DecryptPrompt: 66 installments of LLM research, explained in Chinese

The series as the repository's spine

The 跟着博客读论文 section, following the blog to read papers, indexes sixty-six numbered installments of the Deciphering Prompt series hosted on Tencent Cloud's developer platform. The arc is readable from the titles alone, installments one through three covering tuning-free prompts, freezing prompts while tuning models, and freezing models while tuning prompts, four and five on instruction tuning and automated instruction construction, six on LoRA fine-tuning realities, seven beginning the alignment thread with an OpenAI, DeepMind and Anthropic RLHF comparison, then long inputs, chain of thought across three installments, agents from ReAct through Toolformer to ToolLLM, RAG's recall diversity and density questions, and later entries on MCP from basics to FastMCP, agent memory through Mem0 and LlamaIndex, and context engineering dissected in single and multi-agent code. The later installments show the series keeping pace with the field's shifts rather than repeating its early beats, fifty-four through fifty-seven covering context caching, memory engineering through Mem0 and LlamaIndex, and context engineering code dissected for single and multi-agent systems, fifty-eight and fifty-nine moving to MCP from protocol basics through the low-level API to FastMCP builds, sixty and sixty-one building a Jupiter data-analysis agent from zero and a hand-rolled code sandbox with FastAPI-MCP, and sixty-three covering agent training approaches including RStar2 and early experience.

The archive: forty directories of papers

The repository's directory listing is a topic map of post-2022 LLM research, LLM_agent, LLM_memory, LLM_dialog, LLM_chart, LLM_KG, LLM_ability and LLMS beside task domains like code_generation, nl2sql, image_generation and humanoid. Method domains include RLHF, post_train, instruction_tunning, prompt_tunning, prompt_chain_of_thought, pretrain_data, quantization, MOE, model_edit, model_merge, inference, long_input, long_output, multi-turn, multimodal, adversarial, evaluation, reliablity, self-evolution, skilll_evolve, context_engineer and harness, with survey, others and CS224N_slides rounding out. Slides and PDFs sit at the root, including the Choose Your Weapon survival strategies paper for depressed AI academics, which the README's opening line recommends to anyone discouraged by LLMs' sudden arrival. The archive's granularity is its value, directories like model_edit and model_merge distinguish research areas that general lists collapse, and domain_llms sits beside the general LLMS so a reader following a vertical finds the domain adaptation papers without grep.

Six curated link files

The LLM resource summary section indexes six Chinese-named markdown files. 开源模型, open source models, with evaluation leaderboards. 开源框架, open source frameworks for inference, fine-tuning, agents, RAG and prompt tooling. 开源数据, open SFT, RLHF and pretraining datasets. AIGC各领域应用, applications across AIGC domains. 教程博客会议, tutorials, classic blogs and AI conference interviews. And 推荐SKILLS, recommended agent skills tools. Alongside them sit 值得学习的智能体框架, agent frameworks worth learning, and 几句话聊论文, a few words on papers, the short-form companion to the long series.

A living index, updated to this month

The README promises continuous updating with a star to keep updated, and the record supports the claim, the series runs from installment one in late 2022 through sixty-six covering DeepSeek-OCR's optical compression of visual tokens, with the repository's last push on 2026-09-24, six days before this writing. There are no releases, no license file registered, and no primary language detected, the profile of a repository whose artifact is writing rather than code. The chronological series doubles as a history of the field's attention, each installment marking what the community was arguing about when it was written. The star-to-follow framing positions the repository as a subscription, each push delivering new series installments and archive folders, and the MELLM entry from December 2023 with its Baidu netdisk password shows the resource sharing style, artifacts distributed through whatever channel reaches Chinese readers fastest.

Series threads worth following in order

Several threads reward sequential reading. The COT thread, installments nine through eleven, moves from chain of thought basics through principle exploration to making small models reason. The alignment thread spans seven, sixteen, seventeen, twenty-four, twenty-five and twenty-seven, from RLHF comparisons through data efficiency, WizardLM and back translation, SLiC-HF and DPO, RLAIF, and reducing general capability loss. The RAG thread runs twenty, twenty-one, twenty-two, thirty-seven, thirty-nine and forty-one, from recall diversity through information density, the compression question, pre-retrieval decisions, LLM-refined ranking, and whether GraphRAG is a silver bullet. The agent thread is the longest, twelve through fifteen, eighteen, nineteen, twenty-eight through thirty-one, and fifty. The alignment thread alone reads as a graduate seminar, six installments moving from the three labs' RLHF practices through what LIMA taught about data efficiency, self-alignment, the DPO family, RLAIF's annotation question, and the practical problem of not degrading general capability while specializing.

Who writes it, and for whom

The author writes as a practitioner following the field, and the framing is peer to peer, the opening line empathizes with researchers daunted by LLMs' pace and recommends a survival-strategies paper before the content begins. The audience implied by the choices is Chinese-reading engineers and researchers who want the papers filtered, connected and translated into working understanding, and the pairing of every installment with implementation details, LoRA hour estimates, code examples for structured output and context caching, marks the series as engineering reading rather than press coverage. The CS224N slides directory anchors the archive's fundamentals, giving readers arriving without the background a standard course to reference before the research folders make sense.

Editorial conclusion

Read DecryptPrompt when you prefer Chinese-language explanations of LLM research and want a curated path through prompt tuning, alignment, agents, RAG and reasoning, since each series installment pairs the papers with the author's synthesis rather than bare links. Verify claims against the underlying papers as with any secondary source, and note the reading resources lean on Tencent Cloud articles and Zhihu posts, so access depends on those platforms. Use the top-level directories as a paper archive by topic, the Chinese-named markdown files as the curated link lists, and the series index as the chronological map of how the field moved from prompt tuning to context engineering.

Frequently asked questions

Official sources

1. DSXiangLi/DecryptPrompt on GitHub
2. Issues
3. README