SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.10403 · 2026

KCSAT-ML: Probing Reasoning Models with Nationwide-Cohort Human Difficulty

Geewook Kim, Sanghee Park

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 20 functions out of this paper's own repositories and ran 16 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
naver-ai/KCSAT-ML canonical 16 of 20
FunctionStatusWhere it lives
classify Ran naver-ai/KCSAT-ML/analysis/contamination_analysis.py
pointer only (licence: AGPL-3.0) · get_code("ee089fe2b6308e10")
classify Ran naver-ai/KCSAT-ML/analysis/difficulty_signal_analysis.py
pointer only (licence: AGPL-3.0) · get_code("b786586bab293243")
collect Ran naver-ai/KCSAT-ML/analysis/drg.py
pointer only (licence: AGPL-3.0) · get_code("1b14f78b2b91b280")
crop_and_concat_from_meta Ran naver-ai/KCSAT-ML/src/utils/image_utils.py
pointer only (licence: AGPL-3.0) · get_code("4925c1d16bf123ee")
encode_image Ran naver-ai/KCSAT-ML/src/utils/image_utils.py
pointer only (licence: AGPL-3.0) · get_code("0cf01235db87302e")
encode_image_with_type Ran naver-ai/KCSAT-ML/src/utils/image_utils.py
pointer only (licence: AGPL-3.0) · get_code("2c222ddef9c44201")
evaluation_answer_async Ran naver-ai/KCSAT-ML/src/evaluator.py
pointer only (licence: AGPL-3.0) · get_code("b7bd3ad8d9303c85")
fit_table Ran naver-ai/KCSAT-ML/analysis/difficulty_signal_analysis.py
pointer only (licence: AGPL-3.0) · get_code("f44378d941906ea8")
flatten_with_prefix Ran naver-ai/KCSAT-ML/src/generator.py
pointer only (licence: AGPL-3.0) · get_code("bc09dcbd5da56c1e")
generate_response_async Ran naver-ai/KCSAT-ML/src/runner.py
pointer only (licence: AGPL-3.0) · get_code("7042ee578fb31b8f")
load_json Ran naver-ai/KCSAT-ML/src/evaluator.py
pointer only (licence: AGPL-3.0) · get_code("88816b951ca46a4d")
load_json Ran naver-ai/KCSAT-ML/src/generator.py
pointer only (licence: AGPL-3.0) · get_code("c700c0687c7ab7df")
parse_variant Ran naver-ai/KCSAT-ML/analysis/human_vs_model_error.py
pointer only (licence: AGPL-3.0) · get_code("392350677d773653")
pretty_label Ran naver-ai/KCSAT-ML/analysis/human_vs_model_error.py
pointer only (licence: AGPL-3.0) · get_code("48c63779ce44153d")
regress Ran naver-ai/KCSAT-ML/analysis/difficulty_signal_analysis.py
pointer only (licence: AGPL-3.0) · get_code("4d266da509167653")
variant_score Ran naver-ai/KCSAT-ML/analysis/edge_case_finder.py
pointer only (licence: AGPL-3.0) · get_code("439121706de49e24")
collect_per_item Not yet run naver-ai/KCSAT-ML/analysis/intelligence_emerges.py
pointer only (licence: AGPL-3.0) · get_code("b3586a7111030459")
compute_drg Not yet run naver-ai/KCSAT-ML/analysis/drg.py
pointer only (licence: AGPL-3.0) · get_code("f19200c013e66e91")
download_file Not yet run naver-ai/KCSAT-ML/src/generator.py
pointer only (licence: AGPL-3.0) · get_code("8c4bcc32509a5203")
render Not yet run naver-ai/KCSAT-ML/analysis/intelligence_emerges.py
pointer only (licence: AGPL-3.0) · get_code("c59c23dd3513d56e")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Math reasoning benchmarks have proliferated, yet most lack a per-item difficulty signal grounded in actual human performance. We introduce KCSAT-ML, a decade (2014-2025) of Korean College Scholastic Ability Test (KCSAT; Suneung) mathematics: 664 problems with a 339-item core set carrying official per-item error rates from nationwide cohorts of hundreds of thousands of examinees. We pair the benchmark with Difficulty-aligned Reasoning Gain (DRG): a score-orthogonal metric that asks whether a model's mistakes concentrate on the items humans found hard, or on items humans found easy. Together they expose, across a wide range of VLMs (and LLMs via OCR), three patterns: (i) low-budget accuracy collapses on the high-human-error tail at every model size; (ii) test-time scaling (TTS) raises token use roughly linearly with cohort error rate, while accuracy gains follow a non-monotonic curve; (iii) within a single family, TTS flips between anti-scaling on the hardest items and overthinking on easier ones -two faces of the same alignment failure. On DRG, models with near-identical accuracy can sit at near-opposite values: one model gets wrong what humans also find hard, while another solves the hardest items yet fails on items humans find easy -a contrast that aggregate accuracy hides. Our code and dataset builder will be open-sourced at https://github.com/naver-ai/KCSAT-ML.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2606.10403")
get_code_for_paper("2606.10403")
have("2606.10403")

Connect an agent — have() is free.