Lingfeng Shen (@Lingfeng_nlp) header
Lingfeng Shen (@Lingfeng_nlp) profile picture

Lead Scientist @Apodex_AI | Prev MAI Superintelligence Team, Bytedance Seed, @jhuclsp. An LLM researcher, a trader, and a @MercedesAMGF1 Fan!

Saratoga, CA
Joined February 2020
Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)Photo by Lingfeng Shen (@Lingfeng_nlp)
Lingfeng Shen retweeted
Can LLMs automate frontier LLM research, like pre-training and post-training? In our new paper, LLMs found post-training methods that beat GRPO (69.4% vs 48.0%), and pre-training recipes faster than nanoGPT (19.7 minutes vs 35.9 minutes). 1/
11
139
16
586
118,704
Lingfeng Shen retweeted
Replying to @KobeissiLetter
María Corina Machado is a genius. She understood something the Nobel Committee never will. She understands power, psychology, and how the world actually works, not how moral bureaucrats pretend it works. By openly praising Donald Trump and symbolically giving him the Peace Prize, she demonstrated real intelligence. Not ideology. Not moral vanity. Intelligence. The members of the Nobel Committee, on the other hand, are trapped in a childish worldview - good versus evil, heroes versus villains, black versus white. That is not wisdom. That is simplification for people who cannot handle reality. History does not move forward because of moral purity. It moves forward through paradox, uncomfortable figures, and people willing to act beyond ego and moral theater. If the Nobel Committee had any depth or humility, they would have awarded the prize to Trump themselves - not because he is “good,” but because the world is not a fairy tale. That single decision would have shifted global dynamics more than a thousand empty speeches about peace ever could. But they couldn’t do it. Too much ego. Too little understanding. María Corina Machado saw it clearly. The Nobel Committee did not.
49
14
1
166
29,981
Lingfeng Shen retweeted
Peter Thiel today: “Boomers are strangely uncurious about how the world is not really working for their kids. It’s always hard to know how much bad faith there is or how bad the actors are. I think it’s odd that people thought it was odd that I was complaining about student debt in 2010, when even then the growth in student debt was an exponential process. The national student debt was $300 billion in 2000, and it’s now more than $2 trillion. At some point, that breaks… If all you can say is that Mamdani is a jihadist, communist, ridiculous young person, what that sounds like to me is that you still don’t have any idea what to do about housing or student debt. If that’s the best you can do, you are going to keep losing.”
588
1,089
323
9,947
2,101,777
Lingfeng Shen retweeted
How to start a business on @Polymarket 101 The origin of the new garage-band market-maker 🧵
14
26
1
255
62,048
Lingfeng Shen retweeted
How can we boost LLM agents’ generalizability to OOD tasks and environments? Check out CodeGym, our new project for synthesizing environments for LLM agent RL training. CodeGym is a synthetic environment generation framework for reinforcement learning on multi-turn tool-use tasks. It automatically converts static coding problems into interactive and verifiable RL training environments. Training in CodeGym leads to strong OOD generalization — for example, a Qwen2.5-32B-Instruct model achieved an 8.7-point absolute accuracy gain on τ-Bench! We’ve just released the paper, synthesis pipeline, and dataset: 📄 Paper: arxiv.org/abs/2509.17325 💻 Project: github.com/StigLidu/CodeGym 📊 Dataset: huggingface.co/datasets/Vani… 📷 More details in the thread👇
3
6
2
13
1,980
Intelligence is not the ability to remember the past. 
Intelligence is not the ability to react to the present. 
The best definition of intelligence is the ability to predict the future!
1
10
1,149
Lingfeng Shen retweeted
Finally Someone Looks at the ODTE IRON CONDORS.... In ODTE 3 minutes makes a big difference.. S&P 500 Index (SPX) iron condors executed on zero-days-to-expiration (0DTE) contracts at 3:55 PM EST and 3:58 PM EST, respectively. The evaluation period spans January 2013 through early 2025, encompassing over 3,100 real trade records extracted from structured backtest logs. Each strategy sells one same-day expiring iron condor near the close of trading using a standardized position size and fixed strike width. Both strategies involve selling 45-lot SPX iron condors with 5-point wings on each side, collecting a premium in exchange for accepting defined downside and upside risk. The only difference between the two is entry timing: one initiates at 3:55 PM EST, and the other at 3:58 PM EST—just two minutes before market close. Trades are held until expiration at 4:00 PM or exited early if a stop-loss threshold equal to 130% of the premium collected is breached. All simulations assume an initial capital of $30,000, consistent with the approximate margin required to hold a 45-lot iron condor with a 5-point spread. Key metrics include median terminal equity, maximum drawdown, and probability of account ruin. Both are positively skewed, characterized by a high concentration of small-to-moderate gains and a long left tail of occasional large losses. The 3:58 PM entry strategy (orange) shows a noticeably tighter range of outcomes, with most trades generating between $500 and $3,000 in profit. In contrast, the 3:55 PM entry (blue) displays a broader dispersion in both directions, with slightly more frequent and deeper losses
9
12
1
79
15,666
Lingfeng Shen retweeted
There is no such thing as alpha. Most hedge fund managers were simply in the right place at the right time, found a temporary, repetitive pattern, and rode it all the way up—mistaking it for skill or intelligence. For the lucky ones, the pattern lasted a few more years than for others—but luck always runs out. This is the truth. This is why insider trading exists and why no one consistently beats the index—because alpha doesn’t exist. It’s just luck. The only real alpha is knowing when your luck has run out and finding a new pattern to exploit. That requires having strong opinions—but holding them loosely.
49
47
18
534
199,645
Looking for a better way to fuse speech and text modality with pre-trained large language models? Check out our paper: SSR: Alignment-Aware Modality Connector for Speech Language Models 💡 🔗 arxiv.org/abs/2410.00168 #SpeechLM #ModalityFusion
3
9
39
5,516
Low-resource, high-risk: A @JohnsHopkins study (feat. @DanielKhashabi, @Lingfeng_nlp, @weiting_nlp, @YunmoChen, @jackjingyuzhang, @fe1ixxu, & more) finds that #AI text generators are more likely to create harmful content in certain languages. Learn more: bit.ly/3Y6SacA
2
10
14
4,069
Lingfeng Shen retweeted
Here’s some better news: Combining CPO and SimPO can likely improve the model! Check out more details in our GitHub code: github.com/fe1ixxu/CPO_SIMPO
We recently had multiple rounds of discussions with the SimPO authors regarding the lack of comparison to CPO in their main paper. We both agree that it was an unintentional oversight, and they will update the paper to address it. We appreciate their positive and prompt response in engaging with us to resolve the issue! After my personal experience and reproduction of SimPO, I would like to give a shout out to SimPO—an impressively stable reference-free preference learning method compared to others!
1
13
36
12,309
Will be presenting my work on LLM-MT at #naacl today, at In-Person Poster Session 7 (11am). Please come if you are interested in machien translation and multilinguality! Would also love to discuss multimodal LLM (speech + text) and non-autoregressive decoding (NAT, diffusion)!
📚 LLMs benefit from demonstrations, but can we make zero-shot translations better without access to parallel corpora? Find out how we bridge the gap between zero- and few-shot translation with only monolingual information in our latest paper! 🤓 arxiv.org/abs/2311.02310 @jhuclsp
5
15
2,233
#RLHF is taking the spotlight. We usually focus on boosting reward and policy models to enhance RLHF outcomes. Our paper dives into the interactions between PM and RM from a data-centric way, revealing that their seamlessness is crucial to RLHF outcomes. arxiv.org/abs/2406.07971
2
17
2
70
18,163
Overall, from a data-centric perspective, we conceptualize 'seamlessness' between RM and PM, highlighting its role in RLHF. Additionally, we introduce an automatic method to estimate seamlessness and provide two examples demonstrating how SEAM aids in real-world RLHF scenarios.
1
2
260
So happy to work on this project with @TaimingLu, @Xinyu2ML, @weiting_nlp, @BeidiChen, and @HuaxiuYaoML. Thanks to my amazing collaborators! 🤗
4
236