Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts

Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.

Wideband Compressed-Domain Cramér--Rao Bounds for Near-Field XL-MIMO: Data and Geometric Diversity Decomposition

Wideband orthogonal frequency-division multiplexing (OFDM) over extremely large-scale MIMO (XL-MIMO) arrays in the near-field Fresnel regime suffers from a coupled beam-squint and wavefront-curvature effect that renders single-frequency covariance models severely biased: the per-subcarrier compressed covariance diverges from the center-frequency model by 64\% at $B = 100$~MHz and by 177\% at $B = 400$~MHz. We derive the wideband compressed-domain Cramér--Rao bound (CRB) for hybrid analog--digital architectures and decompose the Fisher information gain into a dominant data-diversity term that scales as $10\log_{10}K_s$~dB and a secondary geometric-diversity term arising from frequency-dependent curvature. At 28~GHz with $M = 256$ antennas, $N_\mathrm{RF} = 16$ RF chains, and $K_s = 512$ subcarriers, wideband processing yields $+27.8$~dB of CRB improvement at $B = 400$~MHz, of which $+0.7$~dB is attributable to geometric diversity.

FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On

Given a person and a garment image, virtual try-on (VTO) aims to synthesize a realistic image of the person wearing the garment, while preserving their original pose and identity. Although recent VTO methods excel at visualizing garment appearance, they largely overlook a crucial aspect of the try-on experience: the accuracy of garment fit -- for example, depicting how an extra-large shirt looks on an extra-small person. A key obstacle is the absence of datasets that provide precise garment and body size information, particularly for "ill-fit" cases, where garments are significantly too large or too small. Consequently, current VTO methods default to generating well-fitted results regardless of the garment or person size. In this paper, we take the first steps towards solving this open problem. We introduce FIT (Fit-Inclusive Try-on), a large-scale VTO dataset comprising over 1.13M try-on image triplets accompanied by precise body and garment measurements. We overcome the challenges of data collection via a scalable synthetic strategy: (1) We programmatically generate 3D garments using GarmentCode and drape them via physics simulation to capture realistic garment fit. (2) We employ a novel re-texturing framework to transform synthetic renderings into photorealistic images while strictly preserving geometry. (3) We introduce person identity preservation into our re-texturing model to generate paired person images (same person, different garments) for supervised training. Finally, we leverage our FIT dataset to train a baseline fit-aware virtual try-on model. Our data and results set the new state-of-the-art for fit-aware virtual try-on, as well as offer a robust benchmark for future research. We will make all data and code publicly available on our project page: https://johannakarras.github.io/FIT.

ClawBench: Can AI Agents Complete Everyday Online Tasks?

AI agents may be able to automate your inbox, but can they automate other routine aspects of your life? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI agents. To this end, we introduce ClawBench, an evaluation framework of 153 simple tasks that people need to accomplish regularly in their lives and work, spanning 144 live platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require demanding capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and challenges of real-world web interaction. A lightweight interception layer captures and blocks only the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 7 frontier models show that both proprietary and open-source models can complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%. Progress on ClawBench brings us closer to AI agents that can function as reliable general-purpose assistants.

Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts

Large language models (LLMs) can struggle to memorize factual knowledge in their parameters, often leading to hallucinations and poor performance on knowledge-intensive tasks. In this paper, we formalize fact memorization from an information-theoretic perspective and study how training data distributions affect fact accuracy. We show that fact accuracy is suboptimal (below the capacity limit) whenever the amount of information contained in the training data facts exceeds model capacity. This is further exacerbated when the fact frequency distribution is skewed (e.g. a power law). We propose data selection schemes based on the training loss alone that aim to limit the number of facts in the training data and flatten their frequency distribution. On semi-synthetic datasets containing high-entropy facts, our selection method effectively boosts fact accuracy to the capacity limit. When pretraining language models from scratch on an annotated Wikipedia corpus, our selection method enables a GPT2-Small model (110m parameters) to memorize 1.3X more entity facts compared to standard training, matching the performance of a 10X larger model (1.3B parameters) pretrained on the full dataset.

Selective Attention System (SAS): Device-Addressed Speech Detection for Real-Time On-Device Voice AI

We study device-addressed speech detection under pre-ASR edge deployment constraints, where systems must decide whether to forward audio before transcription under strict latency and compute limits. We show that, in multi-speaker environments with temporally ambiguous utterances, this task is more effectively modelled as a sequential routing problem over interaction history than as an utterance-local classification task. We formalize this as Sequential Device-Addressed Routing (SDAR) and present the Selective Attention System (SAS), an on-device implementation that instantiates this formulation. On a held-out 60-hour multi-speaker English test set, the primary audio-only configuration achieves F1=0.86 (precision=0.89, recall=0.83); with an optional camera, audio+video fusion raises F1 to 0.95 (precision=0.97, recall=0.93). Removing causal interaction history (Stage~3) reduced F1 from 0.95 to 0.57+/-0.03 in the audio+video configuration under our evaluation protocol. Among the tested components, this was the largest observed ablation effect, indicating that short-horizon interaction history carries substantial decision-relevant information in the evaluated setting. SAS runs fully on-device on ARM Cortex-A class hardware (<150 ms latency, <20 MB footprint). All results are from internal evaluation on a proprietary dataset evaluated primarily in English; a 5-hour evaluation subset may be shared for independent verification (Section 8.8).

A GAN and LLM-Driven Data Augmentation Framework for Dynamic Linguistic Pattern Modeling in Chinese Sarcasm Detection

Sarcasm is a rhetorical device that expresses criticism or emphasizes characteristics of certain individuals or situations through exaggeration, irony, or comparison. Existing methods for Chinese sarcasm detection are constrained by limited datasets and high construction costs, and they mainly focus on textual features, overlooking user-specific linguistic patterns that shape how opinions and emotions are expressed. This paper proposes a Generative Adversarial Network (GAN) and Large Language Model (LLM)-driven data augmentation framework to dynamically model users' linguistic patterns for enhanced Chinese sarcasm detection. First, we collect raw data from various topics on Sina Weibo. Then, we train a GAN on these data and apply a GPT-3.5 based data augmentation technique to synthesize an extended sarcastic comment dataset, named SinaSarc. This dataset contains target comments, contextual information, and user historical behavior. Finally, we extend the BERT architecture to incorporate multi-dimensional information, particularly user historical behavior, enabling the model to capture dynamic linguistic patterns and uncover implicit sarcastic cues in comments. Experimental results demonstrate the effectiveness of our proposed method. Specifically, our model achieves the highest F1-scores on both the non-sarcastic and sarcastic categories, with values of 0.9138 and 0.9151 respectively, which outperforms all existing state-of-the-art (SOTA) approaches. This study presents a novel framework for dynamically modeling users' long-term linguistic patterns in Chinese sarcasm detection, contributing to both dataset construction and methodological advancement in this field.

Don't Overthink It: Inter-Rollout Action Agreement as a Free Adaptive-Compute Signal for LLM Agents

Inference-time compute scaling has emerged as a powerful technique for improving the reliability of large language model (LLM) agents, but existing methods apply compute uniformly: every decision step receives the same budget regardless of its difficulty. We introduce TrACE (Trajectorical Adaptive Compute via agrEement), a training-free controller that allocates LLM calls adaptively across agent timesteps by measuring inter-rollout action agreement. At each step, TrACE samples a small set of candidate next actions and measures how consistently the model commits to the same action. High agreement signals an easy decision; the controller commits immediately. Low agreement signals uncertainty; the controller samples additional rollouts up to a configurable cap before committing to the plurality action. No learned components, no external verifier, and no human labels are required. We evaluate TrACE against greedy decoding and fixed-budget self-consistency (SC-4, SC-8) on two benchmarks spanning single-step reasoning (GSM8K, n=50) and multi-step household navigation (MiniHouse, n=30), using a Qwen 2.5 3B Instruct model running on CPU. TrACE-4 matches SC-4 accuracy while using 33% fewer LLM calls on GSM8K and 39% fewer on MiniHouse. TrACE-8 matches SC-8 accuracy with 55% fewer calls on GSM8K and 65% fewer on MiniHouse. We further show that inter-rollout agreement is a reliable signal of step-level success, validating the core hypothesis that the model's own output consistency encodes difficulty information that can be exploited without training. TrACE is the first training-free, per-timestep adaptive-compute controller for LLM agents to be evaluated on multi-step sequential decision tasks.

DiV-INR: Extreme Low-Bitrate Diffusion Video Compression with INR Conditioning

We present a perceptually-driven video compression framework integrating implicit neural representations (INRs) and pre-trained video diffusion models to address the extremely low bitrate regime (<0.05 bpp). Our approach exploits the complementary strengths of INRs, which provide a compact video representation, and diffusion models, which offer rich generative priors learned from large-scale datasets. The INR-based conditioning replaces traditional intra-coded keyframes with bit-efficient neural representations trained to estimate latent features and guide the diffusion process. Our joint optimization of INR weights and parameter-efficient adapters for diffusion models allows the model to learn reliable conditioning signals while encoding video-specific information with minimal parameter overhead. Our experiments on UVG, MCL-JCV, and JVET Class-B benchmarks demonstrate substantial improvements in perceptual metrics (LPIPS, DISTS, and FID) at extremely low bitrates, including improvements on BD-LPIPS up to 0.214 and BD-FID up to 91.14 relative to HEVC, while also outperforming VVC and previous strong state-of-the-art neural and INR-only video codecs. Moreover, our analysis shows that INR-conditioned diffusion-based video compression first composes the scene layout and object identities before refining textural accuracy, exposing the semantic-to-visual hierarchy that enables perceptually faithful compression at extremely low bitrates.

Temporal Graph Neural Network for ISAC Target Detection and Tracking

Integrated sensing and communication (ISAC) is a key enabler of 6G, supporting environment-aware services. A fundamental sensing task in this setting is reliable multi-target detection and tracking. This paper proposes a temporal graph neural network (TGNN)-based tracking method that exploits delay and Doppler information from the wireless channel. The delay-Doppler map is modeled as a sequence of graphs, and tracking is formulated as a temporal node classification problem, enabling joint clustering and data association of dynamic targets. Using ray-tracing-based channel outputs as ground truth, the method is evaluated across multiple scenes with varying target positions, velocities, and trajectories and is compared with a Kalman filter baseline. Results demonstrate reduced normalized mean squared error (NMSE) in delay and Doppler, leading to more accurate multi-target tracking.