DeepSeek's Deli Chen releases an LLM continual learning survey paper generated autonomously by the DeliAutoResearch system

VIEWS5.5KBOOKMARKS31LIKES60RETWEETS6REPLIES1

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

It's actually pretty cool that Deli is moving the Overton window all by himself. This is not "AI slop", the quality is well above a naive LLM generation or even some "deep research" report that often gets passed for a legit human paper. We should openly share AI surveys.

Deli Chen@victor207755822

The Second Paper from DeliAutoResearch SKILL is here! 🎉✨

https://victorchen96.github.io/continual_learning_survey.pdf

This time the focus is on continual learning and self-iteration, which I believe is an absolutely essential step on the road to AGI. 🧠💡

Future AI systems definitely won't be frozen sets of parameters — they'll be systems that can self-iterate and keep evolving. 🌱🔄 Context management and documented memory can help a model sustain attention, but the attention window will eventually get filled up. When that happens, you need to parameterize knowledge and experience to reduce the cognitive load. 📚🧸

Also, the (simulated) peer review score hit an 8 this round — a huge leap from the 6 last time! 🚀🐣 In this paper, the model tried calling advanced language models to design and run experiments all on its own for the very first time — a capability it didn't have before. 🤖🌸

Btw, the second image includes a production data comparison between Paper 2 and Paper 1. In a nutshell, as the skill itself iterated, the number of interaction rounds dropped dramatically, while the total tokens consumed went up significantly — which is actually a really good sign! It shows SKILL is shifting toward greater autonomy. 📊🐱💕

Really looking forward to the day soon when DeliAutoResearch SKILL can produce truly master-level academic writing. 🎓🍀🌟

31d5.5K6031

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

it is ironic that the paper about continuous learning is obviously biased towards bits preceding the agent's knowledge cutoff. Virtually nothing from 2025-2026 made the cut.

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

It's actually pretty cool that Deli is moving the Overton window all by himself. This is not "AI slop", the quality is well above a naive LLM generation or even some "deep research" report that often gets passed for a legit human paper. We should openly share AI surveys.

31d2.3K111

Deli Chen@victor207755822

Since arXiv doesn't allow AI to be listed as authors, I had to move DeepSeek-V4-Pro and GPT-Image-2 from the author line to the footnotes — even though they literally did 99% of the work on this paper. 🤖🐣✨📝

Deli Chen@victor207755822

The Second Paper from DeliAutoResearch SKILL is here! 🎉✨

https://victorchen96.github.io/continual_learning_survey.pdf

This time the focus is on continual learning and self-iteration, which I believe is an absolutely essential step on the road to AGI. 🧠💡

Future AI systems definitely won't be frozen sets of parameters — they'll be systems that can self-iterate and keep evolving. 🌱🔄 Context management and documented memory can help a model sustain attention, but the attention window will eventually get filled up. When that happens, you need to parameterize knowledge and experience to reduce the cognitive load. 📚🧸

Also, the (simulated) peer review score hit an 8 this round — a huge leap from the 6 last time! 🚀🐣 In this paper, the model tried calling advanced language models to design and run experiments all on its own for the very first time — a capability it didn't have before. 🤖🌸

Btw, the second image includes a production data comparison between Paper 2 and Paper 1. In a nutshell, as the skill itself iterated, the number of interaction rounds dropped dramatically, while the total tokens consumed went up significantly — which is actually a really good sign! It shows SKILL is shifting toward greater autonomy. 📊🐱💕

Really looking forward to the day soon when DeliAutoResearch SKILL can produce truly master-level academic writing. 🎓🍀🌟

31d2.2K230

Sub-account📄🐣🐳@brahim19436933

@victor207755822 @zephyr_z9 My strategy plan !! 🚀 ⬇️

31d168

Ankith 🐋/acc@dhtikna

@victor207755822 @Grad62304977 have you read either this or the previous survey abt auto-research agents. Do you feel they are competent (comprehensive/pareto optimal)?

31d34

Alex UGift@Radipdegen

@victor207755822 glad u admit frozen weights isnt the endgame

think this is the part everyone quietly skips

31d2201

椰奶老师@mimirb

@victor207755822 持续学习才是迈向AGI的必经之路啊

31d1531

Aeron@aeronxbt

@victor207755822 second paper already dropped, love the pace

"never stop learning" as a paper title for an LLM survey is too fitting lol

31d226

Inflectiv AI ⧉@inflectivAI

@victor207755822 This paper advances the path toward truly adaptive AI systems. Self-iteration and continual learning will become standard features in future models.

31d164

Feitong Yang@feitong_yang

@victor207755822 I love the spirit of experimenting here.

However, who is learning the content generated from this process? Other Human? AI? You as the collaborator? Do you recommend us reading this arxiv paper?

31d44

HammerChain@anaphor1c

@victor207755822 can this workflow do more than survey

31d22

Deli Chen@victor207755822

I read some key sections of the paper and I think there's room for improvement 👀📄 (if I wrote it myself, the quality would be higher but the speed would drop massively ⏳🚫). Since my ultimate goal is to iterate DeliAutoResearch SKILL 🧪🔁, I'm choosing to keep it rough and use it as a baseline for progress 📈🐣

31d20