Seminar: Large Language Models and Knowledge-Aware AI

TU Dresden — Summer Term 2026

Overview

Large Language Models are the most powerful knowledge systems ever built — yet they remain, in many ways, black boxes. GPT-4, Llama, and their peers have silently absorbed vast swaths of human knowledge during training, encoding billions of facts across their parameters. But what exactly do they know? How much of it is right? And can we trust it?

This seminar tackles these questions across eight tracks:

Track 1 — Knowledge Auditing & Probing examines how we systematically audit what a model believes, from cloze-style queries to full-scale knowledge materialization, and how we benchmark what LLMs actually know versus what they merely appear to know.

Track 2 — Hallucination & Factual Reliability investigates how and why LLMs fabricate — from invented entities to confident but false claims — and whether detection or prevention is even possible, including recent theoretical arguments that hallucination may be an innate architectural limitation.

Track 3 — Knowledge Editing & Consistency explores whether targeted, surgical correction of LLM knowledge is feasible, or whether cascading inconsistencies and catastrophic forgetting make it a losing battle at scale.

Track 4 — Interpretability & Knowledge Mechanisms digs into how transformers store and recall facts — through feed-forward key-value memories, knowledge neurons, and the architectural constraints that produce phenomena like the Reversal Curse.

Track 5 — LLM Biases, Language Learning & Knowledge Construction addresses systematic distortions in what LLMs learn, including implicit biases introduced by persona assignment, and examines how structured knowledge can be constructed from or around LLMs.

Track 6 — LLM Limitations takes a broader look at fundamental shortcomings of current architectures — including failures in compositionality, long-context reasoning, counterfactual tasks, and self-correction — connecting these limitations back to the core challenge of reliable knowledge retrieval.

Track 7 — Responsible AI: Safety, Privacy & Misuse covers jailbreaking, training data extraction, memorization, and the structural reasons safety training can fail, examining both attacks and defenses.

Track 8 — Can We Afford the Perfect Prompt? examines the economics of prompting, asking whether state-of-the-art techniques like chain-of-thought are worth their computational cost, and what scaling laws apply to compound inference systems.

Logistics

Registration: The number of participants is limited to 12, with priority given to Master students. To express interest, send an email to muhammed.saeed@tu-dresden.de, including a short motivation statement and your transcript. Places will be allocated based on background match (courses taken) and motivation.

Core Papers

These two papers form the backbone of the seminar. All participants should read them as shared reference points.

GPTKB materialized 101 million triples from GPT-4o-mini via recursive prompting — creating the largest LLM-derived knowledge base to date. It revealed that LLMs can serve as massive knowledge bases, but also exposed systemic issues: ~7% false triples, fabricated entities, and deep structural inconsistencies (e.g., only 8K of 318K spouse relations are symmetric). Foundations takes the next step, formally analyzing the theoretical properties of knowledge materialization — when does it terminate, how reproducible is it across runs, and how robust is it to perturbation?

Topics

Papers are organized by thematic track. Each participant selects one paper. Own topic suggestions are welcome.

Track 1: Knowledge Auditing & Probing

What do LLMs know — and how do we find out?

Track 2: Hallucination & Factual Reliability

LLMs fabricate. Can we detect it — and is it fixable?

Track 3: Knowledge Editing & Consistency

When LLMs are wrong, can we fix them?

Track 4: Interpretability & Knowledge Mechanisms

How do transformers store and recall facts — and what are the limits?

A good starting point for getting an overview of interpretability: https://thegradient.pub/explain-yourself/

Track 5: LLM Biases, Language Learning & Knowledge Construction

Broader questions about how LLMs learn, what they get wrong systematically, and how to build structured knowledge.

Track 6: LLM Limitations

A broader look at the fundamental shortcomings of current LLM architectures — in compositionality, long-context reasoning, counterfactual tasks, and self-correction — and what these limitations mean for reliable knowledge use.

Track 7: Responsible AI — Safety, Privacy & Misuse

Track 8: Can We Afford the Perfect Prompt?

Inspired by the EPI paper (McDonald et al.), this track examines the economics and efficiency of prompting — asking whether state-of-the-art prompting techniques are worth their computational cost.

Background Reading

These papers provide excellent overviews for seminar preparation:

• A Review of Knowledge in Language Models — AlKhamissi et al. (arXiv 2022) — Comprehensive survey on how knowledge is stored, probed, and edited in LLMs.
• A Survey on Hallucination in Large Language Models — Huang et al. (arXiv 2023) — Taxonomy, challenges, and open questions.
• Editing Large Language Models: A Survey — Yao et al. (arXiv 2023) — Covers ROME, MEMIT, MEND, and more.
• Knowledge Mechanisms in Large Language Models: A Survey — Wang et al. (EMNLP 2024 Findings) — Theoretical grounding for storage, retrieval, and consistency.
• A Comprehensive Study of Knowledge Editing for LLMs (KnowEdit) — Zhang et al. (arXiv 2024) — Benchmark and survey for knowledge editing.

Grading

The final grade consists of:

• Report (33%): A written report (max. 4 pages, ACL-style)
• Presentation (33%): 20-minute presentation
• Q&A (33%): 15-minute Q&A session (5 min by peers, 10 min by course team). Each participant is assigned to ask questions to two peers, randomly assigned on the seminar days.

Tentative Timeline

Material

• Slides 1st meeting
• Slides 2nd meeting
• Report template