Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Solar Open 2 shows that a 250B-parameter hybrid-attention MoE can match a 1.6T frontier model on Korean officework by combining selective weight transfer, high-value data curation, and multi-teacher distillation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 10:54 UTC pith:AIXD4JOS

load-bearing objection Solid engineering report with a credible English eval, but the flagpole Ko-GDPval parity claim rests on an in-house benchmark that looks worryingly close to the training pipeline — worth refereeing, but the headline needs external validation. the 3 major comments →

arxiv 2607.20062 v2 pith:AIXD4JOS submitted 2026-07-22 cs.CL

Solar Open 2 Technical Report

Sungrae Park , Sanghoon Kim , Gyoungjin Gim , Jungho Cho , Hyunwoong Ko , Minbyul Jeong , Minjeong Kim , Keunwoo Choi
show 45 more authors
Chaehun Shin Chanwoong Yoon Dongjun Kim Eunwon Kim Gyungin Shin Hyeonju Lee Hyungkyu Kang Inseo Song Jisu Bae Jiyoon Han Jiyun Lee Joonkee Kim Junyeop Lee Mikyoung Cha Sangwon Yu Sehwan Joo Seokyoon Kang Seonghoon Yang Seung Shin Seunghyun Lee Seungseop Lim Seungyoun Shin Sukyung Lee Taegyeong Eo Taehwan Oh Taewhoo Lee Wonho Song Wonjun Oh Wonseok Hwang Yunsu Kim Yura Shim Hwalsuk Lee Sunghun Kim Du-Seong Chang Kyunghyun Cho Seungju Han Yejin Choi Junsuk Choe Hwaran Lee Minjeong Ban Yun Taewon Hwanjun Song Jae-Gil Lee KyungTae Lim Alice Oh
This is my paper
classification cs.CL
keywords Mixture-of-Expertshybrid linear attention1M-token contextKorean language modelofficework agentselective weight transfermulti-teacher distillationagentic evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This report claims that a 250B-parameter Mixture-of-Experts model with only 15B active parameters per token can deliver frontier-scale agentic ability, matching a 1.6T model on a 170-task Korean officework benchmark. The authors argue that the bottleneck for long-horizon agents is context and cost, and that their hybrid attention stack—one softmax layer per three linear-attention layers, no positional encoding, and negative-eigenvalue linear cores—cuts both while keeping exact recall. They also argue that selective weight transfer from the previous generation, a value-per-token data curation pipeline, and multi-teacher on-policy distillation make this scale reachable under a fixed compute budget. If the report is right, a modestly sized open-weight model can hold entire agent trajectories and produce deliverable-grade Korean officework at a fraction of the compute of a frontier-scale model, and the same recipe could be reused for other underserved languages.

Core claim

On its own terms, the paper establishes that a 250B-A15B MoE with a hybrid attention stack reaches a 1M-token context at about a quarter of the memory and computation of an all-softmax stack, and that after selective weight transfer from Solar Open 1, a 10T-token pre-training curriculum, and consolidation of twelve domain specialists via multi-teacher on-policy distillation, the model scores 86.8 on the in-house Ko-GDPval Korean officework benchmark, essentially tied with a 1.6T frontier model and ahead of all other measured models. The same model leads comparably sized open-weight models on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and records the highest average across th

What carries the argument

The central mechanism is the hybrid attention stack: three linear-attention layers (KDA with a gated delta rule extended to negative eigenvalues) carry token order and long-range information in a fixed-size recurrent state, keeping the KV cache constant and computation linear, while the interleaved softmax layer provides exact global recall. With no positional encoding, the linear state is the sole carrier of position. Supporting mechanisms are selective weight transfer, which reuses the 5.69B-parameter shared skeleton (embeddings, softmax projections, and the shared expert) across the architectural change to save roughly 1.7x training tokens; the OfficeVerse pipeline, which synthesizes grou

Load-bearing premise

The load-bearing premise is that Ko-GDPval is a fair external measure of Korean officework-agent ability: the report states (Section 4.1) that every pipeline is decontaminated against the Section 5 benchmarks and that Ko-GDPval is in-house (Section 5.4), but it does not document the procedure or show overlap analysis, so if the 170 tasks sit on the OfficeVerse training distribution, the parity claim measures training coverage rather than general officework ability.

What would settle it

Release Ko-GDPval or an independently constructed held-out set of Korean officework tasks built without the OfficeVerse pipeline, and compare Solar Open 2's score: a large drop from 86.8, or clear n-gram and embedding overlap between Ko-GDPval tasks and OfficeVerse training tasks, would demonstrate that the apparent parity with a 1.6T model comes from training-distribution coverage rather than transferred officework competence.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A 1M-token context at about a quarter of the cost of an all-softmax stack makes whole agent trajectories fit in a single window, and the no-positional-encoding design removes the RoPE length-extrapolation limit.
  • Selective weight transfer turns an architectural upgrade between generations into a compute saving, so a model family can evolve without re-training from scratch.
  • The equal-budget data-recipe ablation implies that value-per-token curation is a scalable lever: the same compute buys more capability when the mixture is optimized.
  • MOPD lets capability cultivation be parallelized across domains, so a dozen experts can be trained independently and consolidated without a verifiable-reward term on top of the KL objective.
  • In the sovereign scenario, a language-specialized 250B model can match the 1.6T frontier on deliverable-producing officework, suggesting that domain- and language-specialized models can substitute for much larger general models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the headline parity claim rests on Ko-GDPval, an in-house benchmark whose construction mirrors the OfficeVerse training pipeline; until the stated decontamination is documented or the benchmark is released, the 86.8-vs-86.9 result is best read as a coverage claim rather than proof of general officework ability.
  • Editorial inference: the quarter-cost claim is architectural, reflecting the 1:3 softmax-to-linear layer mix; actual wall-clock savings depend on serving implementation, batch shapes, and hardware.
  • Editorial inference: MOPD's KL-only objective deliberately trades away the chance to beat the teachers; a natural test is whether adding a small verifiable-reward term on top would improve the consolidated model beyond all twelve specialists.
  • Editorial inference: the transfer step depends on the vocabulary being identical across generations; for languages starting with a new tokenizer, the savings from selective weight transfer would shrink considerably.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. Solar Open 2 is a 250B-A15B Mixture-of-Experts language model presented as a Korean sovereign AI for long-horizon agentic tasks. The technical report describes a hybrid attention stack (one softmax layer per three linear-attention layers, NoPE, sigmoid gating, negative eigenvalues) targeting a 1M-token context; a selective weight transfer from Solar Open 1 (2.3% of parameters); a curated 10T-token pre-training mixture; a three-stage curriculum with checkpoint merging; and a post-training pipeline with fully asynchronous RL plus Multi-teacher On-Policy Distillation (MOPD). Evaluation on public English benchmarks shows large generational gains (e.g., SWE-Bench Verified 15.4→70.4, GPQA-Diamond 66.2→86.3) and competitiveness with larger open models. On the in-house Ko-GDPval Korean officework benchmark, Solar Open 2 scores 86.8 versus 86.9 for DeepSeek-V4-Pro (1.6T), which the abstract and conclusion interpret as matching a frontier-scale model at less than a sixth of its size. The central engineering claims are supported by controlled proxy ablations and detailed infrastructure descriptions, but the flagpole Ko-GDPval claim depends on the unverified independence of an in-house benchmark from the in-house training pipeline.

Significance. If the reported results hold, the paper demonstrates a practical recipe for building a competitive long-context agentic model at moderate active scale: hybrid attention, selective cross-generation transfer, value-per-token data curation, and scenario-driven post-training with specialist distillation. The strengths are substantial: the architecture ablations (Figure 4, Section 3.1) are controlled and quantified; the English public-benchmark improvements are large and independently checkable in principle; and the MOPD and asynchronous-RL descriptions contain concrete, falsifiable design choices (e.g., full-vocabulary KL, fresh-token-fraction gateway, length-binned sampling). The main unresolved risk is external validity of Ko-GDPval: it is an in-house, 170-task benchmark built by the same team and with the same 11-domain, occupation-based, rubric-plus-judge structure as the OfficeVerse training pipeline (§4.1.3, §5.4). The 0.1-point gap with DeepSeek-V4-Pro is within sampling noise, and the paper provides no overlap analysis or released artifacts to rule out training-distribution overlap. Thus the headline 'matches 1.6T' is credible only if the benchmark is shown to be independent.

major comments (3)
  1. [§5.4, §4.1, §4.1.3] Ko-GDPval is the sole basis for the abstract's headline claim that Solar Open 2 'essentially matches DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.' This benchmark is in-house, consists of only 170 tasks, and shares its structural skeleton with the OfficeVerse training pipeline: both use 11 industry domains, real-public-data grounding, occupation-based task synthesis, and rubric-plus-LLM-judge grading. The only decontamination statement in the paper is one sentence in §4.1 ('Every pipeline is additionally decontaminated against the benchmarks used in Section 5'), with no procedure, no overlap analysis, and no released task list. The 86.8-vs-86.9 gap is also smaller than the sampling noise expected for 170 tasks. Please provide a concrete overlap analysis between OfficeVerse training data and Ko-GDPval, release the benchmark or a representative subset, and report per-task varian
  2. [§2.2, §5.1, Table 3] The paper's architecture section motivates a 1M-token context window, and Length Expansion (§3.3) is described as extending context to 1M, but no evaluation in Section 5 demonstrates performance at or near 1M tokens. Table 3 lists 'IF / Long' benchmarks (Multi-Challenge, IFBench, AA-LCR), yet no context length or long-context result is reported; the 1M-token claim is supported only by architectural reasoning and the training-stage description. Please add a direct long-context evaluation (e.g., a needle-in-a-haystack test, long-document QA, or passkey at 1M) and report effective context length for the released model. This is load-bearing because the architecture is explicitly designed for this capability.
  3. [§5.2, Table 3] Several English 'leads' are based on very small margins measured with an internal harness: MMLU-Pro 86.2 versus 85.9 for DeepSeek-V4-Flash, LiveCodeBench v6 92.4 versus 92.3, and IFBench 80.0 versus 80.3. The protocol says settings are identical but gives no decoding parameters, number of runs, or run-to-run variance. For a technical report claiming frontier competitiveness, single-run point estimates are not enough when margins are <0.5 points. Please state measurement uncertainty or show that the differences are reproducible across multiple runs; otherwise, the appropriate claim is 'statistically tied,' not 'leads.' This does not affect the large generational jumps (e.g., SWE-Bench 15.4→70.4) but does affect the precise wording of the comparison.
minor comments (5)
  1. [§3.2] The mixture-ratio choices (real:synthetic 4:6, math/code ≥15% each, English ≥80%) are said to be 'optimized through ablation studies,' but no ablation table or budget is shown for these specific ratios. Please cite the supporting experiment or mark the values as chosen by held-out evaluation.
  2. [§4.2.2] MOPD is a central contribution, but there is no ablation isolating its effect from the Specialist stage (e.g., SFT-on-teacher-traces versus MOPD, or MOPD with and without the KL-only objective). A small-scale or proxy ablation would make the claim more convincing.
  3. [Table 4] The table header says 'A vg.' and the caption says 'Avg. is the mean'; please unify the notation. Also, the dagger marker for in-house benchmarks is explained in the table but not in the main text at first mention.
  4. [§2.1, Figure 2] Tokenizer efficiency is measured on Ko-GDPval task prompts, which are an in-house benchmark; this is appropriate as a tokenizer-efficiency sample, but the text should note that the comparison set and prompt source are internal and not yet publicly released.
  5. [§5.4, Table 5] The table reports scores to one decimal place for 170 tasks. Please clarify whether these are averages over multiple runs or single-run scores, and add a column with standard errors if available.

Circularity Check

1 steps flagged

Ko-GDPval, the paper's flagship benchmark, is an in-house benchmark whose construction mirrors the OfficeVerse training pipeline; its independence is only asserted, not demonstrated.

specific steps
  1. other [§4.1 / §4.1.3 / §5.4 / Abstract]
    "Every pipeline is additionally decontaminated against the benchmarks used in Section 5. ... OfficeVerse organizes office work as a matrix of 11 industry domains × 12 task types ... Ko-GDPval is an in-house Korean officework agentic benchmark ... spanning 11 industry domains ... Solar Open 2 stands just behind the strongest model compared, DeepSeek-V4-Pro (86.9)—a 1.6T-parameter model more than six times its size."

    The headline 'essentially matches DeepSeek-V4-Pro (1.6T)' is measured on Ko-GDPval, an in-house benchmark built by the same team that built OfficeVerse, the training pipeline that supplies Solar Open 2's officework skills. The two share the same structural skeleton: both are organized around 11 industry domains, real public data, deliverable-producing Korean officework, and rubric-plus-judge grading. The only stated safeguard is the single unsupported sentence 'Every pipeline is additionally decontaminated against the benchmarks used in Section 5,' with no protocol, overlap analysis, or released artifacts. As presented, the 86.8-versus-86.9 result is not an independent test of general officework ability; it re-scores the model on a task distribution the report has not shown to be outside i

full rationale

The architecture, selective weight transfer, and data-curation claims are largely self-contained: the report gives controlled proxy ablations (10B-A1B architecture runs, a 200B-A15B initialization comparison, and a 300B data-recipe ablation) whose outcomes are not assumed as inputs. The English benchmarks and most Korean public benchmarks (KMMLU-Pro, CLIcK, HAE-RAE, HRM8K, KBL, KorMedMCQA) are externally checkable, and the cited prior work, including the authors' own Solar Open 1 report, is descriptive rather than a load-bearing uniqueness argument. The circularity risk concentrates in the flagship Ko-GDPval comparison: Ko-GDPval is an in-house benchmark, and its construction mirrors the OfficeVerse training pipeline in domain matrix, data grounding, deliverable formats, and grading scheme. Decontamination is asserted in one line in §4.1 but no procedure, contamination measurement, or benchmark artifacts are provided. The abstract and §5.4 then present the 86.8-vs-86.9 score as evidence of frontier-scale Korean officework capability at less than one-sixth the size. That specific claim is therefore partially self-referential, while the remainder of the paper stands on independent empirical content. Score 4.0 reflects one significant self-constructed-evaluation risk without making the entire derivation circular.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

Most of the capability story is carried by engineering choices that are either tuned on proxy models (interleaving ratio, mixture ratios, quality thresholds, checkpoint merge) or are internal workflows with no external artifact (OfficeVerse, Ko-GDPval, decontamination log, MOPD infrastructure). The external-benchmark half (MMLU-Pro, GPQA, SWE-Bench, etc.) relies only on standard evaluation practice and is the most assumption-free part of the report. The 6×-size headline requires the least well-supported premise: that an in-house benchmark built on the same paradigm as the training data is a neutral yardstick. No invented physical or mathematical entities are introduced; the ledger items are benchmarks and pipelines, which are inventions in the engineering sense but not in the theory sense.

free parameters (6)
  • Data mixture ratios (real:synthetic 4:6, math/code ≥15% each, English ≥80%) = 4:6; ≥15%; ≥80%
    Chosen by hand and 'optimized through ablation studies' (§3.2) on a 10B model against downstream benchmark groups; feeds the final model's capability mix.
  • Softmax:linear attention interleaving ratio (1:3, S-L-L-L) = 1 softmax : 3 linear
    Adopted after proxy ablations (§2.2); the reported 82.6%/3%/1.5%/1% token savings are the tuning results, not independent predictions.
  • Staleness hyperparameters smax and fresh-token threshold ρ = not stated
    Central to the asynchronous RL data selection (§4.2.1, Fig 10) but no values or sensitivity analysis are given.
  • Negative-eigenvalue widening β = 2σ(·) = β ∈ (0,2)
    Architectural choice (§2.2) enabling state erasure; its benefit is cited from Grazzi et al. 2025 and proxy loss curves, not verified on the 250B model.
  • Checkpoint merge count and weights (4 checkpoints) = 4 checkpoints, merging scheme unspecified
    The final PLM 'stronger than any single checkpoint' is produced by an unspecified merge (§3.3, Fig 8); the 0.745 score is a tuned artifact.
  • Stage 3/4 quality thresholds and 12-specialist routing = not stated
    Quality thresholds for Intensive Pre-training and Length Expansion are unstated (§3.3); the 12-teacher split and one-teacher-per-prompt routing are design choices (§4.2.2).
axioms (5)
  • domain assumption Proxy-scale ablations (10B-A1B) transfer quantitatively to the 250B model
    The 3.2×/1.8× training-token savings claims (Fig 4) and recipe choices are measured on a proxy; no verification at scale is shown (§2.2, §3.2).
  • domain assumption Negative eigenvalues provably restore state-tracking (Grazzi et al. 2025) and this suffices to carry position in a NoPE design
    §2.2 invokes the cited theorem and asserts the linear state is 'the sole carrier of token order' without direct evidence on the 250B model.
  • domain assumption Decontamination of all training pipelines against all Section 5 benchmarks, including in-house ones, was complete and effective
    Stated in one clause at §4.1 intro ('Every pipeline is additionally decontaminated...'); no procedure, tooling, or holdout analysis is provided, and the claim matters because Ko-GDPval is the headline benchmark.
  • domain assumption LLM-judge and rule-based rubrics over synthesized OfficeVerse tasks yield reward/eval signals aligned with real officework quality
    §4.1.3 relies on verifier-first generation, three-pass self-consistency checks, and judge-graded rubrics; if judges are biased toward the model's own style, both training reward and Ko-GDPval grading are confounded.
  • domain assumption NoPE attention without position signal does not degrade softmax-layer recall at 1M length
    §2.2 removes RoPE entirely; the argument is the ablation figure (Fig 4) and cited work, not a long-context probe shown in the report.
invented entities (2)
  • Ko-GDPval no independent evidence
    purpose: In-house Korean officework-agent benchmark, 170 tasks across 11 domains/58 occupations; supports the headline claim of matching DeepSeek-V4-Pro at <1/6 size
    Not released; tasks, rubrics, and grading code unpublished; constructed by the same team and paradigm as the training pipeline (§5.4 vs §4.1.3), so no external falsifiable handle exists.
  • OfficeVerse no independent evidence
    purpose: In-house pipeline for synthesizing officework training tasks grounded in public data
    Internal-only (§4.1.3); no release, no benchmark tie, and its design closely mirrors Ko-GDPval, creating the train/eval overlap that weakens the headline.

pith-pipeline@v1.3.0-alltime-deepseek · 41677 in / 23926 out tokens · 186499 ms · 2026-08-01T10:54:57.972740+00:00 · methodology

0 comments
read the original abstract

We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.

Figures

Figures reproduced from arXiv: 2607.20062 by Alice Oh, Chaehun Shin, Chanwoong Yoon, Dongjun Kim, Du-Seong Chang, Eunwon Kim, Gyoungjin Gim, Gyungin Shin, Hwalsuk Lee, Hwanjun Song, Hwaran Lee, Hyeonju Lee, Hyungkyu Kang, Hyunwoong Ko, Inseo Song, Jae-Gil Lee, Jisu Bae, Jiyoon Han, Jiyun Lee, Joonkee Kim, Jungho Cho, Junsuk Choe, Junyeop Lee, Keunwoo Choi, Kyunghyun Cho, KyungTae Lim, Mikyoung Cha, Minbyul Jeong, Minjeong Ban, Minjeong Kim, Sanghoon Kim, Sangwon Yu, Sehwan Joo, Seokyoon Kang, Seonghoon Yang, Seunghyun Lee, Seungju Han, Seungseop Lim, Seung Shin, Seungyoun Shin, Sukyung Lee, Sunghun Kim, Sungrae Park, Taegyeong Eo, TaeHwan Oh, Taewhoo Lee, Wonho Song, Wonjun Oh, Wonseok Hwang, Yejin Choi, Yunsu Kim, Yun Taewon, Yura Shim.

Figure 1
Figure 1. Figure 1: Benchmark comparison across three capability groups: knowledge and reasoning [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Tokenizer efficiency in bytes per token (higher is better: fewer tokens for the same text). Solar Open 2 inherits the Solar Open 1 tokenizer unchanged. In the Ko-Agent group (Ko-GDPval task prompts), it ranks first among the 12 tokenizers compared, about 24% above the best global model (MiniMax-M3, 3.54). sovereign models A.X-K1 (4.27) and K-EXAONE (4.12). The margin holds on general Korean text, where it … view at source ↗
Figure 3
Figure 3. Figure 3: Solar Open 2 architecture. The 48-layer stack (left) repeats a four-layer pattern twelve times: one softmax-attention layer followed by three linear-attention layers, each paired with an MoE block, and no positional encoding anywhere in the model. Insets detail the MoE block with 320 routed experts and one shared expert (top), the softmax-attention layer, GQA with an elementwise sigmoid output gate (bottom… view at source ↗
Figure 4
Figure 4. Figure 4: Architecture ablation on a 10B-A1B proxy trained from scratch; only the architecture [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pre-training procedure of Solar Open 2. Stage 1 initializes the model by selective weight transfer from Solar Open 1; Stages 2–4 then train on about 12T tokens in total: General Pre-training (10T), Intensive Pre-training (1T), and Length Expansion (0.9T) with checkpoint merging. Four stages take Solar Open 2 from weight initialization to a 1M-context pre-trained model ( [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 6
Figure 6. Figure 6: Selective weight transfer vs. random initialization on a [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Data-recipe ablation on a 10B MoE model: Solar Open 1 vs. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Average performance over 50+ PLM evaluation metrics as a function of cumulative [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Post-training procedure of Solar Open 2. SFT and Multi-domain RL build the general￾ist base; the Specialist stage then cultivates twelve domain teachers in parallel, and Multi-teacher On-policy Distillation (MOPD) consolidates them into the single released model. of training infrastructure that execute the stages at 250B scale, fully asynchronous RL and MOPD (Section 4.2). 4.1 Agent Scenarios The agent cap… view at source ↗
Figure 10
Figure 10. Figure 10: The fresh-token-fraction gateway. Every token records the policy version that gener [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Length- and staleness-aware batch sampling. Buffered trajectories are partitioned by [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Blind pairwise qualitative evaluation on Korean conversations [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: FATF mutual-evaluation pre-response, delivered in two mutually consistent formats. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Source fidelity and cited grounds across two regulatory-document tasks. Left: a [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Deliverable submitted by Solar Open 2: KMD217_요양급여_등재신청서.pdf. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Deliverable submitted by Solar Open 2: 공장에너지절감_2026연간계획.xlsx. A.2.3 Citizen briefing on national fiscal trends (DOCX) Task prompt 당신은 재정 규모 약 4조 원의 광역지방자치단체 기획조정실 예산담당관실의 재정공시·주민참여예산 담당 주무관입니다. 매년 11월에 개최되는 ‘주민과 함께하는 재정설명회’ 를 2025년 11월 20일(목)에 앞두고, 일반 주민·시민단체·지역 언론을 청중으로 하는 브리핑 문서를 준비해야 합니다. 담당 국장은 ‘숫자 위주 결산서’ 가 아닌 ‘이야기가 있는 브리핑 문서’ 를 요구했고, 최근 10년(2015∼2024년) 추이와 구조적 변화를 함께 보여 달라고 지시했습니다. 첨부된 세 개의 열린재정 데이터… view at source ↗
Figure 17
Figure 17. Figure 17: Deliverable submitted by Solar Open 2: 국가재정흐름_주민설명_브리핑_2025.docx. A.2.4 Municipal budget-committee response deck (PPTX) Task prompt 당신은 한솔시 기획예산담당관실의 예산팀 주무관(예산 분석 담당)입니다. 다음 주 개최되는 한솔시의회 예산결산특별위원회 정례 심사에서, 우리 시의 2024 회계연도 예산 집행 현황을 보고하고 의원 질의에 대비한 부서 답변 자료를 발표해야 합니다. 예결특위 위원들은 단순 집행률 수치를 넘어 ‘중앙정부 재정 흐름과 연계한 지방재정 운용의 합리성’ 을 추궁할 것으로 예상되므로, 국가 재정지출·국세수입 추이를 거시 배경으로 활용 하여 우리 시 사업별 집행 실적과 미집행 사유를 설득력 있게 정리해야 … view at source ↗
Figure 18
Figure 18. Figure 18: Deliverable submitted by Solar Open 2: 한솔시_2024회계연도_예결특위_답변자 료.pptx. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references

  1. [1]

    사회복지_관련_법령_법령.docx — 환자 접근권 보장 근거 인용용 (특히 사회복지사업법 제 5조의2 등)

  2. [2]

    내과( dgsbjtCd=01)·가정의 학과 중심 분석

    의료기관_진료과목별_전문의.xlsx — 진료과별 전문의 수 ( 1,300건). 내과( dgsbjtCd=01)·가정의 학과 중심 분석

  3. [3]

    기관 종별 분포 및 drTotCnt 활용

    의료기관_의료기관_기본정보_인력.xlsx — 요양기관 기본정보 및 의료인력 (10,000건). 기관 종별 분포 및 drTotCnt 활용

  4. [4]

    임상시험_결과_데이터.csv — KMD-217 임상 2상 결과 (Primary/Secondary endpoint, AE 등) 산출물 파일명: KMD217_요양급여_등재신청서.pdf / 8 ∼12페이지 / 표·차트 포함 가능 구성 섹션

  5. [5]

    법적 근거로 국민건강보험법 및 관련 법령 인용, 첨부 법령 문서의 사회복지사업법 제 5조의2(사회복지서비스 제공의 원칙) 및 인권존중 관련 조항을 환자 접근권 보장 근거로 병기

    신청 개요 (1p) 신청 의약품 정보, 신청 구분(신약), 기준일. 법적 근거로 국민건강보험법 및 관련 법령 인용, 첨부 법령 문서의 사회복지사업법 제 5조의2(사회복지서비스 제공의 원칙) 및 인권존중 관련 조항을 환자 접근권 보장 근거로 병기

  6. [6]

    Sitagliptin 100mg, 95% CI 및 p-value 포함 표 정리 • 비열등성 마진 −0.4%(HbA1c 차이) 기준 우월성/비열등성 판정, 충족 항목 플래깅 • AE 요약

    임상적 유용성 평가 (2∼3p) 26 • Primary endpoint: 24 주 후 HbA1c 변화량 • Secondary: HbA1c <7.0% 도달률, 공복혈당 변화, 체중 변화 • 시험군 vs. Sitagliptin 100mg, 95% CI 및 p-value 포함 표 정리 • 비열등성 마진 −0.4%(HbA1c 차이) 기준 우월성/비열등성 판정, 충족 항목 플래깅 • AE 요약

  7. [7]

    경제성 평가 — ICER (1∼2p) • 시험군의 QALY 개선분 대비 추가 비용으로 ICER 산출 (비용·QALY 가정값은 명시) • WTP 임계값: 1QAL Y당 25,000,000원, 미만 시 ‘비용효과적’ 으로 플래깅 • 낙관·비관 민감도 시나리오 2개 이상

  8. [8]

    처방 환경 및 예상 시장 규모 ( 2∼3p) • 진료과목별 전문의 파일에서 내과( dgsbjtCd=01)·가정의학과 전문의 수를 기관 종별·시도별 집계 • 기본정보 파일에서 기관 종별 분포 및 drTotCnt 합계 산출 • 주요 처방 후보 기관 = 기관 종별이 ‘상급종합’ 또는 ‘종합병원’ 이면서 내과 전문의 5인 이상 보유 → 시도별 표로 제시 및 플래깅 • 예상 처방 환자 풀 추정

  9. [9]

    약가 산정 근거 (1p) Sitagliptin 100mg 보험약가 기준 가중평균가 산정 방식 설명(약가 가정값 명시), 임상적 우월성 반영 가산율 근거

  10. [10]

    결론 및 등재 권고 ( 0.5∼1p)

  11. [11]

    작성 지침 • 한국 보건의료 등재 실무 관행 준수 • 수치는 한국식 단위(억 원, 만 원) 사용 HbA1c 변화량(%) -0.94 -0.62 -0.32 (-0.48, -0.16) <0.001 우월성 비열등성 충족 비열등성 판정: 사전에 설정된 비열등성 마진은 HbA1c 차이 -0.4%입니다

    참고문헌 / 출처 첨부 파일 및 웹 보조 검색 출처 명시. 작성 지침 • 한국 보건의료 등재 실무 관행 준수 • 수치는 한국식 단위(억 원, 만 원) 사용 HbA1c 변화량(%) -0.94 -0.62 -0.32 (-0.48, -0.16) <0.001 우월성 비열등성 충족 비열등성 판정: 사전에 설정된 비열등성 마진은 HbA1c 차이 -0.4%입니다. 시험약군과 대조약군의 HbA1c 변화량 차이는 -0.32%로, 비열등성 마진(-0.4%)을 초과하여 충족하였습니다. 나아가 95% 신뢰구간 하한값(-0.48)이 마진(-0.4%)을 초과하여 통계적 우월성도 함께 입증되었습니다(p<0.001). 따라서 Kameglitin은 비열등성 및 우...

  12. [12]

    이 중 내과 전문의가 3,575명으로 가장 많아 전체의 28.9%를 차지하였으며, 외과(1,263명, 10.2%), 영상의학과(1,152명, 9.3%), 소아청소년과(984명, 8.0%), 가정의학과(332명, 2.7%) 순으로 분포하였습니다

    처방 환경 및 예상 시장 규모 4.1 진료과목별 전문의 현황 HIRA 건강보험심사평가원 데이터(데이터셋 15001699, hosp_dtl_dgsbjt)에 따르면, 전국 상급종합병원 47개소의 진료과목별 전문의 수는 총 12,358명으로 집계되었습니다. 이 중 내과 전문의가 3,575명으로 가장 많아 전체의 28.9%를 차지하였으며, 외과(1,263명, 10.2%), 영상의학과(1,152명, 9.3%), 소아청소년과(984명, 8.0%), 가정의학과(332명, 2.7%) 순으로 분포하였습니다. 순위 진료과목 전문의 수(명) 비율(%) 당뇨 처방 관련성 1 내과 3,575 28.9 주요 처방과 2 외과 1,263 10.2 — 3 영상의...

  13. [13]

    현황 진단 (에너지사용.csv + 생산실적.csv) • 공장별(1공장 INJ-A1/INJ-A2, 2 공장 PNT-B1, 3 공장 ASM-C1)·시간대별 (경부하/중간부하/최대 부하) 전력·LNG·스팀 사용량 합계 및 평균 원단위( kWh당) 산출 • 라인별 전력원단위(kWh/개) 및 가동률(가동시간/(가동시간+정지시간)) 계산 • 공장별 최대부하 시간대 전력 사용 비중

  14. [14]

    외부 단가·시장 참조 • 계 약 종 별_ 전 력 사 용 량_2023_ 경 상 남 도.xlsx 에 서 cntr= ‘산 업 용’ 데 이 터 의 unitCost 를 powerUsage 가중평균하여 경상남도 산업용 평균 단가 산출, 계약종별_전력사용량 _2024_경기 도.xlsx 의 산업용 평균 단가와 비교 • SMP_수요예측.xlsx 에서 areaName= ‘육지’ 데이터의 hour별 평균 SMP 곡선을 만들고, 상위 25% 를 피크, 하위 25%를 경부하로 정의

  15. [15]

    2026년 절감 계획 • 라인별(INJ-A1, INJ-A2, PNT-B1, ASM-C1 ) 절감목표( %), 핵심 투자항목, 예상 투자비, 연간 절감 금액, 단순회수기간(투자비÷ 연간절감액), KPI(전력원단위 kWh/개, 피크부하 비중 %) • 전사 합산 절감률 7% 이상 충족 산출물 공장에너지절감_2026연간계획.xlsx (총 2시트, 헤더 고정) • 시트1 현황진단: 공장·라인·시간대별 사용량 /원단위/가동률 + 경상남도·경기도 산업용 단가 비교 + 육지 SMP 시간대별 평균 • 시트2 2026절감계획 : 라인별 절감목표·투자항목·투자비·연간절감액·회수기간· KPI (상단에 전사 합산 요약 행) • 단순회수기간 3년 이하 셀 ...

    2025

  16. [16]

    재정수입구조본예산총수입기준.xlsx — 2015 ∼2024년 회계연도별 본예산 총수입 구조(기금 /예산, 일반회계/특별회계/기금, 사회보험성기금·국세수입·세외수입·융자금회수·기금기타 등 세목별 금액, 단위: 조 원)

    2015

  17. [17]

    재정지출추이총계기준 .xlsx — 2015 ∼2024년 결산 기준 총계 재정지출(일반회계·특별회계·기금 구분, 단위: 조 원)

    2015

  18. [18]

    조세부담률및국민부담률추이.xlsx — 2015 ∼2024년 결산 기준 국세·지방세·사회보장기여금(사회 보장기여금/공무원연금기여금/군인연금기여금/건강보험재정기여금) 추이 및 국민부담률( %) 작성할 산출물 파일명: 국가재정흐름_주민설명_브리핑_2025.docx (한 개 문서, 분량 14∼20쪽) 문서 구성 요구사항 다음 6개 장(章)을 모두 포함하되, 각 장에는 평이한 서술 narrative + 표 또는 차트(이미지 또는 워드 표) + ‘주민 한마디로 정리’ 박스( 2∼3문장)를 함께 배치하십시오. 제1장. 인사말과 브리핑의 목적 ( 1쪽) • 재정설명회의 취지, 본 문서가 다루는 기간( 2015∼2024년)과 데이터 출처(열린재정 ...

    2015

  19. [19]

    결산 기준 국가 총지출은 2015 년 902.6 조 원에서 2024 년 1,537.9조 원으로 10 년간 635.3 조 원 증가 (증감률 70.4%)하였습니다

    10 년간 재정지출 추이 국가 재정지출은 정부가 국민으로부터 조달한 재원을 어떤 방식으로 사용하는지를 보여줍니다. 결산 기준 국가 총지출은 2015 년 902.6 조 원에서 2024 년 1,537.9조 원으로 10 년간 635.3 조 원 증가 (증감률 70.4%)하였습니다. 이는 같은 기간 총수입 증가율(59.1%)을 크게 상회하는 수치로, 국가 재정이 '적자' 방향으로 기울어 왔음을 시사합니다. 구체적으로 일반회계 지출은 257.9 조 원에서 435.4 조 원으로 68.8% 증가했고, 특별회계는 61.5 조 원에서 94.0 조 원으로 52.8%, 기금은 583.2 조 원에서 1,008.5조 원으로 72.9% 증가했습니다. 연도 일...

    2015

  20. [20]

    표지 — 한솔시 기획예산담당관실, 2024 회계연도 예결특위 답변자료, 발표일자( 2025년 6월 17일 기준)

    2024

  21. [21]

    재정 지출추이본예산총지출기준.xlsx 기반

    거시 재정 배경 (1) — 국가 재정지출 추이 (2015∼2024, 일반회계·특별회계·기금 별 추이 그래프). 재정 지출추이본예산총지출기준.xlsx 기반. 31

    2015

  22. [22]

    국세수입추이.xlsx 기반

    거시 재정 배경 ( 2) — 국세수입 추이 및 주요 세목 비중 변화 ( 2020∼2024 집중). 국세수입추이.xlsx 기반. 지방교부세 재원인 내국세 흐름을 강조

    2020

  23. [23]

    재 정수입구조본예산총수입기준.xlsx 기반

    거시 재정 배경 (3) — 국가 재정수입 구조 변화의 시사점 (사회보험성기금·세외수입·국세수입 추이). 재 정수입구조본예산총수입기준.xlsx 기반

  24. [24]

    budget_line_items.xlsx 요약 시트 기반

    한 솔 시 2024 예 산 총 괄 — 분 야 별 예 산 액·집 행 액·집 행 률 ( 10 개 분 야 막 대 그 래 프). budget_line_items.xlsx 요약 시트 기반. 시 전체 평균 집행률과 함께 표기

    2024

  25. [25]

    집행률 우수 분야 / 부진 분야 — 집행률 75% 이상 분야와 65% 미만 분야를 색상 구분(우수: 녹색, 부진: 적색)하여 비교

  26. [26]

    세목 명세 시트 기반

    사회복지비 세부 분석 — SOC-2024-001 산하 세목별 집행 현황. 세목 명세 시트 기반

    2024

  27. [27]

    지역경제·일자리 분야 심층 분석 — ECN-2024-001, ECN-2024-002 세목별 부진 사유 분석

    2024

  28. [28]

    SOC인프라(도로·교통, 상하수도) 미집행 사유 — INF-2024-001, INF-2024-002 분석

    2024

  29. [29]

    거시 재정과 지방 집행 연계 진단 — 국세수입 둔화 → 지방교부세 교부 지연 → 시 사업 집행 부진의 연결 논리를 1장으로 정리

  30. [30]

    향후 집행 계획 및 이월 최소화 대책 — 분야별 4분기 집행 가속화 방안

  31. [31]

    분석 요건 • 집행률은 집행액 / 예산액으로 계산하되, 원본 시트의 집행률 열과 교차검증하십시오

    예상 질의응답(Q&A) — 의원 예상 질의 5개와 답변 요지. 분석 요건 • 집행률은 집행액 / 예산액으로 계산하되, 원본 시트의 집행률 열과 교차검증하십시오. 소수점 첫째 자리(%)까지 표시. • 분야별 잔액(미집행액)을 억 원 단위로 환산하여 표기 (예: 6,555,000,000원 → 65.55억 원). • 거시 재정 추이 그래프는 2015∼2024 전체 10개년을 표시하되, 시사점 도출은 최근 5개년(2020∼2024)에 집중. • 국세수입 분석 시 ‘내국세’ 관련 세목(소득세·법인세·부가가치세 등)을 우선 강조 (지방교부세 재원이기 때문). 제약 조건 • 집행률 70% 미만인 분야는 별도 표시(적색)하고, 미집행 사유를 본문...

    2015