Human-Centric Machine Learning

Max Planck Institute for Software Systems

📚 Featured Research

Test-Time Compute Games Strategic behavior in LLM-as-a-service markets.

• 📄 Read the Paper
• 💻 View Code (GitHub)
• 💾 Download Dataset — Contains full inference traces: sampled answers, token counts, correctness labels, and Best-of-N scores.

Tokenization Multiplicity Leads to Arbitrary Price Variation in LLM-as-a-service Solving tokenization multiplicity with a simple and efficient sampling method.

• 📄 Read the Paper
• 💻 View Code (GitHub)
• 💾 Download Dataset