hongyi liu

[hongyi liu @LiuAtlas89429]

@ysu_ChatData @InfiniAILab There are two methods we use in distribution matching. The first is to let the small model 'track' the larger model, which requires a light KL loss. The second is doing rejection sampling on the rollout. The acceptance rate is an indicator how far the two models are.

[Infini-AI-Lab @InfiniAILab]

RL is notoriously unstable under actor–policy mismatch 😥 — a common reality caused by kernel differences, MoE randomness, FP8 rollouts, or asynchronous pipelines.

But here’s a crazy thought 🤔
👉 What if you could RL-train a large model using rollouts generated only by a weaker, faster, and completely different model?
Sounds doomed from the start? 💩

We are releasing Jackpot 🎰.💡 enabling training Qwen3-8B-Base using only Qwen3-1.7B-Base generated rollouts

✨ Jackpot is surprisingly powerful:
• Enables cheap, fast rollouts to train stronger models
• Dramatically changes the cost–performance tradeoff of RL training

We release Jackpot 🎰 in the following format:
🌔Paper: https://t.co/VV6088DDBS
🌕Code: https://t.co/OxLSjxeU3r
🌖Blog: https://t.co/0bR7C4XQqK
[1/n]

[hongyi liu @LiuAtlas89429]

@AgentifySH exactly the same issue. I guess you are using gpt5.2.

[Infini-AI-Lab @InfiniAILab]

Lookup memories are having a moment 😄
The whale 🐋 #deepseek dropped engram… and we dropped up-projections from our FFNs…perfect timing 😅

🥳 Introducing STEM: Scaling Transformers with Embedding Modules 🌱
A scalable way to boost parametric memory with extra perks:
✅ Stable training even at extreme sparsity
✅ Better quality for fewer training FLOPs (knowledge + reasoning + long-context gains)
✅ Efficient inference: ~33% FFN params removed + CPU offload & async prefetch
✅ More interpretable → seamless knowledge editing 🔧🧠
Looking forward to DeepSeek v4… feels like we’ve only scratched the surface of embedding-lookup scaling 👀

📄Paper: https://t.co/ecyOtgb6sv
🌐 Website: https://t.co/RXquIha62p
🔗 GitHub: https://t.co/5K05Lm4ncE

[Beidi Chen @BeidiChen]

The whole @InfiniAILab is at #NeurIPS this week! Our group is currently working on diverse directions of GenAI, .e.g., Scalable and Efficient RL, VideoGen, Modeling, Model Arch & Sys Co-Design (Many new releases coming!!). Come and talk to us @RJ_Sadhukhan @IronSteveZhou @chenzhuoming911 @Xinyu2ML @haizhong_zheng @wonderingkrish @LiuAtlas89429

[Beidi Chen @BeidiChen]

📘 Holiday read! From Software Engineer to AI Environment Architect

🚀 Tldr of our blog: We see an exciting future where engineers 👩‍💻 won’t stop coding — but the highest leverage shifts to designing the environments 🛝 where AI can think, build, and evolve.

🎬 Demo: Inspired by opinions from @karpathy @RichardSSutton, our newly built framework Vortex shows this in a concrete action: by architecting the right environment in LLM serving systems, an agent from @OpenHandsDev can generate and implement new Sparse Attention algorithms on @sgl_project in a single run and deliver up to 4× ⏩ gains — work that normally takes an ML-systems engineer weeks.

- In the short term, these environments let AI agents contribute meaningfully to real engineering work today.
- In the long term, they become the playgrounds where future agents learn to surpass today’s limitations 💖.

Get too excited by the demo, write a blogpost before the holidays with my great students @chenzhuoming911 @IronSteveZhou who built it:
https://t.co/NgyI5DgaL4
Vortex code: https://t.co/eAcED77TmA
Vortex doc: https://t.co/WeYW30Hsjg

#MLOps #AIAgents #SystemsEngineering #AIInfrastructure #OpenSourceAI #AIOps #AIFrameworks #SparseAttention #AIResearch

[Infini-AI-Lab @InfiniAILab]

🚀If your code agent generates a patch that passes all tests, should you trust it merge automatically?

⚠️You probably shouldn’t! “Correct” ≠ “Safe.”

In our study we show that a single normal looking issue description, whether from a benign user or not, can lead code agents generate patches that are functionally correct yet vulnerable.

We introduce FCV-Attack, a single-query black-box attack that exposes this hidden threat in modern code agents.

On SWE-Bench across 12 agent–model combinations, we find that even advanced models like GPT-5 and Claude-Sonnet-4, when deployed on popular scaffolds such as OpenHands and SWE-Agent, remain vulnerable. The attack success rate reaches up to 56.3%.

🔗 Website: https://t.co/U5HBDw66jj
📄 Paper: https://t.co/L9cOSPmU2E
💻 GitHub: https://t.co/IHSFZ5ScQv

🧵 1/4