arXiv:2604. 12277v2 Announce Type: replace Abstract: Pretrained text encoders are prone to shortcut learning, relying on token-label correlations that fail once the distribution shifts in deployment.
arXiv:2505.19893v2 Announce Type: replace Abstract: Large language model pretraining is compute-intensive, yet many tokens contribute marginally to learning, resulting in inefficiency. We introduce E...
arXiv:2607. 28862v1 Announce Type: cross Abstract: The rapid development of Large Language Models (LLMs) has led to significant advances across a wide range of language tasks, while simultaneously raising growing concerns about unauthorized data exploitation and privacy leakage.
arXiv:2606. 04928v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly deployed across diverse applications, raising critical questions for governance, accountability, and data provenance.
Deep Contrastive Unlearning for Language Models (DeepCUT) is a framework that removes information from fine‑tuned language models by directly optimizing their latent space. It addresses the challenge of machine unlearning in black‑box models, which has been largely overlooked by previous work that only mitigated output effects. Experiments on real‑world datasets show that DeepCUT consistently outperforms baseline methods in both effectiveness and efficiency.
arXiv:2606. 05165v1 Announce Type: new Abstract: Training Data Attribution (TDA) seeks to trace a model's predictions back to its training data.
arXiv:2502.06806v5 Announce Type: replace Abstract: Many commercial Large Language Models (LLMs) are often closed-source, limiting developers to prompt tuning for aligning content generation with spe...
arXiv:2609.38795v1 Announce Type: new Abstract: Block drafters for speculative decoding are commonly trained on corpora written by external models, where a single off-policy token invalidates supervi...
arXiv:2606. 06286v1 Announce Type: cross Abstract: Large language models can reproduce training data, but existing memorization evaluations mostly measure whether models can be forced to do so, rather than whether they do so under ordinary use.
The paper introduces a training‑free approach to detect policy violations in large language models by treating the task as an out‑of‑distribution problem in the model’s activation space. It uses whitening‑inspired techniques to compute policy‑violation scores directly from normalized hidden activations, requiring only the policy text and a few illustrative examples. Experiments on several LLMs and policy benchmarks show the method achieves up to 86.0% F1, outperforming fine‑tuned and LLM‑as‑a‑judge baselines while being computationally lightweight.
The paper introduces MOSAIC, a large adversarial benchmark for detecting AI-generated text, and presents NeuroStat, a new framework that combines token‑level probabilistic logits with deep semantic hidden states from a single language model. NeuroStat fuses these signals via Macro‑State Residual Modulation and uses orthogonal and contrastive losses to learn complementary representations. Experiments show that NeuroStat outperforms existing methods on MOSAIC, achieving superior robustness against adversarial attacks.
arXiv:2609.00082v1 Announce Type: cross Abstract: LLMs acquire vast amounts of knowledge during pre-training, but often lack the specialized knowledge needed to answer questions from niche sources su...
DECSELFMASK is a decoder‑only classification method that uses unlabeled clinical text to improve performance. It creates self‑supervised training examples by masking portions of the text identified as relevant through relevance attribution, then trains the model to reconstruct the masked tokens via next‑token prediction. Experiments on 136 tasks from 1.9 M Italian hospital notes show consistent gains across five models, outperforming base models (+9.1 Macro F1), continual pretraining (+6.3), and synthetic label generation (+12.5).