aiqbench analysis
This research addresses the critical challenge of maintaining high-quality reasoning capabilities in low-resource languages like Cantonese during model fine-tuning. The study demonstrates that standard supervised fine-tuning can inadvertently strip away native reasoning traces, a problem effectively mitigated through Reinforcement Learning from Verifiable Rewards (RLVR) with specific language constraints. By releasing these checkpoints and training environments, the work provides a vital resource for advancing multilingual AI beyond English-centric benchmarks.
CantoneseLLM v2: Reasoning in a Low-Resource Language
arXiv:2609.06970v1 Announce Type: new Abstract: Cantonese is widely spoken but remains low-resource in written data, with no large corpus of native Cantonese reasoning traces available for model training. We develop and release CantoneseLLM v2, comprising models based on Qwen3 8B and 30B-A3B. The models are trained through CPT on 784 million Cantonese and Hong Kong-related tokens, chat-vector merging, SFT, DPO, and RLVR. Evaluation across the training stages shows that chat-vector merging transfers instruction following but preserves the donor model's reasoning language, while SFT with limited Cantonese reasoning data substantially shortens or removes reasoning traces and reduces benchmark performance. DPO restores the reasoning-block format, particularly for the 8B model, but recovers only part of the lost performance. The RLVR training with Cantonese language and Traditional Chinese scripts as multiplicative constraints introduced Cantonese language alignment and restored the lost performance. The 30B-A3B model reaches 73.16 on HKCanto-Eval, within 1.20 points of its merged checkpoint, while retaining the Cantonese reasoning behaviour absent from that checkpoint. We release the model checkpoints, the training environments, and a thirteen-year Traditional Chinese Common Crawl dataset. The models can be accessed at https://huggingface.co/collections/hon9kon9ize/cantonesellm-v20
Abstract:Cantonese is widely spoken but remains low-resource in written data, with no large corpus of native Cantonese reasoning traces available for model training. We develop and release CantoneseLLM v2, comprising models based on Qwen3 8B and 30B-A3B. The models are trained through CPT on 784 million Cantonese and Hong Kong-related tokens, chat-vector merging, SFT, DPO, and RLVR. Evaluation across the training stages shows that chat-vector merging transfers instruction following but preserves the donor model's reasoning language, while SFT with limited Cantonese reasoning data substantially shortens or removes reasoning traces and reduces benchmark performance. DPO restores the reasoning-block format, particularly for the 8B model, but recovers only part of the lost performance. The RLVR training with Cantonese language and Traditional Chinese scripts as multiplicative constraints introduced Cantonese language alignment and restored the lost performance. The 30B-A3B model reaches 73.16 on HKCanto-Eval, within 1.20 points of its merged checkpoint, while retaining the Cantonese reasoning behaviour absent from that checkpoint. We release the model checkpoints, the training environments, and a thirteen-year Traditional Chinese Common Crawl dataset. The models can be accessed at this https URL
Computation and Language (cs.CL)
arXiv:2609.06970 [cs.CL]
(or arXiv:2609.06970v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.06970 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.
Full text under the publisher's open licence · 277 words
Related coverage
Multilingual Bridges: How Data Mixing Unlocks In-Language Reasoning
Cohere · Oct 6, 2026
AI Futures: Racing to Lose, A sermon
LessWrong · Sep 28, 2026
Enhancing Low-Resource Language Reasoning via High-Resource Language Feature Transfer
arXiv · Sep 1, 2026
Academia is for Ambition — Alex Zhang, MIT
Latent Space · Oct 2, 2026