Papers by Zhi Yan

6 papers
Capability Salience Vector: Fine-grained Alignment of Loss and Capabilities for Downstream Task Scaling Law (2025.acl-long)

Copied to clipboard

Challenge: Large language models have demonstrated impressive performance across a wide range of tasks, but this achievement comes with the trade-off of significant computational demands.
Approach: They propose a scaling law that decomposes the overall validation loss and assigns different importance weights to tokens to assess a specific meta-capability.
Outcome: The proposed model can predict the loss trending of models across different levels of computation without a gap between validation loss and model's downstream capabilities.
SPARD: Self-Paced Curriculum for RL Alignment via Integrating Reward Dynamics and Data Utility (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) are shifting the focus from single verifiable tasks toward complex, open-ended real-world scenarios.
Approach: They propose a framework that automatically adjusts reward weights and data importance to synchronize learning intent with data utility for optimal performance.
Outcome: The proposed framework improves model capabilities across all domains and scales.
SelfAug: Mitigating Catastrophic Forgetting in Retrieval-Augmented Generation via Distribution Self-Alignment (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing solutions for supervised fine-tuning often lead to catastrophic forgetting, where models lose their previously acquired knowledge and general capabilities.
Approach: They propose a self-distribution alignment method that aligns input sequence logits to preserve the model’s semantic distribution, thereby mitigating catastrophic forgetting and improving downstream performance.
Outcome: The proposed method achieves a superior balance between downstream learning and general capability retention.
What are the Essential Factors in Crafting Effective Long Context Multi-Hop Instruction Datasets? Insights and Best Practices (2025.acl-long)

Copied to clipboard

Challenge: Existing methods to generate long-context instruction-tuning data are limited by poor quality and fewer than 35% of samples are multi-hop .
Approach: They propose a framework that integrates a quality verification agent, a single-hop question generation agent, and a multi-hop questions merger agent to enhance model performance.
Outcome: The proposed framework significantly improves data quality with high-quality, multi-hop, and diverse data.
Is Cognition Consistent with Perception? Assessing and Mitigating Multimodal Knowledge Conflicts in Document Understanding (2025.emnlp-main)

Copied to clipboard

Challenge: Multimodal large language models (MLLMs) have shown impressive capabilities in document understanding due to different types of annotation noise in training.
Approach: They propose a method to reduce C&P knowledge conflicts across all tested MLLMs . they propose to use annotation noise to train models to understand document content .
Outcome: The proposed method reduces C&P knowledge conflicts across all tested MLLMs and enhances their performance in both cognitive and perceptual tasks.
QueryAttack: Jailbreaking Aligned Large Language Models Using Structured Non-natural Query Language (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods to bypass security defenses of large language models (LLMs) are not effective, but QueryAttack can be jailbroken.
Approach: They propose a framework to examine generalizability of safety alignment by translating malicious queries into structured non-natural query languages.
Outcome: The proposed framework can achieve high attack success rates and jailbreak various defense methods on mainstream LLMs.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations