🏢 ByteDance Research

Boosting LLM collaboration, this research introduces a novel theoretical framework for multi-LLM debate, revealing key principles like the effect of similar models and interventions to enhance accurac…

ADVPO, a novel method, tackles reward overoptimization in RLHF via a lightweight uncertainty quantification approach, resulting in enhanced LLM performance and alignment with human values.

A novel Decomposition-based Policy Learning (DPPL) method optimally balances short-term and long-term rewards, even with interrelated objectives, by transforming the problem into intuitive subproblems…

SuperClass, a novel vision-language pre-training method, achieves superior performance on various downstream tasks by directly using tokenized raw text as supervised classification labels, eliminating…

Revolutionizing antibody design, ABDPO uses direct energy-based preference optimization and a pre-trained diffusion model to generate high-quality antibodies with low energy and strong binding affinit…

AGILE, a novel reinforcement learning framework, significantly enhances LLM agents’ performance on complex conversational tasks by integrating memory, tools, expert interactions, and reflection, outpe…

This paper introduces a computationally efficient method to approximate the optimal accuracy-fairness trade-off curve for various datasets, providing rigorous statistical guarantees and quantifying un…