayaz, Shiqiang Wang · 2026-10-13

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Against fixed opponents in five games, a simple bandit beat the tested reflection and embedding selectors on net gain, but still lost to the best fixed model.

What it does

Using ten open models and five games, this study asks whether a router can learn which model to field against an opponent. The simple bandit reported higher net gain than the tested embedding and reflection routers, yet still trailed the best fixed model. Selector trials used one seed per matchup; broader replication and free-form negotiation remain open.

Why it matters

The authors argue that ordinary quality/cost routing misses the effect of an opponent's behavior. Their experiments expose a routing problem, not a proof that adaptive selection beats a good fixed-model baseline.

How it applies

If your agents face changing opponents, compare your existing selector, a stationary bandit and a fixed model chosen on separate evaluation data. Use the same payoff/cost measure: the paper calls that trade-off net gain. This is a proposed evaluation, not production validation.

The short version

Make your router beat a fixed-model baseline first.

My take

My takeaway is to earn the extra routing machinery with measured gains. I'd compare the existing router with a stationary bandit and a fixed model before choosing one; this study doesn't establish a universal winner for deployed agents.

How it connects

• For a proposed agent evaluation, treat the opponent's behavior as part of the task, rather than assuming ordinary question-answer benchmarks capture it.
• Keep the simplified game setup and limited selector replication separate from production safety or negotiation claims.

Bottom line

For a strategic-agent experiment, record payoffs and compare selectors with a fixed-model baseline before adopting extra routing complexity.

Takeaways

• Strategic competence is highly dependent on the specific game setup and opponent.
• The study finds that reliable strategic evaluation needs game-specific replication; it does not establish a universal seed count.
• The tested embedding and reflection selectors did not consistently capture strategic signals in these game experiments.
• The stationary bandit beat the tested embedding and reflection routers on reported net gain, but still lost to the best fixed model. Selector trials used one seed per matchup.
• In the Cournot Duopoly adaptive-opponent experiment, the tested non-stationary and adversarial variants did not beat the stationary baseline.

Topics: LLM Routing, Game Theory, Strategic AI, Multi-Armed Bandits