KOREAson/KO-REAson-AX3_1-7B-0831

KO-REAson

KO-REAson is a series of Korean-centric reasoning language models developed in collaboration with OneLineAI, KISTI-KONI, HAE-RAE and ORACLE.

We use the Language-Mixed Chain-of-Thought (CoT) approach, which allows the model to alternate between English and Korean during the “Think” stage of reasoning, preserving key Korean terms while leveraging English for logical scaffolding.

Top-performing models of our series KO-REAson-AX3_1-7B-0831 (KONI-7B-R-20250831) and KO-REAson-7B-Q2_5-0831 show performance comparable to models trained on closed-source datasets such as Exaone-Deep-7.8B.

Left: Average performance (Held-out-Ko) of open models trained on closed or open data; our models are highlighted in green.

Model Details

The KO-REAson-0831 family comes in six variants based on the base model used.

Model (link)	Base	Notes
KO-REAson-L3_1-8B-0831	Llama-3.1-8B	L3_1 → Llama-3.1-8B
KO-REAson-KL3_1-8B-0831	Koni-Llama-3.1-8B	KL3_1 → Koni-Llama-3.1-8B; also called KONI-Llama3.1-8B-R-20250831
KO-REAson-G3-4B-0831	Gemma-3 4B	G3 → Gemma-3-4B
KO-REAson-AX3_1-7B-0831	A.X.-3.1-Light (≈7B)	AX3_1 → A.X.-3.1-Light; also called KONI-7B-R-20250831
KO-REAson-K2505_8B-0831	Kanana-2505 (8B)	K2505 → Kanana-2505
KO-REAson-7B-Q2_5-0831	Qwen-2.5 (7B)	Q2_5 → Qwen-2.5

Performance

Evaluation Datasets

The model's performance was evaluated across a total of 11 benchmarks, and the evaluation suite is divided into two parts: (You can check these benchmarks in HAERAE-HUB/KoSimpleEval)

• Held-in: This set of benchmarks is used for routine monitoring of the model's performance during the training and ablation study phases.
• Held-out: This set is used only once to evaluate the final model after all training and ablations are complete.

This separation is designed to prevent inadvertent overfitting to the benchmarks during the iterative training process and to provide a more accurate measure of the model's generalization capabilities.

Category	Held-in	Held-out
General Knowledge	KMMLU-Redux	KMMLU-HARD, KMMLU-Pro
Reasoning	MCLM	KSM, GPQA, AIME2024, AIME2025
Korean-specific	HAE-RAE Bench	CLIcK, KoBALT-700

Comparison with models trained on public datasets

Models	# Instances	Methodology	Held-Out (Ko)	Held-Out (En)	Total
KO-REASon-AX3_1-7B-0831(KONI-7B-R-20250831; Ours)	260k	SFT	44.6	41.2	43.3
KO-REASon-7B-Q2_5-0831(Ours)	260k	SFT	45.10	38.75	49.95
KO-REAson-KL3_1-8B-0831(KONI-Llama3.1-8B-R-20250831)	260k	SFT	40.13	30.57	43.66
Open Recipe (En)
OpenThinker3-7B	1.2M	SFT	33.6	55.5	41.8
s1.1-7B	1k	SFT	35.6	23.4	31.1
Llama-3.1-Nemotron-Nano-8B-v1	>3M	SFT & RL	27.0	44.1	33.4
Open Recipe (Ko)
Ko-R1-14B	45k	SFT	43.7	46.3	44.7
Ko-R1-7B	45k	SFT	27.3	36.1	30.6
LLaMa-3.1-Ko-Reasoning-8B	63k	SFT	17.7	7.7	14.0

Held-out benchmark performance

Model	Model Size	General		Reasoning				Korean-Specific		Average (Held-out)	Average (Held-out-Ko)
		KMMLU-HARD	KMMLU-Pro	KSM	AIME 2024	AIME 2025	GPQA	CLIcK	KoBALT-700
Llama-3.1-Nemotron-Nano-8B	8.03	21.47	22.89	47.06	56.67	43.33	32.32	34.54	9.29	33.45	27.05
Llama-3.1-Korean-Reasoning-8B-Instruct	8.03	14.91	21.72	6.09	0.00	0.00	23.23	39.65	6.14	13.97	17.70
EXAONE-Deep-7.8B	7.82	40.96	37.35	70.80	70.00	63.33	64.65	54.24	18.86	52.52	44.44
DeepSeek-R1-Distill-Qwen-7B	7.62	0.00	23.00	56.09	60.00	40.00	43.43	0.00	8.29	28.85	17.48
DeepSeek-R1-Distill-Llama-8B	8.03	23.22	26.26	29.97	33.33	20.00	46.46	39.05	13.29	28.95	26.36
s1.1-7B	7.62	31.16	37.70	30.60	16.67	23.33	30.30	56.84	21.86	31.06	35.63
OpenThinker3-7B	7.62	30.31	26.26	63.59	66.67	53.33	46.46	47.69	10.14	35.63	30.60
Ko-R1-7B	7.61	28.46	19.31	51.61	46.67	33.33	28.28	32.48	4.71	30.61	27.31
KO-REAson-KL3_1-8B-0831(KONI-Llama3.1-8B-R-20250831)	8.03	44.64	40.08	37.96	23.33	30.00	38.38	56.39	21.57	30.57	40.13
KO-REASon-AX3_1-7B-0831 (KONI-7B-R-20250831)	7.26	45.57	38.13	52.80	53.33	33.33	36.87	62.86	23.43	43.29	44.56
KO-REASon-7B-Q2_5-0831	7.26	46.81	44.93	48.11	43.33	30.00	42.93	60.65	25.00	42.72	45.10

Citation

The paper will be released soon!

Contact

For any questions contact us via the following email :)

spthsrbwls123@yonsei.ac.kr

Acknowlegments

This research was supported by the Korea Institute of Science and Technology Information (KISTI) (No.(KISTI) K25L1M1C1), aimed at developing KONI (KISTI Open Neural Intelligence), a large language model specialized in science and technology.