Tom Zehle

Tom Zehle

PhD Candidate researching Optimization of Agentic Systems

Download PDF

My research enables agentic systems to learn through prompt optimization instead of weight updates. My optimizers significantly outperform the strongest baselines in both performance and cost. I seek a research internship opportunity in Agentic AI bringing my research to frontier models.

Research experience

PhD Candidate · University of Freiburg & ELLIS

Oct 2025 – present

Supervisor: Prof. Frank Hutter · Expected end date: Late 2027

  • Developed methods for automated optimization of Multi-Agent Systems (MAS) prompts and topologies using (multi-objective) system-level reward
  • Conceived and developed CANTANTE as sole author: contrastive credit attribution improved accuracy by 12–19 pp over the strongest evaluated baselines on coding and mathematical reasoning with fewer inference tokens
  • Contributed to MO-CAPO, which handles an arbitrary number of objectives (e.g. performance, cost, wall time) while exploring up to 5.7× as many candidate prompts as an NSGA-II baseline within the same token budget
  • Core contributor to the promptolution open-source framework
  • Supervising students in projects on MAS Optimization
  • Organizing seminars and lecturing at the University of Freiburg
  • Task leader in EU-Project ELLIOT: coordinating research direction, collaborations and deliverables; communicating with the European Commission
  • Reviewer for NeurIPS (2026) and co-reviewer for Nature (2025)

Industry experience

Data Scientist (part-time) · Airbus

Munich · 2023 – 2025

  • Developed a fine-tuning algorithm for similarity search models deployed in aircraft maintenance, and implemented evaluation and backend logic
  • Built a prompt optimization proof of concept and an AI-assisted export-classification pipeline for sensitive documents

Dual Student · Airbus

Munich/Ingolstadt/Madrid · 2020 – 2023

  • Worked on data science projects in six internships at various departments
  • Built Transformer-Tabular-Hybrid to predict flight arrival times
  • Developed a computer vision algorithm for ExoMars Rover project
  • Developed circuit simulations for electric aircraft, including superconducting components

Education

M.Sc. Statistics and Data Science · LMU Munich

2023 – 2025

Machine Learning Track · Grade 1.1, Thesis 1.0 (1.0 is the best possible grade)

  • Developed CAPO, combining evolutionary prompt optimization with racing: over 20 pp higher accuracy than the strongest evaluated baseline on mathematical reasoning and 44% lower optimization cost on average
  • Thesis (CALIOPE): studied positional biases in long-context LLMs and developed training-free inference-time calibration
  • Initiated the open-source framework promptolution

B.Sc. Business Informatics · DHBW Ravensburg

2020 – 2023

Dual programme with Airbus · Grade 1.4, Thesis 1.0

  • Thesis: fine-tuning similarity search for aircraft incident reports
  • Project: Neural Swabian dialect translator

Publications

  • AutoML 2025 · Methods Track YouTube GitHub Paper
    CAPO: Cost-Aware Prompt Optimization Tom Zehle, Moritz Schlager, Timo Heiß, Matthias Feurer. CAPO jointly optimizes instructions and few-shot examples via evolutionary search and racing under limited budgets.
  • EACL 2026 · System Demonstrations YouTube GitHub Paper
    promptolution: A Unified, Modular Framework for Prompt Optimization Tom Zehle, Timo Heiß, Moritz Schlager, Matthias Aßenmacher, Matthias Feurer. promptolution unifies prompt optimizers, tasks, and LLM backends through a modular interface.
  • EACL 2026 · Findings YouTube GitHub Paper
    Can Calibration of Positional Encodings Enhance Long Context Utilization? Tom Zehle, Matthias Aßenmacher. CALIOPE improves long-context retrieval and cross-chunk reasoning by calibrating positional encodings at inference time.
  • Under review · 2026 GitHub Paper
    CANTANTE: Optimizing Agentic Systems via Contrastive Credit Attribution Tom Zehle. CANTANTE introduces contrastive credit attribution to optimize individual agents from system-level rewards.
  • Under review · 2026 GitHub Paper
    MO-CAPO: Multi-Objective Cost-Aware Prompt Optimization Jan Büssing, Moritz Schlager, Timo Heiß, Tom Zehle, Matthias Feurer. MO-CAPO extends CAPO to an arbitrary number of objectives, e.g. performance, cost and wall time.
  • IJCAI-ECAI 2026 · AutoAI-FM Workshop GitHub Paper
    Towards Benchmarking Agentic Data Scientists Kartik Nayak, Nitishkumar Solpure, Niladri Mitra, Tom Zehle, Alexander Pfefferle, Omar Swelam, Frank Hutter. Evaluates LLM-based data science agents against non-LLM AutoML baselines.
  • Under review · 2026
    Design and Tuning of Multi-Agent Systems Using Bayesian Evolutionary Optimization Dominik Jehle, Tom Zehle. Uses Bayesian evolutionary optimization to search over how agents are connected in Multi-Agent Systems.
  • Under review · 2026
    From Supergraphs to Small Teams: Contrastive Attribution for Multi-Agent Topology Search Fabian Tritthart, Tom Zehle. Combines neural architecture search with contrastive credit attribution to optimize the topology of Multi-Agent Systems.