The paper investigates how interventions on intermediate concepts in concept‑based models can sometimes worsen downstream predictions, a problem termed intervention harm. It shows that the previously proposed leakage‑mitigation objective, which optimises for full‑concept interventions, does not reliably prevent harm when only a subset of concepts is altered. The authors propose a new monotonic intervention objective that penalises increases in task loss during partial interventions, and demonstrate that combining this with the endpoint objective substantially reduces both the frequency and severity of partial‑intervention harm while maintaining overall model performance.
arXiv:2608. 20338v1 Announce Type: new Abstract: Large Language Models (LLMs) increasingly require selective removal of harmful or sensitive knowledge, called unlearning, yet existing methods and benchmarks fail to evaluate this capability completely.
arXiv:2606. 10669v1 Announce Type: cross Abstract: Concept-based models (CMs), deep neural networks that ground their predictions on representations aligned with human-understandable concepts (e.
arXiv:2505. 18672v2 Announce Type: replace Abstract: Representation intervention aims to localize and modify the representations that encode the underlying concepts in large language models (LLMs) to elicit the aligned and expected behaviors.
The paper introduces CREAMs, a flexible framework for Concept Bottleneck Models that allows explicit encoding of concept‑concept and concept‑task relationships such as mutual exclusivity, hierarchy, and correlations. CREAMs can optionally use a regularized side‑channel to compensate for incomplete concept sets, maintaining competitive performance while encouraging concept‑grounded predictions. The authors also propose a new metric to assess interpretability when predictions partially rely on the side‑channel, and demonstrate that CREAMs support efficient interventions, avoid concept leakage, and match black‑box performance even with missing concepts.
arXiv:2605. 28591v2 Announce Type: replace-cross Abstract: The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings.
GRACE is a new framework for concept erasure in text-to-image diffusion models that uses a semantically weighted sensitive subspace to guide localized interventions and lightweight subspace-constrained adapters to avoid global semantic disruption. It replaces manual counterfactual prompts with an automatically decoupled safe-anchor mechanism and controls intervention strength through an energy-driven dynamic gating system. Experiments show GRACE improves NSFW reduction by 17.86% over five state-of-the-art methods while also reducing target CLIP Score and FID, indicating stronger concept suppression with better preservation of generative quality.
arXiv:2607. 03932v1 Announce Type: cross Abstract: LLMs can be conveniently adapted to a diverse set of tasks, e.
The paper introduces a concept-guided poisoning unlearning framework that explicitly identifies semantic concepts most strongly linked to a target class or knowledge. For image classification, it uses a Post-hoc Concept Bottleneck Model to locate relevant image regions and creates poisoned samples; for large language models, it aggregates Integrated Gradients over multiple question responses to mask important content and generate poisoned training targets. Experiments across several datasets demonstrate effective unlearning while largely preserving model utility.
arXiv:2610.15365v1 Announce Type: new Abstract: Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk during genera...
arXiv:2606. 16535v1 Announce Type: new Abstract: Concept Bottleneck Models (CBMs) are a relevant tool for explainable Artificial Intelligence because they make their predictions through human-interpretable symbols.
arXiv:2510. 15770v4 Announce Type: replace-cross Abstract: Although deep neural networks achieve strong predictive performance, their internal reasoning often remains difficult to inspect and control.
arXiv:2601.21702v4 Announce Type: replace Abstract: We consider Representation Misdirection (RM), a class of large language model (LLM) unlearning methods that achieve forgetting by redirecting the l...