LMG

[LMG @lmgAboveFog]

3. Model's capacity to remember and process long-term context is a limitation. This probably motivates the LSTM and attention mechanism.

[LMG @lmgAboveFog]

After working on LLM research and applications for about two years, I decided to go back in time and revisit some the critical works from time to time now. Takeaways in comments.

Starting with The Unreasonable Effectiveness of Recurrent Neural Networks: https://t.co/qXiZuoPF7o

[LMG @lmgAboveFog]

2. Emergent understandings without explicit supervision. From predicting the next character, the model learns generating complex texts implicitly.

[LMG @lmgAboveFog]

1. Context is critical to forming emergent understandings in models. The whole point of RNNs is that state carries forward, letting the model remember long-range dependencies.

[Muyu He @HeMuyu0327]

Our EMNLP main paper presents a fun but very challenging benchmark for LLMs: to solve over 300+ Ace Attorney detective cases that are ultra-long-context and can only be solved by long chain-of-thoughts. Here're the lessons we learned:

- Brute test time scaling does not work for long logical reasoning tasks, as GPT-4.1 can one shot the example as good as QwQ-32B

- Even when the space of possible evidences is super large, models of different sizes are still doing okay at identifying the questionable evidences

- The real struggle happens when models perform the long deduction steps, not identifying the correct facts, setting a upper bound of models' reasoning abilities

- Models tend to think a lot more for questions they are eventually wrong at, because they are stuck in a loop and just keep bringing up new points

- Surprisingly, by given the complete story context which has no additional clues, larger models do a lot better, suggesting the importance of reasoning in context

As probably zero models are benchmaxxing on deductive stories, our leaderboard faithfully shows how reasoning abilities generalize well out of domain (spoil: DeepSeek rocks).

Check it out on your model of choice!

[LMG @lmgAboveFog]

Congrats on the wonderful work! If anyone is interested in LLM Personalization, take a look at this!

[Lauren Jiang @laurenbjiang]

🚀 How well can LLMs know you and personalize your response? Turns out, not so much!

Introducing the PersonaMem Benchmark --
👩🏻‍💻Evaluate LLM's ability to understand evolving persona from 180+ multi-session user-chatbot conversation history
🎯Latest models (GPT-4.1, GPT-4.5, o4-mini, Llama-4, Gemini 2.0, Deepseek-R1, Claude-3.7) all struggle in personalization!
🎨7 personalization skills tested in 15 scenarios
🌟Realistic long-context evaluation up to 1M tokens

👇 Check out what we discovered… (1/6)

[Lauren Jiang @laurenbjiang]

🚀 How well can LLMs know you and personalize your response? Turns out, not so much!

Introducing the PersonaMem Benchmark --
👩🏻‍💻Evaluate LLM's ability to understand evolving persona from 180+ multi-session user-chatbot conversation history
🎯Latest models (GPT-4.1, GPT-4.5, o4-mini, Llama-4, Gemini 2.0, Deepseek-R1, Claude-3.7) all struggle in personalization!
🎨7 personalization skills tested in 15 scenarios
🌟Realistic long-context evaluation up to 1M tokens

👇 Check out what we discovered… (1/6)

[Lauren Jiang @laurenbjiang]

👥 Authors: Bowen Jiang* @laurenbjiang, Zhuoqun Hao* @ZhuoqunHao, Young-Min Cho @jeffrey_ch0, Bryan Li @bryanlics, Yuan Yuan @lmgyuan, Dr. Sihao Chen @soshsihao, Prof. Lyle Ungar @LyleUngar, Prof. Camillo J. Taylor, @camillo_taylor, Prof. Dan Roth @DanRothNLP (*co-first authors) 🏛️ University of Pennsylvania @PennEngineers @PennEngAI @CIS_Penn @upennnlp @GRASPlab (6/6)

[LMG @lmgAboveFog]

@yuyy614893671 @turingou 我也同意你的说法。现在btc是1.3万亿。如果能达到黄金的水平（13万亿），那就是现在的十倍。我们打个对折6万亿，那就是现在的5倍左右。我再保守一点，估计在可预见的未来里比特币应该能到20万。再��上就看天意了。

[LMG @lmgAboveFog]

@lidangzzz 每次看到这些私信就觉得国内的教育太sb了。先是pua你到相信成绩至上，然后不给人去尝试别的东西和了解这个世界的机会。最后导致除了天赋异禀或家庭有资源，几乎没人知道自己喜欢什么，应该追求什么。

[LMG @lmgAboveFog]

@jackli727 报名

[LMG @lmgAboveFog]

@jackli727 求拉群

[LMG @lmgAboveFog]

@sleepinyourhat May I ask if you take any interns for the summer? Thanks!

[LMG @lmgAboveFog]

@0xcryptowizard 怎么找到头部有上交易所潜力的标的呢？

[LMG @lmgAboveFog]

@sunlc_crypto 想知道套利和打新有什么策略或者方法可以用

[LMG @lmgAboveFog]

@wangfeng_0128 峰哥您好！不好意思打扰您，一直看您的推文学习，受益良多。我想请教一下，作为一个刚入web3对比特币生态感兴趣的计算机学生，我应该怎么找到愿意一起尝试做bitcoin defi的伙伴呢？现在比特币的应用生态还没有成型，有很多机会，我想尝试一下做一个lending platform。希望峰哥能指个方向。十分感谢！

[LMG @lmgAboveFog]

@whyyoutouzhele 能割韭菜而不割韭菜实在太可惜

[LMG @lmgAboveFog]

@ihateUSPS The most stupid service and worst fucking staff