8 papers
Asynchrony Is Nearly Free for Evolution Strategies on Long-Horizon LLM Agentic Tasks
William Hoy, Jingxuan Fan, Nurcin Celik +1
Synchronous policy updates can limit the efficiency of post-training for long-horizon LLM agents. When trajectories span multiple interaction turns, completion times can vary subst…
Matching Accuracy, Different Geometry: Evolution Strategies vs GRPO in LLM Post-Training
William Hoy, Binxu Wang, Xu Pan
Evolution Strategies (ES) have emerged as a scalable gradient-free alternative to reinforcement learning based LLM fine-tuning, but it remains unclear whether comparable task perfo…
Circuit Mechanisms for Spatial Relation Generation in Diffusion Transformers
Binxu Wang, Jingxuan Fan, Xu Pan
Diffusion Transformers (DiTs) have greatly advanced text-to-image generation, but models still struggle to generate the correct spatial relations between objects as specified in th…
Diffusion-Inspired Masked Fine-Tuning for Knowledge Injection in Autoregressive LLMs
Xu Pan, Ely Hahami, Jingxuan Fan +2
Large language models (LLMs) are often used in environments where facts evolve, yet factual knowledge updates via fine-tuning on unstructured text often suffer from 1) reliance on…
User-Assistant Bias in LLMs
Xu Pan, Jingxuan Fan, Zidi Xiong +3
Modern large language models (LLMs) are typically trained and deployed using structured role tags (e.g. system, user, assistant, tool) that explicitly mark the source of each piece…
Memorization and Knowledge Injection in Gated LLMs
Xu Pan, Ely Hahami, Zechen Zhang +1
Large Language Models (LLMs) currently struggle to sequentially add new memories and integrate new knowledge. These limitations contrast with the human ability to continuously lear…
CF-CAM: Cluster Filter Class Activation Mapping for Reliable Gradient-Based Interpretability
Hongjie He, Xu Pan, Yudong Yao
As deep learning continues to advance, the transparency of neural network decision-making remains a critical challenge, limiting trust and applicability in high-stakes domains. Cla…
TBGC: Task-level Backbone-Oriented Gradient Clip for Multi-Task Foundation Model Learning
Zelun Zhang, Xue Pan
The AllInOne training paradigm squeezes a wide range of tasks into a unified model in a multi-task learning manner. However, optimization in multi-task learning is more challenge t…