Analysing Moral Bias in Finetuned LLMs through Mechanistic Interpretability

A new conceptual reasoning layer named CORE has been proposed to enhance the performance of large language models (LLMs) in multi-turn interactions. CORE aims to address the limitations of existing models, which struggle to maintain user intent and task state across conversations, leading to inconsistencies and prompt drift. By utilizing a compact semantic state and cognitive operators, CORE reduces the need for extensive token history, resulting in a significant decrease in cumulative prompt tokens.

Large language models (LLMs) face a modern variant of the symbol grounding problem, questioning whether their outputs can represent extra-linguistic reality without human interpretation. The research emphasizes the necessity of referential grounding, which connects internal states to the world through causal relations and historical selection.

Large language models (LLMs) are increasingly utilized for factual inquiries, yet their internal representations of truth remain inadequately understood. A recent study introduces the concept of representational stability, assessing how robustly LLMs differentiate between true, false, and ambiguous statements through controlled experiments involving linear probes and model activations.

Large language models (LLMs) exhibit two mechanisms of value expression: intrinsic, based on learned values, and prompted, based on explicit prompts. This study analyzes these mechanisms at a mechanistic level, revealing both shared and unique components in their operation.

Large language models (LLMs) are increasingly being integrated into multi-agent systems (MAS), where peer interactions significantly influence decision-making. A recent study introduces KAIROS, a benchmark designed to simulate collaborative quiz-style interactions among peer agents, allowing for a detailed analysis of how rapport and peer behaviors affect LLMs' decision-making processes.

A recent study titled 'RL-MTJail' explores the vulnerabilities of large language models (LLMs) to jailbreak attacks, focusing on black-box multi-turn jailbreaks. The research proposes a reinforcement learning framework to optimize the harmfulness of outputs through a series of prompt-output interactions, addressing the limitations of existing single-turn optimization methods.

A new framework called LUNE has been introduced, enabling efficient unlearning in large language models (LLMs) through LoRA fine-tuning with negative examples. This method allows for targeted suppression of specific knowledge without the need for extensive computational resources, addressing challenges related to privacy and bias mitigation.

A recent study highlights the importance of incorporating multiple generations in the evaluation of large language models (LLMs) to enhance benchmark accuracy. The proposed hierarchical statistical model addresses the randomness inherent in LLMs, which traditional evaluation methods often overlook. This approach aims to provide a more reliable assessment of LLM capabilities by reducing variance in benchmark score estimates.