Knowledge Conflicts in Large Language Models
Module Description
| Course | Module Abbreviation | Credit Points |
|---|---|---|
| BA-2010[100%|75%] | CS-CL | 6 LP |
| BA-2010[50%] | BS-CL | 6 LP |
| BA-2010[25%] | BS-AC | 4 LP |
| BA-2010 | AS-CL | 8 LP |
| Master | SS-CL-TAC | 8 LP |
| Lecturer | Katja Markert |
| Module Type | Proseminar / Hauptseminar |
| Language | English |
| First Session | 13.04.2026 |
| Time and Place | Mo, 15:15 - 16:45, SR 10 / INF 346 |
| Commitment Period | tbd. |
Participants
All advanced CL Bachelor students and all CL master students. Students from MSc Data and Computer Science or MSc Scientific Computing with Field of Application Computational Linguistics are welcome after getting permission from the lecturer. MSc Scientific Computing students can only take the course as HS for 8 LP. If the seminar should be oversubscribed, CL students will have priority.
Prerequisites for Participation
- MA: none.
- BA: ECL, Programming I, Statistical Methods
Assessment
- Active Participation
- Presentation
- Project that is orally demonstrated.
No term paper or written project report required.
Content
Large language models are being used in a variety of knowledge-intensive tasks such as question answering, summarization and fact checking. They rely both on their own parametric knowledge (derived from pretraining data as well as RHLF and instruction tuning) as well as contextual knowledge provided by the user or web search (RAG systems, for example). Unfortunately these diverse knowledge sources can be in conflict with each other, for example due to temporal updates, different points of view, disinformation campaigns or simple errors in a knowledge source. This seminar will investigate sources of such knowledge conflicts as well as methods for conflict resolution. It will cover:
- All three main types of knowledge conflicts (between parametric knowledge and context, between two or more different context sources, and within different pieces of parametric knowledge).
- Impact of knowledge conflicts on LLM behaviour such as increase of hallucinations, self-inconsistency, and lowering performance.
- Causes of knowledge conflicts (temporal misalignment, misinformation pollution etc)
- Solutions, including faithful-to-context of faithful-to-memory strategies (via fine-tuning, prompting, decoding or disentangling and then synthesizing sources), query augmentation, improving consistency (for example with output ensembles) and improving factuality (for example via constrastive decoding).
