Research map / Large language models / AI and society
Computational Arbitrage in AI Model Markets
Ricardo Olmedo, Bernhard Schölkopf, Moritz Hardt
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Cascade GPT-5 mini (up to $0.08 per issue) then DeepSeek v3.2 instead of either alone | Arbitrageur · SWE-bench Verified | cost to reach 75% solve rate / profit margin: $150 and $120 -> $80; margin up to 40% | Improved |
| Add a second arbitrageur competing on price | AI model market · SWE-bench Verified | arbitrage profit margin | Hurt |
| Introduce an arbitrageur into the market | Model providers' revenue · SWE-bench Verified | marginal revenue: decreases by up to 40% | Hurt |
| Introduce an arbitrageur into the market | Market segmentation · SWE-bench Verified | revenue spread across providers | Improved |
| Increase arbitrage fit search budget to $10 (ablation) | Arbitrage policy fitting · SWE-bench Verified | profit margin | Improved |
| Fit arbitrage policy on Django issues, evaluate on non-Django (and vice versa) (ablation) | Arbitrage policy · SWE-bench Verified Django split | profit margin | No clear effect |
| Add four more providers (Qwen Coder 30B/480B, Claude Sonnet 4.5, mini-coder 4B) to the market (ablation) | Arbitrage opportunities · SWE-bench Verified | max profit margin: 45% -> 58% | Improved |
| Increase distillation tokens from 70M to 5.4B for Qwen 3 1.7B (ablation) | Arbitrage profitability with Qwen Coder 480B · SWE-bench Verified, 61% to 71% | profit margin: nearly 30% at 5.4B tokens | Improved |
| Train on all trajectories including incorrect ones (only 20% correct) (ablation) | Distillation SFT data · SWE-bench Verified | pass@k | Improved |
| Enter mini-coder 4B into a GPT-5 mini + Qwen Coder 30B market | Qwen Coder 30B revenue · SWE-bench Verified | revenue share | Hurt |
Closest papers by meaning
- Pricing and Competition for Generative AI (NeurIPS 2024 · similarity 0.87)
- Conformal Arbitrage: Risk-Controlled Balancing of Competing Objectives in Language Models (NeurIPS 2025 · similarity 0.87)
- MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model (ICLR 2025 · similarity 0.84)
- Impacts of Aggregation on Model Diversity and Consumer Utility (NeurIPS 2026 · similarity 0.83)
- R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.83)
- Matching Markets Meet LLMs: Algorithmic Reasoning with Ranked Preferences (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
