About

This site is an independent, unofficial English-language mirror of marxism.cass.cn (马克思主义研究网, the Marxism Research Network). It is not affiliated with the Chinese Academy of Social Sciences or its Institute for Marxism Studies, and nothing here is endorsed by them.

What is being mirrored

The Chinese Academy of Social Sciences (CASS) is the PRC's top state research organisation in the humanities and social sciences. Its Institute for Marxism Studies (马克思主义研究院) runs marxism.cass.cn, which publishes research articles, commentary, and institutional news across Marxist theory, CPC history, Party building, discipline inspection, and related fields. The material is rarely translated, so most of it is invisible to readers who do not read Chinese. This mirror exists to make it readable, searchable, and citable in English.

Method

1. Crawl. Section listings on marxism.cass.cn are fetched and each article is stored as plain text (title, author, date, source, paragraphs). Article ids and URLs mirror the original path so a page here can always be traced to its source.
2. Translate. Each article is translated with a large language model via OpenRouter. The prompt enforces a fixed glossary of Party and Marxist terminology (below), keeps paragraphs aligned one-to-one with the Chinese, and asks the model to add numbered endnotes for allusions, idioms, internal Party terminology, and historical references that would be opaque to a reader with a Western Marxist background. Endnotes appear as sidenotes on wide screens and as a list at the end of each article.
3. Build. A small script renders the translations to static HTML with one stylesheet and no JavaScript on article pages. Search is provided by Pagefind, built at deploy time. New articles are crawled, translated, and published on a weekly schedule.

Model choice

Model comparison

Date: 2026-09-08. Total spend: $1.9720 (translation $0.3680, judging $1.6040) of a $6 cap.

Method

• Sample: 8 articles from data/raw/, stratified by Chinese character count (3 short < 2000, 3 medium 2000–6000, 2 long > 6000; when a stratum lacked enough articles the nearest by length were taken) across 5 sections; ids fixed in bench/sample.json.
• Each candidate translated every article with the production prompt (translateArticle), cached under bench/out/. Deterministic metrics per (model, article): OpenRouter-reported cost, wall time, parse OK, paragraph alignment, length ratio (en/zh chars; < 1.2 or > 4 flagged), glossary hit rate, note count, truncation.
• Blind judging: anthropic/claude-sonnet-5 and x-ai/grok-4.3 (families absent from the candidate list) each saw the Chinese source plus every candidate's output under shuffled labels A, B, C… (seed = FNV-1a of judge:id:part, recorded per judgment in bench/judge/), scored accuracy, terminology, fluency and notes 1–10 against the source only, and were told not to reward length. Articles judged: 8 of 8.
• Composite = mean over judges and articles of the 4-score mean. Spearman rank correlation between the two judges' model rankings: 0.88.
• Candidates ran cheapest-first (baseline second) and stopped launching once spend reached the cap.

Sample

Results (sorted by composite)

J1 = anthropic/claude-sonnet-5, J2 = x-ai/grok-4.3.

Skipped: z-ai/glm-5.3-flash (repeated 120 s timeouts after 3/8 articles); qwen/qwen3.5-plus-20260420 (2 consecutive errors: OpenRouter: qwen/qwen3.5-plus-20260420 failed after 4 attempts); moonshotai/kimi-k2.5 (2 consecutive errors: HTTP 200: OpenRouter: empty completion).

Worst error per model (judge's one-line note on the lowest-accuracy article)

• openai/gpt-5-mini: anthropic/claude-sonnet-5 on ddzg__202605__t20260507_5987761: Consistently mistranslates 习近平总书记 as 'President Xi Jinping' rather than 'General Secretary,' and leaves 成果 untranslated twice in the prope
• google/gemini-3-flash-preview: x-ai/grok-4.3 on mksjbyl__202607__t20260702_6057182: Inconsistent rendering of 现实世界 as 'objective world' instead of 'real world'
• deepseek/deepseek-v4-flash: x-ai/grok-4.3 on ddzg__202607__t20260702_6057180: Factual error: opens by citing the '14th Five-Year Plan' instead of the 15th.
• google/gemini-3.1-flash-lite: x-ai/grok-4.3 on ddzg__202605__t20260507_5987761: states '14th Five-Year Plan' instead of the correct '15th Five-Year Plan' for 十五五
• qwen/qwen3.7-flash: anthropic/claude-sonnet-5 on shoufa__202606__t20260611_6053364: Translation is drastically truncated, cutting off after the introductory paragraph and omitting roughly 90% of the source article's cont
• openai/gpt-oss-120b: x-ai/grok-4.3 on dsydj__202607__t20260720_6059948: mostly untranslated Chinese source text
• mistralai/mistral-small-2603: x-ai/grok-4.3 on ddzg__202605__t20260507_5987761: provides the original Chinese source text rather than a full English translation

Recommendation

Rule: highest composite among models whose mean cost per article is ≤ 25% of the baseline's (google/gemini-3-flash-preview), unless the baseline beats it by > 1.0 composite points; ties broken by glossary hit rate, then latency.

Eligible: deepseek/deepseek-v4-flash, qwen/qwen3.7-flash.

deepseek/deepseek-v4-flash — highest composite (7.48) among 2 model(s) at ≤ 25% of the baseline's $0.0120/article; baseline leads by 0.06 ≤ 1.0.

Glossary

Terms the translation model is required to render consistently.

Disclaimer

Every text on this site is a machine translation. It has not been reviewed by a human translator. Names, dates, quotations, and technical terms can be wrong; where accuracy matters, verify against the original Chinese, linked at the foot of every article. The endnotes are also model-generated and may contain errors.

Source code

The crawler, translation pipeline, benchmark, and site builder are at https://github.com/jtrv/cassm-en. Subscribe via RSS.