Jianghao Lin

Coordinated evidence acquisition in multi-hop RAG can boost performance by nearly 6 F1 points, revealing the critical role of learned control in retrieval processes.

GraphBU achieves a remarkable 96.7% feasibility rate while preserving structural integrity, revolutionizing how MILP instances are generated for solver development.

Skill organization can dramatically enhance agent performance, with a 4.1% increase in successful outcomes when using Progressive Disclosure over traditional methods.

DiffCold shatters the seesaw dilemma in item recommendation, enabling accurate cold-start predictions without sacrificing the performance of warm items.

Nüshu-PitchVITS not only revives an endangered script but also sets a new benchmark in low-resource TTS systems by achieving superior speech synthesis quality through innovative pitch conditioning.

LatentSkill achieves a 21.4-point increase in task success while slashing prefill token usage by over 64%, revolutionizing how LLM agents utilize skills.

LLMs' final layers might be holding your recommendation system back: representations from middle layers actually perform *better*, and a modular compression approach can unlock significant gains.

LMM-based GUI agents stick out like a sore thumb in human-centric mobile environments, but simple techniques can make them blend in without sacrificing utility.

Function calling gets a serious upgrade: a new reward model and inference scaling technique boosts performance by focusing on the *process* of tool use, not just the outcome.

Current LLM evaluation benchmarks often conflate chatbots and true AI agents, leading to misaligned research efforts, but this survey provides a framework for targeted evaluation based on environmental complexity and agent capabilities.