TurnWise: The Gap between Single- and Multi-turn Language Model Capabilities

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Graf, Victoria, Pyatkin, Valentina, Dziri, Nouha, Lambert, Nathan, Hajishirzi, Hannaneh
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866914402517647360
author Graf, Victoria
Pyatkin, Valentina
Dziri, Nouha
Lambert, Nathan
Hajishirzi, Hannaneh
author_facet Graf, Victoria
Pyatkin, Valentina
Dziri, Nouha
Lambert, Nathan
Hajishirzi, Hannaneh
contents Multi-turn conversations are a common and critical mode of language model interaction. However, current open training and evaluation data focus on single-turn settings, failing to capture the additional dimension of these longer interactions. To understand this multi-/single-turn gap, we first introduce a new benchmark, TurnWiseEval, for multi-turn capabilities that is directly comparable to single-turn chat evaluation. Our evaluation isolates multi-turn specific conversational ability through pairwise comparison to equivalent single-turn settings. We additionally introduce our synthetic multi-turn data pipeline TurnWiseData which allows the scalable generation of multi-turn training data. Our experiments with Olmo 3 show that training with multi-turn data is vital to achieving strong multi-turn chat performance, and that including as little as 10k multi-turn conversations during post-training can lead to a 12% improvement on TurnWiseEval.
format Preprint
id arxiv_https___arxiv_org_abs_2603_16759
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle TurnWise: The Gap between Single- and Multi-turn Language Model Capabilities
Graf, Victoria
Pyatkin, Valentina
Dziri, Nouha
Lambert, Nathan
Hajishirzi, Hannaneh
Computation and Language
Artificial Intelligence
Multi-turn conversations are a common and critical mode of language model interaction. However, current open training and evaluation data focus on single-turn settings, failing to capture the additional dimension of these longer interactions. To understand this multi-/single-turn gap, we first introduce a new benchmark, TurnWiseEval, for multi-turn capabilities that is directly comparable to single-turn chat evaluation. Our evaluation isolates multi-turn specific conversational ability through pairwise comparison to equivalent single-turn settings. We additionally introduce our synthetic multi-turn data pipeline TurnWiseData which allows the scalable generation of multi-turn training data. Our experiments with Olmo 3 show that training with multi-turn data is vital to achieving strong multi-turn chat performance, and that including as little as 10k multi-turn conversations during post-training can lead to a 12% improvement on TurnWiseEval.
title TurnWise: The Gap between Single- and Multi-turn Language Model Capabilities
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2603.16759