The paper proposes a reinforcement learning approach that rewards language models for producing high‑quality moral arguments, under the hypothesis that better arguing leads to better judging. Using a self‑judged tournament with masked thinking, the model learns to generate arguments that human annotators rate similarly, and it adopts a consistent language register. Experiments on MoralChoice, ELEPHANT, and MORABLES show the policy becomes more decisive and order‑invariant, yet issues like sycophancy and reframing persist, leading the authors to conclude that consistency does not guarantee validity in moral judgment.
arXiv:2607. 17946v1 Announce Type: cross Abstract: Large Language Models (LLMs) often struggle to navigate value conflicts when trained with the compressed scalar rewards of Reinforcement Learning from Human Feedback (RLHF).
The paper proposes a new method for evaluating AI accountability by analyzing the structural quality of a model’s defense for its decisions, using a four‑phase dialectical protocol based on Walton’s argumentation schemes and Govier’s criteria. Applied to nine large language models and 200 ambiguous moral-choice items, the study finds that models generally defend their reasoning well above the rubric minimum, though failures cluster on grounds and sufficiency and correlate with epistemic hedging. The protocol also reveals that models often present different argument schemes in justification than in reasoning, detects indefensible defenses, and highlights challenges in assessing retraction in AI alignment.
Large Language Models (LLMs) often struggle to navigate value conflicts when trained with the compressed scalar rewards of Reinforcement Learning from Human Feedback (RLHF). To address this challenge, we investigate how chain-of-thought (CoT) reasoning can help improve performance in this domain.
The paper shows that fine‑tuning a large language model (LLM) with a debate framework—where a generator and a critic compete and a weaker LLM judge adjudicates—reduces reward hacking compared to standard reinforcement learning from AI feedback (RLAIF). In experiments on mathematics tasks, the debate approach keeps the judge’s performance stable, achieving a 45% higher peak validation accuracy than the RLAIF baseline and mitigating the rapid exploitation of judge errors. Additional findings indicate that weakening the judge speeds hacking unless countered by extra debate rounds, that debate can override misalignment prompts, and that word‑limit constraints on critiques help balance the game and prevent judge hacking. whyItMatters":"The study demonstrates a practical method to curb reward hacking in RL‑based AI systems, addressing a key obstacle for safely scaling AI oversight."
arXiv:2606. 18327v1 Announce Type: cross Abstract: Language models (LMs) that faithfully describe their own behavior can more easily be audited, understood, and trusted by users.
arXiv:2609.08016v1 Announce Type: new Abstract: Multi-agent debate, in which several LLMs exchange arguments before answering, is widely assumed to improve answer quality by surfacing genuine disagre...
The paper introduces DeCert, a framework that evaluates the validity and verdict consistency of rationales generated by generative reward models (GenRMs) without needing reference explanations. DeCert translates these evaluations into preference supervision that considers both verdict correctness and justification quality, enabling the construction of rationale preference pairs that differentiate stronger from weaker justifications. Experiments across multiple model families and benchmarks show that DeCert outperforms strong generative judge baselines, improves rationale quality, and mitigates positional bias, raising position-consistent accuracy by up to 14.8 points.
The paper argues that AI alignment depends on a system’s ability to exhibit a coherent moral policy—stable, monotonic, decisive, and Pareto‑viable—rather than on any specific moral standard. The authors test nine large language models across varied moral scenarios and find that none maintain consistent verdicts, with surface‑form changes causing up to 99% shifts in outcomes. This indicates that current LLM agents lack the structural moral competence required for meaningful alignment.
arXiv:2511.11500v3 Announce Type: replace Abstract: Modern language models fail a fundamental requirement of trustworthy intelligence: knowing when not to answer. Despite achieving impressive accurac...
EvoRubric introduces a co‑evolutionary reinforcement learning framework that learns rubric‑driven evaluation criteria for open‑ended text generation. The system uses a shared policy that simultaneously generates responses and rubrics, while a frozen meta‑verifier and grader provide validity checks and scoring. Experiments on medical, writing, and science benchmarks show that EvoRubric outperforms static and other evolving‑rubric baselines, achieving higher average scores on five tasks.
arXiv:2510. 10813v2 Announce Type: replace Abstract: Large Language Models (LLMs) are increasingly applied to domains that require reasoning about other agents' behavior, such as negotiation, policy design, and market simulation.
arXiv:2606. 31748v1 Announce Type: new Abstract: Safety training on language models often induces over-refusal: improved safety on harmful prompts at the cost of increased refusal on harmless ones.