Yi-Ke Guo

Key contribution not extracted.

Physical robots can finally sustain coherent, multi-turn creative collaboration by decoupling evolving semantic intent from spatial grounding and stroke execution rather than relying on reactive single-shot rendering.

Legal agents frequently reach the correct legal outcome through entirely fabricated citations and broken tool-use trajectories, revealing a deceptive "Right-Answer-Wrong-Reason" failure mode invisible to standard benchmarks.

In-context memory plateaus fast during interactive 3D segmentation, but turning clinician clicks into meta-learned test-time gradient steps halves human correction effort while cutting volumetric failure rates by over 40%.

As human oversight wanes, LRMs could autonomously evolve, but this shift introduces significant risks like reward hacking and feedback drift.

Automated data curation and imbalance-aware training strategies significantly enhance LALMs' performance on culturally diverse folk music, yet deep musical understanding remains elusive.

Augmenting images from a model's own failures can lead to substantial performance boosts in multimodal tasks, outperforming conventional augmentation techniques.

HOMIE redefines video personalization by seamlessly integrating inter- and intra-subject inputs, achieving unprecedented fidelity and interaction accuracy.

Embedding reference tokens at semantic positions allows for unprecedented precision in multi-reference video editing, setting a new benchmark for instruction quality.

Language corrections in PhysClaw-0 not only enhance robot autonomy but also boost success rates by over 35% while slashing human oversight time.

Harnesses can evolve in real-time during evaluation, leading to significant performance gains without retraining the underlying model.

UniSAE enables seamless, granular editing of speech attributes, allowing for precise control over speaker, emotion, and content in a unified framework.

LUNA achieves realistic 3D human animation from 2D inputs without the limitations of traditional skinning methods, enabling unprecedented flexibility and expressivity.

Regularizing in the activation space with Sparse Autoencoders leads to superior continual learning performance in large language models, outperforming traditional weight-space methods.

Achieving high-fidelity audio generation with just four sampling steps, AudioX-Turbo dramatically cuts inference costs while enhancing performance across multimodal tasks.

LLMs can maintain long-context performance even with aggressive KV-cache eviction by learning to predict token importance and compressing evicted tokens into a latent memory.

Forget prompt engineering: MOSS lets autonomous agents rewrite their own source code to fix bugs and improve performance in production.

Forget solitary AI assistants; ClawNet envisions a future where your agent collaborates with *other people's* agents, securely and autonomously.