SoM-MTM: Synesthesia of Machines (SoM)-Driven Masked Token Model for Cooperative Perception over Packet Loss Channel

To support the large-scale and heterogeneous visual cooperative perception (CP) demands in next-generation mobile networks, intelligent and efficient sensory data transmission is a critical challenge. Under the emerging convergence of communication networks and agentic artificial intelligence (AI), existing research emphasizes utilizing end-to-end neural networks to simplify communication modules, which has shown promising potential for CP. However, these studies are still limited to specific channel models, cooperation modes, and perception tasks, failing to fully leverage powerful visual processing approaches to enhance universality. To address this, we propose a Synesthesia of Machines (SoM)-driven Masked Token Model, referred to as SoM-MTM, as a plug-and-play paradigm for generic visual CP. Inspired by masked image modeling methods such as MAE, it possesses great perceptual context learning capabilities to recover distorted features over packet loss channels, thereby improving information carrying efficiency. Building upon Swin Transformer, SoM-MTM further embeds prior masked information through an External Routing MoE mechanism, maximally repairing and enhancing environmental perception features during cooperation. Comprehensive experimental results confirm that SoM-MTM can consistently enhance perception performances on various tasks, especially strong generalization to unseen scenarios, while maintaining favorable model cost and scalability.

WiFo-INR: A Wireless Foundation Model Based on Implicit Neural Representations

Wireless foundation models are emerging as a promising paradigm for AI-native physical-layer design. However, existing methods typically model channel state information (CSI) as image-like discrete tensors with generic token decoders that may struggle to capture complex high-frequency variations efficiently and often produce high-dimensional, size-dependent representations. In this paper, we propose WiFo-INR, an implicit neural representation (INR)-based wireless foundation model that represents CSI as a coordinate-conditioned neural function. A Transformer encoder maps partial or coarse CSI to fixed-dimensional modulation tokens that adapt a SIREN-based decoder, and a compression autoencoder enables quantized CSI feedback. It adopts a two-stage self-supervised pretraining scheme, where mixed masking and denoising improve channel reconstruction and compression-enhanced pretraining enables accurate CSI feedback at low compression ratios. Extensive experiments demonstrate that WiFo-INR learns efficient, compact, and CSI-size-independent implicit wireless representations. Compared with existing foundation models, WiFo-INR improves channel reconstruction and CSI feedback performance while substantially reducing inference latency. It also transfers efficiently to diverse wireless tasks with minimal fine-tuning overhead and achieves zero-shot generalization to unseen CSI sizes.

Semantic Sampling via Learnable Observation Front Ends

Sampling determines the form of information available to downstream reconstruction systems. Conventional lowrate sampling forms finite-dimensional observations directly from the raw waveform, with the sampling rule mainly guided by bandwidth, sparsity, or fixed signal-level structures. For acoustic signals such as speech, however, reconstruction-relevant information is often expressed through content-related spectral-temporal structures rather than waveform samples alone. This paper proposes semantic sampling via learnable observation front ends, where finite-dimensional observations are generated from learned signal responses instead of directly subsampled waveform points. The proposed front end consists of a semantic feature filterbank, a constrained semantic observation matrix, and a low-rate readout module. The filterbank maps the input waveform into multiple acoustic response channels, the observation matrix combines these responses into a small number of observation channels, and the readout module produces low-rate finite-dimensional samples. A reconstruction network is then used to recover the signal from the resulting observations. Experiments on low-rate speech reconstruction show that, under the same observation budget, the proposed semantic sampling front end provides more informative observations than fixed low-rate sampling and neural restoration methods based on predetermined low-rate waveforms. The improvements in waveform fidelity, spectral consistency, and perceptual quality show that learnable observation front ends preserve more useful information for acoustic signal reconstruction under the same observation budget.

Learning When to Denoise: Optimizing Asynchronous Schedules for Latent Diffusion

Multi-representation diffusion models can improve visual synthesis by denoising complementary views of an image, but their performance depends critically on the asynchronous schedule that determines when each representation is denoised. We propose to learn this schedule. Our method formulates asynchronous flow matching over multiple representation spaces and uses a schedule-corrected objective that keeps each representation's local noising-time weights fixed as the schedule changes. We instantiate the schedule with a flexible parametric class that is convex and monotone by construction, and learn it using a fast joint probe with less than 1% additional training compute. On ImageNet 256x256, the learned schedule substantially improves both convergence speed and final quality under a matched 675M-parameter XL backbone. With AutoGuidance, our 200-epoch model reaches FID 1.05, matching the 800-epoch SFD-XL baseline with 4x less training. Training to 600 epochs further improves to FID 1.02, outperforming the 1B-parameter SFD-XXL result of FID 1.04 while using a smaller model. In the unguided setting, our 200-epoch model reaches FID 2.37, already below the best 800-epoch SFD-XL result (2.54) at 4x less training, and improves to FID 2.14 at 600 epochs. Code is available at https://github.com/bsq532087/LWD

Transformers Can Learn Posterior Predictive Distributions In-Context

Prior-data fitted networks (PFNs) have recently emerged as a powerful approach for Bayesian prediction tasks, approximating the posterior predictive distribution (PPD) through in-context learning. Despite their strong empirical performance and ability to go beyond point predictions, theoretical understandings of the algorithmic capability of transformers to learn distributions in context are still lacking. Focusing on Gaussian process regression problems, we show by construction that transformers can implement a gradient descent algorithm targeting the posterior predictive mean and variance, followed by nonlinear mappings that yield binned probabilities of PPD. We study the error bounds of the approximated PPD in terms of attention depth and bin resolution. Based on these results, we further demonstrate the key role of normalization and the choice of attention depth in enabling the extrapolation abilities of transformers beyond the pretraining sample size range. We conduct simulations that corroborate our findings, providing insight into the expressivity of PFNs targeting PPDs and how architectural choices may influence generalization capabilities.

GroupTravelBench: Benchmarking LLM Agents on Multi-Person Travel Planning

Travel planning is a realistic task for evaluating the planning and tool-use abilities of LLM agents. However, existing benchmarks typically assume only a single user, thereby avoiding one of the most challenging aspects of real-world scenarios: an agent's ability to identify and resolve conflicts among multiple users. To address this gap, we introduce \textbf{GroupTravelBench}, the first benchmark for \textbf{multi-user, multi-turn} travel planning. Based on real user profiles, POI data, and ticket price data, we synthesize 650 tasks and divide them into three difficulty levels. Beyond standard abilities in single-user itinerary planning, such as multi-step reasoning and tool use, our benchmark further evaluates three key capabilities required for travel agents: \emph{(i) elicitation} -- proactively engaging in multi-turn dialogue to gather preferences from each user; \emph{(ii) coordination} -- resolving conflicts among users through compromise or subgrouping strategies; and \emph{(iii) planning} -- searching for travel plans that maximize overall group utility while maintaining fairness and feasibility. To simulate real-world conversational itinerary planning while enabling reliable tool use and offline evaluation, we build an interactive sandbox environment with cached real-world tool data. We evaluate a wide range of LLMs and find that even frontier models still show substantial weaknesses in preference coverage and group fairness. \textit{GroupTravelBench} provides a practical and reproducible benchmark for advancing research on LLM agents for real-world travel planning.

A Multi-Modal Intelligent U2V Channel Model for 6G Sensing-Communication Integration

This paper proposes a novel UAV-to-Vehicle (U2V) channel model for sixth-generation (6G) intelligent sensing-communication integration, based on three-dimensional (3D) scatterer prediction. To explore the mapping relationship between physical environment and electromagnetic space, a new high-fidelity mixed sensing-communication integration U2V simulation dataset under wide-lane scenarios with different vehicular traffic densities (VTDs) and UAV heights is constructed. Based on the constructed dataset, a novel 3D Scatterer Prediction and Distribution Estimation (3D-SPADE) algorithm is proposed, which leverages LiDAR point clouds to accurately predict the spatial distribution of scatterers. Furthermore, the clustering of scatterers and the subsequent classification into dynamic and static types are meticulously designed for highly dynamic U2V scenarios, while reducing computational complexity and improving modeling accuracy. As LiDAR point clouds vary over time, dynamic and static clusters evolve via 3D-SPADE, enabling precise modeling of channel non-stationarity and consistency. Simulation results demonstrate that, in the wide-lane scenario with varying VTDs and UAV heights, the proposed 3D-SPADE consistently achieves high scatterer occupancy detection performance within the voxel grid. In particular, under favorable configurations, recall reaches 93.26%, and precision reaches 95.74%, highlighting the reliability of 3D-SPADE. Key channel statistical characteristics are simulated and analyzed. These characteristics from the simulation experiments are highly consistent with ray-tracing results and exhibit better agreement than with the standardized model and inconsistent model, validating the necessity of exploring the mapping relationship and the effectiveness of the proposed model.

Sensing-Assisted LoS/NLoS Identification in Dynamic UAV Positioning Systems

In this paper, a sensing-assisted non-line-of-sight (NLoS) identification method for dynamic uncrewed aerial vehicle (UAV) positioning is proposed for the first time. For urban UAV-to-ground scenarios, a new multi-modal sensing-communication integrated dataset is constructed to support line-of-sight (LoS)/NLoS identification, covering two typical urban scenarios and a wide range of flight altitudes. Based on the constructed dataset, a novel dual-input feature fusion network is proposed, which addresses the challenge of heterogeneous representations between RGB images and channel impulse response (CIR) data to enable the joint extraction and fusion of sensing and communication features for LoS/NLoS identification. Simulation results show that the identification accuracy can reach up to 97.69%, while achieving an improvement of at least 3.59% compared to traditional CIR-only and RGB-only methods. Moreover, strong few-shot generalization is observed, as the proposed method stabilizes and approaches full-sample performance with fewer than 200 target samples and exceeds traditional CIR-only and RGB-only methods with fewer than 100 target samples in all cross-scenario and cross-altitude experiments. Even under Gaussian noise with a variance of 0.35 applied to RGB images, the accuracy degradation remains approximately 0.5%. By utilizing the proposed LoS/NLoS identification method, the error of trilateration positioning can be reduced by approximately 70% in a crossroad scenario, verifying the utility of the proposed method.

From Generic Correlation to Input-Specific Credit in On-Policy Self Distillation

On-policy self-distillation has emerged as a promising paradigm for post-training language models, in which the model conditions on environment feedback to serve as its own teacher, providing dense token-level rewards without external teacher models or step-level annotations. Despite its empirical success, what this reward actually measures and what kind of credit it assigns remain unclear. Under a posterior-compatibility interpretation of feedback conditioning, standard in the implicit-reward literature, we show that the self-distillation token reward is a Bayesian filtering increment whose trajectory sum is exactly the pointwise mutual information between the response and the feedback given the input. This pMI can be raised by input-specific reasoning or by input-generic shortcuts, so we further decompose the teacher log-probability along the input axis. Based on this analysis, we propose CREDIT (Contrastive REward from DIsTillation), which isolates the input-specific component with a batch-contrastive baseline. At the sequence level, CREDIT is a teacher-side surrogate for a contrastive pMI objective that also penalizes responses remaining likely under unrelated inputs. Across coding, scientific reasoning, and tool-use benchmarks on two model families, CREDIT delivers the strongest aggregate performance at negligible additional compute.

Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information

On-policy self-distillation, where a student is pulled toward a copy of itself conditioned on privileged context (e.g., a verified solution or feedback), offers a promising direction for advancing reasoning capability without a stronger external teacher. Yet in math reasoning the gains are inconsistent, even when the same approach succeeds elsewhere. A pointwise mutual information analysis traces the failure to the privileged context itself: it inflates the teacher's confidence on tokens already implied by the solution (structural connectives, verifiable claims) and deflates it on deliberation tokens ("Wait", "Let", "Maybe") that drive multi-step search. We propose Anti-Self-Distillation (AntiSD), which ascends a divergence between student and teacher rather than descending it: this reverses the per-token sign and yields a naturally bounded advantage in one step. An entropy-triggered gate disables the term once the teacher entropy collapses, completing a drop-in replacement for default self-distillation. Across five models from 4B to 30B parameters on math reasoning benchmarks, AntiSD reaches the GRPO baseline's accuracy in 2 to 10x fewer training steps and improves final accuracy by up to 11.5 points. AntiSD opens a path to scalable self-improvement, where a language model bootstraps its own reasoning through its training signal.