Benjamin Van Durme

Training LLMs with long contexts can paradoxically weaken their ability to retain knowledge, leading to poorer performance when context is not available.

Contextual tunneling in LLMs can be overcome, leading to more reliable and physically grounded materials discovery through the innovative ARIA framework.

Document LoRA can recover up to 21 ROUGE-L points when context is scarce, redefining its role in memory architecture for question answering.

Agentic harnesses can significantly enhance LLM performance on deontic reasoning tasks, but not without introducing risks of degradation in numerical accuracy.

Even state-of-the-art LLMs struggle to follow complex instruction hierarchies, achieving only ~40% accuracy when navigating conflicts across a dozen privilege levels in agentic tasks.

LLMs struggle to navigate the nuances of real-world rules, achieving only ~45% accuracy on a new benchmark of legal and policy reasoning tasks.

Reasoning rerankers don't magically fix fairness issues in search, preserving the biases of their input rankings despite boosting relevance.

Stop blindly optimizing for retrieval relevance in RAG pipelines: coverage-based retrieval metrics are better early indicators of the final generated response's information coverage.

Attention-guided clustering slashes the storage costs of multi-vector document representations for retrieval across text, images, and video, often *improving* performance compared to uncompressed indexes.