Papers by Gjergji Kasneci
Analysing the Safety Pitfalls of Steering Vectors (2026.findings-acl)
Copied to clipboard
| Challenge: | Activation steering has emerged as a powerful tool to shape LLM behaviour without the need for weight updates. |
| Approach: | They propose to audit steering vectors obtained with Contrastive Activation Addition (CAA) and propose a mechanistic explanation for this finding. |
| Outcome: | The proposed approach significantly improves the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. |
CURE: Controlled Unlearning for Robust Embeddings — Mitigating Conceptual Shortcuts in Pre-Trained Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained language models are susceptible to spurious, concept-driven correlations that impair robustness and fairness. |
| Approach: | They propose a framework that disentangles and suppresses conceptual shortcuts while preserving essential content information. |
| Outcome: | The proposed framework improves on IMDB and Yelp datasets with minimal computational overhead. |
Where Paths Split: Localized, Calibrated Control of Moral Reasoning in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models display heterogeneous moral preferences across settings. |
| Approach: | They propose a method for steering toward a desired ethical framework while preserving general competence. |
| Outcome: | The proposed method outperforms baselines while providing interpretable mechanism. |
Probabilistic Aggregation and Targeted Embedding Optimization for Collective Moral Reasoning in Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have impressive moral reasoning abilities, yet they often diverge when confronted with complex, multi-factor moral dilemmas. |
| Approach: | They propose a framework that synthesizes multiple LLMs’ moral judgments into a collectively formulated moral judgment, realigning models that deviate significantly from this consensus. |
| Outcome: | The proposed framework synthesizes multiple LLMs’ moral judgments into a collectively formulated moral judgment, realigning models that deviate significantly from this consensus. |
Is Crowdsourcing Breaking Your Bank? Cost-Effective Fine-Tuning of Pre-trained Language Models with Proximal Policy Optimization (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods to improve language models require manual ranking and annotators. |
| Approach: | They propose a self-supervised text ranking approach for applying Proximal-Policy-Optimization to fine-tune language models while eliminating the need for human annotators. |
| Outcome: | The proposed method significantly outperforms baselines regarding BLEU, GLEU, and METEOR scores on three tasks and is consistent with humans. |
From Confidence to Collapse in LLM Factual Robustness (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing evaluation methods focus on performance-based metrics, often investigating from the perspective of prompt perturbations, which captures only the externally triggered side of knowledge robustness. |
| Approach: | They propose a method to measure factual robustness from the perspective of the generation process by analyzing token distribution entropy and temperature scaling sensitivity. |
| Outcome: | The proposed method measures factual robustness from the perspective of the generation process and entropy and temperature scaling sensitivity. |
The Language of Trauma: Modeling Traumatic Event Descriptions Across Domains with Explainable AI (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Psychological trauma can manifest following various distressing events, but studies focus on a single aspect of trauma, often neglecting the transferability of findings across different scenarios. |
| Approach: | They propose a language model that fine-tunes a single aspect of trauma to better predict traumatic events across domains. |
| Outcome: | The proposed model outperforms large language models on trauma-related datasets . it also outperformed models on court data, counseling conversations, and forum posts . |
Not All Features Deserve Attention: Graph-Guided Dependency Learning for Tabular Data Generation with Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have shown strong potential for tabular data generation by modeling textualized feature-value pairs. |
| Approach: | They propose a method that explicitly integrates sparse dependency graphs into LLMs’ attention mechanism. |
| Outcome: | The proposed method outperforms existing LLM-based approaches by up to 12% on complex datasets while achieving competitive results with state-of-the-art approaches in synthetic data quality. |
P-TA: Using Proximal Policy Optimization to Enhance Tabular Data Augmentation via Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Contemporary approaches to generate tabular data are limited due to the lack of external knowledge. |
| Approach: | They propose to use proximal policy optimization to apply GANs and fine-tune Large Language Models to enhance the probability distribution of tabular features. |
| Outcome: | The proposed method improves accuracy of GANs and LLMs over state-of-the-art over three real-world datasets. |
Doubling Your Data in Minutes: Ultra-fast Tabular Data Generation via LLM-Induced Dependency Graphs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Tabular data is critical across diverse domains, yet high-quality tabular datasets remain scarce due to privacy concerns and the cost of collection. |
| Approach: | They propose a lightweight generative framework that captures sparse dependencies via an LLM-induced graph. |
| Outcome: | The proposed framework reduces constraint violations by 4% and accelerates generation by nearly 9,500 over diffusion-based methods. |
SAGE: Sparse Adaptive Guidance for Dependency-Aware Tabular Data Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Recent approaches to generate tabular data are limited by their static dependences and lack of fidelity. |
| Approach: | They propose a novel LLM-based generation framework that enforces sparse and dynamic dependency guidance. |
| Outcome: | The proposed framework boosts F1 scores by 10% and reduces policy violations by one point. |