TY - GEN
T1 - Beyond the Leaderboard
T2 - 30th Conference on Empirical Methods in Natural Language Processing, EMNLP 2025
AU - Boughorbel, Sabri
AU - Dalvi, Fahim
AU - Durrani, Nadir
AU - Hawasly, Majd
N1 - Publisher Copyright:
© 2025 Association for Computational Linguistics.
PY - 2025
Y1 - 2025
N2 - As fine-tuning becomes the dominant paradigm for improving large language models (LLMs), understanding what changes during this process is increasingly important. Traditional benchmarking often fails to explain why one model outperforms another. In this work, we use model diffing, a mechanistic interpretability approach, to analyze the specific capability differences between Gemma-2-9b-it and a SimPO-enhanced variant. Using crosscoders, we identify and categorize latent representations that differentiate the two models. We find that SimPO acquired latent concepts predominantly enhance safety mechanisms (+32.8%), multilingual capabilities (+43.8%), and instruction-following (+151.7%), while its additional training also reduces emphasis on model self-reference (-44.1%) and hallucination management (-68.5%). Our analysis shows that model diffing can yield fine-grained insights beyond leaderboard metrics, attributing performance gaps to concrete mechanistic capabilities. This approach offers a transparent and targeted framework for comparing LLMs.
AB - As fine-tuning becomes the dominant paradigm for improving large language models (LLMs), understanding what changes during this process is increasingly important. Traditional benchmarking often fails to explain why one model outperforms another. In this work, we use model diffing, a mechanistic interpretability approach, to analyze the specific capability differences between Gemma-2-9b-it and a SimPO-enhanced variant. Using crosscoders, we identify and categorize latent representations that differentiate the two models. We find that SimPO acquired latent concepts predominantly enhance safety mechanisms (+32.8%), multilingual capabilities (+43.8%), and instruction-following (+151.7%), while its additional training also reduces emphasis on model self-reference (-44.1%) and hallucination management (-68.5%). Our analysis shows that model diffing can yield fine-grained insights beyond leaderboard metrics, attributing performance gaps to concrete mechanistic capabilities. This approach offers a transparent and targeted framework for comparing LLMs.
UR - https://www.scopus.com/pages/publications/105040231368
U2 - 10.18653/v1/2025.emnlp-main.1598
DO - 10.18653/v1/2025.emnlp-main.1598
M3 - Conference contribution
AN - SCOPUS:105040231368
T3 - EMNLP 2025 - 2025 Conference on Empirical Methods in Natural Language Processing, Proceedings of the Conference
SP - 31360
EP - 31371
BT - EMNLP 2025 - 2025 Conference on Empirical Methods in Natural Language Processing, Proceedings of the Conference
A2 - Christodoulopoulos, Christos
A2 - Chakraborty, Tanmoy
A2 - Rose, Carolyn
A2 - Peng, Violet
PB - Association for Computational Linguistics (ACL)
Y2 - 4 November 2025 through 9 November 2025
ER -