7 papers
Ask, Condition or Abstain: Reinforcement Learning for Missing-Premise Reasoning
Yongqi Tong, Zhenyu Zhang, Zimi Liu +8
Answer-only reinforcement learning (RL) trains reasoning models to solve fully specified problems, but many realistic queries omit a premise needed for a unique answer. In this set…
STAGE: Controlled Objective Admission for Multi-Preference LLM Alignment
Yongqi Tong, Zhenyu Zhang, Ruirui Wang +6
Multi-preference alignment is often framed as scalarization: combine reward dimensions, then optimize. This leaves a temporal decision underspecified: when should each preference d…
ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction
Yongqi Tong, Tan Li Hui Faith, Choy Zhen Wen Marcus +5
Open-ended real-world interaction admits multiple valid behaviors: an agent may answer directly, ask for clarification, provide progress updates, or confirm before acting. This fle…
How to Train a Real-World Silicon Concierge? Internalizing Complex Business Workflow to Only OneModel
Chang Liu, Chaoyang Ning, Dayi Jiang +32
Traditional industrial agents rely on modular pipelines, including Router, Retriever, Planner, Executor, Responder, Reviewer, and other components. These systems often fracture int…
LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion
Guanghao Zhou, Panjia Qiu, Cen Chen +4
The safety mechanisms of large language models (LLMs) exhibit notable fragility, as even fine-tuning on datasets without harmful content may still undermine their safety capabiliti…
REWARD CONSISTENCY: Improving Multi-Objective Alignment from a Data-Centric Perspective
Zhihao Xu, Yongqi Tong, Xin Zhang +2
Multi-objective preference alignment in language models often encounters a challenging trade-off: optimizing for one human preference (e.g., helpfulness) frequently compromises oth…
CCJA: Context-Coherent Jailbreak Attack for Aligned Large Language Models
Guanghao Zhou, Panjia Qiu, Mingyuan Fan +4
Despite explicit alignment efforts for large language models (LLMs), they can still be exploited to trigger unintended behaviors, a phenomenon known as "jailbreaking." Current jail…