Xin Liu

MTS @FireworksAI_HQ
| Prev. @Amazon @HKUST @MSFTResearch

[Xin Liu @SeanLiu0x60]

@tourzhao @ShiyangLi6 I’m grateful for your trust and the team support, and proud of what we’ve built from scratch since 2023. Everyone is an explorer.

[Xin Liu @SeanLiu0x60]

A great way to close out my three-year journey with Rufus!

I’ll never forget the days and nights spent brainstorming ideas, tackling challenges, and pushing through launches together. Over these years, I’ve learned so much from everyone on the team — from research to production, from technical skills to personal philosophies, and from senior leaders to interns.

Grateful for all the people I’ve worked with and everything we’ve built together.

[Tuo Zhao @tourzhao]

1/13. Rufus-Air (https://t.co/NgbCbDx12t) 🧵
Sep 2025–Apr 2026, minus a 2-month production pause: a small research team at Amazon Search built an open post-training recipe from scratch on GLM-4.5-Air-Base, using public data and open-source components, rarely over 400 H200s.

[Xin Liu @SeanLiu0x60]

@hongming110 2333

[Xin Liu @SeanLiu0x60]

Agentic judges are one of the most promising directions for scaling RL training. Pure exploration is ideal, but well-designed supervision can help steer models in the right direction.

@dzy97 is one of the strongest interns and is now on the job market. Don’t miss out!

[Zoey Sha Li @ZoeyLi20]

Scaling test-time compute for competitive programming usually means over-generate and rank: sample thousands of programs, cluster, submit the best.

What if a verifier could tell the solver exactly why a program fails?

Meet ARBITER 🧵

[Xin Liu @SeanLiu0x60]

When scaling hits the memory wall, borrowing RAM is an alternative way to go. Incredible work on the architecture ablation and performance analysis by @RJ_Sadhukhan. Kudos!

[Infini-AI-Lab @InfiniAILab]

🧠 How can we keep scaling LLMs with minimal changes in their serving energy?

DeepSeek’s Engram and our concurrent work STEM showed the promise of static sparse memory: large embedding tables with efficient access. But can mature pretrained models learn to use a newly added memory within a limited training budget, avoiding costly pretraining from scratch?

🚀 Introducing STAPLE: a recipe to extend pretrained models' abilities with static sparse memory. Serving energy per token stays close to dense and >50% below MoE upcycling in high-throughput regime.

Scalable capacity, interpretable memory, and compatibility with MoE upcycling - without any meaningful increase in serving cost. 🧵

Paper: https://t.co/qwVh1k9KAC
Code: https://t.co/7nn5b08Qpv
Website: https://t.co/m5CCP0gw9l

[Xin Liu @SeanLiu0x60]

@shuyanzhxyc @DukeU @dukecompsci Congrats!

[AAlanYJ @alanyyjj]

🛍️Large Language Models (LLMs) can be your AI shopping assistant that navigates online stores with ease. Amazon KDD Cup 2024 is here with 57 tasks & ~20K questions to level-up online shopping. https://t.co/qmOBZQT77j

[Jacky ChunKit Chan @Jacky_chunkit]

🚨Our new paper🚨has been accepted to Findings of #ACL2023: "DiscoPrompt: Path Prediction Prompt Tuning for Implicit Discourse Relation Recognition". We try to address how to inject labels’ hierarchical structure information into pre-trained language models via prompt tuning.

[Xin Liu @SeanLiu0x60]

We hope that our evaluation and analysis can help the research community!

[Jacky ChunKit Chan @Jacky_chunkit]

Thanks for @yqsong @SeanLiu0x60 @TFang229 @MightyWeaver2 and other co-authors to complete this work on #ChatGPT.