From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial?

Multi-agent systems (MAS) tackle complex tasks by distributing expertise, though this often comes at the cost of heavy coordination overhead, context fragmentation, and brittle phase ordering. Distilling a MAS into a single-agent skill can bypass these costs, but this conversion lacks a principled answer for when and what to distill. Instead, the empirical outcome is surprisingly inconsistent: skill lift ranges from a 28% improvement to a 2% degradation across metrics of the exact same task. In this work, we reveal that skill utility is governed not by the task, but by the evaluation metric. We introduce Metric Freedom ($F$), the first a priori predictor of skill utility. $F$ measures the topological rigidity of a metric's scoring landscape by quantifying how output diversity couples with score variance via a Mantel test. Guided by $F$, we propose a two-stage adaptive distillation framework. Stage 1 acts as a selective extraction mechanism, extracting tools and knowledge while discarding restrictive structures on "free" metrics to preserve exploration. Stage 2 targets computationally intensive iterative refinement exclusively toward "rigid" metrics ($F \lesssim 0.6$) to eliminate trajectory-local overfitting. Evaluating across 4 tasks, 11 datasets, and 6 metrics, $F$ strongly predicts skill utility ($ρ= -0.62$, $p < 0.05$). Strikingly, identical agent trajectories yield diametrically opposite skill lifts under rigid versus free metrics, demonstrating that skill utility is fundamentally a metric-level property. Driven by this signal, our adaptive agent matches or exceeds the original MAS while reducing cost up to 8$\times$ and latency by up to 15$\times$.

SHIFT: Stochastic Hidden-Trajectory Deflection for Removing Diffusion-based Watermark

Diffusion-based watermarking methods embed verifiable marks by manipulating the initial noise or the reverse diffusion trajectory. However, these methods share a critical assumption: verification can succeed only if the diffusion trajectory can be faithfully reconstructed. This reliance on trajectory recovery constitutes a fundamental and exploitable vulnerability. We propose $\underline{\mathbf{S}}$tochastic $\underline{\mathbf{Hi}}$dden-Trajectory De$\underline{\mathbf{f}}$lec$\underline{\mathbf{t}}$ion ($\mathbf{SHIFT}$), a training-free attack that exploits this common weakness across diverse watermarking paradigms. SHIFT leverages stochastic diffusion resampling to deflect the generative trajectory in latent space, making the reconstructed image statistically decoupled from the original watermark-embedded trajectory while preserving strong visual quality and semantic consistency. Extensive experiments on nine representative watermarking methods spanning noise-space, frequency-domain, and optimization-based paradigms show that SHIFT achieves 95%--100% attack success rates with nearly no loss in semantic quality, without requiring any watermark-specific knowledge or model retraining.

The 1st Winner for 5th PVUW MeViS-Text Challenge: Strong MLLMs Meet SAM3 for Referring Video Object Segmentation

This report presents our winning solution to the 5th PVUW MeViS-Text Challenge. The track studies referring video object segmentation under motion-centric language expressions, where the model must jointly understand appearance, temporal behavior, and object interactions. To address this problem, we build a fully training-free pipeline that combines strong multimodal large language models with SAM3. Our method contains three stages. First, Gemini-3.1 Pro decomposes each target event into instance-level grounding targets, selects the frame where the target is most clearly visible, and generates a discriminative description. Second, SAM3-agent produces a precise seed mask on the selected frame, and the official SAM3 tracker propagates the mask through the whole video. Third, a refinement stage uses Qwen3.5-Plus and behavior-level verification to correct ambiguous or semantically inconsistent predictions. Without task-specific fine-tuning, our method ranks first on the PVUW 2026 MeViS-Text test set, achieving a Final score of 0.909064 and a J&F score of 0.7897. The code is available at https://github.com/Moujuruo/MeViSv2_Track_Solution_2026.

A Survey of On-Policy Distillation for Large Language Models

Knowledge distillation has become a primary mechanism for transferring reasoning and domain expertise from frontier Large Language Models (LLMs) to smaller, deployable students. However, the dominant paradigm remains \textit{off-policy}: students train on static teacher-generated data and never encounter their own errors during learning. This train--test mismatch, an instance of \textit{exposure bias}, causes prediction errors to compound autoregressively at inference time. On-Policy Distillation (OPD) addresses this by letting the student generate its own trajectories and receive teacher feedback on these self-generated outputs, grounding distillation in the theory of interactive imitation learning. Despite rapid growth spanning divergence minimization, reward-guided learning, and self-play, the OPD literature remains fragmented with no unified treatment. This survey provides the first comprehensive overview of OPD for LLMs. We introduce a unified $f$-divergence framework over on-policy samples and organize the landscape along three orthogonal dimensions: \emph{feedback signal} (logit-based, outcome-based, or self-play), \emph{teacher access} (white-box, black-box, or teacher-free), and \emph{loss granularity} (token-level, sequence-level, or hybrid). We systematically analyze representative methods, examine industrial deployments, and identify open problems including distillation scaling laws, uncertainty-aware feedback, and agent-level distillation.

Activation Saturation and Floquet Spectrum Collapse in Neural ODEs

We prove that activation saturation imposes a structural dynamical limitation on autonomous Neural ODEs $\dot{h}=f_θ(h)$ with saturating activations ($\tanh$, sigmoid, etc.): if $q$ hidden layers of the MLP $f_θ$ satisfy $|σ'|\leδ$ on a region~$U$, the input Jacobian is attenuated as $\norm{Df_θ(x)}\le C(U)$ (for activations with $\sup_{x}|σ'(x)|\le 1$, e.g.\ $\tanh$ and sigmoid, this reduces to $C_Wδ^q$), forcing every Floquet (Lyapunov) exponen along any $T$-periodic orbit $γ\subset U$ into the interval $[-C(U),\;C(U)]$. This is a collapse of the Floquet spectrum: as saturation deepens ($δ\to 0$), all exponents are driven to zero, limiting both strong contraction and chaotic sensitivity. The obstruction is structural -- it constrains the learned vector field at inference time, independent of training quality. As a secondary contribution, for activations with $σ'>0$, a saturation-weighted spectral factorisation yields a refined bound $\widetilde{C}(U)\le C(U)$ whose improvement is amplified exponentially in~$T$ at the flow level. All results are numerically illustrated on the Stuart--Landau oscillator; the bounds provide a theoretical explanation for the empirically observed failure of $\tanh$-NODEs on the Morris--Lecar neuron model.

Enhancing Gradient Inversion Attacks in Federated Learning via Hierarchical Feature Optimization

Federated Learning (FL) has emerged as a compelling paradigm for privacy-preserving distributed machine learning, allowing multiple clients to collaboratively train a global model by transmitting locally computed gradients to a central server without exposing their private data. Nonetheless, recent studies find that the gradients exchanged in the FL system are also vulnerable to privacy leakage, e.g., an attacker can invert shared gradients to reconstruct sensitive data by leveraging pre-trained generative adversarial networks (GAN) as prior knowledge. However, existing attacks simply perform gradient inversion in the latent space of the GAN model, which limits their expression ability and generalizability. To tackle these challenges, we propose \textbf{G}radient \textbf{I}nversion over \textbf{F}eature \textbf{D}omains (GIFD), which disassembles the GAN model and searches the hierarchical features of the intermediate layers. Instead of optimizing only over the initial latent code, we progressively change the optimized layer, from the initial latent space to intermediate layers closer to the output images. In addition, we design a regularizer to avoid unreal image generation by adding a small ${l_1}$ ball constraint to the searching range. We also extend GIFD to the out-of-distribution (OOD) setting, which weakens the assumption that the training sets of GANs and FL tasks obey the same data distribution. Furthermore, we consider the challenging OOD scenario of label inconsistency and propose a label mapping technique as an effective solution. Extensive experiments demonstrate that our method can achieve pixel-level reconstruction and outperform competitive baselines across a variety of FL scenarios.

Harmonization mitigates diffusion MRI scanner effects in infancy: insights from the HEALthy Brain and Childhood Development (HBCD) study

The HEALthy Brain and Childhood Development (HBCD) Study is an ongoing longitudinal initiative to understand population-level brain maturation; however, large-scale studies must overcome site-related variance and preserve biologically relevant signal. In addition to diffusion-weighted magnetic resonance imaging images, the HBCD dataset offers analysis-ready derivatives for scientists to conduct their analysis, including scalar diffusion tensor (DTI) metrics in a predetermined set of bundles. The purpose of this study is to characterize HBCD-specific site effects in diffusion MRI data, which have not been systematically reported. In this work, we investigate the sensitivity of HBCD bundle metrics to scanner model-related variance and address these variations with ComBat-GAM harmonization within the current HBCD data release 1.1 across six scanner models. Following ComBat-GAM, we observe zero statistically significant differences between the distributions from any scanner model following FDR correction and reduce Cohen's f effect sizes across all metrics. Our work underscores the importance of rigorous harmonization efforts in large-scale studies, and we encourage future investigations of HBCD data to control for these effects.

Offline Constrained RLHF with Multiple Preference Oracles

We study offline constrained reinforcement learning from human feedback with multiple preference oracles. Motivated by applications that trade off performance with safety or fairness, we aim to maximize target population utility subject to a minimum protected group welfare constraint. From pairwise comparisons collected under a reference policy, we estimate oracle-specific rewards via maximum likelihood and analyze how statistical uncertainty propagates through the dual program. We cast the constrained objective as a KL-regularized Lagrangian whose primal optimizer is a Gibbs policy, reducing learning to a convex dual problem. We propose a dual-only algorithm that ensures high-probability constraint satisfaction and provide the first finite-sample performance guarantees for offline constrained preference learning. Finally, we extend our theoretical analysis to accommodate multiple constraints and general f-divergence regularization.

Feature-level Site Leakage Reduction for Cross-Hospital Chest X-ray Transfer via Self-Supervised Learning

Cross-hospital failure in chest X-ray models is often attributed to domain shift, yet most work assumes invariance without measuring it. This paper studies how to measure site leakage directly and how that measurement changes conclusions about transfer methods. We study multi-site self-supervised learning (SSL) and feature-level adversarial site confusion for cross-hospital transfer. We pretrain a ResNet-18 on NIH and CheXpert without pathology labels. We then freeze the encoder and train a linear pneumonia classifier on NIH only, evaluating transfer to RSNA. We quantify site leakage using a post hoc linear probe that predicts acquisition site from frozen backbone features $f$ and projection features $z$. Across 3 random seeds, multi-site SSL improves RSNA AUC from 0.6736 $\pm$ 0.0148 (ImageNet initialization) to 0.7804 $\pm$ 0.0197. Adding adversarial site confusion on $f$ reduces measured leakage but does not reliably improve AUC and increases variance. On $f$, site probe accuracy drops from 0.9890 $\pm$ 0.0021 (SSL-only) to 0.8504 $\pm$ 0.0051 (CanonicalF), where chance is 0.50. On $z$, probe accuracy drops from 0.8912 $\pm$ 0.0092 to 0.7810 $\pm$ 0.0250. These results show that measuring leakage changes how transfer methods should be interpreted: multi-site SSL drives transfer, while adversarial confusion exposes the limits of invariance assumptions.

Training deep learning based dynamic MR image reconstruction using synthetic fractals

Purpose: To investigate whether synthetically generated fractal data can be used to train deep learning (DL) models for dynamic MRI reconstruction, thereby avoiding the privacy, licensing, and availability limitations associated with cardiac MR training datasets. Methods: A training dataset was generated using quaternion Julia fractals to produce 2D+time images. Multi-coil MRI acquisition was simulated to generate paired fully sampled and radially undersampled k-space data. A 3D UNet deep artefact suppression model was trained using these fractal data (F-DL) and compared with an identical model trained on cardiac MRI data (CMR-DL). Both models were evaluated on prospectively acquired radial real-time cardiac MRI from 10 patients. Reconstructions were compared against compressed sensing(CS) and low-rank deep image prior (LR-DIP). All reconstrctuions were ranked for image quality, while ventricular volumes and ejection fraction were compared with reference breath-hold cine MRI. Results: There was no significant difference in qualitative ranking between F-DL and CMR-DL (p=0.9), while both outperformed CS and LR-DIP (p<0.001). Ventricular volumes and function derived from F-DL were similar to CMR-DL, showing no significant bias and accptable limits of agreement compared to reference cine imaging. However, LR-DIP had a signifcant bias (p=0.016) and wider lmits of agreement. Conclusion: DL models trained using synthetic fractal data can reconstruct real-time cardiac MRI with image quality and clinical measurements comparable to models trained on true cardiac MRI data. Fractal training data provide an open, scalable alternative to clinical datasets and may enable development of more generalisable DL reconstruction models for dynamic MRI.