The paper introduces C-Instrument, a constitution‑grid data instrument designed to automate reinforcement learning (RL) data generation and hillclimbing. It also presents C‑LIM, a per‑cell learnability score that guides actions such as prune, densify, amend, or expand, and identifies dead‑weight data regions before training. The authors report that their method improves over‑refusal rates from 22.4% to 12.8% and reduces under‑refusal on adversarial attacks from 0.27 to 0.33, while lifting learning impact in a previously ineffective data region from 0.733 to 0.80.
By Xianling Zhang
The paper investigates how different post‑training methods—supervised fine‑tuning, reasoning‑augmented fine‑tuning, and preference optimization (ORPO)—affect the internal computation of refusal behavior in language models. Experiments on Llama‑3.1‑8B, Gemma‑2‑9B, and Qwen3‑8B show that reasoning‑augmented training consistently creates a distinct refusal computation across models, while the architecture influences the internal structure and steerability of refusal. None of the studied methods simultaneously achieve a distributed refusal mechanism, preserve general capability, and allow easy corrective edits, indicating that current post‑training approaches are not a fully reliable defense for safety-critical applications.
By Hoang Cuong Nguyen, Mark Dras, Usman Naseem
The paper proposes probe-guided fine-tuning, a method that uses probes detecting undesired properties in model activations as a direct training signal. Experiments show that continuously updated probes reduce harmfulness and improve honesty while preserving utility, outperforming DPO and inference-time steering in safety-utility trade-offs and robustness to jailbreak and abliteration attacks. Importantly, the concepts remain linearly encoded after fine-tuning, maintaining monitorability.
By Lena Libon, Alexander Panfilov, Ben Rank, Xin Chen, Jonas Geiping, Maksym Andriushchenko
arXiv:2606. 02530v1 Announce Type: new Abstract: Aligning Large Language Models (LLMs) with human values often degrades their general capabilities, termed the alignment tax.
By Hao Li, Jingkun An, Zijun Song, Pengyu Zhu, Rui Li, Hao Wang, Wendi Feng, Yesheng Liu, Lijun Li, Jin-Ge Yao, Lei Sha
arXiv:2610.16814v1 Announce Type: cross
Abstract: Large language models (LLMs) are increasingly used as judges to assign scalar scores that drive downstream decision-making, including tasks such as p...
By Anning Hu, Yangyu Li, Yuqi Huang, Shengchao Hu, Junchi Yan
arXiv:2609.36657v1 Announce Type: cross
Abstract: Training models to act in accordance with an explicitly defined set of principles, or "constitution," has shown promise as a robust and transparent m...
By Adam S. Lowet, Mark Kurzeja
arXiv:2606. 00738v1 Announce Type: cross Abstract: Adversarial Training (AT) is a leading defense against adversarial examples but often suffers from Catastrophic Overfitting (CO) in efficient single-step variants, where robustness to multi-step attacks collapses despite high single-step performance.
By Mazdak Teymourian, Ramtin Moslemi, Farzan Rahmani, Mohammad Hossein Rohban
arXiv:2512. 20806v3 Announce Type: replace Abstract: Ensuring the safety of language models (LMs) while maintaining their usefulness remains a critical challenge in AI alignment.
By Anselm Paulus, Ilia Kulikov, Brandon Amos, R\'emi Munos, Ivan Evtimov, Kamalika Chaudhuri, Arman Zharmagambetov
arXiv:2603. 13026v2 Announce Type: replace Abstract: Prompt injection poses serious security risks to real-world LLM applications, particularly autonomous agents.
By Chenlong Yin, Runpeng Geng, Yanting Wang, Jinyuan Jia
arXiv:2602. 12124v2 Announce Type: replace Abstract: While most AI alignment research focuses on preventing models from generating explicitly harmful content, a more subtle risk arises from capability-seeking RL training in vulnerable environments.
By Yujun Zhou, Yue Huang, Han Bao, Kehan Guo, Zhenwen Liang, Pin-Yu Chen, Tian Gao, Werner Geyer, Nuno Moniz, Nitesh V Chawla, Xiangliang Zhang
The paper examines the vulnerability of open‑weight large language model safeguards that aim to block harmful usage by assuming new harmful behavior arises only through fine‑tuning. It demonstrates that simple, low‑cost attacks—abliteration and prefilling—can significantly increase attack success rates across multiple benchmarks, and proposes an abliteration‑resistant tuning (ART) method that mitigates these attacks. The study shows that safeguarding defenses must be evaluated against a broader range of attack strategies beyond fine‑tuning.
By Kevin Kuo, Virginia Smith, Chhavi Yadav
The paper introduces Jev, a reinforcement‑learning‑trained model that provides calibrated probability answers to typed questions about a single input in one call. Jev is evaluated on RLCDAlignBench, a benchmark covering ten alignment failures across 44 tests and five target models, achieving a median AUROC of 0.886 zero‑shot and outperforming supervised baselines on most tasks. The study shows that question wording has little impact, while contextual fields that encode labels are more influential, and that Jev matches human‑label agreement while being 63× cheaper than LLM‑judge scorers.
By Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li, Lida Zhao, Yutao Wu, Simin Chen, Ying Zhang, Leo Yu Zhang