Preference Orchestrator: Prompt-Aware Multi-Objective Alignment for Large Language Models

Safety-aligned large language models (LLMs) are increasingly used in sensitive applications where fairness is crucial. Evaluating their fairness is complex, often relying on standard question-answer schemes that may misinterpret refusal responses as indicators of fairness. This paper introduces the concept of silenced biases, which are unfair preferences hidden within the models' latent space, masked by safety-alignment. Previous methods have limitations, prompting the need for a new approach to assess these biases effectively.

The paper titled 'Fair In-Context Learning via Latent Concept Variables' explores the in-context learning (ICL) capabilities of large language models (LLMs) and their potential biases when applied to tabular data. It emphasizes an optimal demonstration selection method that leverages latent concept variables to enhance task adaptation while promoting fairness. The study introduces data augmentation strategies aimed at minimizing correlations between sensitive variables and predictive outcomes, ultimately striving for equitable predictions.

The paper titled 'Pre-Attention Expert Prediction and Prefetching for Mixture-of-Experts Large Language Models' introduces a method to enhance the efficiency of Mixture-of-Experts (MoE) Large Language Models (LLMs). The authors propose a pre-attention expert prediction technique that improves accuracy and reduces computational overhead by utilizing activations before the attention block. This approach aims to optimize expert prefetching, achieving about a 15% improvement in accuracy over existing methods.

The article discusses a new theoretical framework for training multi-agent systems using large language models (LLMs). It aims to connect system-level evaluations with agent-level learning by integrating cooperative game-theoretic attribution and process reward modeling. This approach produces local, signed, and credit-conserving signals, enhancing cooperation among agents while penalizing harmful actions in failure scenarios.

The paper titled 'Identifying and Analyzing Performance-Critical Tokens in Large Language Models' explores how large language models (LLMs) utilize in-context learning (ICL) for few-shot learning. It categorizes tokens in ICL prompts into content, stopword, and template tokens, aiming to identify those that significantly impact LLM performance. The study reveals that template and stopword tokens have a greater influence on performance than informative content tokens, challenging existing assumptions about human attention to informative words.

Recent advancements in large language models (LLMs) highlight their potential in automating scientific research ideation. Current methods often produce ideas that do not meet expert standards of novelty, feasibility, and effectiveness. To address these issues, a new framework is proposed that combines Supervised Fine-Tuning (SFT) and controllable Reinforcement Learning (RL) to enhance the quality of generated research ideas through a two-stage approach.

Large language models (LLMs) exhibit distinct political leanings, but their consistency in representing these orientations varies. This study introduces the concept of ideological depth, defined by a model's ability to follow political instructions reliably and the richness of its internal political representations, assessed using sparse autoencoders. The research compares Llama-3.1-8B-Instruct and Gemma-2-9B-IT, revealing that Gemma is significantly more steerable and activates approximately 7.3 times more distinct political features than Llama.

A recent study published on arXiv examines the phenomenon of negative bias in large language models (LLMs), which refers to their tendency to generate negative responses in binary decision tasks. The research highlights that previous studies have primarily focused on identifying negative attention heads that contribute to this bias. The authors introduce a new evaluation pipeline that categorizes responses based on the model's parametric knowledge, revealing that the format of prompts significantly influences the responses more than the semantics of the content itself.