Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Springer, Jacob Mitchell, Advani, Madhu, Aichberger, Lukas, Bradley, Arwen, Malach, Eran, Saremi, Omid, Williamson, Sinead, Nakkiran, Preetum, Littwin, Etai, Raghunathan, Aditi
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.09995
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866918494007721984
author Springer, Jacob Mitchell
Advani, Madhu
Aichberger, Lukas
Bradley, Arwen
Malach, Eran
Saremi, Omid
Williamson, Sinead
Nakkiran, Preetum
Littwin, Etai
Raghunathan, Aditi
author_facet Springer, Jacob Mitchell
Advani, Madhu
Aichberger, Lukas
Bradley, Arwen
Malach, Eran
Saremi, Omid
Williamson, Sinead
Nakkiran, Preetum
Littwin, Etai
Raghunathan, Aditi
contents Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.
format Preprint
id arxiv_https___arxiv_org_abs_2605_09995
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Annotations Mitigate Post-Training Mode Collapse
Springer, Jacob Mitchell
Advani, Madhu
Aichberger, Lukas
Bradley, Arwen
Malach, Eran
Saremi, Omid
Williamson, Sinead
Nakkiran, Preetum
Littwin, Etai
Raghunathan, Aditi
Computation and Language
Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.
title Annotations Mitigate Post-Training Mode Collapse
topic Computation and Language
url https://arxiv.org/abs/2605.09995