Blog CRC1646

Florian Kankowski had the opportunity to present his research at the CogSci 2025 conference in San Francisco. Organized by The Cognitive Science Society, the conference, titled "Theories of the past, theories of the future," took place from July 30th to August 2nd. The conference was structured with a diverse program of talks, symposia, and poster sessions. As is tradition for CogSci, the topics language, communication and cooperation were strongly represented at the conference, majorly intersecting with the research topics of the CRC.

Florian Kankowski (Project Ö) presented two posters at the conference, both situated within the domain of discourse. The first, "Instruction tuning modulates discourse biases in language models" (Project Ö), co-authored with Torgrim Solstad, Sina Zarrieß and Oliver Bott, investigated how instruction tuning and model size affect the emergence of Implicit Causality biases in different language models. Their findings revealed that instruction tuning can induce biases that align more closely with human-like preferences, even exceeding them in some cases.

His second poster, "Linguistic Creativity affects Discourse Expectations related to Contiguity Relations but not Implicit Causality" (Projects B01 and Ö), was co-authored with Oliver Bott and Torgrim Solstad. This poster presented findings from an experiment on human participants, showing that while a "creativity on demand" task can lead to more elaborate discourse, it does not override strong lexical-semantic biases like Implicit Causality. This suggests that even in creative language use, humans tend to produce language that adheres to their preferred discourse structure and instead elect to express linguistic creativity through different means.

Attending a large, interdisciplinary conference like CogSci was incredibly valuable for Florian. It provided a unique opportunity to connect with researchers from different sub-fields of cognitive science and get a broader perspective on his work. The feedback he received on his posters was insightful and has already given him new ideas for future research. In line with the overarching questions of the CRC, he is particularly interested in extending the LLM research to develop a more robust benchmark for comparing human and machine creative language use.

alle © SFB 1646