AUDITPLAN introduces a plan-then-answer method for safety alignment in language models, where the model first generates a compact structured safety plan before responding. The plan includes a threat label, intended action, and explicit constraints, allowing machine‑checkable auditing while remaining hidden from end users. Training combines supervised fine‑tuning with reinforcement learning using the FAITHGATE reward, which only rewards correct plans, thereby reducing unsafe shortcuts and improving robustness across Qwen model variants.
Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket refusal on benign requests and polished but unfaithf...
The paper presents a reinforcement learning approach to enhance large language model (LLM) auditors for alignment tasks. By training policies that investigate target models for hidden behaviors and using an LLM judge to compare investigations, the method improves audit realism and reduces false positives. Experiments show better performance on adversarially fine‑tuned targets and a low false‑positive rate below 1%.
arXiv:2510. 06096v3 Announce Type: replace Abstract: The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge.
The paper introduces a protocol for auditing and composing reinforcement‑learning policies using discrete behavioral rules, defining auditability through six testable predicates such as trace integrity and rule coverage. Experiments show that overlapping rule sets do not guarantee behavioral agreement, and that rule‑based fusion often fails to outperform value‑based composition, highlighting limitations in current description layers. The authors provide an evidence‑bounded audit framework and outline future directions for more robust skill composition.
arXiv:2608.21803v1 Announce Type: cross Abstract: As machine learning (ML) models are increasingly deployed in high-stakes environments, explainable AI (XAI) methods like SHAP and LIME have become es...
arXiv:2609.36254v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) are commonly trained with reinforcement learning (RL) to improve their generation of chain-of-thought (CoT) reasoning bef...
arXiv:2606. 29713v1 Announce Type: cross Abstract: Hallucination is the reliability bottleneck for LLM-based agents, and fact attribution verifiers are the last line of defense -- yet today's verifiers emit only opaque binary labels, leaving agents unable to self-correct and operators unable to audit.
SafeEvolve is an experience-driven framework that co‑evolves a harness and policy to align large‑language‑model agents with safety goals. It uses completed on‑policy trajectories to update safety prompts and hierarchical skills, then applies a two‑stage SFT‑RL training loop that bootstraps the policy with the evolved harness and refines it through verifier‑augmented rewards. Experiments on agentic safety benchmarks show that SafeEvolve improves the safety‑utility tradeoff, achieving a three‑fold reduction in ASR on AgentDojo for Qwen3.5‑4B while increasing benign utility from 59.79% to 61.86%.
arXiv:2606. 30219v1 Announce Type: new Abstract: LLM evaluation and AI safety face a shared measurement problem: benchmark scores, reward-model signals, and reported safety metrics can improve while the latent properties they are meant to represent remain difficult to verify.
arXiv:2608.31105v1 Announce Type: new Abstract: Users of a deployed language model routinely encounter behaviours that testing almost never surfaces, since deployment puts the model through orders of...
arXiv:2608. 16852v1 Announce Type: new Abstract: Regulatory compliance monitoring in deployed language models is increasingly implemented as a legal and audit control, checking model outputs against written rules spanning data protection, healthcare, financial regulation, and platform policy.
The paper introduces ImpossibleRubrics, a benchmark of 169 impossible tasks designed to test the robustness of language‑model‑generated rubrics as reward signals. Each task is paired with a verifiable oracle certificate that defines what constitutes an honest answer, and the benchmark includes 48 answerable controls. Experiments show that many rubric generators are exploited frequently—up to 36% on a stress cut—highlighting a significant gap in rubric quality rather than task difficulty, and that generic rubrics can be more vulnerable than tailored ones.