Student Thesis: Finding gaps in expert reasoning

The Bundesanstalt für Materialforschung und -prüfung (BAM) is a senior scientific and technical federal institute with responsibility to the Federal Ministry for Economic Affairs and Energy. It tests, researches, and advises to protect people, the environment and material goods.

The division of Transport Container Safety is concerned with research, testing and assessment of transport containers for radioactive materials. This includes mechanical and thermal safety verifications and the development of experimental and computational test methods for the analysis of transport containers.

Topic of the Thesis

Experienced engineers at BAM review the safety of transport packages for radioactive material. Their reasoning is being recorded and structured into chain of steps, in an open knowledge format. Experts skip steps that are obvious to them. A recorded chain is therefore often incomplete. A step may presuppose a check that was never stated. A conclusion may rest on premises that are not on record. Or the chain jumps in a way a later reader cannot follow. Recent research shows that large language models tend to fill such gaps silently instead of flagging them. A dedicated detection tool is therefore needed. Your thesis builds it.

Task of the Thesis

You develop and evaluate a gap-finding tool. The tool scans structured reasoning chains for gaps in the logic. It draws on three published mechanisms. Premise linking exposes steps whose support is missing. Backward reasoning detects information that was never stated. Consistency checks test formalised steps for contradictions. One fixed rule applies. Every detected gap becomes a question to the expert, never a suggestion. You evaluate the tool on reasoning chains with seeded gaps, measuring detection rate, false alarms, and how useful reviewers find the generated questions.

Agenda

• The thesis can be written in English or German
• Literature research on premise linking, missing-information detection, and logical consistency checking
• Design the gap taxonomy together with your supervisors. Unstated check, missing premise, unfollowable jump
• Build a prototype of the detection tool and the question-generation loop
• Evaluate on reasoning chains with seeded gaps. Detection rate, false alarms, and the usefulness of the questions

Qualifications

• Studies in computer science, data science, mathematics, engineering, cognitive science, or a related field
• You can build working prototypes with AI coding assistants (“vibe coding”). Solid skills in a language like Python are a plus, not a must
• Interest in reasoning evaluation, logic, and human-in-the-loop tools
• Good German reading comprehension, because the reasoning chains are in German
• Helpful but not required: experience with LLM APIs, prompt engineering, formal logic, or SMT solvers

Contact

Mohamed Tababi , phone: 030 8104-3674, E-Mail: Mohamed.Tababi@bam.de
Kutlualp Tazefidan, phone: 030 8104-3827, E-Mail: kutlualp.tazefidan@bam.de
Dr. Tobias Gleim, phone: 030 8104-3168, E-Mail: Tobias.Gleim@bam.de

Address

Bundesanstalt für Materialforschung und -prüfung (BAM)
Division 3.3 Safety of Transport Containers
Unter den Eichen 44-46, 12203 Berlin