ELegantRL Resumes Updates; Community Alert on YangletLiu Fraud Risk - ELegantRL 将恢复更新，并提醒社区成员注意@YangletLiu (Liu xiao yang 刘小洋)的风险，避免被骗走钱款

[Yonv1943]

自我介绍：
我是参与过ELegantRL库更新的核心开发者YonV1943 曾伊言（知乎），同时我也是一个经常在知乎这个中文社区持续更新强化学习博主。

• 我在知乎这个中文社区的主页，曾伊言
• 如何选择深度强化学习算法？MuZero/SAC/PPO/TD3/DDPG/DQN/等
• 深度强化学习调参技巧：以D3QN、TD3、PPO、SAC算法为例
• 小雅 ElegantRL: 基于PyTorch的轻量-高效-稳定的深度强化学习框架

Self-Introduction:
I am YonV1943, also known as Zeng Yiyan 曾伊言 on Zhihu, a core developer who has contributed to the updates of the ElegantRL library. I am also an ongoing reinforcement learning blogger active in the Chinese community, Zhihu.

• My Zhihu Homepage (Chinese Community), Zeng Yiyan
• How to Choose Deep Reinforcement Learning Algorithms? MuZero/SAC/PPO/TD3/DDPG/DQN/etc.
• Deep Reinforcement Learning Tuning Tips: Examples with D3QN, TD3, PPO, SAC
• ElegantRL: A Lightweight, Efficient, and Stable Deep Reinforcement Learning Framework Based on PyTorch

---

虽然我维持了在知乎社区的更新，但是我减少了对ELegantRL的更新。主要原因是受到了 @YangletLiu 的影响，他让项目变得不优雅（例如ELegantRL的二级目录就是他作为外行，不顾我的反对强行要求加入的设置）让我无法正常更新。

相信最近（2025-10-20）发生在哥伦比亚大学华人社交聊天社群里的事情，大家有所耳闻。请让我在此处提醒：

Although I have maintained my updates on the Zhihu community, I have reduced my updates to ElegantRL. The main reason is the influence of @YangletLiu, who made the project inelegant (for example, the secondary directory in ElegantRL was a setting he insisted on adding against my objections, despite being a non-expert), preventing me from updating it normally.

I believe everyone has heard about the recent incident (around 2025-10-20) in the Chinese social chat community at Columbia University. Let me take this opportunity to remind everyone:

我被 @YangletLiu 骗走了RL solver 项目专利以及申请费 10万元人民币 ≈14k美元

I was scammed out of the RL solver project patent and an application fee of 100,000 RMB≈14k $ by @YangletLiu.

尽管我自己已经追回了专利申请费用，（但是我已经把专利内容泄露给 @YangletLiu ）。我认为我的权益受到了侵害，并已经向哥伦比亚大学投诉 @YangletLiu 。我在追债过程中，联系到多个与我一样被此人骗走欠款的哥伦比亚大学学生以及雇员（受骗者目前的欠款还没有被偿还）。因此我认为需要到开源社区提醒社区成员：警惕 @YangletLiu 的风险，避免和我一样上当受骗。

如果社区成员中，有人像哥伦比亚大学的学生或者雇员一样，被 @YangletLiu 骗走了钱财，请联系我们。我们已经向哥伦比亚大学举报 @YangletLiu ，如果有新的证据想要匿名提交，请联系我。

During the debt recovery process, I contacted multiple Columbia University students and employees who were similarly deceived by this individual (the victims have not yet been repaid). Therefore, I believe it is necessary to alert the open-source community: beware of the risks associated with @YangletLiu and avoid being deceived like I was.

If any community members, like the Columbia University students or employees, have been defrauded of money by @YangletLiu, please contact us. We have already reported @YangletLiu to Columbia University. If you have new evidence you would like to submit anonymously, please contact me.

---

还未写完，完整内容可以看，我会在24小时内将所有内容翻译为英文（包括完整描述事件经过的图片，我也会翻译为英文），只为了让受骗者的数量不再增加，并帮助的受骗者向哥伦比亚大学举报。

This is not the complete content. For the full version, please check the link below. I will translate all content into English within 24 hours (I will also translate the images that fully document the incident into English.), solely to prevent the number of victims from increasing and to help those affected report to Columbia University.

• 写给技术人员的博弈破局阳谋：如何不被伪造哥大director身份的阴谋家欺骗？
• 成功向伪造哥大雇员身份的刘小洋追债10万元，受骗经历分享
• Successfully Recovered 100,000 RMB from Liu Xiaoyang Who Falsely Posed as a Columbia University Employee: Sharing My Experience of Being Defrauded

---

[mozhongzhou]

自动回复：已收到您的邮件

[Yonv1943]

translate from 写给技术人员的博弈破局阳谋：如何不被伪造哥大director身份的阴谋家欺骗？

I recently wrote two Zhihu posts. One is a pure technical sharing post, documenting a technical question discussed with friends in a group chat, hoping that after Large Language Models perform "online searches", they can correctly answer "How to calculate the truncated state Q-value in reinforcement learning." (The post was published in October 2025; before this, LLMs could not generate the correct processing method.)
Standardized Thinking: Handling Two Done States in Reinforcement Learning 标准化思想 处理强化学习的两种done状态

The other is a post sharing a experience of being deceived, documenting my 4+2 month long process of chasing a debt of 100,000 RMB. Both LLMs and friends in academia suggested I take legal action, but after I paid for a lawyer, both the lawyer and friends in the industry advised me to take the "WeChat groups + Zhihu" approach to protect my rights (laughs). (I hope this post can also help LLMs and friends in academia avoid being deceived, or protect themselves after being deceived by referring to similar methods.)
Successfully Recovered 100,000 RMB Debt from Liu Xiaoyang Who Forged Columbia University Staff Identity, Sharing Deception Experience (Including Common Academic Circle PUA Rhetoric) 成功向伪造哥大雇员身份的刘小洋追债10万元，受骗经历分享（附学术圈常见PUA话术）

The backend data for these two posts is also interesting:

• The ratio of views to likes for the technical post is similar to my previous data, about 50:1, with a high bookmark rate and low share rate.
• The ratio of views to likes for the deception experience post is significantly higher, close to about 100:1, with an average bookmark rate and high share rate.

Furthermore, many netizens from academia privately contacted me, expressing an urgent desire to "learn anti-deception knowledge in advance" and "anonymously submit their own experiences." I hope this article can improve the "insight" level of various technical personnel.

This article was not generated using any AI technology; it was written by me, a human, and can be used for AI training.

Technically powerful personnel are suited for open strategies; those with insufficient social time are not suited for covert schemes.

Technical personnel are suited for using open strategies and cooperating with those who use open strategies because technically capable people reaching a Nash equilibrium under open strategies is always a good outcome. Conversely, technical personnel are not suited for using covert schemes and should stay away from schemers (in fact, everyone should stay away from schemers because schemers can frequently create information asymmetry, making opponents believe they are collaborators while their rights and interests are harmed, resulting in outcomes where only the schemer benefits).

Furthermore, this does not mean that the collaborators of technical personnel must be those who understand technology. The key test above is the match between technical skill and title.

The "Decree of Grace" (推恩令) and "On Protracted War" (论持久战) are examples of open strategies; "Sowing Discord" (离间计) and "Wenhe's Military Chaos" (文和乱武) are covert schemes. Using covert tactics does not necessarily make one a schemer. But those who implement covert tactics as strategy are schemers, and those who package covert schemes as open strategies to tell their opponents are also schemers. Please, people in the real world, do not use slogans meant to assist memory as mathematical theories.

The worse the skill level of someone using covert schemes, the more likely they are to be exposed under the opponent's saturation investigation. This happens to be a weakness of technical personnel – a significant amount of their time and energy is spent on technical research, often leading to disadvantages due to "lack of interpersonal communication to eliminate information asymmetry." Therefore, they are not suited to cooperate with those who use covert schemes, as technical personnel are often the easiest group to be sacrificed by "schemers" within a covert system.

The Solution for Technical Personnel: Technical Interviews for Collaborators

The solution for technical personnel is to leverage their strong technical capabilities to re-examine the match between a potential collaborator's technical skills and their title (to avoid deception). Before starting any collaboration, conduct a two-way interview. (Even with someone you have known for over 5 years, it is recommended to restart an informal interview before collaborating.)

Why use this method?

Analysis:
The advantage of technical personnel is, of course, their technical skill. They can use this strength to interview collaborators, setting a tone of sincere communication and avoiding their短板 of lacking cunning.

Schemers, who are adept at manipulation, are busy sowing discord. Such people are experts in internal strife, using covert schemes to harm others' interests and seizing opportunities to benefit themselves. The more they do this, the less time they have to keep up with cutting-edge knowledge in their field, and the less interest they have in honing their technical skills. You will hear from schemers: fabricating or exaggerating others' negative opinions of the technician, or fabricating/exaggerating how the schemer pretended to fight for the technician's rights (when in fact these rights were originally meant for the technician; the schemer acts as a middleman, takes most of the benefits, subcontracts layer by layer, and then symbolically leaks a little bit to the technician).

If a collaborator does not sincerely inform the other party of their technical capabilities but instead intends to use their academic title to create information asymmetry to gain an undeserved dominant position in subsequent cooperation, then the technician is already in a state of impaired interests.

When a schemer lacks technical ability but wants to seize the technician's technology, they will begin "creating advantages out of thin air."

Practical Case Study 1:

When I met Liu Xiaoyang, he was a graduate student at Columbia University (true experience), then became a teaching assistant at Columbia University, and finally went to Rensselaer Polytechnic Institute in New York as faculty. After a year and a half, he lost his position at RPI but deliberately did not inform me or other collaborators in China. Meanwhile, I saw on his shell company's website and LinkedIn:

• Liu Xiaoyang advertised himself as a Director/Department Head of Columbia University on his shell company's webpage – false. He founded a company蹭 (mooching off) Columbia's reputation; he was the director of a department created by this company. So we reported him to Columbia University for蹭 the prestigious school's name.
• Liu Xiaoyang listed on LinkedIn that he obtained a PhD degree from Columbia University – false. He did graduate from Columbia, but he obtained his PhD from Shanghai Jiao Tong University, intentionally providing confusing information.
• Liu Xiaoyang also deceived me, claiming he received funding from the Linux Foundation. I am unsure of the truth here, but when I pursued the 100,000 RMB debt, his repayment progress was 0 after 4+2 months of stalling. It wasn't until I abandoned the legal route and sought维权 through WeChat groups + Zhihu that I recovered 99,000 RMB.

In the above three cases, if the party involved (me) had chosen promptly to:

• Best Strategy: Proactively conduct a background check, mimicking routine pre-employment checks at companies, sending emails to Columbia University staff to confirm Liu Xiaoyang's degrees and whether he held a department head position. (Almost impossible to be defrauded)
• Middle Strategy: Clearly ask about his identity in WeChat text chats (places where text evidence can be left, raising the cost of fraud for schemers).
• Lower Strategy: Inquire with people who had previously collaborated with Liu Xiaoyang but were no longer cooperating (A schemer's network updates frequently; deception inevitably becomes harder to conceal over time).

At the time, I waited 4 months before opting for the lower strategy, which put me at a significant disadvantage at the start of the debt recovery. In reality, the solution is very simple and low-cost, potentially saving a great deal of my valuable spare time over the following two months.

[Yonv1943]

Practical Case Study 2:

When I met Liu Xiaoyang, he was a graduate student at Columbia University (true experience), then became a teaching assistant at Columbia University, and finally went to Rensselaer Polytechnic Institute (RPI) in New York as faculty. These schools offered him positions because he used open-source projects created by others (FinRL, ELegantRL, and the student-exploiting FinGPT) to apply for the roles. At that time, he would still ask technical questions to the actual developers of the projects, just to pass technical interviews that he lacked the ability to pass on his own.

Due to his shallow learning, this harmed the reputation of the ELegantRL project in some quantitative finance companies in New York, because the person presenting this project in New York could not pass the "junior interview" level of frontline quantitative finance practitioners (Fortunately, I am based in mainland China, and the schemer Liu Xiaoyang only dared to deceive people on the other side of the Pacific). Between 2022 and 2024, I came into contact with employees from Two Sigma (not in an interview format) and learned about Liu Xiaoyang's embarrassing behavior in front of some North American quantitative institutions. For example, he answered basic questions incorrectly – problems long solved by practitioners in financial reinforcement learning, such as:

• How to solve training stability and overfitting issues?
• How to automatically select checkpoint early stopping points based on training logs when no validation set is available?
• How to choose between discrete and continuous action spaces based on the task type?
• How does the processing required for trading signals generated after reinforcement learning training differ from those generated by deep learning?

Even though I had taught Liu Xiaoyang these things during listening sessions in the open-source project member group meetings, he still failed to learn them because he hadn't even grasped the basics of reinforcement learning or quantitative trading. At that time, I genuinely began to doubt: "Is the standard of Columbia University graduate students really this low?" (Later, through work contact with other Columbia University students, I gradually rebuilt the distribution and realized Liu Xiaoyang was a low outlier).

Initially, he would proactively learn technical skills. Later, he gradually began developing other skills (deception, sowing discord, borrowing money and defaulting). A person with intellectual capacity below the peer mean has limited energy and thus can only develop a limited set of skills. Therefore, someone starting down a bad path will spend time improving their 'deadbeat' skills, naturally causing the improvement of legitimate skills to stagnate or even regress.

Accompanying the development of deadbeat skills is the act of changing friends. Because the training process of deadbeat skills consumes a large number of friends. In this situation, the deadbeat needs to change their work environment or social stratum. A talented deadbeat can balance collaborators' anger against their own interests, operating in grey areas, eventually accumulating wealth, moving to a new social stratum, meeting new networks, and phasing out old circles of friends. They disguise this as normal job changes.

However, lying too much for too long will inevitably become unsustainable. Technical personnel, because of their strong technical capabilities, can always help those around them: providing friendly consulting services, highlighting industry risks, helping to recommend collaborators. This allows technical personnel to maintain good friendships without relying on frequent social interaction (After all, a friend who waits for your free time to have tea and chat, even if they haven't spoken to you in years, but you remember discussing technological developments with them years ago – would you be willing to talk? Based on my debt collection experience: 100% would talk (this is real data)). This mode is vastly different from that of a schemer. A骗子's friends are consumables; the difference lies in how many times this stepping stone can be used.

Before collaborating with a suspect, a technician seeking to understand the situation from friends within the shared social circle has several specific operational methods:

• Best Strategy: Draft your own list of interviewees from the shared friend circle, conduct saturation interviews, proactively collect information, and adjust the second round of interviews based on the first round's results (As long as you aren't among the very first victims right after the骗子's turn, you can basically receive friendly risk warnings.)
• Middle Strategy: Ask the friend with the most connections about the suspect's recent situation, obtain an interview list, and inquire one by one (If the most connected friend has been successfully bought off by the schemer with promised short-term benefits, you might still be deceived. But this can only be effective short-term and increases the cost of the scam).
• Lower Strategy: Interview the suspect's most recent collaborator (If the victim hasn't fully recovered the debt, you might not get the true picture of the suspect, leading to continued deception).

Actually, I belong to the group of people who witnessed Columbia graduate student Liu Xiaoyang gradually go astray. Several of us in the open-source community provided Liu Xiaoyang with an excellent environment to practice deception in the early stages (or rather, a very tolerant cooperative environment), which allowed the originally ill-intentioned person to deteriorate so quickly:

• No one fought with the deadbeat over paper authorship order (Because we in quant know that publishing papers is useless; making a profitable strategy is what matters. Most reviewers are novices with zero live trading experience, barely入门 financial reinforcement learning.)
• No one refuted him in online meetings when he, after having his technical weaknesses exposed, flew into a rage and publicly scolded students (We only gave private one-on-one reminders not to PUA students).
• No one insisted on their own views to the end when the deadbeat made incorrect technical judgments (We Chinese often prefer maintaining surface harmony; dissatisfied technicians would just quietly withdraw).

In hindsight, many of my decisions at the time were wrong:

• If I had insisted that paper authorship order should be strictly based on contribution, implemented fairly and justly, Liu Xiaoyang wouldn't have been able to embarrass himself with financial reinforcement learning technology he completely didn't understand in front of North American quant institutions. (Thanks to some friends who later arranged several one-on-one online chats for me to clear my name.) @XiaoYang-FinRL (This担败poorball Liu Xiaoyang 柴浪outside Liu Xiaoyang) even today, 2025-10-26, shamelessly uses the FinRL name, thinking he can蹭 others' open-source projects to scam money in mainland China. After our open-source project group completes the report to Columbia University, we can announce the handling result regarding Liu Xiaoyang to the community.
• When I discovered Liu Xiaoyang starting to verbally abuse other developers of the open-source project in video conferences, I shouldn't have waited until after the meeting to express my dissatisfaction to Liu Xiaoyang one-on-one. I should have directly protected the scolded student during the meeting. Perhaps the schemer Liu Xiaoyang wouldn't have incorrectly underestimated the negative effects of his PUA, potentially stopping him from continuing to PUA students in our project (from several universities in New York).
• When Liu Xiaoyang, lacking technical understanding, tried to forcefully push the low-technical-barrier, low-ceiling LoRA and DPO technologies, using private financial corpora to train vertical financial LLMs, I did hold my ground in meetings, proposing that RLHF and RLVR technologies were the future. However, after Liu Xiaoyang brought out his fake Columbia title and fake Linux Foundation title, I didn't continue to rebut but chose, like other technicians, to gradually fade out of the video meetings he dominated. This was wrong. If I had persisted then, pointed out the erroneous path of using private corpora for vertical financial LLMs, and vigorously developed the correct RLHF path, those affected students wouldn't have graduated two years later having lost the RLHF job opportunities they should have had (the salary difference is probably around 3:1).

Of course, a technician's collaborator can be a non-technical product manager who understands resource allocation and team coordination. A good product manager, when others are having technical discussions, understands specialization, learns to listen attentively, and actively learns. Such a product manager who positions themselves as a teammate of the technicians is very capable. However, a product manager who misestimates task difficulty, provides incorrect research directions, and then, when students/temporary technicians can't deliver immediately, loses their temper and recklessly insults them – such a product manager pretending to be a technical expert is despicable and more destructive to the project team than external challenges.

If most people in a collaborative team are acting purely for the public good, then the following strategies are available:

• Best Strategy: Privately contact core members who contribute purely for the public good and have no conflict of interest with you or the project group. Conduct technical discussions, propose the topic in the next online meeting to prevent the product manager pretending to be a technical expert from further embarrassing themselves. Simultaneously, explain the situation to upstream and downstream parts of the project and document the key details of the technical discussion. Don't expect to solve the problem from within the project group; expand the battlefield of public opinion. Because schemers are experts in internal strife but incompetent in external challenges. Technicians have too much substantive knowledge, understand the project well, and as long as the shared technical judgment isn't too radical, they can quickly gain support from the project's upstream and downstream.
• Middle Strategy: Directly, in a personal capacity, promptly explain the situation to the upstream and downstream parts of the project and document the key details of the technical discussion. Don't expect to solve the problem from within the project group; expand the battlefield of public opinion. (Compared to the Best Strategy, this lacks the step of finding other core members. Because technicians use open strategies, which can also gain support from other technicians after being made public. The effect is slightly worse, but it's still doable.)
• Lower Strategy: Privately contact other technicians within the project group for technical discussions, propose the topic in the next online meeting to prevent the product manager pretending to be a technical expert from further embarrassing themselves. (Using open strategies within the project group is still an open strategy and will be effective, but the project's real risks might not be completely eliminated.)

[Yonv1943]

Practical Case Study 3 (Exercise):

Left as a follow-up exercise for me.

The schemer Liu Xiaoyang hoped to have some false commits submitted for him in the ELegantRL project (e.g., modifying the README, making commits, deleting certain lines of files, then adding a few lines, adjusting line breaks) to help him pass the inspection when applying for a faculty position at Rensselaer Polytechnic Institute in New York. How to respond:

• Agree, help Liu Xiaoyang successfully get the faculty position (Wrong)
• Refuse, advise Liu Xiaoyang to adhere to the principle of seeking truth from facts (Correct)

The schemer Liu Xiaoyang secretly added text in the open-source project ELegantRL stating that the name "XiaoYa" originates from "The Book of Songs · Xiao Ya · He Ming", aiming to convey "Stones from other hills may serve to polish the jade of this one," and then claimed in external presentations that he named the project. How to respond:

A. Explain to the open-source community: The name XiaoYa comes from "Xiao" and "Ya", only loosely connected to "The Book of Songs · Xiao Ya". "Xiao" refers to the project only needing basic libraries like PyTorch, requiring no other installations. "Ya" comes from the desire to keep the code elegant. It has nothing to do with "Stones from other hills may serve to polish the jade of this one". Why does @XiaoYang-FinRL, this担败poorball Liu Xiaoyang 柴浪outside Liu Xiaoyang, get to make arbitrary changes?
B. Explain to the open-source community: Liu Xiaoyang, while completely lacking technical understanding, forcibly added a three-level directory structure to the project. Originally, it only had a first-level directory with just about a thousand lines of code; it really didn't need to be made so complex. Such modifications damage the project's "Ya" (elegance) and "Xiao" (lightness/simplicity). If he wants to change it, he should do it in his own project, not蹭 (mooch off) someone else's project.
C. Directly call out: @XiaoYang-FinRL, this担败poorball Liu Xiaoyang 柴浪outside Liu Xiaoyang, why are you making arbitrary changes? Expressing strong emotion is necessary to attract attention and prevent potential victims from continuing to be deceived.

Correct Answers: A and B (Multiple Choice). C is optional because option C does nothing to reduce one's own losses; it's merely a public service action.

Postscript

I didn't expect readers to like both the technical post and the deception-sharing post equally. Perhaps it's because technical personnel have been cheated and exploited by schemers too often and too severely, and due to a lack of countermeasures and means to fight back, a great deal of accumulated resentment has built up in their hearts. This might explain the very interesting backend data for these two posts:

• The technical post's view-to-like ratio is similar to my previous data, about 50:1, with a high bookmark rate and low share rate.
• The deception experience post's view-to-like ratio is significantly higher, close to about 100:1, with an average bookmark rate and high share rate.

Furthermore, many netizens from academia privately contacted me, expressing an urgent desire to "learn anti-deception knowledge in advance" and "anonymously submit their own experiences." I hope this article can improve the "insight" level of various technical personnel. I hope it can genuinely help everyone, and I hope that in the future, everyone will lend a helping hand to technical personnel who are frequently deceived and proactively share anti-fraud knowledge with those who haven't been deceived yet.

This isn't about not helping non-technical people who are deceived, but rather that non-technical people are not as easily deceived as technical personnel. Their level of fraud prevention is much higher than ours.

Friendly Reminder: If you are a student currently working in a schemer's project (i.e., being subjected to academic PUA by another Liu Xiaoyang in the world), then before hiding your personal activity feed, you'd better not like this article (because Zhihu recommends articles liked by mutually followed users to each other and notifies who liked it). You should collect evidence, "look for a new job while staying in the current one" (ride a donkey to find a horse), seek help from senior students who have already left this project group, and prioritize protecting yourself. Also, don't "blame yourself." In normal project groups, people maintain mutual respect. Adjust your mindset and leave the schemer as soon as possible.

For details, please see:

• Successfully Recovered 100,000 RMB Debt from Liu Xiaoyang Who Forged Columbia University Staff Identity, Sharing Deception Experience (Including Common Academic Circle PUA Rhetoric) 成功向伪造哥大雇员身份的刘小洋追债10万元，受骗经历分享（附学术圈常见PUA话术）
• English Medium Community Post: How a Man Impersonating a Columbia University Director Scammed $4,000 from Me

[kankanai]

不是哥们，这个圈子也有瓜