Ji-rong Wen

Attention-guided denoising can dramatically enhance reasoning performance in diffusion language models, outperforming traditional post-training methods.

FORT-Searcher achieves superior performance by synthesizing training tasks that actively resist shortcut exploitation, transforming how we train deep search agents.

Fine-tuning on DeNovoSWE catapults LLM performance in generating entire software repositories, achieving nearly an 8x improvement on a challenging benchmark.

LLM agents are shockingly vulnerable to multi-stage "trojan" attacks that inject malicious instructions into their workspace, achieving near-perfect success rates where standard prompt injection defenses fail.

Forget hand-crafting mobile benchmarks – PhoneWorld lets you automatically generate them from real-world GUI trajectories, leading to massive performance gains for phone-use agents.

An AI agent autonomously discovered four new superconductors, shrinking the discovery timeline from years to GPU hours.

Agent-World reveals that self-evolving environments can dramatically boost agent performance, outperforming established models by leveraging dynamic task synthesis.

Autonomous ML research agents achieve significantly better long-horizon performance by maintaining durable state through a shared workspace, suggesting that orchestration and memory are more critical than raw reasoning power.

LLMs can now navigate 100-turn multimodal search tasks without context explosion, thanks to a file-based visual representation that slashes token costs.

LLMs' training trajectories in RLVR are more predictable than you think: modeling the non-linear evolution of a rank-1 subspace lets you extrapolate parameters and cut compute by 37.5%.

A small, outcome-driven proxy model can outperform complex indexing methods for LLM memory retrieval, offering a more efficient and scalable solution for long-term tasks.