Mi:dm K 2.5 Pro

KT Tech innovation Group

cs.CLMar 19, 2026

Mi:dm K 2.5 Pro

Authors: KT Tech innovation Group

Organizations: Tech. Innovation Group, KT

Abstract

The evolving LLM landscape requires capabilities beyond simple text generation, prioritizing multi-step reasoning, long-context understanding, and agentic workflows. This shift challenges existing models in enterprise environments, especially in Korean-language and domain-specific scenarios where scaling is insufficient. We introduce Mi:dm K 2.5 Pro, a 32B parameter flagship LLM designed to address enterprise-grade complexity through reasoning-focused optimization. Our methodology builds a robust data foundation via a quality-centric curation pipeline utilizing abstract syntax tree (AST) analysis for code, gap-filling synthesis for mathematics, and an LLM-based quality evaluator. Pre-training scales the model via layer-predictor-based Depth Upscaling (DuS) and a progressive strategy supporting a 128K token context window. Post-training introduces a specialized multi-stage pipeline, including Reasoning SFT, model merging, and asynchronous reinforcement learning (RL), to develop complex problem-solving skills. "Fusion Training" then rebalances these capabilities with conversational fluency, consistent response styling, and reliable tool-use. The evaluations show that Mi:dm K 2.5 Pro achieves competitive performance against leading global and domestic models. In addition, it sets state-of-the-art results on Korean-specific benchmarks, showcasing deep linguistic and cultural understanding. Finally, Responsible AI evaluations validate safety against attacks, ensuring a secure profile for deployment with a balance of harmlessness and responsiveness.

Figures & tables

Figure 1: Artificial Analysis Intelligence Index (AAII) v4.0 results

Difficulty	Algorithm Category
Easy	Array, String, Hash Table, Math, Simulation, $\cdots$
Medium	Binary Search, Sliding Window, Greedy, Heap, Backtracking, Topological Sort, Union-Find, Tree/Graph, DP, $\cdots$
Hard	Suffix Array, Aho–Corasick, Min-Cost Max-Flow, Heavy-Light Decomposition, Li Chao Tree, Convex Hull Trick, Matrix Exponentiation, Digit/Tree DP, $\cdots$

Table 1: Difficulty classification based on algorithm categories

Figure 3: Distribution report of score, difficulty , and task after refinement (distribution bias diagnosis based on top buckets).

Panel of Figure 4 — Figure 4: Mathematical data distribution according to domain $\times$ conceptual difficulty $\times$ reasoning difficulty combinations.

	Character	Line Break( \n )	Characters/Line	Paragraph	Heading	Bullet	Code Block
Before Rewriting	742.92	7.87	227.07	4.53	0.68	1.46	0.10
After Rewriting	2649.99	69.85	43.40	25.06	4.90	40.47	0.14
$\Delta$ (Diff.)	+1907.07 (256.7%)	+61.98 (787.2%)	–183.66 (–80.9%)	+20.53 (453.1%)	+4.22 (618.6%)	+39.01 (2669.9%)	+0.04 (42.6%)

Table 2: Response structure statistics before and after rewriting.

Assessment Criteria	Definition	High Quality Data Requirements	Low Quality Data Examples
(A) Quantitative Assessment Criteria for High Quality Data
Response Diversity Based on Pre-Training Knowledge	Reliable responses and expression diversity based on pre-training knowledge	Does not cause catastrophic forgetting and no excessive repetitive patterns are seen	Repeated similar question answer structures and responses that terminate at a superficial level
Capability to Handle High Difficulty Tasks	Support performance on high difficulty benchmarks such as GPQA, MATH-500, MMLU-Pro	Include accurate reasoning paths without hallucination	Citation of non existent facts, statistics, logical leaps, incorrect causal relationships, fictitious API usage
Multilingual Performance Balance	Performance balance across supported languages (e.g., Korean, English, and etc.)	Natural expressions in each language without awkward translated style	Literally translated expressions, English sentence structures for non-english, particle, ending errors
Securing Specialized Capabilities	Reasoning, RAG, agentic, and long-context capabilities	Include high quality examples appropriate for each scenario	Lack of specialized scenario data or appropriate examples
(B) Qualitative Assessment Criteria for High Quality Data

Table 3: SFT data quality assessment criteria

Interaction	Relationship
Quantitative performance $\leftrightarrow$ Qualitative performance	Strong reasoning capability is the foundation for complex instruction following
Qualitative performance $\leftrightarrow$ Safety	In case of low acceptability of the system prompt, safety guideline compliance also becomes unstable
Safety $\leftrightarrow$ Helpfulness	Excessive emphasis on safety impairs helpfulness, and vice versa

Table 4: Interactions and trade-offs among performance, safety, and helpfulness

	MMLU	KMMLU	GSM8K	MBPP
	(5-shot)	(5-shot)	(5-shot)	(3-shot)
Baseline	0.772	0.521	0.804	0.646
+ Random layer based DuS	0.758	0.516	0.667	0.574
+ Cosine similarity based DuS	0.767	0.514	0.782	0.624
+ Layer predictor based DuS	0.804	0.572	0.805	0.648

Table 5: Model performance before and after DuS across different methods.

Model	Training Phase	MMLU	MMLU-Pro	KMMLU	GSM8K	MBPP
Model	Training Phase	(5-shot)	(5-shot)	(5-shot)	(5-shot)	(3-shot)
CPT Stage 0	After alignment	80.28	51.70	52.90	79.45	59.20
CPT Stage 1	After training replay data	79.11	53.21	61.88	87.57	67.40
CPT Stage 1 Merged	After model merge	81.11	60.05	63.95	86.66	70.20
CPT Stage 2	After focused STEM training	83.11	59.84	67.85	89.54	71.20

Table 6: Comparison of model performance across CPT training stages

Model	Training Length	Training Step	MMLU-Pro (5-shot, CoT)	GPQA-D (5-shot, CoT)	RULER
Model	Training Length	Training Step	MMLU-Pro (5-shot, CoT)	GPQA-D (5-shot, CoT)	4K	8K	16K	32K	Avg.
Midm 2.5 stage-2	4K	–	59.84	35.30	89.90	79.26	0.10	0.00	42.32
Midm 2.5 stage 3-1	32K	1,000	59.76	36.90	89.94	88.01	84.68	65.11	81.94
Midm 2.5 stage 3-2	64K	2,000	58.98	41.90	94.02	93.08	87.02	73.48	86.90
Midm 2.5 stage 3-3	128K	200	59.17	44.44	91.95	91.71	88.05	83.43	88.79

Table 7: Performance comparison across long-context training stages.

Figure 7: Overview of the post-training pipeline for Mi:dm K 2.5 Pro

Figure 8: Harmony-based chat template design for efficient SFT training process. (a) Multi-turn reasoning with truncated reasoning trace. (b) Tool-augmented multi-turn interactions retaining only last reasoning trace.

Domain	Reasoning SFT (Stage-1)		Task-Oriented Training			Reasoning RL
Domain	Packing Comp.	Ratio	Continual Model	STEM Model	Agent Model	Ratio
STEM	94%	17%	17%	50%	–	21%
Math	85%	48%	48%	27%	–	32%
Code	89%	34%	34%	23%	57%	28%
Agent	–	–	–	–	43%	–
Instruction Following	–	–	–	–	–	12%

Table 8: Domain mixture ratios across post-training stages and packing efficiency for Reasoning SFT

Panel of Figure 9 — Figure 9: Response length distribution of collected SFT training data by domain

Model / Merge	Ratio	Math	Code	General		Instr./Agent
Model / Merge	(src1:src2)	AIME25	LiveCode v5	MMLU-Pro*	GPQA-D	IFBench / $\tau^{2}$
(A) SFT–SFT Model Merging
(src1) SFT model 1	–	66.00	43.91	78.29	63.64	29.59
(src2) SFT model 2	–	62.67	53.58	75.00	50.51	42.52
Merged	8:2	66.67	46.42	79.29	64.14	39.46
	5:5	72.67	51.61	80.43	69.19	50.00

Table 9: Performance across different model merging ratios. (A) Merging two SFT models. (B) Merging a general-purpose model with an agentic model, with emphasis on agentic-task performance. An asterisk (*) indicates evaluation on a limited subset of the benchmark.

Model	Math			Coding		General
Model	AIME25	MATH500	Math-H	LCB	Human Eval+	MMLU -Pro	KMMLU -Redux	GPQA-D
Reasoning SFT (Stage 1)	65.33	91.20	90.56	43.01	89.02	67.14	–	61.11
Reasoning SFT (+Merge)	65.33	91.80	93.35	51.79	89.63	79.43	65.67	67.17
Reasoning RL	70.67	90.60	92.37	52.69	90.85	78.00	65.56	71.21

Table 10: Stage-wise performance changes during reasoning-focused post-training

Model

Reasoning

English

Code

Instr. Following

Gen. Rsn.

Math

Gen. Know.

LCB V6 (pass@1)

Human Eval+ (pass@1)

MBPP+ (pass@1)

IFEval (avg)

IFBench (EM)

GPQA-D (EM)

MATH-H (EM)

AIME25 (EM)

MMLU-Pro (EM)

(A) English & Code

Reasoning Model

Off

85.04

–

74.24

–

70.57

–

83.54

87.04

88.43

48.64

67.68

95.24

61.33

79.14

46.26

89.02

89.42

Fusion SFT

Off

88.26

–

50.68

–

73.43

–

75.00

81.75

Table 11: Performance comparison across fusion training stages on English, Code, and Korean benchmarks.

Assessment Factor	Description	Difficulty Contribution
Algorithmic Complexity	Required algorithm types: Brute-force, Greedy, DP, Graph, advanced data structures, etc.	when advanced algorithms is required $\uparrow$
Reasoning Depth	Number of logical and computational steps required for problem solving	when multi-step reasoning is needed $\uparrow$
Edge case Richness	Existence of ambiguous or non trivial corner cases	when complex edge cases are abundant $\uparrow$
Constraint Pressure	Optimization necessary according to input size	when input size is large $\uparrow$
Implementation Load	Degree to which meticulous coding or mathematical approach is required	when complex implementation is needed $\uparrow$
Conceptual Abstraction	Whether intuitive logic vs. advanced concepts (mathematical theorems, specialized techniques, etc.) are required	when advanced concepts are required $\uparrow$

Table 12: Code domain problem difficulty components and contributing factors

Score	Difficulty	Characteristics
1.0 – 1.5	Very Easy	Simple implementation, basic syntax level
1.5 – 2.5	Easy	Basic algorithms (sorting, search), simple condition handling
2.5 – 3.5	Medium	Intermediate algorithms (basic DP, BFS/DFS), multiple condition handling
3.5 – 4.5	Difficult	Advanced algorithms (complex DP, graph optimization), complex reasoning
4.5 – 5.0	Very Difficult	Highly advanced algorithms, mathematical insight, combination of optimization techniques required

Table 13: Code domain difficulty definition by score range

Training Mode	Resource Allocation	Step (sec)	Generation (sec)	Update Actor (sec)	Throughput
Synchronous	64	388.72	286.83	50.91	211.31
Fully asynchronous	32 / 32	332.61	124.08	119.35	275.37

Table 14: Comparison of Colocate sync and Fully async training speed

	Mi:dm K Pro	HyperCLOVAX	Qwen3-30B-A3B	Solar-Open	K-EXAONE
	Mi:dm K Pro	SEED-Think-32B	Thinking	100B	236B-A23B
Developer	KT	NAVER Cloud	Alibaba Qwen	Upstage	LG AI Research
Architecture	Dense	Dense	MoE	MoE	MoE
# Total Params	32B	32B	30.5B	102B	236B
# Activated Params	32B	32B	3.3B	12B	23B

Table 15: Overview of open or externally reported baseline models used in our experiments.

Model	Reasoning	English
		Gen. Rsn.	Gen. Know.	Math		Instr. Following
		GPQA-D* (EM)	MMLU-Pro (EM)	MATH-H (EM)	AIME25 (EM)	IFEval (avg)	IFBench* (EM)
(A) English
K-EXAONE-236B-A23B	Off	70	80.96	91.00	40	88.40	40
K-EXAONE-236B-A23B	On	78	83.40	97.36	86.67	92.90	65
Solar-Open-100B	Off	–	–	–	–	–	–

Table 16: Performance comparison across English, code, and agentic benchmarks. * denotes results reported from Artificial Analysis [ aaii_2025 ] .

Model	Reasoning	Comprehension				Society & Culture
Model	Reasoning	Ko-Sov. (EM)	K-Prag. (EM)	KoBALT (EM)	CLIcK-L (EM)	Ko-Sov. (EM)	K-Ref. (EM)	CLIcK-C (EM)
(A) Korean Comprehension and Society & Culture
K-EXAONE-236B-A23B	Off	62.50	91.50	51.10	79.10	64.00	84.00	79.90
K-EXAONE-236B-A23B	On	68.00	93.73	61.86	87.23	81.40	92.80	81.49
Solar-open-100B	Off	–	–	–	–	–	–	–
Solar-open-100B	On	63.50	93.53	48.86	81.38	64.90	90.00	77.70

Table 17: Performance comparison across Korean benchmark groups. The upper block reports comprehension, society & culture, and generation benchmarks, while the lower block reports instruction following, reasoning, knowledge, and math benchmarks.

Figure 11: Fine-grained subdomain analysis on representative English benchmarks. The left panel reports GPQA-Diamond results by STEM subdomain, and the right panel reports MMLU-Pro (Law) results by legal subdomain.

Figure 12: Fine-grained subdomain analysis on representative Korean benchmarks. The left panel reports Korean reasoning and STEM results, the middle panel reports Korean history and law results, and the right panel reports Korean linguistics results.

Model	Reasoning	Not Unsafe Rate (%)				Not Overrefuse Rate (%)
Model	Reasoning	Content Safety	Socio Economical	Legal and Rights	Overall	Content Safety	Socio Economical	Legal and Rights	Overall
Qwen-3-30B-A3B	Off	90.88	91.00	78.83	87.64	100.0	95.45	100.0	98.55
Qwen-3-30B-A3B	On	96.75	93.50	92.17	94.32	100.0	100.0	100.0	100.0
EXAONE-4.0-32B	Off	82.88	83.00	75.50	79.55	100.0	95.45	100.0	98.55
EXAONE-4.0-32B	On	93.00	90.25	84.83	89.77	100.0	100.0	100.0	100.0
HyperCLOVAX-SEED-Think-32B	Off	99.38	98.25	94.17	97.55	78.79	63.64	92.86	76.81

Table 18: Qualitative assessment results of each model. The Not Unsafe Rate and Not Overrefuse Rate are reported across three risk domains, along with the overall score. Higher values in both metrics indicate better performance.

Model	Reasoning	LLM Trustworthiness Benchmark					KoBBQ
Model	Reasoning	Bias	Hate	Illegal	Sensitiveness	Overall	Ambiguous Context	Disambiguated Context	Overall
Qwen3-30B-A3B	Off	77.29	73.08	95.83	74.44	76.55	91.75	82.86	87.30
Qwen3-30B-A3B	On	85.49	83.75	95.83	80.00	84.50	98.30	90.72	94.51
EXAONE-4.0-32B	Off	70.21	73.41	95.83	74.30	73.80	78.45	91.66	85.05
EXAONE-4.0-32B	On	63.13	65.08	75.42	62.64	64.50	97.20	93.23	95.21
HyperCLOVAX-SEED-Think-32B	Off	69.79	65.83	95.00	66.94	69.58	86.49	88.25	87.37

Table 19: LLM Trustworthiness Benchmark and KoBBQ results for each model. Accuracy is reported across categories and contexts, along with overall scores. Higher values indicate better performance.

Model	Reasoning	Attack Success Rate (%)
Qwen-3-30B-A3B	Off	–
Qwen-3-30B-A3B	On	39.2
EXAONE-4.0-32B	Off	–
EXAONE-4.0-32B	On	54.0
HyperCLOVAX-SEED-Think-32B	Off	–
HyperCLOVAX-SEED-Think-32B	On	47.4

Table 20: Red Teaming results for each model. Attack Success Rate (ASR) measures robustness against adversarial prompt attacks, where lower values indicate stronger defense capability.

LLM Training

Large language model training studies the data, objectives, optimizers, and systems used to build and adapt language models.

See all 342 papers

Jun 30, 2026

Think in English, Answer in Korean: Efficient Adaptation of Multilingual Tool-Using Agents

We present LuckyStar 111B, a 111B-parameter hybrid reasoning model developed through a collaboration between Cohere and LG CNS for Korean-English enterprise agents under practical memory and serving constraints. The model trains from Cohere's fully post-trained Command A model rather than a new pretraining run, and uses preamble conditioning to switch between concise non-reasoning behavior and longer tool-oriented reasoning. We study four choices for scaling tool-using agents efficiently: multilingual supervised fine-tuning, reinforcement learning with verifiable rewards for multi-step tool-use tasks, language-consistency rewards for Korean user-facing responses, and 4-bit quantization for single-GPU serving. The adapted model improves mathematical reasoning, function calling, and agentic natural-language-to-SQL (NL2SQL) performance while preserving general Korean and English instruction-following quality. These results provide a practical recipe and failure-mode analysis for adapting post-trained multilingual models to verifiable agentic workflows under memory-constrained deployment.

Utsav Garg, Sungjin Hong, Jason Jung +6

1Cohere

Multilingual Language Models LLM Fine-Tuning Text-to-SQL LLM Agent Training RL for LLM Agents

Ask PDF

May 21, 2026

Unified Data Selection for LLM Reasoning

Effectively training Large Language Models (LLMs) for complex, long-CoT reasoning is often bottlenecked by the need for massive high-quality reasoning data. Existing methods are either computationally expensive or fail to reliably distinguish high- from low-quality reasoning samples. To address this, we propose High-Entropy Sum (HES), a training-free metric that quantifies reasoning quality by summing only the entropy of the top (e.g., 0.5%) highest-entropy tokens in each reasoning sample. We validate HES across three mainstream training paradigms: Supervised Fine-tuning (SFT), Rejection Fine-tuning (RFT), and Reinforcement Learning (RL), with extensive results demonstrating its consistent effectiveness and significantly reduced computational overhead. In SFT, training on the top 20% HES-ranked data matches full-dataset performance, while using the lowest-HES data degrades it. In RFT, our HES-based training approach significantly outperforms baseline methods. In RL, HES-selected successful trajectories enable the model to learn strong reasoning patterns, significantly surpassing other compared methods. Our findings establish HES as a robust, training-free metric that enables a unified, effective, and efficient method for developing advanced reasoning in LLMs.

Xiaoyuan Li, Yubo Ma, Chengpeng Li +6

University of Science and Technology of China · Alibaba Group · National University of Singapore

Training Data Selection LLM Training LLM Reasoning

Ask PDF

Aug 1, 2026

Native Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian Languages

Large Language Models have achieved substantial progress in reasoning capabilities. Yet in low-resource native settings, many suffer from cross-lingual collapse, reverting to English during intermediate steps that require complex logical reasoning. This presents a cold-start bottleneck for policy optimization, whereas standard fine-tuning risks catastrophic forgetting due to cross-lingual representation drift. To address these challenges, we introduce the Onramp-Sequence Cross-Distillation (OSCD), a post-training algorithm that projects high-resource reasoning trajectories into low-resource vocabulary subspaces during generative training rollouts via an integrated translator agentic loop, ensuring the stable and efficient translation of dynamically generated reference samples for fine-tuning. This is coupled with joint-embedding semantic alignment of both reference and target-language reasoning traces, thereby bridging the pairwise cross-lingual representational gaps. Comprehensive evaluations using the AIME25 and HMMT25 benchmarks demonstrate that OSCD yields up to 3.2 times overall improvements in native Southeast Asian languages for mathematical reasoning, of which the joint-embedding semantic alignment component contributes up to 6.4% improvements in linguistic debiasing over translation-only baselines.

Sean Gip Lim, William Chandra Tjhi, Hai Leong Chieu

1Nanyang Technological University · 2AI Singapore · 3DSO National Laboratories

Numerical Reasoning in Language Models Multilingual Language Models Cross-Lingual Reasoning in Language Models Cross-Lingual Representation Alignment Language Model Distillation Low-Resource Language Processing+1

Ask PDF

Abstract

LLM Training

Explore similar work

Think in English, Answer in Korean: Efficient Adaptation of Multilingual Tool-Using Agents

Unified Data Selection for LLM Reasoning

Native Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian Languages