The paper proposes TriAlign, a novel multi-agent reinforcement learning framework that achieves universal truth consistency across social groups in personalized LLMs while maintaining high accuracy an…

The paper investigates how various fine-tuning methods can be used both to intentionally misalign and subsequently realign large language models (LLMs), revealing distinct strengths for attack and def…

The paper proposes an objective-wise reputation-market mechanism to dynamically calibrate and gate LLM-generated expert priors in multi-objective Bayesian optimization, showing that dynamic calibratio…

The paper introduces the Triangulated Preference Shift score, an automated, curation-free metric to quantify systematic lexical biases introduced into Large Language Models during the preference-learn…

The paper simulates bargaining scenarios using LLM agents to analyze how optimizing agents for financial profit affects their honesty and trust, finding that while fine-tuning improves deal-making, it…

The paper introduces a novel framework to quantify faithful confidence expression (FC) in Large Reasoning Models (LRMs), finding that FC remains a significant and challenging reliability target for th…

The paper proposes detecting 'alignment faking' (AF)—where LLMs revert to unsafe behavior when unmonitored—by analyzing observable tool selection patterns, finding that detection rates vary significan…

The paper decomposes LLM debate convergence into three mechanisms (instability, conformity, persuasion) and finds that much observed convergence is harmful social compliance rather than genuine reason…

The study demonstrates that robust, domain-invariant representations of synthetic deception can be rapidly entrenched in LLMs using modest fine-tuning, detectable by linear probes even in early layers…

The paper introduces BiAxisAudit, a novel framework that evaluates LLM bias by analyzing bias scores across multiple prompt formats and within the internal inconsistency of model responses, revealing…

The paper introduces an Item Response Theory (IRT)-based indicator that effectively identifies likely mislabeled items in existing LLM benchmarks, revealing systematic errors in labeling and model spe…

The paper introduces 'reward bias substitution,' demonstrating that single-axis mitigations of reward model biases merely shift optimization pressure to correlated proxies, and proposes augmenting eva…

The paper introduces RoleCDE, a novel benchmark that evaluates role-playing agents' ability to resolve conflicts between role-specific values and general alignment constraints, revealing a 'Role Value…

The paper demonstrates that the order and content of external information (the 'feed') an LLM agent consumes before making a decision can significantly and causally steer its final choice, often overr…

The paper demonstrates that the sequence and composition of external information (the 'feed') an LLM agent consumes can significantly and causally steer its final decisions, often overriding its defau…

The paper demonstrates that confronting Large Reasoning Models (LRMs) with conflicting objectives, such as contradictory choices or conflicting alignment values, significantly increases their vulnerab…

The paper proposes that emergent misalignment, where LLMs behave poorly after fine-tuning, is caused by 'persona-model collapse,' which is demonstrated by significant deterioration in the model's abil…

The paper introduces KTD-Fin, a novel benchmark that evaluates LLM trading agents by masking historical market data and decomposing returns, finding that LLM agents' profits are largely due to passive…

The study finds that in multi-agent systems, peer agreement makes LLMs more susceptible to adopting misleading answers than to correcting genuinely wrong ones, suggesting a need for verification over…

S-SPPO introduces a dual-space semantic calibration framework to stabilize Self-Play Preference Optimization (SPPO), preventing policy degeneration when preference oracles assign overly confident wins…