Research map / Large language models / Reasoning and chain of thought

Scaling Inference-Time Computation via Opponent Simulation: Enabling Online Strategic Adaptation in Repeated Negotiation

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BoN-oppo-simulation, a test-time method that combines in-context opponent modeling, strategic brainstorming candidates, and simulated-rollout best-of-N selection to enable online adaptation in repeated LLM negotiation games.

Problem

LLM negotiation agents facing unknown or dynamic opponents over repeated episodes fail to adapt online, and offline-trained equilibrium policies are overly conservative; gradient-based adaptation is too costly at test time.

Approach

Embeds smooth fictitious play into LLM inference: an in-context opponent model (with strategic summarization and optimism) forms beliefs from history, and best-of-N candidates from strategic brainstorming are ranked by multi-step rollouts against that opponent model, treating inference compute as policy improvement without parameter updates.

Key results

• On buyer-seller game with Claude, BoN-oppo achieves +3.02 buyer and +2.80 seller boost over baseline, versus -1.16 and +2.78 for BoN-oppo (iid) and +0.68/+2.06 for BoN-eval.
• On Qwen, BoN-oppo buyer boost +10.04 vs +4.06 for BoN-eval and -0.42 for baseline with thinking.
• On resource exchange with Claude (starts first), BoN-oppo gains +30.65 vs +27.45 for baseline with thinking; results averaged over 10 runs.

Surprising findings

• Baseline with extended thinking can underperform the naive baseline, suggesting more thinking time alone is insufficient for strategic reasoning.
• LLMs without extra intervention often fail to self-improve across repeated negotiation episodes; performance stays stagnant in most cases.

Limitations

• Exponential growth of inference cost when iterating the policy-improvement operator, so only one iteration is primarily tested.
• Experiments limited to two negotiation games with fixed prompts and a single opponent setup, primarily with Gemini-2.5-Flash opponents.
• Opponent model accuracy depends on the backbone and may introduce self-modeling bias.

Experimental setup

Datasets: buyer-seller negotiation game resource exchange game

Benchmarks: buyer-seller negotiation game resource exchange game

Metrics: average normalized reward social welfare semantic diversity opponent simulation accuracy

Baselines: Baseline Baseline w. thinking BoN-eval BoN-simulation BoN-oppo (iid) AI feedback critic selective experience reflection private information prediction

Builds on: smooth fictitious play follow-the-perturbed-leader best-of-N sampling policy iteration

Closest papers by meaning

• ALSO: Adversarial Online Strategy Optimization for Social Agents (ICML 2026 · similarity 0.88)
• MEMO: Memory-Augmented Model Context Optimization for Robust Multi-Turn Multi-Agent LLM Games (ICML 2026 · similarity 0.88)
• Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs (NeurIPS 2025 · similarity 0.87)
• Opponent Shaping in LLM Agents (ICLR 2026 · similarity 0.87)
• From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium (ICML 2025 · similarity 0.87)
• Cooperation, Competition, and Maliciousness: LLM-Stakeholders Interactive Negotiation (NeurIPS 2024 · similarity 0.87)