Leqi Zheng

SciLENS outperforms existing models in scientific literature synthesis by leveraging structural visualization to enhance reasoning and context retention.

Gradient-Aligned Reward transforms how we leverage expert knowledge in reinforcement learning, enabling LLMs to achieve superior reasoning performance with minimal overhead.

Soft reweighting in FiRe-OPD not only preserves critical information but also boosts optimization stability, leading to unprecedented performance improvements in on-policy distillation.

VLMs still struggle to generate code for complex, multi-panel charts from real-world data, despite excelling on simpler benchmarks.