The paper presents Context-Aware Attention-based Gaussian Mixture Models (CAA‑GMM) for predicting vehicular trajectories in complex, uncertain environments. CAA‑GMM models future motion as a probabilistic mixture conditioned on scene context and agent dynamics, using a lightweight attention mechanism to encode inter‑agent interactions and contextual salience. Experiments on nuScenes and Argoverse 2 show competitive or superior accuracy to raster‑based baselines while keeping computational complexity low, and ablation studies confirm the attention module’s importance for robust contextual reasoning.
By Arash Raftari, Babak Ebrahimi Soorchaei, Yaser P. Fallah
TR-PTQ is a post‑training quantization method that enables high‑accuracy, integer‑only inference for transformer models. It identifies learned scale parameters in LayerNorm and compounded approximations in GELU as the main sources of accuracy loss, while SoftMax is robust to aggressive quantization. By reformulating these operations with shared Taylor Region exponential and logarithm primitives and a calibration‑free, outlier‑aware optimization for LayerNorm, TR‑PTQ eliminates the need for floating‑point units for nonlinearities and achieves less than 1.5% absolute accuracy degradation on vision and language benchmarks.
By Eliyahu Levy, Adam Teman, Yoni Pugachov
PVSync is a unified model that estimates audio‑visual offset and scores phoneme‑level articulation for lip‑sync tasks. It uses window‑level contrastive learning for synchronisation and a phoneme‑level articulation objective that aligns audio and video embeddings of the same viseme class across clips. The model achieves higher correlation with human rankings and better performance on offset recovery and viseme matching benchmarks compared to existing methods.
By Kevin Stephen, Varun Menon, Timo Mertens, Nikita Drobyshev
The paper introduces Δ‑MOPD, a multi‑teacher on‑policy distillation method that transfers each teacher’s logit shift relative to its base policy, re‑anchored at the student’s frozen initialization. Δ‑MOPD is evaluated in two settings: common‑domain composition, where multiple teachers score a single student rollout, and routed‑domain distillation, where prompts are assigned to specialist teachers. Across experiments, Δ‑MOPD outperforms traditional endpoint supervision, especially when teacher signals are combined at a state, achieving higher accuracy and reducing performance gaps in phased routing scenarios.
By Hejian Sang, Zhengze Zhou, Shayan Mohajer Hamidi, Xiaomin Li, Rohit Jain, Alborz Geramifard
The paper introduces Continual Graph Multi-Agent Reinforcement Learning (CGMARL), a framework that maps sequences of tasks into attributed graphs to capture task-specific structures such as network topologies or formation arrangements. It presents the Graph-based Formation (GRAFO) benchmark to illustrate forgetting in graph-based CMARL settings and proposes the Frozen Graph Encoder (FROG) to preserve past structural information using a frozen graph backbone. Experiments on GRAFO demonstrate that combining FROG with existing continual learning methods significantly improves performance across multiple CGMARL scenarios.
By Tommaso Marzi, Ahmed Hendawy, Jan Peters, Carlo D'Eramo, Andrea Cini, Cesare Alippi
The paper presents a bias-aware digital‑twin calibration and control framework for multiscale bioprocess models, using stochastic differential equations and quasi‑likelihood estimation with adjoint sensitivity analysis. It introduces SDE generator‑based moment expansions to quantify truncation‑induced parameter bias and forward‑backward adjoints to trace how calibration uncertainty affects value functions and policy performance. An Actor‑Simulator algorithm is proposed to jointly update model parameters, select informative experiments, and optimize control policies, achieving improved calibration accuracy, sample efficiency, and control performance in numerical studies.
By Keilung Choy, Wei Xie
The paper introduces a method to estimate a model’s vulnerability to the Likelihood Ratio Attack (LiRA) without training reference models, using only the target model’s train and test loss distributions. It shows that LiRA’s per‑sample signal can be decomposed into a variance‑ratio term and a residual mean‑shift term, and that different loss‑distribution shapes dictate which reference‑free proxy to use. Two proxies are presented: the LOSS attack TNR for heavy‑tailed losses and the LOSS attack AUC for symmetric losses, both achieving low RMSE in predicting LiRA TPR across multiple architectures and datasets.
By Euodia Dodd, Nata\v{s}a Kr\v{c}o, Igor Shilov, Matthew Wicker, Yves-Alexandre de Montjoye
PrismSSL is a Python library that consolidates leading self‑supervised learning methods for audio, vision, graphs, and cross‑modal tasks into a single, modular codebase. It allows users to install, configure, and run pretext training with minimal code, reproduce compact benchmarks, and extend the framework with new modalities or methods via clean trainer and dataset abstractions. The library, available on PyPI under an MIT license, integrates with HuggingFace Transformers and offers features such as distributed PyTorch training, Optuna hyperparameter search, LoRA fine‑tuning, animated embedding visualizations, Weights & Biases logging, and a Flask‑based graphical dashboard for easy pipeline configuration.
By Melika Shirian, Kianoosh Vadaei, Kian Majlessi, Audrina Ebrahimi, Peyman Adibi, Hossein Karshenas
Continuous-Utility Direct Preference Optimization (CU‑DPO) replaces binary preference labels with continuous scores to better capture fine‑grained reasoning quality. The framework aligns language models to multiple prompt‑based cognitive strategies, proving a θ(K log K) sample‑complexity improvement and convergence to an entropy‑regularized utility‑maximizing policy. A two‑stage pipeline—strategy selection via best‑vs‑all comparisons and execution refinement using margin‑stratified pairs—yields significant gains on mathematical reasoning, code generation, and causal reasoning benchmarks.
By Muhammad Ahmed Mohsin, Muhammad Umer, Emily Fox
The paper introduces RASA, a routing‑aware safety alignment framework for Mixture‑of‑Experts (MoE) language models. RASA repairs Safety‑Critical Experts by selectively fine‑tuning only those experts that are disproportionately activated during jailbreak attacks, while keeping routing fixed and enforcing routing consistency with safety‑aligned contexts. Experiments on two MoE architectures show that RASA achieves near‑perfect robustness against diverse jailbreak attacks, generalizes across attacks, reduces over‑refusal, and maintains performance on standard benchmarks such as MMLU, GSM8K, and TruthfulQA.
By Jiacheng Liang, Yuhui Wang, Tanqiu Jiang, Ting Wang
SkillRL is a framework that enhances large language model agents by automatically discovering and evolving skills from raw experience. It builds a hierarchical skill library called SkillBank, uses an adaptive retrieval strategy for heuristics, and allows the skill library to co‑evolve with the agent’s policy during reinforcement learning. These techniques reduce token usage and improve reasoning, achieving state‑of‑the‑art results on ALFWorld, WebShop, and seven search‑augmented tasks, outperforming baselines by 15.3% and remaining robust as task complexity grows.
By Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, Zeyu Zheng, Cihang Xie, Huaxiu Yao
The paper introduces beignet, a neural field architecture that replaces the random Fourier feature embedding in physics-informed neural networks with a trainable multi‑resolution Fourier feature pyramid. By using Fourier interpolation and spectral derivatives, beignet efficiently computes spatial derivatives and scales accuracy through the feature pyramid rather than the neural network size. Experiments show that beignet achieves more accurate PDE solutions with fewer parameters than existing PINN methods and can reach machine‑precision residuals on the inviscid Burgers blowup problem using the Adam optimizer.
By Brandon Zhao, Yixuan Wang, Jonathan T. Barron, Katherine L. Bouman, Dor Verbin, Pratul P. Srinivasan
The paper demonstrates that a fixed‑weight transformer can exactly execute one update step of a standard continuous causal discovery algorithm, and by stacking such blocks it reproduces the algorithm’s optimization trajectory. The transformer maintains both the current graph and the algorithm’s multiplier, which is shown to be essential for accurate updates. Experiments confirm that the constructed block matches the reference solver to floating‑point precision and that arithmetic replay on synthetic and benchmark networks inherits the solver’s performance, while ordinary attention models fail to reliably execute the update or scale to larger graphs.
By Amartya Roy, Sayar Karmakar
The paper introduces Exploration-Distillation (ExpDis), a method that separates exploration from optimization in reinforcement learning with verifiable rewards (RLVR) for language models. ExpDis trains explorer policies with a novelty bonus, filters their outputs for correctness, and distills the results into a student policy that is optimized without the novelty bonus. This iterative process enables aggressive exploration while preserving model quality, outperforming DAPO on multiple mathematical reasoning benchmarks and improving pass@$k$ scaling.
By Saif Punjwani, Micah Goldblum
WaveScat introduces a new family of feature extractors for speech deepfake detection that merges wavelet scattering transform (WST) with self‑supervised learning. By cascading wavelet convolutions with modulus nonlinearities, WaveScat produces deformation‑stable, multi‑scale features that capture both low‑level spectral details and higher‑level information. Experiments on Deepfake‑Eval‑2024 and cross‑dataset tests (SpoofCeleb, In‑the‑Wild, ASVspoof 5) demonstrate that WaveScat significantly outperforms existing front‑ends, with a small averaging scale and high‑frequency directional resolution proving crucial for detecting subtle artifacts.
By Xi Xuan, Davide Carbone, Wenxin Zhang, Tomi H. Kinnunen
TiTok is an audio‑visual large language model designed for multi‑segment temporal grounding in untrimmed videos. It introduces Time Token Interleaving (TTI) to align temporal perception with prediction, and uses decoupled reinforcement‑learning rewards (global, local, count, precision, format) optimized via GDPO. Evaluated on a new UnAV‑100 protocol with the CountF1 metric, TiTok achieves state‑of‑the‑art results (65.7 mIoU, 0.58 CountF1).
By Eunji Shin, Dahyun Choi, Seungyeon Jo, Yejin Hong, Jiyoung Lee
The paper introduces Ctrl-CWM, a multi‑agent controllable crowd world model that merges crowd generation with run‑time control. It uses an encoder to learn human motion dynamics, an actor to propose pedestrian displacements, a critic to evaluate imagined trajectories, and a planner that selects actions based on critic scores and user‑defined costs. Experiments show that Ctrl‑CWM outperforms state‑of‑the‑art methods in realism and collision metrics while allowing new control objectives to be added without retraining.
By JunGyu Lee, Jisu Shin, Seunghyun Shin, Hae-Gon Jeon
SoccerNet‑FoulRet is a new benchmark for retrieving semantically similar soccer foul videos, addressing inconsistencies in refereeing decisions by allowing comparison of a contentious foul with past precedents. The task requires retrieving relevant fouls regardless of camera angle, teams, or appearance, differing from prior video‑to‑video retrieval that focuses on visual similarity. The benchmark, built from the SoccerNet‑MVFoul dataset, evaluates zero‑shot video and vision‑language embeddings and a fine‑tuned baseline on 693 human‑verified queries, revealing that even the best zero‑shot model achieves less than 5% HitRate@10, highlighting the challenge of semantic foul retrieval.
By Jacobus Arthur, Ahmad Sait, Batool Hani, Merey Ramazanova, Jan Held, Marc Van Droogenbroeck, Bernard Ghanem, Anthony Cioppa, Silvio Giancola
Consistent Distribution Matching is a data‑free distillation technique that accelerates diffusion and flow models by unifying sample generation and score estimation in a single student network. The method requires only a frozen teacher and a trainable student, optimizing a single objective that provably leads to Wasserstein convergence of the student’s flow‑map to the teacher’s marginals. On ImageNet 256×256, it achieves an FID of 2.04 with one function evaluation and 1.37 with four evaluations after just 40 training epochs, outperforming existing data‑free baselines.
By Yuxiang Fu, Qi Yan, Zike Wu, Yongxing Zhang, Purang Abolmaesumi, Lele Wang, Renjie Liao
arXiv:2610. 09651v1 Announce Type: new Abstract: DP-SGD protects training data by adding Gaussian noise to clipped gradients.
By Murat Bilgehan Ertan, Marten van Dijk