Monday, 14th April - 11am: Omri Abend Speaker: Omri Abend (Associate Professor, Hebrew University)Talk Title: LLMs as Crime Fiction Writers: A Formal ApproachAbstract: Storytelling is often viewed as one of the defining characteristics of humans and human cultures. It requires an intricate set of abilities, including (partially) adhering to the conventions and schemata of their domain. We focus on the domain of golden age detective fiction, and study LLM-generated short stories through the process by which the culprit’s identity is revealed. We formalize some conventional properties of such processes, and study generated stories in terms of their compliance with these conventions. We argue that from a literary point of view, the coherence and the unpredictability of such fiction are in inherent tension, and should be studied jointly. We empirically show that while LLM-generated stories tend to be unpredictable, they fail at balancing this trade-off, which greatly contributes to their poor quality. Time permitting, I will discuss a related line of work on formalizing different interpretations of LLM output distributions. Joint work with Eitan Wagner and Renana Keydar.Bio: I am an associate professor at the Hebrew University, working in the departments of Computer Science and Cognitive Science. Previously I was a research associate (post-doc) with Mark Steedman under the GramPlus project. I earned my PhD from the Hebrew University, where my supervisor was Ari Rappoport. During my PhD studies, I was a member of the Azrieli Fellows Program. Before that, I studied mathematics and cognitive sciences in the Hebrew University (BSc, summa cum laude). Apr 14 2025 11.00 - 11.00 Monday, 14th April - 11am: Omri Abend This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. G.03, Bayes and Teams
Monday, 14th April - 11am: Omri Abend Speaker: Omri Abend (Associate Professor, Hebrew University)Talk Title: LLMs as Crime Fiction Writers: A Formal ApproachAbstract: Storytelling is often viewed as one of the defining characteristics of humans and human cultures. It requires an intricate set of abilities, including (partially) adhering to the conventions and schemata of their domain. We focus on the domain of golden age detective fiction, and study LLM-generated short stories through the process by which the culprit’s identity is revealed. We formalize some conventional properties of such processes, and study generated stories in terms of their compliance with these conventions. We argue that from a literary point of view, the coherence and the unpredictability of such fiction are in inherent tension, and should be studied jointly. We empirically show that while LLM-generated stories tend to be unpredictable, they fail at balancing this trade-off, which greatly contributes to their poor quality. Time permitting, I will discuss a related line of work on formalizing different interpretations of LLM output distributions. Joint work with Eitan Wagner and Renana Keydar.Bio: I am an associate professor at the Hebrew University, working in the departments of Computer Science and Cognitive Science. Previously I was a research associate (post-doc) with Mark Steedman under the GramPlus project. I earned my PhD from the Hebrew University, where my supervisor was Ari Rappoport. During my PhD studies, I was a member of the Azrieli Fellows Program. Before that, I studied mathematics and cognitive sciences in the Hebrew University (BSc, summa cum laude). Apr 14 2025 11.00 - 11.00 Monday, 14th April - 11am: Omri Abend This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. G.03, Bayes and Teams
Apr 14 2025 11.00 - 11.00 Monday, 14th April - 11am: Omri Abend This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk.
