Yi-Sang: Science, Corpus & Poetry

• Yi-Sang is a term used in disparate fields, defining a neutrino mixing angle (θ13) in physics, a large-scale Korean reasoning corpus in NLP, and representing modernist poetic works.
• The neutrino oscillation study measured sin²2θ13 using reactor-based experiments, employing rate-plus-shape fits and rigorous statistical methods to quantify uncertainties.
• The Korean NLP research developed Yi-Sang via extensive web curation and Language-Mixed Chain-of-Thought techniques, while computational literary work used it to analyze modern poetry's emotional nuances.

Yi-Sang is not a single fixed referent across recent technical literature. In the Double Chooz neutrino-oscillation analysis, “Yi-Sang” denotes the mixing angle $\theta_{13}$, operationalized through the reactor-disappearance parameter $\sin^2 2\theta_{13}$ (Abe et al., 2011). In multilingual reasoning-model research, “Yi-Sang” denotes a large-scale Korean post-training corpus, together with the filtered subset Yi-Sang-HQ, used to train the KO-REAson model family (Son et al., 5 Oct 2025). The name of that corpus refers to Yi Sang (1910–1937), the Korean modernist writer and architect; the poet is also one of the five canonical authors in KPoEM, a dataset for computational emotion analysis of modern Korean poetry (Lim et al., 4 Sep 2025).

1. Referential scope and nomenclature

The term has three distinct technical uses in the supplied literature. The first is a physics usage internal to the Double Chooz paper, where “Yi-Sang” is explicitly identified with the neutrino mixing angle $\theta_{13}$. The second is a data-resource usage in Korean NLP, where Yi-Sang denotes a supervised fine-tuning corpus built from native Korean prompts and teacher-generated long reasoning traces. The third is the literary-historical referent Yi Sang, the modernist poet whose work is directly included in KPoEM and whose name was intentionally adopted for the reasoning corpus.

This tripartite usage is not accidental. In the NLP paper, the corpus name is motivated both by Yi Sang’s association with mathematically structured, experimental Korean writing and by an echo of a Korean noun meaning “the most complete state” (Son et al., 5 Oct 2025). In consequence, the term links particle physics notation, language-model post-training, and computational literary analysis, but each usage remains domain-specific and should be interpreted locally rather than conflated.

2. Yi-Sang as $\theta_{13}$ in reactor neutrino oscillation physics

Within the three-flavor neutrino framework, $\theta_{13}$ is the smallest of the PMNS mixing angles and controls the amplitude of reactor $\bar\nu_e$ disappearance driven by the atmospheric mass-squared splitting. Double Chooz measures this quantity through $\sin^2 2\theta_{13}$ using a far-detector-only reactor-disappearance analysis. The main quoted result, obtained from a rate-plus-shape fit, is

$$\sin^2 2\theta_{13} = 0.086 \pm 0.041~(\text{stat}) \pm 0.030~(\text{syst}),$$

with a Feldman–Cousins 90% confidence interval

$0.017 < \sin^2 2\theta_{13} < 0.16.$

The corresponding no-oscillation hypothesis, $\sin^2 2\theta_{13}=0$, is excluded at 94.6% confidence level, which the paper characterizes as an indication of non-zero $\sin^2 2\theta_{13}$0 rather than a high-significance discovery (Abe et al., 2011).

The rate-only extraction uses the total inverse beta decay deficit. Double Chooz reports an observed-to-predicted ratio

$\sin^2 2\theta_{13}$1

which corresponds, for $\sin^2 2\theta_{13}$2, to

$\sin^2 2\theta_{13}$3

At the time, this central value was consistent with accelerator appearance hints from T2K and MINOS and with then-current global fits favoring $\sin^2 2\theta_{13}$4–0.1. The result therefore occupied an important intermediate status: stronger than a pure upper limit, but still statistically limited.

3. Experimental architecture of the Double Chooz measurement

The Double Chooz data set used 101 days of running at the Chooz Nuclear Power Plant in France, with two $\sin^2 2\theta_{13}$5 reactor cores and a single $\sin^2 2\theta_{13}$6 gadolinium-loaded liquid-scintillator detector centered at approximately $\sin^2 2\theta_{13}$7 from the cores under an overburden of about $\sin^2 2\theta_{13}$8. Reactor antineutrinos were detected through inverse beta decay,

$\sin^2 2\theta_{13}$9

with a prompt positron signal and a delayed neutron-capture signal on Gd. Double Chooz selected candidates using a delayed-coincidence window defined by $\theta_{13}$0, $\theta_{13}$1, $\theta_{13}$2, a multiplicity veto, and PMT-quality cuts. The Gd-capture requirement effectively defined the fiducial volume and yielded a signal-to-background ratio of about 11:1 (Abe et al., 2011).

The oscillation inference used the approximate survival probability

$\theta_{13}$3

together with a binned prompt-energy fit over 18 bins between 0.7 and 12.2 MeV. The expected no-oscillation count, including background, was $\theta_{13}$4, while 4121 IBD candidates were observed in 96.8 live days. The spectral fit incorporated covariance matrices and pull terms for reactor flux, detector response, and background components, with the 8–12.2 MeV region used to constrain correlated backgrounds. A key methodological choice was to anchor the overall reactor-flux normalization to the Bugey4 short-baseline measurement, after a relative correction of $\theta_{13}$5, thereby reducing vulnerability to reactor-flux normalization uncertainty and possible very-short-baseline effects. Background estimates were $\theta_{13}$6 for accidentals, $\theta_{13}$7 for fast neutrons and stopping muons, and $\theta_{13}$8 for cosmogenic $\theta_{13}$9Li, for a total of $\theta_{13}$0. The detector systematic on absolute normalization was 2.1%, the reactor-related systematic 1.8%, and the total normalization systematic about $\theta_{13}$1–3%. The paper further notes that a planned near detector was expected to reduce the $\theta_{13}$2 precision on $\theta_{13}$3 to about 0.02.

4. Yi-Sang as a Korean reasoning corpus

In the multilingual reasoning paper, Yi-Sang is the central data resource for Korean reasoning post-training. It consists of 5.79M native Korean prompts harvested from web Q&A, exam archives, STEM communities, and code forums; 3.7M long chain-of-thought trajectories generated for a subset of those prompts by Qwen3-32B under a Language-Mixed CoT protocol; and Yi-Sang-HQ, a filtered 260k high-yield subset selected through ablation-driven curation (Son et al., 5 Oct 2025).

The construction pipeline deliberately prioritizes native Korean over translated English instructions. The authors begin from 54 candidate websites, classify them by crawling and redistribution status, exclude category C sites, low-volume or heavily obfuscated sources, and near-duplicates, and crawl the remainder with one crawler per site. Prompts are preserved verbatim rather than LLM-normalized in order to retain “in-the-wild” artifacts such as typos, mixed scripts, and internet slang. Filtering is intentionally light: prompts with Korean-character ratio below 30% are discarded except for mixed-language coding prompts, and prompts shorter than 50 characters or longer than 8,192 characters are dropped. To supplement underrepresented competition-style reasoning, the instruction pool is augmented with Korean translations of OpenThought, translated with Gemini-2.5-Flash after observed instabilities with other translators.

The full resource therefore combines native Korean prompt collection, teacher-generated long reasoning supervision, and domain-selective refinement. Yi-Sang-HQ is produced from an initial 1.8M-instance pool containing OpenThought, Code, Science, and Exams; the curation process includes a one-epoch “shakedown” run on a small proxy model, manual inspection of loss spikes, rule-based removal of degeneration loops, multiple > ... blocks, and non-Korean final answers, followed by removal of traces longer than 16k tokens and $\theta_{13}$4-gram decontamination with $\theta_{13}$5. The resulting 260k examples comprise 62k OpenThought, 86k Code, 37k Science, and 66k Exams. The paper presents Yi-Sang and Yi-Sang-HQ, together with the curation pipeline and model releases, as the nucleus of a broader open recipe for Korean reasoning models, and characterizes the resource as the largest publicly documented Korean post-training corpus at 6.77B tokens.

5. Language-Mixed CoT and the KO-REAson model family

A central methodological contribution of the Yi-Sang project is Language-Mixed Chain-of-Thought. The motivation is explicitly trifold: English-only reasoning can introduce translation noise and weaken fidelity to Korean prompts; Korean-only reasoning empirically degrades reasoning performance because many base models reason better in English; and extended Korean-only training can induce cross-lingual drift. The proposed schema therefore keeps the prompt in Korean, allows the hidden reasoning process to be mostly in English while preserving key Korean names, terms, and quoted spans, and requires the final answer to be fluent Korean. A post-hoc filter retains only traces whose reasoning has a Korean-character ratio between 5% and 20%, enforcing the intended code-switched regime (Son et al., 5 Oct 2025).

The ablation results are highly specific. For Gemma-3-4B, English-only CoT yields HRB 50.3, MCLM 48.1, and KMMLU-Redux 52.2; Korean-only CoT yields 40.6, 25.6, and 42.5; Language-Mixed CoT yields 54.9, 55.8, and 53.0, the best result on all three. For Kanana-1.5-8B, English-only CoT yields 66.2, 60.5, and 64.0; Korean-only CoT yields 67.2, 31.8, and 53.4; Language-Mixed CoT yields 74.6, 57.4, and 64.4, the best result on HRB and KMMLU-Redux and still competitive on MCLM. These numbers support the paper’s narrower claim that language mixing produces the best overall balance between deep reasoning and Korean-specific fidelity, rather than uniformly maximizing every benchmark.

Yi-Sang-HQ is then used for standard SFT, without reinforcement learning, across nine models from six families spanning 4B to 35B parameters: Gemma-3, Qwen2.5, A.X-3.1, Llama-3.1, KONI-Llama-3.1, and Kanana-1.5. Training uses AdamW with learning rate $\theta_{13}$6, 10% warmup, weight decay $\theta_{13}$7, bfloat16 precision, FSDP, Liger kernels, FlashAttention-2, and no sequence packing in final runs. The flagship model, KO-REAson-35B, is built from A.X-3.1-35B and trained for three epochs on Yi-Sang-HQ. On the nine-benchmark evaluation suite, it achieves the highest overall average score, $\theta_{13}$8, ranks first on 5 of 9 benchmarks, and ranks second on the remaining 4. Across all nine benchmarks and all trained models, the reported average improvement is +18.6 points. The paper also reports consistent gains on English reasoning benchmarks such as AIME24, AIME25, and GPQA, and selective multimodal gains on reasoning-heavy Korean VLM benchmarks. At the same time, several limitations are explicitly retained: competition-style math remains comparatively underrepresented at about 60k examples, all long reasoning traces inherit the style and biases of Qwen3-32B, and held-in ablation benchmarks create a residual overfitting risk partly mitigated by decontamination.

6. Yi Sang in computational literary analysis

In KPoEM, Yi Sang is one of five canonical modern Korean poets selected on the basis of scholarly prominence, public canonicity, and copyright expiration, alongside Han Yong-un, Im Hwa, Kim So-wol, and Yun Dong-ju. The dataset contains 7,662 entries in total, comprising 7,007 line-level entries from 483 poems and 615 work-level entries. Yi Sang contributes 464 lines from 46 poems. The corpus explicitly focuses on his Korean-language series poems such as “Critical Condition” ($\theta_{13}$9), “Crow’s Eye View” ($\theta_{13}$0), and “Reverse” ($\theta_{13}$1); however, poems with strong visual formatting, including “Crow’s Eye View” No. 4 and No. 5, are excluded because their layout cannot be accurately reproduced for quantitative analysis (Lim et al., 4 Sep 2025).

KPoEM uses a 44-category multi-label emotion schema adopted from KOTE. Annotation is performed by five trained researchers specializing in Korean literature and digital humanities, with up to 10 labels allowed per line or work. The most frequent labels in the full corpus are anxiety, sadness, expectancy, disappointment, sorrow, interest, resolute, care, admiration, and embarrassment. For Yi Sang in particular, the paper identifies categories such as anxiety, despair, fear, shock, exhaustion, self-hatred, preposterousness, realization, and resolute as especially relevant to themes of urban alienation, bodily disintegration, fragmentation, surreal violence, and ambivalent epiphany. The dataset includes concrete Yi Sang examples, such as lines from “A TOY BRIDE” and “Critical Condition,” that receive dense multi-label profiles combining fear, shock, anxiety, guilt, realization, sadness, preposterousness, contempt, exhaustion, despair, and self-hatred.

The modeling results are framed as a correction to the inadequacy of generic sentiment models for poetry. KPoEM fine-tunes KcELECTRA in three variants: KOTE only, KPoEM only, and sequential KOTE $\theta_{13}$2 KPoEM. The sequential model performs best, with accuracy 0.79, precision$\theta_{13}$3 0.53, recall$\theta_{13}$4 0.69, $\theta_{13}$5, $\theta_{13}$6, and MCC 0.47, compared with $\theta_{13}$7 and $\theta_{13}$8 for the KOTE-only model. This makes KPoEM directly relevant to Yi Sang not merely as a background corpus but as a model partially trained on his own emotional idiom. The paper nonetheless preserves several caveats: visual poems are structurally underrepresented, a finite 44-label scheme cannot exhaust the ambiguity of experimental modernism, and the corpus is limited to five canonical poets. A plausible implication is that KPoEM functions best as a structured co-reading instrument for Yi Sang rather than as a substitute for literary interpretation.

7. Synthesis across domains

Taken together, the three uses of Yi-Sang illustrate how a single label can occupy unrelated but technically rigorous positions in contemporary research. In neutrino physics, Yi-Sang denotes $\theta_{13}$9, a small PMNS angle isolated through reactor $\bar\nu_e$0 disappearance and quantified in Double Chooz through $\bar\nu_e$1. In Korean NLP, Yi-Sang denotes a large supervised post-training corpus whose design centers on Language-Mixed CoT and whose high-yield subset supports strong SFT-only reasoning models. In computational literary analysis, Yi Sang denotes the poet whose formal experimentalism motivates the corpus name and whose texts are used directly to study emotion under a culturally grounded 44-label schema.

The commonality is therefore not substantive identity but methodological density. Each usage is attached to a precise inferential apparatus: covariance-matrix oscillation fitting in reactor physics, large-scale web curation plus SFT ablations in reasoning-model training, and expert multi-label annotation with sequential fine-tuning in literary emotion modeling. The term “Yi-Sang” accordingly names, in different disciplinary settings, either a measured latent variable, a curated training resource, or a modernist poetic corpus. Understanding which of these is meant is a prerequisite for interpreting any technical claim attached to the name.