When Reasoning Models Hurt Behavioral Simulation: A Solver-Sampler Mismatch in Multi-Agent LLM Negotiation
📰 ArXiv cs.AI
Stronger reasoning in LLMs can hurt behavioral simulation in multi-agent negotiation, learn how to identify and address solver-sampler mismatch
Action Steps
- Identify the objective of your simulation: is it to solve a strategic problem or to sample plausible boundedly rational behavior?
- Assess the potential for solver-sampler mismatch in your LLM negotiation model
- Evaluate the trade-off between reasoning strength and simulation fidelity in your model
- Consider using techniques such as regularization or noise injection to mitigate over-optimization
- Test and validate your model's performance in simulations with varying levels of complexity and uncertainty
Who Needs to Know This
Researchers and developers working on multi-agent systems and LLM negotiation can benefit from understanding the limitations of reasoning models in behavioral simulation, to improve the fidelity of their simulations
Key Insight
💡 Reasoning-enhanced models can become better solvers and worse simulators, highlighting the need for a nuanced approach to modeling boundedly rational behavior
Share This
🚨 Stronger reasoning in LLMs can hurt behavioral simulation in multi-agent negotiation! 🤖 Learn how to identify and address solver-sampler mismatch #LLMs #MultiAgentSystems
Key Takeaways
Stronger reasoning in LLMs can hurt behavioral simulation in multi-agent negotiation, learn how to identify and address solver-sampler mismatch
Full Article
Title: When Reasoning Models Hurt Behavioral Simulation: A Solver-Sampler Mismatch in Multi-Agent LLM Negotiation
Abstract:
arXiv:2604.11840v1 Announce Type: cross Abstract: Large language models are increasingly used as agents in social, economic, and policy simulations. A common assumption is that stronger reasoning should improve simulation fidelity. We argue that this assumption can fail when the objective is not to solve a strategic problem, but to sample plausible boundedly rational behavior. In such settings, reasoning-enhanced models can become better solvers and worse simulators: they can over-optimize for s
Abstract:
arXiv:2604.11840v1 Announce Type: cross Abstract: Large language models are increasingly used as agents in social, economic, and policy simulations. A common assumption is that stronger reasoning should improve simulation fidelity. We argue that this assumption can fail when the objective is not to solve a strategic problem, but to sample plausible boundedly rational behavior. In such settings, reasoning-enhanced models can become better solvers and worse simulators: they can over-optimize for s
DeepCamp AI