Modeling Rapid Contextual Learning in the Visual Cortex with Fast-Weight Deep Autoencoder Networks

Recent neurophysiological studies have revealed that the early visual cortex can rapidly learn global image context, as evidenced by a sparsification of population responses and a reduction in mean activity when exposed to familiar versus novel image contexts. This phenomenon has been attributed primarily to local recurrent interactions, rather than changes in feedforward or feedback pathways, supported by both empirical findings and circuit-level modeling. Recurrent neural circuits capable of simulating these effects have been shown to reshape the geometry of neural manifolds, enhancing robustness and invariance to irrelevant variations. In this study, we employ a Vision Transformer (ViT)-based autoencoder to investigate, from a functional perspective, how familiarity training can induce sensitivity to global context in the early layers of a deep neural network. We hypothesize that rapid learning operates via fast weights, which encode transient or short-term memory traces, and we explore the use of Low-Rank Adaptation (LoRA) to implement such fast weights within each Transformer layer. Our results show that (1) The proposed ViT-based autoencoder's self-attention circuit performs a manifold transform similar to a neural circuit model of the familiarity effect. (2) Familiarity training aligns latent representations in early layers with those in the top layer that contains global context information. (3) Familiarity training broadens the self-attention scope within the remembered image context. (4) These effects are significantly amplified by LoRA-based fast weights. Together, these findings suggest that familiarity training introduces global sensitivity to earlier layers in a hierarchical network, and that a hybrid fast-and-slow weight architecture may provide a viable computational model for studying rapid global context learning in the brain.

DISCOVERSE: Efficient Robot Simulation in Complex High-Fidelity Environments

We present the first unified, modular, open-source 3DGS-based simulation framework for Real2Sim2Real robot learning. It features a holistic Real2Sim pipeline that synthesizes hyper-realistic geometry and appearance of complex real-world scenarios, paving the way for analyzing and bridging the Sim2Real gap. Powered by Gaussian Splatting and MuJoCo, Discoverse enables massively parallel simulation of multiple sensor modalities and accurate physics, with inclusive supports for existing 3D assets, robot models, and ROS plugins, empowering large-scale robot learning and complex robotic benchmarks. Through extensive experiments on imitation learning, Discoverse demonstrates state-of-the-art zero-shot Sim2Real transfer performance compared to existing simulators. For code and demos: https://air-discoverse.github.io/.

SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting

3D Gaussian Splatting (3DGS) serves as a highly performant and efficient encoding of scene geometry, appearance, and semantics. Moreover, grounding language in 3D scenes has proven to be an effective strategy for 3D scene understanding. Current Language Gaussian Splatting line of work fall into three main groups: (i) per-scene optimization-based, (ii) per-scene optimization-free, and (iii) generalizable approach. However, most of them are evaluated only on rendered 2D views of a handful of scenes and viewpoints close to the training views, limiting ability and insight into holistic 3D understanding. To address this gap, we propose the first large-scale benchmark that systematically assesses these three groups of methods directly in 3D space, evaluating on 1060 scenes across three indoor datasets and one outdoor dataset. Benchmark results demonstrate a clear advantage of the generalizable paradigm, particularly in relaxing the scene-specific limitation, enabling fast feed-forward inference on novel scenes, and achieving superior segmentation performance. We further introduce GaussianWorld-49K a carefully curated 3DGS dataset comprising around 49K diverse indoor and outdoor scenes obtained from multiple sources, with which we demonstrate the generalizable approach could harness strong data priors. Our codes, benchmark, and datasets will be made public to accelerate research in generalizable 3DGS scene understanding.

Bridging the Gap: In-Context Learning for Modeling Human Disagreement

Large Language Models (LLMs) have shown strong performance on NLP classification tasks. However, they typically rely on aggregated labels-often via majority voting-which can obscure the human disagreement inherent in subjective annotations. This study examines whether LLMs can capture multiple perspectives and reflect annotator disagreement in subjective tasks such as hate speech and offensive language detection. We use in-context learning (ICL) in zero-shot and few-shot settings, evaluating four open-source LLMs across three label modeling strategies: aggregated hard labels, and disaggregated hard and soft labels. In few-shot prompting, we assess demonstration selection methods based on textual similarity (BM25, PLM-based), annotation disagreement (entropy), a combined ranking, and example ordering strategies (random vs. curriculum-based). Results show that multi-perspective generation is viable in zero-shot settings, while few-shot setups often fail to capture the full spectrum of human judgments. Prompt design and demonstration selection notably affect performance, though example ordering has limited impact. These findings highlight the challenges of modeling subjectivity with LLMs and the importance of building more perspective-aware, socially intelligent models.

SCRum-9: Multilingual Stance Classification over Rumours on Social Media

We introduce SCRum-9, a multilingual dataset for Rumour Stance Classification, containing 7,516 tweet-reply pairs from X. SCRum-9 goes beyond existing stance classification datasets by covering more languages (9), linking examples to more fact-checked claims (2.1k), and including complex annotations from multiple annotators to account for intra- and inter-annotator variability. Annotations were made by at least three native speakers per language, totalling around 405 hours of annotation and 8,150 dollars in compensation. Experiments on SCRum-9 show that it is a challenging benchmark for both state-of-the-art LLMs (e.g. Deepseek) as well as fine-tuned pre-trained models, motivating future work in this area.

Hierarchical Safety Realignment: Lightweight Restoration of Safety in Pruned Large Vision-Language Models

With the increasing size of Large Vision-Language Models (LVLMs), network pruning techniques aimed at compressing models for deployment in resource-constrained environments have garnered significant attention. However, we observe that pruning often leads to a degradation in safety performance. To address this issue, we present a novel and lightweight approach, termed Hierarchical Safety Realignment (HSR). HSR operates by first quantifying the contribution of each attention head to safety, identifying the most critical ones, and then selectively restoring neurons directly within these attention heads that play a pivotal role in maintaining safety. This process hierarchically realigns the safety of pruned LVLMs, progressing from the attention head level to the neuron level. We validate HSR across various models and pruning strategies, consistently achieving notable improvements in safety performance. To our knowledge, this is the first work explicitly focused on restoring safety in LVLMs post-pruning.

RAGSynth: Synthetic Data for Robust and Faithful RAG Component Optimization

RAG can enhance the performance of LLMs on knowledge-intensive tasks. Various RAG paradigms, including vanilla, planning-based, and iterative RAG, are built upon 2 cores: the retriever, which should robustly select relevant documents across complex queries, and the generator, which should faithfully synthesize responses. However, existing retrievers rely heavily on public knowledge and struggle with queries of varying logical complexity and clue completeness, while generators frequently face fidelity problems. In this work, we introduce RAGSynth, a framework that includes a data construction modeling and a corresponding synthetic data generation implementation, designed to optimize retriever robustness and generator fidelity. Additionally, we present SynthBench, a benchmark encompassing 8 domain-specific documents across 4 domains, featuring diverse query complexities, clue completeness, and fine-grained citation granularity. Leveraging RAGSynth, we generate a large-scale synthetic dataset, including single and multi-hop. Extensive experiments demonstrate that the synthetic data significantly improves the robustness of the retrievers and the fidelity of the generators. Additional evaluations confirm that RAGSynth can also generalize well across different domains. By integrating the optimized retrievers into various RAG paradigms, we consistently observe enhanced RAG system performance. We have open-sourced the implementation on https://github.com/EachSheep/RAGSynth.

Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

TriniMark: A Robust Generative Speech Watermarking Method for Trinity-Level Attribution

The emergence of diffusion models has facilitated the generation of speech with reinforced fidelity and naturalness. While deepfake detection technologies have manifested the ability to identify AI-generated content, their efficacy decreases as generative models become increasingly sophisticated. Furthermore, current research in the field has not adequately addressed the necessity for robust watermarking to safeguard the intellectual property rights associated with synthetic speech and generative models. To remedy this deficiency, we propose a \textbf{ro}bust generative \textbf{s}peech wat\textbf{e}rmarking method (TriniMark) for authenticating the generated content and safeguarding the copyrights by enabling the traceability of the diffusion model. We first design a structure-lightweight watermark encoder that embeds watermarks into the time-domain features of speech and reconstructs the waveform directly. A temporal-aware gated convolutional network is meticulously designed in the watermark decoder for bit-wise watermark recovery. Subsequently, the waveform-guided fine-tuning strategy is proposed for fine-tuning the diffusion model, which leverages the transferability of watermarks and enables the diffusion model to incorporate watermark knowledge effectively. When an attacker trains a surrogate model using the outputs of the target model, the embedded watermark can still be learned by the surrogate model and correctly extracted. Comparative experiments with state-of-the-art methods demonstrate the superior robustness of our method, particularly in countering compound attacks.

Nearly isotropic segmentation for medial temporal lobe subregions in multi-modality MRI

Morphometry of medial temporal lobe (MTL) subregions in brain MRI is sensitive biomarker to Alzheimers Disease and other related conditions. While T2-weighted (T2w) MRI with high in-plane resolution is widely used to segment hippocampal subfields due to its higher contrast in hippocampus, its lower out-of-plane resolution reduces the accuracy of subregion thickness measurements. To address this issue, we developed a nearly isotropic segmentation pipeline that incorporates image and label upsampling and high-resolution segmentation in T2w MRI. First, a high-resolution atlas was created based on an existing anisotropic atlas derived from 29 individuals. Both T1-weighted and T2w images in the atlas were upsampled from their original resolution to a nearly isotropic resolution 0.4x0.4x0.52mm3 using a non-local means approach. Manual segmentations within the atlas were also upsampled to match this resolution using a UNet-based neural network, which was trained on a cohort consisting of both high-resolution ex vivo and low-resolution anisotropic in vivo MRI with manual segmentations. Second, a multi-modality deep learning-based segmentation model was trained within this nearly isotropic atlas. Finally, experiments showed the nearly isotropic subregion segmentation improved the accuracy of cortical thickness as an imaging biomarker for neurodegeneration in T2w MRI.