analiza aiqbench
This research introduces ICLHerd, a novel data selection framework that mitigates mode-seeking bias in LLM training by balancing similarity with diversity using the von Mises-Fisher kernel and in-context learning as a validation proxy. The study demonstrates that this approach outperforms existing algorithms, underscoring the critical need to avoid clustering training data too tightly around development samples to improve downstream model performance.
ICLHerd: Herding-Based Data Selection for LLMs with In-Context Learning as a Proxy
arXiv:2610.27142v1 Announce Type: cross Abstract: Data selection plays a central role in modern large language model training, as the quality of training data strongly influences downstream performance. A common approach is to curate a development set that encodes the desired skills (e.g., reasoning or multilinguality) and select data that are similar to them. Existing methods compute similarity using cosine distance over either the model's hidden states or gradients. However, we observe that this max-similarity strategy induces a mode-seeking bias: the selected data clusters tightly around the development samples. To address this, we propose ICLHerd, a data selection framework that leverages the von Mises-Fisher kernel to balance mode-seeking with mode-averaging. We further introduce in-context learning as a validation proxy, enabling efficient hyperparameter search over the balance trade-off. Results show that ICLHerd outperforms existing LLM data selection algorithms, highlighting the importance of balancing mode-seeking and mode-averaging.
Abstract:Data selection plays a central role in modern large language model training, as the quality of training data strongly influences downstream performance. A common approach is to curate a development set that encodes the desired skills (e.g., reasoning or multilinguality) and select data that are similar to them. Existing methods compute similarity using cosine distance over either the model's hidden states or gradients. However, we observe that this max-similarity strategy induces a mode-seeking bias: the selected data clusters tightly around the development samples. To address this, we propose ICLHerd, a data selection framework that leverages the von Mises-Fisher kernel to balance mode-seeking with mode-averaging. We further introduce in-context learning as a validation proxy, enabling efficient hyperparameter search over the balance trade-off. Results show that ICLHerd outperforms existing LLM data selection algorithms, highlighting the importance of balancing mode-seeking and mode-averaging.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
I.2.7; I.2.6; I.2.0
arXiv:2610.27142 [cs.CL]
(or arXiv:2610.27142v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.27142 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 243 cuvinte