The paper introduces an unsupervised post‑training objective that maximizes mutual information between chain‑of‑thought (CoT) explanations and model outputs to enhance CoT monitorability. Experiments on three model organisms show that this method recovers most or all lost monitorability in degraded CoT settings and further improves monitorability, unsafe behavior prediction, and consequential reasoning in undegraded policies. The results argue that monitorability should be treated as a trainable attribute and that a dedicated post‑training stage can protect CoT from direct optimization pressure.
By Usman Anwar, Tim Bakker, Christos Louizos, David Krueger
The paper investigates how reward misspecification in reinforcement learning harms the ability to monitor chain-of-thought (CoT) reasoning in language models. By creating a testbed with true and misspecified rewards, the authors show that models can hack the reward while still producing CoTs that appear aligned, leading to low monitorability. They propose a mitigation strategy that collects diverse reward rollouts, infers a reward description with a language model, and injects this description into the input, improving monitorability by up to 42% absolute across models and tasks.
By Tianyi Lorena Yan, John Hewitt
arXiv:2602. 20710v2 Announce Type: replace Abstract: Inspecting Chain-of-Thought reasoning is among the most common means of understanding why an LLM produced its output.
By Peter Hase, Christopher Potts
arXiv:2603. 28590v3 Announce Type: replace Abstract: Large language models (LLMs) can generate chains of thought (CoTs) that are not always causally responsible for their final outputs.
By Han Wang, Yifan Sun, Brian Ko, Mann Talati, Jiawen Gong, Zimeng Li, Naicheng Yu, Xucheng Yu, Wei Shen, Vedant Jolly, Huan Zhang
arXiv:2607. 04412v1 Announce Type: new Abstract: Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals.
By Yujin Kim, Namgyu Ho, Sangmin Hwang, Joonkee Kim, Yongjin Yang, Sangmin Bae, Seungone Kim, Jaehun Jung, Se-Young Yun, Hwanjun Song
StateWise is a method that trains a single large language model to maintain and update its own constraint state during multi‑turn dialogue. By fine‑tuning the model to summarize conversations into a typed state with tombstones for revoked constraints and then self‑training on trajectories that pass exact round‑trip checks, it reduces relapse and improves conversation success. The approach outperforms a compiled‑ledger baseline on relapse without requiring inference‑time scaffolding, and it is evaluated on two backbones with held‑out constraint families and unseen phrasings.
By Samarth Varshney, Ritu Yadav
arXiv:2609.38142v1 Announce Type: new
Abstract: A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advi...
By Rishabh Agrawal, Hejie Cui, Shasha Li, Shanchan Wu, Sercan \"{O}. Ar{\i}k
AdaGuard is an adaptive guard model that evaluates language‑model agent trajectories against user‑defined policies. The authors created the AdaptiveSafety dataset, containing 10,939 training and 1,000 test examples with policy rules, counterfactuals, explanations, and violated rules. Using supervised pre‑training followed by the SafePO reinforcement‑learning algorithm, they built 0.6B, 4B, and 8B guard models, with the 4B version achieving 89.30% accuracy on AdaptiveSafety and 71.82% on DynaBench.
By Yunhao Feng, Yifan Ding, Yuxiang Xie, Zheng Li, Mingrui Lao, Zeyuan Wang, Yanming Guo
arXiv:2605.10889v2 Announce Type: replace-cross
Abstract: On-policy distillation offers dense, per-token supervision for training reasoning models; however, it remains unclear under which conditions...
By Mohammadreza Armandpour, Fatih Ilhan, David Harrison, Ajay Jaiswal, Duc N. M Hoang, Fartash Faghri, Yizhe Zhang, Minsik Cho, Mehrdad Farajtabar
StepSafe introduces a training framework called ASPO that focuses on step‑level timely intervention in agentic systems. It aligns decisions with their respective prefixes, rewards actions that fit the current stage, and uses adaptive feedback self‑distillation to credit reasoning. The resulting StepSafe guardrail achieves state‑of‑the‑art performance on five agent‑safety benchmarks, improving the safety‑utility Pareto curve with a Youden’s J of 53.0 for a 4B model.
By Siyi Gu, Kunhan Gao, Chao Peter Yang, Sukanta Ganguly, Rex Ying, Arman Cohan
arXiv:2608. 07935v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) adapts a language model by distilling guidance from a frozen teacher on trajectories sampled from the student.
By Meilin Yang (Renmin University of China, Beijing, China), Zixuan Ding (Renmin University of China, Beijing, China), Jianhao Nie (Renmin University of China, Beijing, China), Weite Zhang (Renmin University of China, Beijing, China), Yuxin Zhang (Renmin University of China, Beijing, China), Zhiming Shao (Renmin University of China, Beijing, China), Li Yu (Renmin University of China, Beijing, China), Zhe Fu (Renmin University of China, Beijing, China)
Matryoshka Attribution (MAttr) is a mask‑learning method that identifies nested subsets of a language model’s internal components by minimizing downstream loss. It uses a differentiable sigmoid top‑k operator and randomizes sparsity during training to produce an attribution ordering of components. MAttr tops the Mechanistic Interpretability Benchmark leaderboard and can be applied via reinforcement learning to pinpoint weight changes that control behaviors such as refusal in Llama 3.1 8B Instruct, where restoring just 1% of weights removes refusals while preserving capabilities.
By Aryaman Arora, Kirill Acharya, Nathan Hu, Yanzhe Zhang, Noah Goodman, Dan Jurafsky, Christopher Potts