AI RESEARCH IN CHINA — READING AND COMPARISON RECORD
Version: 2026-09-11
Guide: https://yenra.com/china-technology-standards/

Purpose: trace one AI research claim from its source to the evidence that
supports it. Copy this record for each paper or model version. Keep unknown
fields explicit. Preserve primary sources, raw outputs and dated revisions.

RESEARCH QUESTION
Topic and intended task:
Team / authors / affiliations:
Paper URL / title / version / publication status:
Model identifier / release date / repository URL and commit:
Model artifact: weights / hosted API / complete application / other:
What is available: code, data, configurations, evaluation outputs:
Terms associated with each artifact / source of those terms:
One sentence describing the proposed contribution:

METHOD AND EVIDENCE
Training or architectural change under study:
Comparison model and exact version:
Dataset and version / selection method / language and domain:
Prompts / translation and review / response format:
Reasoning or generation budget / number of attempts / tool access:
Scoring method / evaluator version / ambiguous cases:
Hardware or service configuration / latency or resource measurement:
Author-reported result:
Independent reproduction, if available, with source:
Raw outputs / failures / uncertainty / repeated runs:
Evidence needed to connect the result to the intended application:

FICTIONAL CALCULATION — not results for any named model
A: 80 correct answers out of 100 = 80% accuracy.
B: 86 correct answers on the same 100 questions = 86% accuracy.
Difference: 6 percentage points.
Relative increase in accuracy: (86 - 80) / 80 = 7.5%.
Check matched conditions, changed answers, scoring reliability, test-set
selection and repeated-run variation before interpreting the improvement.

STANDARDS, WHEN RELEVANT
Exact standard number / edition / official catalogue URL:
Catalogue status and date checked:
Document scope and applicable provisions actually read:
Assessed system and version / assessment report and issuer:
Which result is a benchmark score, which is a conformity assessment,
and which is a measurement from an actual deployment:
Replacement or withdrawal information:

NEXT STEP
Supported conclusion, with its conditions:
Main uncertainty or failure case:
One follow-up test or source to inspect:
Changes found at the next review / date / effect on the conclusion:

Starting points: the article links DeepSeek-R1, Qwen3, Intern-S,
Tsinghua AIR, OpenCompass and official Chinese standards records.
Research contributions, accessible artifacts, independently reproduced
results and successful deployments are distinct things to document.