The paper introduces StereoBind, a framework that links visual source motion to stereo audio generation, addressing the need for dynamic spatial correspondence in immersive applications like AR/VR and gaming. It employs three mechanisms—Visual Motion Binding, Spatial Track Encoder, and Residual Track RoPE—to coordinate visual motion with stereo sound. The authors also present StereoWorld-29K, a large-scale dataset with paired motion tracks, and StereoWorldBench for evaluating audiovisual spatial consistency, demonstrating that StereoBind improves spatial alignment while maintaining overall quality.
By Hanmo Chen, Chengcheng Liu, Tianxiao Chen, Zheyu Zhang, Siming Zheng, Jinwei Chen, Xu Yang, Cheng Deng, Bo Li, Peng-tao Jiang
LAIR-Net is a Leaky Alignment-Impulse Residual Network designed for tabular regression that incorporates a shallow learned anchor into each hidden state via a leaky residual transition. The authors provide a depth‑uniform bound on input‑perturbation sensitivity and show through controlled simulations that performance gains over randomized baselines stem from the anchor rather than recursion or added capacity. On 23 benchmark datasets, LAIR‑Net achieves the best average rank among eight randomized networks and twelve conventional models, with its success linked to the presence of nonlinear target structure and manageable noise levels.
By Rahul Goswami, Aryan Bhambu, Bittu Karmakar
The paper introduces Gaussian mechanisms for Newton updates that enable continual machine unlearning—updating a model after record deletions without full retraining. Using Gaussian differential privacy and adaptive composition, it shows that the sequence of released models is statistically indistinguishable from exact retraining. The authors derive error bounds for the Newton approximation, calibrate independent and random‑walk Gaussian noise, and demonstrate minimax optimality under various residual and increment bounds, with empirical validation on simulations and credit‑default data.
By Qi Kuang, Yin Xia
Guard models are used to prevent LLM agents from performing forbidden actions, but this alone does not guarantee safety. The paper shows that many agent trajectories contain unfulfilled safety‑critical obligations—actions that should have been performed—highlighting a new source of risk. To address this, the authors introduce ObligationBench, a benchmark of 240 expert‑validated trajectories, and develop ObligationGuard, a model that improves obligation detection recall and exact‑match rates over existing models.
By Youwei Feng, Yitong Zhang, Yuetong Liu, Jia Li
The paper tackles learning structured linear dynamical systems when only a few observations are available at each time step. It proposes an estimator that minimizes a bias‑corrected, potentially non‑convex objective and derives non‑asymptotic error bounds that depend on the local complexity of the constraint set, the trajectory length, and the sub‑sampling probability. The authors also prove convergence of projected gradient descent and demonstrate the theory for three specific constraint sets, showing that accurate recovery of the transition matrix is achievable with much shorter trajectories and very low observation rates.
By Aravinda Kanchana Ruwanpathirana, Hemant Tyagi, Sunny G. W. Wang
InfiFPO is a new preference optimization technique for implicit model fusion that replaces the reference model in Direct Preference Optimization with a fused source model, preserving probability information and avoiding complex vocabulary alignment. By incorporating probability clipping and max‑margin fusion, it aligns a pivot model with human preferences while distilling knowledge from multiple source models. Experiments on 11 benchmarks show that InfiFPO consistently outperforms existing fusion and preference optimization methods, raising Phi‑4’s average score from 79.95 to 83.33 and improving performance in mathematics, coding, and reasoning tasks.
By Yanggan Gu, Yuanyi Wang, Zhaoyi Yan, Yiming Zhang, Qi Zhou, Fei Wu, Hongxia Yang
The paper investigates why neural networks trained on modular addition tasks develop Fourier-structured representations that enable exact generalization. Using probability signatures, the authors show that leading gradient interactions are cyclic shift operators, which are diagonalized by the discrete Fourier transform, leading to decoupled Fourier-mode dynamics that explain Fourier sparsity, frequency matching, and phase alignment. The framework also clarifies how label noise can accelerate early loss decrease by increasing conditional label collisions, and it is demonstrated to predict frequency behavior in other operators such as XOR.
By Yunji Wang, Junjie Yao, Linyu Liu, Pinyan Lu, Zhi-Qin John Xu
The paper introduces Multi-Agent Self-Supervision (MASS), a recursive self‑improvement technique that alternates between evolutionary workflow optimization and supervised fine‑tuning on self‑generated multi‑agent trajectories. MASS enables a single base model to iteratively propose, execute, and self‑evaluate complex multi‑agent workflows, discovering effective roles and information routing. Experiments with Qwen3.6‑27B show 1.2–1.6× higher performance per output token on four open‑ended benchmarks, and student models trained on MASS traces outperform single‑agent students trained on 1.4× more tokens.
By Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia, Yujin Tang
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing proposes the Superposed Specialisation Hypothesis (SSH), suggesting that experts in Sparse Mixture of Experts models specialise in a disjoint union of fine‑grained features rather than a single domain. The paper introduces RouterInterp, a method that identifies Sparse Autoencoder features predictive of routing decisions and generates unified natural‑language explanations. On the gpt‑oss‑20b model, RouterInterp achieves approximately 65% higher detection accuracy than previous token‑statistics based methods, offering a scalable way to explain expert routing.
By Ilya Lasy, Nora Yinuo Cai, Kola Ayonrinde
The study examines whether large language models truly base legal verdicts on the statutes or precedents they cite. By fixing case facts and swapping the cited authority with an unrelated one, researchers found that while models correctly name the authority 66.7%-100% of the time, the resulting verdict often remains unchanged—only 0.0%-21.7% of the time on CaseHOLD, 30.0%-76.7% on ECHR and SCOTUS, and 43.3%-50.0% on ContractNLI. Even a purpose‑built legal‑reasoning model and larger scales did not close this gap, and adversarial instructions further exposed vulnerabilities in verdict consistency.
By Saisab Sadhu, Shreeyans Arora, Pratinav Seth
MRCert is a masking‑based certified recovery defender that aims to maintain high prediction accuracy while verifying the benignity of labels for adversarially patched inputs in post‑deployment scenarios. It introduces type‑specific necessary properties for both benign and patched samples, enabling a novel label recovery and certification function pair that avoids the accuracy degradation typical of smoothing methods. Experiments show MRCert achieves 35.1% adversarial certified accuracy on ImageNet with 16‑pixel patches, outperforming the state‑of‑the‑art PatchCURE.
By Qilin Zhou, Zhengyuan Wei, Haipeng Wang, Zhuo Wang, Shuo Liu, W. K. Chan
Nullify is a training‑free method for unlearning sensitive information from large language models by steering activations during inference. It uses steering vectors to redirect privacy‑related activations away from memorized answers while enforcing a null‑space constraint that preserves activations for retained queries, thereby maintaining model utility. Experiments on TOFU and MUSE demonstrate that Nullify matches or outperforms existing baselines in forget quality while keeping utility nearly intact, all without any weight updates.
By Wei Zhai, Xiang Liu, Qiang Huang, Rui Qian, Lemao Liu, Ziwei Li, Ziqi Wang, Zhitao Huang, Dejing Dou
The paper introduces a benchmark for evaluating large language model (LLM) agents in closed‑loop embedded software development. It provides five embedded‑control tasks with realistic build‑and‑runtime feedback, testing agents that must translate requirements into code, self‑verify, and iterate until device behavior meets specifications. Seven GPT‑family and Qwen‑family models were evaluated across four feedback scenarios, revealing that gpt‑5.4 leads in pass rate but does not fully saturate the benchmark, while qwen3.5‑27B is the strongest local model and smaller local models perform poorly.
By Jorge Garc\'ia-Carrasco, Sergio Garc\'ia-Carrasco, Alejandro Mat\'e, Juan Trujillo
The paper evaluates a three‑part control stack— a 5‑Pillar runtime constitution, a 4‑tier swarm (Director, majority vote, Monitor, hard gate), and Cognitive Annealing—for multi‑agent LLM negotiation against an adversarial Gatekeeper. Experiments with Gemini 2.5 Pro agents and Claude Haiku 4.5 show that the constitution and Director improve framing but not reliability, the Monitor and hard gate do not reduce unlocks, and deterministic recovery via atomic purge and a canonical strike successfully resolves deadlocks while keeping call budgets unchanged. The study demonstrates that the stack can recover from deadlocks without additional model cost, though it does not reduce overall call or token usage as initially hypothesized.
By Masaaki Nakatsu (AO, Inc. / OrbLabs AG), Reno Wang (AO, Inc.)
The paper investigates how adversarial prompt‑injection attacks can shift the scaling of attack success from a slow polynomial growth to an exponential growth as the number of inference‑time samples increases. By proposing a minimal statistical framework and a spin‑glass generative model, the authors explain that short injected prompts act as a weak magnetic field leading to power‑law scaling, while long prompts act as a strong field producing exponential scaling. The phenomenon is observed consistently across large language models ranging from 3B to 70B parameters and across various attack methods and benchmark datasets.
By Indranil Halder, Annesya Banerjee, Cengiz Pehlevan
DCVD (Dual-Channel Cross-Modal Vulnerability Detection) is a unified framework that jointly performs function-level vulnerability detection and statement-level localization. It extracts control-dependency and semantic features via two parallel branches and fuses them using contrastive alignment and bidirectional cross-attention. The model incorporates explicit supervision at both granularities, achieving superior performance on a large real-world vulnerability benchmark compared to state-of-the-art methods.
By Wenxin Tang, Junliang Liu, Wenbin Li, Jingyu Xiao, Xi Xiao, Mingzhe Liu, Jinlong Yang, Xuan Liu, Yuehe Ma, Wang Luo, Qing Li, Lei Wang, Peng Xiangli
The paper introduces Neural Conjugate Aggregation (NCAM), a hierarchical Bayesian model that fuses multiple noisy, biased sensor measurements without ground‑truth labels. NCAM learns sensor‑specific bias and reliability conditioned on context, producing an analytically tractable posterior over a latent target with decomposed epistemic and aleatoric uncertainty. The method resolves non‑identifiability via sensor anchoring and variance regularization, and adds locally adaptive Monte Carlo conformal prediction to provide heteroscedastic prediction intervals with coverage guarantees. Experiments on synthetic and real‑world air‑quality data show that NCAM outperforms unsupervised baselines such as mean aggregation, probabilistic PCA, and Kalman filtering in predictive accuracy and uncertainty calibration.
By Muhammed Faruk Aytin, Zehra Demir, Alper Unal, Julian Marshall, Gozde Unal
CPU-Auth is a novel authentication method that uses unique variations in a device’s CPU, measured via the Dynamic Voltage and Frequency Scaling (DVFS) governor from a browser, to create a hardware-based device fingerprint. The approach is evaluated on over 50,000 data traces using distance‑based and deep learning techniques. The study highlights how CPU-Auth could serve as a secondary authentication factor, addressing flaws in existing methods such as SMS, biometrics, and token‑based systems.
By Ryan Swift
The study investigates whether Modern Standard Arabic (MSA) dominates internal representations in large language models (LLMs) when generating Arabic, even if prompted with dialectal forms. By extending a language-dominance framework to 26 Arabic varieties, the authors find no evidence that MSA serves as a dominant internal representation; instead, dialect representations are densely overlapping, with mutual information dropping sharply compared to more distinct languages. The strongest separability effects vary across layers and architectures, indicating that generation bias toward MSA does not necessarily reflect internal representational dominance.
By Abdu Sallouh, Nicholas Popovi\v{c}, Michael F\"{a}rber
Hybrid Cinematography is a workflow that combines physical capture with generative video reshooting to manage hallucination risk. It uses an editable 3D shot plan and a proxy of the take to evaluate hallucination risk in real time, allowing filmmakers to adjust plans, shoot guided pickups, or accept hallucination. The approach is demonstrated through a mobile AR app for on‑set planning and an offline pipeline for existing video, and a study with experienced filmmakers shows how risk previsualization informs camera decisions and highlights tensions between creative intent and generative hallucination.
By Nhan (Nathan), Tran, Neal Wadhwa, Abe Davis, Stefan Stojanov