Huseyin A. Inan

Principal researcher at Microsoft

[Huseyin A. Inan @HuseyinAInan]

New blog post: observations on training instability in RL for long-horizon agentic LLMs, with a focus on retokenization drift.
Feedback and discussion welcome.
👉 https://t.co/uK8O5hdYby

[Yuetai Li @yuetai12575]

Free your hands 🙌 from endless agent environment engineering!

Tired of building envs to train agents? We ditched them entirely.

Turns out, LLMs can reason to generate realistic env feedback from just tool specs + reference examples.

So we trained agents without any real envs: synthesizing simulated SFT trajectories + RL on simulated env.

One prompt = one new environment.
No env code, no pain.

We burned through GPT-5 APIs to curate 90K+ SFT trajectories, now open-sourced here 👉 https://t.co/EEObGsMQ5v

Result? Strong perf on τ²-Bench 💪

Imagine a future where agent training scales across thousands of simulated envs — powered purely by LLM.

➡️ https://t.co/D4nmL5F6te 🔥🔥

[Huseyin A. Inan @HuseyinAInan]

There's much more to explore, but for all the details, check out our paper and code! 🚀📄💻

📜 Paper: https://t.co/DwhRpzzIvk
💾 Code: https://t.co/DIoNDOPvIg

[Huseyin A. Inan @HuseyinAInan]

Let me share our fair share of emergent behaviors! 🚀 Our Reinforcement Learning Self-Play (RLSP) framework enables the emergence of complex reasoning abilities in LLMs—without relying on stronger teacher models. Check out our latest paper! 👇📄

[Huseyin A. Inan @HuseyinAInan]

We can achieve this by the simplest exploration reward that encourages the model to take more intermediate steps before arriving at a solution. Our key innovation? Decoupling exploration and correctness signals during PPO training! 🔄⚖️

[Huseyin A. Inan @HuseyinAInan]

@dilarakkl @sabanciu Congrats!!

[Huseyin A. Inan @HuseyinAInan]

@niloofar_mire hahahaha super informative :D

[Chulin Xie @ChulinXie]

Some text data is private & cannot be shared... Can we generate synthetic replicas with privacy guarantees?🤔

Instead of DP-SGD finetuning, use Aug-PE with inference APIs! Compatible with strong LLMs (GPT-3.5, Mistral), where DP-SGD is infeasible.
🔗https://t.co/qJeSQ0XJES [1/n]

[Xiang Yue @xiangyue96]

📢📢Want to share your textual data outside your org or team but worry about privacy leakage? Check out our new preprint "Synthetic Text Generation with Differential Privacy: A Simple and Practical Recipe"(https://t.co/7jpDR2KoFr) 👇👇👇

[Huseyin A. Inan @HuseyinAInan]

And thanks to the robustness to post-processing
property of DP, we are free to build any model on the generated data or inspect/debug/investigate the samples with no additional privacy loss :)

[Huseyin A. Inan @HuseyinAInan]

As LLMs finetuned with DP show impressive performance on many NLP benchmarks, it is exciting to see that they can also be used to generate synthetic version of text datasets in a private manner: https://t.co/SXFdRWka0G

[Huseyin A. Inan @HuseyinAInan]

It includes language modeling examples with sample-level and user-level differential privacy. We'd like to actively continue adding more examples from different tasks (and also update it with more recent versions). We'd love to engage with community :)

[Huseyin A. Inan @HuseyinAInan]

Dear ppml community, we are happy to announce our repo https://t.co/FzslOwnRTd for training transformer models with differential privacy. It's based on two of my favorite libraries, integrating Opacus to the Hugging Face platform.

[Huseyin A. Inan @HuseyinAInan]

@abidlabs @Gradio Awesome news, congrats! 🙂

[Golnoosh Farnadi @NeurIPS2024 @gfarnadi]

I have openings for Ph.D. students who are interested to join my group at @Mila_Quebec to work on #fairness and #privacy in machine learning.
Info: https://t.co/VgUFucbVNN
Deadline: Dec 1
Application form: https://t.co/6OdLXAaGWu
I appreciate retweets to help spread the word.

[Huseyin A. Inan @HuseyinAInan]

@TheCurlyBread I can see some future opportunities in that case 🙂

[Huseyin A. Inan @HuseyinAInan]

Privacy in AI team at Microsoft Research is looking for interns for Summer 2022 to be working on privacy-preserving ML and/or Federated Learning projects. Help spread the word please 🙂 https://t.co/Q7U8mrG2DH

[Huseyin A. Inan @HuseyinAInan]

@OpenPosition_ML Huge thanks!

[Huseyin A. Inan @HuseyinAInan]

@limufar @MSFTResearch (and did fantastic work in both times) 🤩