17 papers
EnactOpt: Diagnosis-Guided Harness Evolution for Embodied Agents
ZhiYuan Feng, Zhuoshi Pan, Zhaolu Kang +5
Embodied agents built on foundation models rely on an execution harness to organize observations, guide actions, and incorporate execution feedback into subsequent decisions. Failu…
Blind Men and the Elephant: Probing the Epistemic Myopia of LLMs under Long-Tail Divergent Knowledge
Zhuoshi Pan, Junru Lu, Yan Qian +3
Factual question answering (QA) typically assumes a single canonical answer, obscuring whether large language models (LLMs) retain divergent accounts of long-tail facts. To address…
ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL
Zhuoshi Pan, Qizhi Pei, Junru Lu +4
Long-horizon agentic tasks require large language models (LLMs) to iteratively retrieve, integrate, and maintain dispersed information across multi-turn interactions, but preservin…
CausalMix: Data Mixture as Causal Inference for Language Model Training
Zinan Tang, Yukun Zhang, Shaomian Zheng +7
In Large Language Model (LLM) training, data mixing plays a pivotal role in determining model performance. Recent methods optimize mixture weights via proxy models, but they rely o…
Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs
Yu Li, Xiaoran Shang, Qizhi Pei +11
Post-training data plays a pivotal role in shaping the capabilities of Large Language Models (LLMs), yet datasets are often treated as isolated artifacts, overlooking the systemic…
ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch
Zheng Liu, Honglin Lin, Chonghan Qin +13
Chart reasoning is a critical capability for Vision Language Models (VLMs). However, the development of open-source models is severely hindered by the lack of high-quality training…
OpenDataArena: A Fair and Open Arena for Benchmarking Post-Training Dataset Value
Mengzhang Cai, Xin Gao, Yu Li +13
The rapid evolution of Large Language Models (LLMs) is predicated on the quality and diversity of post-training datasets. However, a critical dichotomy persists: while models are r…
Scaling Code-Assisted Chain-of-Thoughts and Instructions for Model Reasoning
Honglin Lin, Qizhi Pei, Xin Gao +5
Reasoning capability is pivotal for Large Language Models (LLMs) to solve complex tasks, yet achieving reliable and scalable reasoning remains challenging. While Chain-of-Thought (…
ScaleDiff: Scaling Difficult Problems for Advanced Mathematical Reasoning
Qizhi Pei, Zhuoshi Pan, Honglin Lin +6
Large Reasoning Models (LRMs) have shown impressive capabilities in complex problem-solving, often benefiting from training on difficult mathematical problems that stimulate intric…
Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning
Zinan Tang, Xin Gao, Qizhi Pei +5
Supervised Fine-Tuning (SFT) Large Language Models (LLM) fundamentally rely on high-quality training data. While data selection and data synthesis are two common strategies to impr…
Can One Domain Help Others? A Data-Centric Study on Multi-Domain Reasoning via Reinforcement Learning
Yu Li, Zhuoshi Pan, Honglin Lin +3
Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for enhancing the reasoning capabilities of LLMs. Existing research has predominantly conce…
REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once
Zhuoshi Pan, Qizhi Pei, Yu Li +5
Recent Large Reasoning Models (LRMs) have achieved remarkable progress on task-specific benchmarks, yet their evaluation methods remain constrained by isolated problem-solving para…
IDEAL: Data Equilibrium Adaptation for Multi-Capability Language Model Alignment
Chenlin Ming, Chendi Qu, Mengzhang Cai +6
Large Language Models (LLMs) have achieved impressive performance through Supervised Fine-tuning (SFT) on diverse instructional datasets. When training on multiple capabilities sim…
MetaLadder: Ascending Mathematical Solution Quality via Analogical-Problem Reasoning Transfer
Honglin Lin, Zhuoshi Pan, Yu Li +5
Large Language Models (LLMs) have demonstrated promising capabilities in solving mathematical reasoning tasks, leveraging Chain-of-Thought (CoT) data as a vital component in guidin…
LEMMA: Learning from Errors for MatheMatical Advancement in LLMs
Zhuoshi Pan, Yu Li, Honglin Lin +7
Large language models (LLMs) have demonstrated remarkable reasoning capability in solving mathematical problems. However, existing approaches primarily focus on improving the quali…
MathFusion: Enhancing Mathematical Problem-solving of LLM through Instruction Fusion
Qizhi Pei, Lijun Wu, Zhuoshi Pan +6
Large Language Models (LLMs) have shown impressive progress in mathematical reasoning. While data augmentation is promising to enhance mathematical problem-solving ability, current…
TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection
Wei Wu, Zhuoshi Pan, Chao Wang +6
Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challeng…