Talks & Presentations
A selection of talks and presentations I’ve given during seminars and reading groups at the University of Heidelberg.
Most were short academic discussions — sharing recent papers, experiments, or reproduction results with classmates and professors.
I enjoy breaking down complex research into clear ideas, connecting theory with implementation, and learning from the conversations that follow.
🧠 Large Language Models & Interpretability
LLMint8: Quantization-Aware Fine-Tuning for Efficient LLMs
📅 June 11, 2025 · Preview Slides (PDF)
Presented and analyzed the LLMint8 approach for quantization-aware fine-tuning (QAT) of large language models.
The talk discussed how mixed-precision training and INT8 quantization maintain reasoning quality while cutting GPU memory use by ~60%.
Also compared QLoRA, post-training quantization, and QAT strategies — highlighting the trade-off between efficiency and alignment stability.
LMKG-GOFA: Language Models as Knowledge Graph Builders
📅 July 7, 2025 · Preview Slides (PDF)
Discussed the LMKG-GOFA framework for LLM-based ontology and knowledge graph construction.
Explained how large models extract entities and relations through instruction prompting and structured templates.
Reviewed experimental results showing that language models can dynamically expand KGs, bridging symbolic reasoning and text generation.
Character-Level Language Modeling: Hierarchies in Text Understanding
📅 May 21, 2025 · Preview Slides (PDF)
Presented a critical review of hierarchical character-level language modeling techniques.
Explored how multi-layer recurrent networks capture structure and morphology from raw text sequences.
Highlighted how these architectures relate to current subword tokenization and interpretability research.
Self-Conditioning: Teaching Models to Listen to Themselves
📅 * December 20, 2024 * · Preview Slides (PDF)
Reviewed the paper Self-Conditioned Pretrained Language Models (ICML 2022), focusing on mechanistic interpretability and concept-level control.
Explained how “expert neurons” can be reactivated to guide text generation without additional fine-tuning, enabling fast and low-overhead conditioning.
Compared the approach with FUDGE and PPLM, emphasizing improvements in speed, perplexity, and semantic precision.
🩺 Multimodal & Clinical Language Understanding
Research on Multimodal Fusion of Temporal Electronic Medical Records
📅 November 10, 2024 · Preview Slides (PDF)
Reviewed a paper proposing T-MAG (Time-series Multimodal Adaptation Gate) — a fusion model for heterogeneous EMR data.
Explained how LSTM and Transformer-XL encoders integrate structured and textual time-series data, while MAG dynamically balances multiple modalities and attention-backtracking captures long-term dependencies.
Highlighted that clinical notes as the main modality achieved the best predictive performance (AUROC ≈ 0.95 on stroke datasets).
Discussed implications for clinical outcome prediction and temporal representation learning in healthcare AI.
Clinical Language Understanding — Part I & II
📅 February 03, 2025 · Part I PDF | Part II PDF
Analyzed systems from the 2024 Chemotherapy Treatment Timeline Extraction shared task.
Summarized LAILab’s Flan-T5 + LoRA model for instruction-tuned temporal relation extraction and KCLab’s PubMedBERT pipeline integrating UMLS knowledge.
Compared end-to-end and pipeline approaches, discussing trade-offs between precision, recall, and interpretability in EHR processing.
Concluded with insights on low-frequency relation handling, semi-supervised data augmentation, and the limits of LLM generalization in clinical text.
⚖️ AI Ethics, Bias & Security
Data Poisoning in Large Models: Risks and Defenses
📅 November 25, 2024 · Preview Slides (PDF)
Presented a review of Carlini et al., 2023 on the practical feasibility of large-scale data poisoning.
Explained the Split-View and Frontrunning attacks that exploit expired or predictable URLs in distributed and centralized datasets such as LAION-400M and Wikipedia.
Showed that poisoning as little as 0.01 % of data (~ $60 cost) can inject targeted vulnerabilities into downstream LLMs.
Discussed defense mechanisms including cryptographic integrity checks, snapshot randomization, and consensus-based dataset verification to strengthen the security of open web-scale corpora.
AI Generates Covertly Racist Decisions Based on Dialect
📅 *November 12, 2024 * · Preview Slides (PDF)
Analyzed a sociolinguistic study revealing dialect-driven bias in LLM and speech-to-text systems.
The presentation examined how models encode implicit stereotypes when processing African-American Vernacular English (AAVE) and other non-standard dialects.
Highlighted findings that classifiers produce less favorable judgments and higher toxicity scores for identical content differing only in dialectal style.
Discussed mitigation through prompt calibration, representation balancing, and contrastive fine-tuning to reduce latent sociocultural bias.