Le Chen

PhD student at MPI-IS and ETH Zurich

[Weiyang Liu @Besteuler]

🚀 Excited to introduce PEFT-Arena! This project grew directly out of the challenges we faced when evaluating existing PEFT methods.

We argue that PEFT evaluation should not ask only what improves. It should also ask what is forgotten. For too long, parameter-efficient finetuning methods have been evaluated by downstream accuracy. But target performance alone is not enough. A method can adapt well while silently losing pretrained capabilities.

PEFT-Arena evaluates existing popular PEFT methods through the stability-plasticity trade-off, i.e., how much the model learns on the target task, and how much general ability it still preserves after finetuning.

A good PEFT method should adapt without forgetting and achieve a strong stability-plasticity trade-off. In our evaluation, we find that Orthogonal Finetuning achieves the strongest trade-off among all methods.

Beyond task-driven benchmark scores, PEFT-Arena also provides geometry-based internal diagnostics in weight space, activation space, and interpolation paths, helping us understand not only which methods work, but why they forget.

What PEFT-Arena has done:
🌟 Adaptation + preservation: evaluates both target gains and forgetting.
🌟 Trade-off frontier: reveals stability-plasticity patterns, with OFT often leading.
🌟 Geometry diagnosis: explains forgetting via weight spectra, activation distortion, and interpolation paths.

Welcome to try PEFT-Arena and evaluate your own PEFT method!

🌐 Project page: https://t.co/y8TsHWtHSZ
📝 Paper: https://t.co/AD4SguO4sj
💻 Code: https://t.co/JdtQkiF9IP

[NBA @NBA]

THE SAN ANTONIO SPURS ARE HEADED TO THEIR FIRST NBA FINALS SINCE 2014 🔥

The Spurs win Game 7 on the road to defeat OKC 4-3 in the Western Conference Finals, and advance to the NBA Finals for the 1st time in 12 years.

They will take on the Knicks in a rematch of the 1999 NBA Finals. Game 1 is in San Antonio on Wednesday, June 3 at 8:30pm/et on ABC!

[NBA @NBA]

WEMBY: 2025-26 WEST FINALS MVP!

🏆 27.3 PPG
🏆 10.9 RPG
🏆 2.7 BPG
🏆 48.1 FG%
🏆 40.0 3P%
🏆 Earvin "Magic" Johnson Trophy winner

The Spurs will seek their first NBA title since 2014 starting Wednesday, June 3 at 8:30pm/et on ABC!

[Le Chen @clthegoat]

Happy to be part of it :)

[Quankai Gao @UUUUUsher]

Great to see our work LOME cited in NVIDIA's Cosmos 3 report as a specialized baseline 🎉 LOME generates realistic hand–object interaction video from an image, text, and per-frame body & hand actions. Glad the direction is gaining momentum in open physical AI!

[Le Chen @clthegoat]

Orbit enables RL post-training for trillion-parameter LLMs on a single GPU node, with an extremely small train-rollout gap. A big step toward making post-training for frontier models accessible!

[Weiyang Liu @Besteuler]

🚀 Meet Orbit: OFT-based RL infrastructure for stable, efficient post-training of trillion-parameter LLMs.

Orbit can train 1T+ LLMs (e.g. Kimi-2.6, DeepSeek-V4-Pro) on a single GPU node (8xB200) with extremely small train-rollout gap!

Code: https://t.co/pyyOg6s7RQ
Blog: https://t.co/Rc7S1zQUel
Blog in Chinese: https://t.co/rvToBFG4Iq

[Le Chen @clthegoat]

Great work!

[Yunke Ao @YunkeAo]

PPO is an incredible algorithm, but its heuristic loss function leaves a gap between RL theory and practice.

We introduce the Bounded Ratio RL (BRRL) framework, which reveals that PPO approximately approaches an analytical solution of a constrained policy optimization problem.

With little refinement to PPO, we propose BPO: an algorithm that directly minimizes a divergence from this analytical solution.
The result? Better theoretical foundations + on-par or superior performance in MuJoCo, NVIDIA IsaacLab, and TTRL.

🌐Project website: https://t.co/jxlR3VoxgX
📄Paper: https://t.co/ygrTLLLQ6h
📦Code: https://t.co/zUL8vIv6Eb

[Quankai Gao @UUUUUsher]

Introducing LOME: Learning Human-Object Manipulation with Action-Conditioned Egocentric World Model

📖 Paper: https://t.co/83xokdkJbl
🌐 Project: https://t.co/aAlje6lOMU

[Weiyang Liu @Besteuler]

🚀 Meet OFTv2 — Orthogonal Finetuning made scalable, finally.

⚡️ 10× faster
💾 3× less GPU memory
🤖 Quantized OFT: plug-and-play on quantized LLMs, better than QLoRA

Try it now on Hugging face PEFT: https://t.co/IHIEbFPuCw
Website: https://t.co/4QRHWBAxlr

#AI #LLM

🧵1/6

[Le Chen @clthegoat]

@lgraesser3 Thanks a lot!

[Le Chen @clthegoat]

Learning to play the piano with two robot hands is super challenging, even in simulation! It requires coping with bimanual coordination at high speed to achieve human-level dexterity. We introduce RP1M, a large-scale robot piano-playing motion dataset, featuring ~1M trajectories over 2k music pieces.

Website: https://t.co/Bvk90MV6KH
Paper: https://t.co/45j6y1NncI

[Le Chen @clthegoat]

@kevin_zakka Thank you so much, Kevin!

[Le Chen @clthegoat]

@BlockyBard 🔥

[Le Chen @clthegoat]

@YagaoDirac For each RL agent, we use DroQ to train an MLP.

[Le Chen @clthegoat]

@AuroraDevEco Thanks a lot!

[Le Chen @clthegoat]

@yaru_niu yes finally🤣

[Le Chen @clthegoat]

RP1M contains various piano-playing motions, including highly dynamic motions and long-distance movements. Here is an example of the song Flight of the Bumblebee:

[Le Chen @clthegoat]

🌟Our OT-based method:
1️⃣ Low-cost: No need for human demonstrations or annotations.
2️⃣ Cross embodiments: Supports diverse hand morphologies and robot platforms.
3️⃣ Superior performance: Allows robots to discover optimal fingering aligned with their unique morphology.

[Jan Schneider @JanS1854]

Gradient subspace optimization unlocked for RL 🔒➡️🔓
Used only for supervised learning so far, our #ICLR2024 paper illustrates that policy gradients evolve in a small, slowly-changing subspace, opening up many opportunities for more efficient RL.

https://t.co/jVwdOL9cOw