The paper investigates whether representational entanglement—shared structure between knowledge domains—impedes unlearning in neural networks. Using Selective Gradient Masking, the authors train six 254M‑parameter language models with varying degrees of disentanglement between biology and non‑biology knowledge, then apply three standard unlearning methods to each. Results show that more disentangled models consistently achieve better retain‑forget trade‑offs, with up to four‑fold lower retain cost at the same forgetting level, providing direct evidence that entanglement contributes to collateral damage in unlearning.
By Ev\v{z}en Wybitul, Tim G. J. Rudner, Christian Schroeder de Witt
The paper investigates whether representational entanglement—shared structure between knowledge domains—impedes unlearning in neural networks. By training six 254M‑parameter language models with varying degrees of disentanglement between biology and non‑biology knowledge and applying three unlearning methods, the authors find that more disentangled models consistently achieve better retain‑forget trade‑offs, with up to four‑fold lower retain cost. This controlled experiment provides direct evidence that entanglement contributes to collateral damage during unlearning, supporting a long‑standing hypothesis in interpretability research.
arXiv:2607. 18114v1 Announce Type: cross Abstract: Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer.
By Prakhar Gupta, Terry Jingchen Zhang, Florent Draye, Bernhard Sch\"olkopf, Zhijing Jin
arXiv:2604. 07650v2 Announce Type: replace Abstract: The rapid growth of the large language model (LLM) ecosystem raises a critical question: are seemingly diverse models truly independent?
By Chenchen Kuai, Jiwan Jiang, Zihao Zhu, Hao Wang, Keshu Wu, Zihao Li, Yunlong Zhang, Chenxi Liu, Zhengzhong Tu, Zhiwen Fan, Yang Zhou
arXiv:2607. 08173v1 Announce Type: new Abstract: Black box auditing of language models is an essential pre-deployment tool, but it may miss subtle forms of misalignment and hidden information.
By Jack Hopkins, Dipika Khullar, Fabien Roger
Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model.
arXiv:2607. 16741v1 Announce Type: new Abstract: B\"urger et al.
By Francesco Karim Vicidomini
The paper investigates how machine unlearning for large language models (LLMs) can unintentionally erase related knowledge, even in distant domains. By analyzing the propagation of unlearning effects before any model updates, the authors discover a consistent decay pattern where collateral damage is strongest near the targeted forget set and diminishes with semantic distance but never fully disappears at domain boundaries. They propose a pre-unlearning prediction task—forget-set auditing—to identify potential collateral damage early, finding that interaction features between the forget set and evaluation set are the most predictive signals. This approach offers an early warning system for risky unlearning runs and guides the design of more reliable unlearning procedures.
By Bo Su, Ankit Shah, Thai Le
The paper introduces PACT, a method for unlearning deceptive behaviors in large language models by using contrastive forget sets that compare a model’s responses under deceptive and neutral contexts. PACT trains the model to produce pressure‑aware counterfactual targets, preserving benign system‑prompt adherence and reasoning traces while dramatically reducing deception rates from over 50% to under 3% on 32B reasoning models.
By Haoran Tang, Rajiv Khanna
arXiv:2607. 24797v1 Announce Type: cross Abstract: In the literate human brain, reading and writing are two doubly-dissociable systems: a ventral decoding route (impaired in pure alexia) and a fronto-parietal encoding route (impaired in pure agraphia), sharing a partial orthographic core.
By Diego Salda\~na Ulloa
arXiv:2609.22219v1 Announce Type: cross
Abstract: Evaluating an aligned language model by reading its answers assumes the answers carry the distinction the evaluator cares about. We introduce LLM end...
By Fabrizio Davide, Francesco Collova
The paper investigates whether language models can identify sentences from their training data by using exact duplication counts from publicly released corpora for two model families, OLMo‑2 and Pythia. It finds that for typical duplication levels, models show only a weak trace of exposure, with a rank correlation near –0.08, and that strong signals only appear when a sentence appears roughly a thousand times, at which point fame rather than memory dominates. The study also demonstrates that common membership tests can be misleading, as changing a single word does not alter the model’s preference, and that controlling for register can significantly improve detector performance.
By Arman Nik Khah