Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
We propose LF-MultiDiffusion, a training-free panorama generation method that extends MultiDiffusion to support linear projections between target and reference image spaces. Our key idea is to reformu...
Activation steering suppresses undesired behaviors in language models by adding a steering vector to the hidden state during generation. Recent conditional methods such as CAST and DSAS improve the be...
Future 6G networks will rely on Large Language Model (LLM) agents to manage the Radio Access Network (RAN). However, current architectures assume inter-agent messages convey objective facts. A message...
In many settings, studying causal questions based on text data requires adjusting for confounding information within texts. Yet there is a tradeoff in constructing text representations for adjustment:...
Vision Transformers (ViTs) are increasingly used in split-inference systems, where edge devices transmit intermediate token representations to a remote cloud. In this setting, token reduction lowers c...
Safety benchmarks for large language models often assess the risk of a user query, although the outcome of question answering depends on whether the response violates a policy. This distinction is cri...
Recent advancements in low-light image enhancement have leveraged diffusion models for their strong ability to generate perceptually realistic, detailed images. Patch diffusion models further offer a...
Crystal generators can now propose periodic structures, but their control interfaces remain poorly matched to the mixed descriptors used in materials design. Text provides a compact way to combine com...
Production LLMs must handle inputs that attempt to override system instructions, bypass safety policies or elicit harmful responses. A common mitigation is a separate guardrail model. Existing reports...
Whether large language models (LLMs) can control their own internal representations matters for both machine metacognition and AI safety. A recent study applied neurofeedback to LLMs and claimed that...
The rapid advancement of vision-language models (VLMs) has accelerated progress in computational pathology; however, whole-slide image (WSI)-based pathology report generation remains limited by the sc...
Graph data across diverse domains can expose valuable relational information to unauthorized representation learning, creating a pressing need for protection against such misuse. Unlearnable examples...
Motus2 is a self‑evolving general world model designed for dexterous manipulation. It integrates a shared‑weight model that offers three control interfaces—a policy, a simulator, and an evaluator—forming a closed decision‑and‑learning loop for policy improvement. The system scales both model size and data, progressing from large‑scale monocular egocentric data to synchronized stereo data and robot‑domain adaptation, while also incorporating tactile feedback and a biomimetic platform with dual arms and hands.
arXiv:2608.30811v1 Announce Type: cross Abstract: Long-context compression is essential for reducing the cost and latency of large language model inference. However, existing methods can fragment imp...
The paper introduces CLAIMPROBE, a claim-level audit tool that breaks down deep-research reports into individual claims and evaluates them for hallucination, misattribution, citation hygiene, and necessary-fact recall against retrieved evidence. Using CLAIMPROBE, the authors show that even high-scoring deep-research pipelines can omit key evidence and misattribute claims. They also propose CLAIMWRITER, a hierarchical claim-based writer that extracts source facts, maps them to an outline, and drafts sections from a source-linked claim representation, which reduces hallucination by 2.6 to 4.5 times and improves necessary-fact recall by 1.2 to 1.7 times while preserving overall report quality and enabling efficient localized revisions.
The study examined whether providing full prompt-level context to a large multimodal model would improve speech transcription accuracy on a production oral‑history corpus. Using a preregistered within‑item paired ablation, the authors found that adding context did not produce a detectable change in side‑level word error rate (WER) for either gpt‑4o‑transcribe or gemini‑2.5‑flash. The results suggest that context alone may not be sufficient to enhance aggregate transcription accuracy, and that finer‑grained, sequence‑aligned metrics are needed to evaluate such mechanisms.
The study shows that hallucination detection in large language models is largely driven by a single mean‑shift component in hidden states. Across three 7B‑scale models and multiple datasets, removing this direction reduces detection to chance, while a simple L2‑regularized logistic regression achieves high AUROC (0.952) and outperforms more complex probe architectures. The authors introduce LayerMix, a multi‑layer aggregation method that matches oracle‑layer performance without requiring oracle access, demonstrating that apparent probe complexity stems from high‑dimensional covariance estimation rather than non‑linearity.
Team FME submitted a method for the MAMA-MIA Challenge that tackles primary tumor segmentation and pathological complete response (pCR) prediction using dynamic contrast‑enhanced breast MRI. For segmentation, they employed a five‑fold residual‑encoder nnU‑Net ensemble trained on the first post‑contrast minus pre‑contrast image, augmented with mirroring test‑time augmentation and largest‑connected‑component filtering, achieving a Dice score of 0.713 and a normalized Hausdorff distance of 0.099. For pCR prediction, they ensembled 25 pretrained 3D video classifiers on lesion‑centred crops from the pre‑contrast and first two post‑contrast volumes, reaching a balanced accuracy of 0.541 and an equalized‑odds disparity of 0.212, and ranked second in both tasks.
Background-Free Objectness Learning (B-FOR) is a dense, class‑agnostic detection framework that learns objectness without treating unlabeled regions as background. It predicts multi‑scale object‑center and scale fields, using spatially structured soft targets to supervise only reliable annotated areas and introduces displacement‑aware scale fields to model object extent. Experiments on PASCAL VOC, MS‑COCO, and Open Images show B‑FOR improves recall by over +10 AR points compared to prior class‑agnostic baselines, with ablation studies confirming the importance of localized supervision and displacement‑aware scaling.
QCell is a query‑based model designed to improve overlapping cell instance segmentation in microscopy images. It introduces an instance recombination module that decomposes and recombines query representations in latent space, allowing the model to reason about entire cell structures even when they overlap. Additionally, a contrastive query alignment objective is used to learn distinctive instance features and separate overlapping cell queries. The authors also present a new Organoid dataset benchmark and demonstrate that QCell surpasses state‑of‑the‑art methods, achieving +2.2 AP and +2.7 AJI on the ISBI2014 benchmark.