PaperlayerView exact revision on arXiv

Source-linked AI summary

K-EXAONE 2.0 Technical Report

Eunbi Choi, Kibong Choi, Sehyun Chun, Seokhee Hong, Junwon Hwang, Hyojin Jeon, Ahra Jo, Hyunjik Jo, Yeonsik Jo, Minhyeok Jung, Doyoung Kim, Heegyu Kim, Joonkee Kim, Seonghwan Kim, Soyeon Kim, Sunkyoung Kim, Yireun Kim, Yongil Kim, Byungoh Ko, Changhun Lee, Dohaeng Lee, Haeju Lee, Jinsik Lee, Kyungmin Lee, Minwoo Lee, Wonkee Lee, Sangha Park, Sungjune Park, Kwangrok Ryoo, Kijung Seo, Minju Seo, Yongwoo Song, Sejong Yang, Heuiyeen Yeen, Stanley Jungkyu Choi, Yemuk Choi, Yongchan Chun, Jiwon Ham, Dasol Hong, Sujeong Im, Kijeong Jeon, Gerrard Jeongwon Jo, Hyeongjun Jo, Yujin Jo, Jiyeon Jung, Naeun Kang, Daeseong Kim, Euisoon Kim, Hayeon Kim, Hyosang Kim, Myoungshin Kim, Unsol Kim, Youchul Kim, Chaeeun Lee, ChaeYoon Lee, Edward Hwayoung Lee, Honglak Lee, Hwansoo Lee, Minkyung Lee, Sangeun Lee, Solji Lim, Woohyung Lim, Chanwoo Moon, Jueun Mun, Jimin Park, Seojeong Park, Yongmin Park, Hyerin Seo, Donghyeon Shin, Donghyun Son, Eunyong Son, Kaehyun Um, Sihoon Yang, Chang En Yea, Sihyuk Yi, Kyungjae Yoo, Chansik Yoon

arXiv:2608.04505v1cs.CL

TL;DR

K-EXAONE demonstrated domestic development of a globally competitive model, while extending its scale and deployment-oriented capabilities toward the global frontier remained a challenge. K-EXAONE 2.0 upcycles that model and shows more than 10% average improvement over its predecessor across 24 benchmarks in nine categories.

  • Problem

    K-EXAONE demonstrated domestic development of a globally competitive model, while extending its scale and deployment-oriented capabilities toward the global frontier remained a challenge.

    Source: §1 · p. 2

  • Method

    The report upcycles K-EXAONE by expanding its depth and expert capacity, then continues large-scale pre-training.

    Source: §1 · p. 2, §2.2 · p. 4

  • Results

    More than 10% average improvement over K-EXAONE across 24 benchmarks in nine categories, with largest gains in coding and agentic coding and long-context understanding, and clearest open-weight advantages in retrieval and safety.

    Source: §9 · p. 14

  • Takeaways & Limitations

    The Apache 2.0 release lets the wider AI ecosystem independently evaluate, deploy, adapt, and build upon K-EXAONE 2.0.

    Source: §9 · p. 15

    Problematic content may remain in training data and potentially lead K-EXAONE 2.0 to generate undesirable responses.

    Source: §7 · p. 14

Abstract

from arXiv · show

This technical report presents K-EXAONE 2.0, an open-weight multilingual foundation model developed by LG AI Research as a step in our effort toward global frontier-scale foundation models. Rather than training from scratch, we upcycle K-EXAONE and expand its architecture, yielding a Mixture-of-Experts (MoE) model with 750B total parameters and approximately 37B activated per token---more than three times the capacity of its predecessor. K-EXAONE 2.0 supports context lengths of up to 256K tokens and expands multilingual coverage from six to ten languages. Its training pipeline combines continual pre-training, difficulty-focused mid-training, and post-training to strengthen reasoning, agentic coding, multilingual capability, and safety grounded in Korean sociocultural contexts. Across nine evaluation categories selected to reflect the conditions of practical use, K-EXAONE 2.0 improves over K-EXAONE and remains competitive with open-weight models, showing its largest gains in agentic coding and long-context understanding and its clearest strengths in long-context retrieval and safety. Released under the Apache 2.0 license, K-EXAONE 2.0 enables the wider AI ecosystem to evaluate, deploy, adapt, and build upon it, while marking the beginning---rather than the endpoint---of our challenge toward the global frontier.

1 Introduction · 2 Modeling

K-EXAONE 2.0 upcycles K-EXAONE into a 750B-parameter MoE model with about 37B parameters activated per token, retaining support for 256K-token contexts while expanding its architecture and inference options. The report describes targeted training and evaluation, with improvements over K-EXAONE and strengths in long-context retrieval and safety.

2 Modeling

  • K-EXAONE 2.0 expands K-EXAONE from 236B to 750B total parameters while increasing activated parameters from 23B to 37B per token, using upcycling rather than training from scratch.It retains hybrid sliding-window and global attention for contexts up to 256K tokens and uses MTP and DSpark for speculative decoding.Source: §1 · p. 2, §2 · pp. 2–3

1 Introduction

  • K-EXAONE 2.0 improves consistently over K-EXAONE, with its largest gains in agentic coding and long-context understanding and clearest advantages among comparable open-weight models in long-context retrieval and safety.It was evaluated across nine practical-use categories and released under the Apache 2.0 license.Source: §1 · p. 2

2.1 Model Configurations

  • The model has 78 layers, with two dense layers followed by 76 MoE layers; each MoE layer selects eight routed experts from 256 alongside one shared expert.It retains hybrid attention and uses a 4096-token sliding window in the second layer and 128 tokens in other sliding-window layers.Source: §2.1 · p. 3, Figure 2 · p. 3

2.2 Model Upcycling

  • Upcycling expands depth from 48 to 78 layers and width from 128 to 256 experts, repeating middle-stack LLLG blocks and adding norm-preserving random rotation noise to duplicated experts.The activated experts, sliding-window size, routing policy, and tokenizer are inherited unchanged.Source: §2.2 · p. 4
  • Clamping SwiGLU branches at 7.0 limits peak activation to 48.96, whereas without the clamp it reaches 6862 at the last layer.The clamp addresses activation growth in deeper layers that degraded low-precision training and serving.Source: Figure 3 · p. 4, §2.2 · pp. 4–5

2.3 Inference Optimization via Speculative Decoding

  • MTP is trained jointly with the backbone and refined using total-variation loss before RL, while DSpark uses a five-layer block-diffusion drafter to generate seven-token draft blocks in one pass.A lightweight sequential module then refines each DSpark block left to right; DSpark is trained on target-generated data.Source: §2.3 · p. 5
  • With the same draft budget, DSpark improves acceptance length by 32–66% over MTP and raises end-to-end speedup from 1.27–1.77× to 1.81–2.57×.The comparison used the same K-EXAONE 2.0 FP8 target and draft budget γ = 7.Source: Table 2 · p. 5, §2.3 · p. 5

3 Continual Pre-training

K-EXAONE 2.0’s continual pre-training combines scaled training with new synthetic-data methods and targeted Korean and multilingual data. The reported findings favor Active Reading for knowledge-intensive domains, institution-sourced data for Korean culture and history, and show decreasing Korean-subset loss across the K-EXAONE lineage.

3 Continual Pre-training

  • K-EXAONE 2.0 applies depth upscaling, a healing stage, and an additional 8T tokens of training using its pre-training data mixture.The recipe retains K-EXAONE’s data composition and filtering approach while adding Active Reading [33] and thinking-augmented pre-training with latent intermediate reasoning [46].Source: §3 · p. 5

3.1 Knowledge-Oriented Synthetic Data

  • Active Reading achieved the largest average improvement over the initial checkpoint, while textbook-style generation performed best on MMLU and GSM8K.The comparison used independently trained models from the same late-stage K-EXAONE checkpoint; the results motivated selective Active Reading and reusable prompts in a lightweight pseudo-AR pipeline.Source: §3.1 · pp. 5–6, Table 3 · p. 6

3.2 Korean Data

  • Korean institutional data performed better on culture and history benchmarks, while open-source educational data scored higher on knowledge-intensive benchmarks.The findings suggest general knowledge transfers across languages, whereas country-specific historical and cultural knowledge benefits from high-quality local sources.Source: Table 4 · p. 6, §3.2 · p. 6

3.3 Multilingual Data

  • Continual pre-training adds French, Italian, Polish, and Portuguese to the six languages supported by K-EXAONE, and incorporates translation and synthetic QA data.High-quality organic corpora are built from FineWeb2 using an in-house filtering pipeline to strengthen language capabilities and cross-lingual knowledge transfer.Source: §3.3 · p. 6

3.4 Compute Scaling on Korean Data

  • Korean-subset pre-training loss decreases consistently from the small-scale K-EXAONE model through K-EXAONE 2.0.The trend may reflect increased model capacity and compute as well as the growing volume, coverage, and diversity of Korean data; thinking-augmented and latent-thought data may contribute, but their effect is not isolable.Source: §3.4 · p. 7, Figure 4 · p. 7

4 Mid-training

Mid-training progressively extends context from 8K to 256K tokens while targeting reasoning, long-context understanding, and agentic capabilities. It combines longer-context data with advanced reasoning and agent-workflow training, and reports gains in HLE and perfect NIAH retrieval across the evaluated range.

4 Mid-training

  • Mid-training uses two sequential 400B-token stages, moving from a 64K context window to windows exceeding 64K and ultimately up to 256K tokens.The longer contexts support complete reasoning trajectories, repository-level code dependencies, and multi-step tool-use workflows; Mid-Stage 2 increases emphasis on long-form reasoning, repository code, and agentic workflows.Source: §4 · p. 7

4.1 Long-Context Adaptation Data

  • K-EXAONE 2.0 achieves perfect NIAH retrieval across evaluated needle positions and context lengths up to 256K tokens.The model is trained on complete documents, large code repositories, and extended tool-use and interaction trajectories with minimal truncation; the share of long-context examples increases from Mid-Stage 1 to Mid-Stage 2.Source: §4.1 · pp. 7–8, Figure 5 · p. 8

4.2 Reasoning-Centric Data

  • 4.71 points on HLE [42] separate Mid-Stage 2 from Mid-Stage 1; SFT adds 0.95 after Mid-Stage 2, totaling 5.66 points over Mid-Stage 1.SFT directly after Mid-Stage 1 yields a 3.15-point gain. Mid-Stage 2 adds data identified by retrained classifiers and an internal search agent for advanced knowledge and reasoning, particularly in STEM.Source: §4.2 · p. 8, Table 5 · p. 8

4.3 Agent Workflow Data

  • Agent-workflow data targets planning, reflection, summarization, and information aggregation, while coding-agent data draws on repository-level GitHub pull requests.The coding examples combine PR descriptions, patch diffs, reviewer comments, relevant files, and commit history across bug-fixing and feature-implementation scenarios; expert models generate agentic tool-use trajectories.Source: §4.3 · p. 9

4.4 Tool-Calling Formats

  • Tool-calling training uses XML-style invocation and diverse templates and response formats to promote format-invariant tool use.The report states that exposure to varied tool-calling forms is intended to improve generalization across heterogeneous real-world tool environments.Source: §4.4 · p. 9

5 Post-training

Post-training combines broad supervised fine-tuning with targeted preference learning to strengthen reasoning, agentic capabilities, instruction following, and safety. Its design spans data construction, preserved reasoning behavior, and compliance review.

5.1 Supervised Fine-Tuning

  • SFT aggregates 350B tokens across reasoning, world and expert knowledge, instruction following, and agentic systems, while freezing router parameters and supporting thinking and non-thinking modes.For agent training, the model’s thinking behavior is preserved to avoid suppressing reasoning capabilities.Source: §5 · p. 9, §5.1 · p. 9

5.1.1 Reasoning

  • Reasoning SFT covers mathematics, code, long-context, knowledge, and general chat, with filtering and validation to improve supervision quality and reduce artifacts and benchmark leakage.Checks include answer- or execution-based verification where available, removal of templated or duplicate trajectories, and lexical-overlap decontamination against evaluation benchmarks and public datasets.Source: §5.1.1 · pp. 9–10

5.1.2 Agent

  • Agent SFT standardizes tool calls and preserves reasoning blocks across turns; balancing depth- and breadth-oriented search queries and adding hybrids improved both capabilities in internal ablations.Breadth queries require exhaustive candidate identification, verification, de-duplication, and structured results; depth queries use indirect clues to identify hidden targets.Source: §5.1.2 · p. 10

5.2 Preference Learning

  • Preference learning uses sequential multi-task and safety-aware GROUPER optimization to improve instruction following and reasoning, then promote appropriate abstention, harmful-request refusal, and jailbreak robustness.Safety data follow K-AUT-V2, a taxonomy grounded in global ethical standards and Korean-specific considerations; reward criteria and mixture ratios are tuned against refusal accuracy and retention of prior capabilities.Source: §5.2 · p. 11
  • Multi-task preference data cover reasoning, agentic, and chat tasks, using verifiable signals, instance-specific rubrics, agent-action and answer quality, and similar-length response pairs where possible.Proof-oriented mathematics also uses an LLM-as-a-judge approach to encourage valid reasoning and accurate mathematical knowledge.Source: §5.2 · p. 11

5.3 Data Compliance

  • LG AI Research conducts AI Compliance reviews throughout data collection, model training, and information provision to minimize legal risks involving copyright, intellectual property, and personal information.The report refers readers to the EXAONE 3.0 Technical Report and LG AI Ethics Principles for further information.Source: §5.3 · p. 11

6 Evaluation

K-EXAONE 2.0 improves over K-EXAONE across several practical evaluation areas while maintaining strong performance in world knowledge, mathematics, and Korean. Results span reasoning, agentic abilities, general abilities, Korean and multilingual capabilities, and safety.

6.1 Benchmarks and Setup

  • The evaluation spans nine benchmark categories; when official baseline scores are unavailable, baselines are evaluated internally using each model’s recommended inference configuration.The detailed benchmark setup is in Appendix C.Source: §6.1 · pp. 11–12

6.2 Results

  • K-EXAONE 2.0 scores 83.5 on MMLU-PRO, 82.2 on GPQA-DIAMOND, and 18.3 on the text-only HUMANITY’S LAST EXAM subset; it improves over K-EXAONE by 3.1 and 4.7 points on the latter two benchmarks.It scores 92.3 on AIME 2026, 78.4 on HMMT FEB 2026, and 78.6 on IMO-ANSWERBENCH, improving by 2.3 points on IMO-ANSWERBENCH.Source: §6.2 · p. 12
  • K-EXAONE 2.0 scores 40.1 on SCICODE, 68.2 on SWE-BENCH VERIFIED, and 43.8 on TERMINAL-BENCH 2.1, improving over K-EXAONE by 18.8 and 13.5 points on the latter two benchmarks.It scores 80.0 on CLAW-EVAL, a 5.7-point improvement, and 14.2 on τ3-BANKING.Source: §6.2 · p. 12
  • K-EXAONE 2.0 improves over K-EXAONE by 2.7 points on IFEVAL and 5.3 points on IFBENCH, and raises its OPENAI-MRCR score from 52.3 to 94.4.Its scores on AA-LCR and KO-LONGBENCH improve to 56.2 and 89.6, respectively, indicating stronger retrieval and reasoning over long English and Korean inputs.Source: §6.2 · p. 12
  • K-EXAONE 2.0 scores 69.1 on KMMLU-PRO, 84.2 on CLICK, and 91.1 on HRM8K-KSM; across nine evaluated non-English languages, it scores 86.6 on both MMMLU and GLOBALMMLU-LITE.Its POLYMATH score is 71.3, 13.9 points above K-EXAONE, indicating enhanced multilingual mathematical reasoning.Source: §6.2 · p. 12
  • K-EXAONE 2.0 demonstrates strong performance on ROK-FORTRESS and KGC-SAFETY, while its extensible safety framework supports refining evaluation criteria and adapting to emerging attack patterns.The report identifies continued advancement of the evaluation framework as future work for addressing increasingly sophisticated safety risks.Source: §6.2 · p. 12
  • Overall, K-EXAONE 2.0 substantially improves on K-EXAONE in agentic coding, tool use, instruction following, long-context understanding, and multilingual mathematical reasoning, while maintaining strength in knowledge, mathematics, and Korean.The main results are reported for REASONING mode.Source: §6.2 · p. 12, Table 6 · p. 13

7 Limitations

K-EXAONE 2.0 may produce inappropriate or biased responses, errors, and outdated information because its outputs depend on training-data statistics. LG AI Research seeks to reduce risks and prohibits malicious use that could induce inappropriate outputs.

7 Limitations

  • Training-data-based generation may produce personal, harmful, biased, or otherwise inappropriate content, and its responses do not reflect LG AI Research’s views.Problematic content may remain in training data despite efforts to exclude it.Source: §7 · p. 14
  • Reliance on training-data statistics can yield semantically or syntactically incorrect sentences, while missing latest information can make responses false or contradictory.Source: §7 · p. 14
  • LG AI Research strives to reduce potential risks and prohibits malicious user activity that may induce outputs violating its ethical principles.The passage gives entering illegal information as an example of malicious activity.Source: §7 · p. 14

8 Deployment

Section 8 directs readers to Appendix B for K-EXAONE 2.0 model license information, which is essential for legal use.

8 Deployment

  • Appendix B provides license information essential for the legal use of K-EXAONE 2.0 models.Source: §8 · p. 14

9 Conclusion · B Model License

K-EXAONE 2.0 upcycles its predecessor into a larger, long-context model and translates that scale into gains across practical evaluations. Its Apache 2.0 release begins, rather than ends, the effort toward global frontier models.

9 Conclusion

  • Upcycling K-EXAONE expanded it to 750B total parameters with approximately 37B activated per token, retaining hybrid attention for contexts up to 256K tokens.It provides two self-speculative decoding paths on the same target weights: jointly trained Multi-Token Prediction and a DSpark drafter trained after the final weights were fixed.Source: §9 · p. 14
  • Continual pre-training, difficulty-focused mid-training, and post-training expanded coverage from six to ten languages while targeting reasoning, long-context understanding, agentic coding, tool use, reliability, and Korean-context safety.The stages stabilized the expanded architecture and extended knowledge, Korean proficiency, and multilingual foundations before refining instruction following, reasoning, factual reliability, agentic behavior, and alignment.Source: §9 · p. 14
  • Across 24 benchmarks in nine practical-deployment categories, K-EXAONE 2.0 improved more than 10% on average over K-EXAONE, with approximately 30% gains on three coding and agentic-coding benchmarks.Long-context understanding was also among the largest gains over its predecessor.Source: §9 · p. 14
  • K-EXAONE 2.0 showed its clearest advantages over comparable open-weight models in long-context retrieval and safety, but did not lead on every benchmark.Closing remaining gaps will require advances in model scale, data, training methodology, inference efficiency, safety, and evaluation.Source: §9 · p. 14
  • The Apache 2.0 release lets the wider AI ecosystem evaluate, deploy, adapt, and build on the model; the next phase targets models at and beyond trillion-parameter scale.The report frames K-EXAONE 2.0 as the beginning, not the endpoint, of its challenge toward global frontier models.Source: §9 · p. 15

Apache 2.0

  • The model is licensed under Apache License 2.0, whose terms disclaim warranties for distributed software, subject to stated legal or agreement exceptions.The license states that software is distributed on an “AS IS” basis without warranties or conditions, express or implied.Source: §9 · p. 15, Apache 2.0 · p. 17

C Evaluation Setup Details · D Preserved Thinking

This section specifies benchmark-specific evaluation protocols and describes preserved thinking, which carries reasoning traces across turns to support difficult or long-horizon agentic tasks.

C Evaluation Setup Details

  • For unspecified benchmarks, evaluation follows the official setup; the report specifies internal settings for selected benchmarks.The report says it aims to follow each benchmark’s official evaluation setup.Source: C Evaluation Setup Details · p. 18
  • Claw-Eval uses latest commit d3f02d4, with missing mock_today fields patched to make date-dependent grading deterministic across runs.Source: C Evaluation Setup Details · p. 18
  • Terminal-Bench 2.1 uses Terminus 2, temperature 1.0, top-p 0.95, up to 32,768 generated tokens per step, and a 256K context window.Tasks run with 8 CPU cores, 16GB of memory, 20GB of disk, and a 4-hour timeout.Source: C Evaluation Setup Details · p. 18
  • SWE-Bench Verified uses mini-SWE-agent [64], temperature 1.0, top-p 0.95, up to 32,768 generated tokens per step, and a 256K context window.Each instance runs with 4 CPU cores, 8GB of memory, 10GB of disk, and a 4-hour timeout.Source: C Evaluation Setup Details · p. 18
  • OpenAI-MRCR follows its official protocol, averages 2-, 4-, and 8-needle scores within bins through 128K tokens, then reports the macro-average across bins.Scores use difflib.SequenceMatcher ratio after the model prepends the provided alphanumeric hash; MRCR supports contexts up to 1M tokens.Source: C Evaluation Setup Details · p. 18
  • ROK-Fortress uses Kimi K2 Thinking, o4-mini, and GPT-5.2 as judges across four categories spanning English or Korean and US context or Korean cultural adaptation.The categories are VEn,US, VKo,US, VEn,KR, and VKo,KR, with the last representing full transcreation.Source: C Evaluation Setup Details · p. 18
  • ROK-Fortress safety score subtracts the mean TRS across four categories from 100, so higher scores indicate better safety performance.Source: C Evaluation Setup Details · p. 18

D Preserved Thinking

  • Preserved thinking carries the full reasoning trace across turn boundaries, conditioning later reasoning on prior tool results and supporting consistent long-horizon agentic reasoning.Without it, intermediate reasoning blocks are typically dropped after a turn and must be re-derived.Source: D Preserved Thinking · p. 19, Figure 6 · p. 19

E Multilingual · F Safety

K-EXAONE 2.0 expands multilingual support to ten languages while reporting balanced performance gains. Its safety work broadens language and attack coverage, and addresses the limits of known-risk red teaming by involving domain experts in risk discovery.

E Multilingual

  • K-EXAONE 2.0 supports ten languages and reports higher average performance than K-EXAONE, with gains distributed evenly across languages.The reported multilingual comparisons cover POLYMATH and GLOBALMMLU-LITE; the supplied passages provide no specific scores.Source: E Multilingual · p. 20, Table 7 · p. 20, Table 8 · p. 20
  • Translation performance is reported on WMT24++ for English and nine additional languages in both English-to-target and target-to-English directions.Detailed results are presented in Table 9, but no scores are supplied here.Source: E Multilingual · p. 20, Table 9 · p. 20

F Safety

  • K-EXAONE introduced K-AUT, a taxonomy with four domains and 226 risk areas grounded in universal human values and adapted to Korean context.It also covers foreseeable future risks, and its methodology and principal categories were disclosed in the 2025 Accountability Report on AI Ethics.Source: F Safety · p. 21
  • Safety training adds French, Italian, Polish, and Portuguese, incorporates defenses against previously reported attack methods, and tests adversarial multi-turn interactions.Source: F Safety · p. 21
  • Established-method red teaming tests whether known risks can be elicited but is comparatively weak at surfacing risks not previously considered or defined.The report attributes this blind spot chiefly to limited perspectives in risk discovery, rather than insufficient technical sophistication.Source: F Safety · p. 21
  • To address risks missed by general-purpose frameworks, the report expands risk discovery to field experts with direct experience in specialized professional domains.The aim is to improve safety in domain-specific contexts as well as against broad societal harms.Source: F Safety · p. 21

F.1 Expanding the Locus of Risk Discovery · F.2 Consultation Process with the Safety Teacher Advisory Council · F.3 K-AUT-V2 : Expanded Korea-Augmented Universal Taxonomy

K-EXAONE 2.0’s safety taxonomy was expanded through a four-week consultation with a diverse council of UNESCO-trained teachers, broadening risk coverage and clarifying the model behaviors the taxonomy governs.

F.1 Expanding the Locus of Risk Discovery

  • The 46-member council comprises teachers trained in UNESCO Global Citizenship Education who teach it in their classrooms.Members combine sensitivity to universal values such as human rights, diversity, and inclusion with familiarity with the Korean context.Source: §F.1 · p. 21
  • The final selection balances six dimensions—gender, age, school level, teaching experience, subject area, and region—to maximize diversity of perspectives among applicants who meet the expertise requirements.Source: §F.1 · p. 21

F.2 Consultation Process with the Safety Teacher Advisory Council

  • Across four weeks, council red-teaming and proposed criteria changes go through research-team curation and a second human-review round before adoption into K-AUT for training and evaluation.Source: §F.2 · p. 22

F.3 K-AUT-V2 : Expanded Korea-Augmented Universal Taxonomy

  • 70 new risk areas expand K-AUT from 226 to 296, while the council refines criteria for more than 100 existing areas.Table 10 provides detailed revision statistics.Source: Table 10 · p. 22, §F.3 · p. 22
  • An example revises judgment criteria for reverse discrimination in a multicultural context through the type-identification process.Source: §F.3 · p. 22, Figure 8 · p. 23
  • K-AUT-V2 strengthens protections for vulnerable parties and adds Korea-specific, frontier, multi-turn, and interaction-based risks.Examples include the North Korean nuclear issue, AI goal misalignment, model sycophancy, misplaced empathy, and complicity with harmful premises.Source: §F.3 · p. 23
  • K-AUT specifies what models should refuse, how much assistance they may provide, and how refusals should be delivered.The taxonomy is intended as a comprehensive behavioral policy, not merely a list of prohibitions.Source: §F.3 · p. 23

F.4 Making the Judgments Trustworthy

The section frames trustworthy safety judgments as a three-stage process: expert selection, resolution of divergent views, and verification that decisions change model behavior. It reports higher Safe Rates across all evaluated KGC-SAFETY dimensions for K-EXAONE 2.0 than for previous K-EXAONE models.

Selection

  • APCEIU provides a pre-qualified expert pool, and the team directly applies the stated balancing criteria during final selection.Source: §F.4 · p. 23

Handling disagreement

  • Teacher disagreement is treated as evidence of sensitivity: multiple teachers review each response, and the team examines proposed revisions without averaging judgments or discarding minority views.Teachers identify grounds for risk, while the team makes the final determination and retains responsibility for the criterion.Source: §F.4 · pp. 23–24

Verification

  • Revised criteria are turned into training data and verified through complementary human and model-based evaluation, with human judgment anchoring continuous validation of the judge model.Human evaluation is reserved for safety-critical and contested areas, while model-based evaluation supports broad coverage; AI ethics staff who designed K-AUT conduct the human evaluations.Source: §F.4 · p. 24

Conclusion and Analysis

  • Iterative refinement of evaluation criteria helps mitigate safety vulnerabilities, and K-EXAONE 2.0 achieves consistently higher Safe Rates across all evaluated KGC-SAFETY dimensions than previous K-EXAONE models.The report frames this as advancing a robust social-safety framework and reliability in the Korean context.Source: Table 11 · p. 24, §F.4 · p. 24

Related papers

Loading 2608.04505v1…