%0 Conference Paper %1 kobs-etal-2024-pollice %A Kobs, Konstantin %A Pfister, Jan %A Hotho, Andreas %B Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024) %C Mexico City, Mexico %D 2024 %E Ojha, Atul Kr. %E Dogruöz, A. Seza %E Tayyar Madabushi, Harish %E Da San Martino, Giovanni %E Rosenthal, Sara %E Rosá, Aiala %I Association for Computational Linguistics %K author:hotho author:pfister from:janpf motiv myown %P 1529--1536 %T Pollice Verso at SemEval-2024 Task 6: The Roman Empire Strikes Back %U https://aclanthology.org/2024.semeval-1.219 %X We present an intuitive approach for hallucination detection in LLM outputs that is modeled after how humans would go about this task. We engage several LLM ``experts'' to independently assess whether a response is hallucinated. For this we select recent and popular LLMs smaller than 7B parameters. By analyzing the log probabilities for tokens that signal a positive or negative judgment, we can determine the likelihood of hallucination. Additionally, we enhance the performance of our ``experts'' by automatically refining their prompts using the recently introduced OPRO framework. Furthermore, we ensemble the replies of the different experts in a uniform or weighted manner, which builds a quorum from the expert replies. Overall this leads to accuracy improvements of up to 10.6 p.p. compared to the challenge baseline. We show that a Zephyr 3B model is well suited for the task. Our approach can be applied in the model-agnostic and model-aware subtasks without modification and is flexible and easily extendable to related tasks.

@inproceedings{kobs-etal-2024-pollice, abstract = {We present an intuitive approach for hallucination detection in LLM outputs that is modeled after how humans would go about this task. We engage several LLM {``}experts{''} to independently assess whether a response is hallucinated. For this we select recent and popular LLMs smaller than 7B parameters. By analyzing the log probabilities for tokens that signal a positive or negative judgment, we can determine the likelihood of hallucination. Additionally, we enhance the performance of our {``}experts{''} by automatically refining their prompts using the recently introduced OPRO framework. Furthermore, we ensemble the replies of the different experts in a uniform or weighted manner, which builds a quorum from the expert replies. Overall this leads to accuracy improvements of up to 10.6 p.p. compared to the challenge baseline. We show that a Zephyr 3B model is well suited for the task. Our approach can be applied in the model-agnostic and model-aware subtasks without modification and is flexible and easily extendable to related tasks.}, added-at = {2024-07-01T03:27:07.000+0200}, address = {Mexico City, Mexico}, author = {Kobs, Konstantin and Pfister, Jan and Hotho, Andreas}, biburl = {https://www.bibsonomy.org/bibtex/2e2234c8b039b7b82e96fa381736be0ae/dmir}, booktitle = {Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024)}, editor = {Ojha, Atul Kr. and Do{\u{g}}ru\"oz, A. Seza and Tayyar Madabushi, Harish and Da San Martino, Giovanni and Rosenthal, Sara and Ros\'a, Aiala}, interhash = {7e73d61fab4f1359294e4365e97d39d8}, intrahash = {e2234c8b039b7b82e96fa381736be0ae}, keywords = {author:hotho author:pfister from:janpf motiv myown}, month = {06}, pages = {1529--1536}, publisher = {Association for Computational Linguistics}, timestamp = {2024-11-06T14:19:51.000+0100}, title = {Pollice Verso at SemEval-2024 Task 6: The Roman Empire Strikes Back}, url = {https://aclanthology.org/2024.semeval-1.219}, year = 2024 }