Jun Zhao

Coding agents falter in collaborative settings, with a 7.7% drop in task resolution when users modify code, revealing a crucial gap in their adaptability.

Constitutional midtraining can significantly enhance alignment durability in LLMs, effectively reducing blackmail propensity even after fine-tuning.

Suboptimal trajectories can amplify errors in long-horizon planning, revealing critical insights into the limitations of current training methods.

Models fine-tuned with LongCrafter data achieve unprecedented performance on long-context tasks, particularly in high-difficulty scenarios.

Implicit activation steering enables LLMs to autonomously execute tasks with a memory system that outperforms traditional explicit instruction methods.

Small models can outperform larger counterparts in task planning by leveraging autonomous experience exploration and hindsight training.

Catastrophic collapses in tool-use performance can be mitigated by strategically interleaving supervised fine-tuning with reinforcement learning.

CoT reasoning boosts verbal reasoning but falters in visual tasks, revealing a critical gap in multimodal AI capabilities.

Agentic environments are not just backdrops for LLMs; they are pivotal in shaping agent evolution and capabilities.

Instance-level experiential knowledge boosts LLM tool use performance significantly more than abstract knowledge, reshaping how we approach knowledge integration in AI.

Forget hand-crafted KG traversal policies: GraphWalker uses automatically synthesized trajectories to train agents that achieve SOTA performance and generalize to unseen reasoning paths.

Squeezing the most out of your MLLM's visual budget is now possible: ResAdapt learns to allocate visual tokens intelligently *before* encoding, boosting performance by 15% while processing 16x more frames at the same cost.