Computation and Language
From Structure to Validity: Mechanism-Guided Generation and Refinement of LLM Rubrics
Feifei Qian, Xuan Chen, Lixin Cui et al. (+6 more)
Rubrics have been widely used to evaluate and optimize
language models on open-ended tasks. However, existing work primarily focuses on generating better rubrics, with limited understanding of what makes rubric design effective. In this work, we conduct a systematic mechanistic analysis of rubric design across criterion organization, criterion operationalization, and score aggregation. Our analysis shows that structural organization enables more comprehensive criterion construction, semantic specification alone does not ensure effective criterion-level judgments, and sophisticated score aggregation provides limited additional gains. Guided by these findings, we introduce Adaptive Rubric Tree Search with Hybrid refinement (ARTS-Hybrid), which combines adaptive tree construction with semantic-first, response-calibrated refinement. Experiments across
benchmarks show that ARTS-Hybrid improves both end-to-end evaluation performance and intrinsic rubric quality. Further experiments suggest that, beyond evaluation, rubrics generated by ARTS-Hybrid provide effective signals for model optimization.
Read the full abstract ▾
All authors Feifei Qian, Xuan Chen, Lixin Cui, Chaoyang Mei, Fanlin Meng, Shuyao Xiao, Ming Cui, Junxi Yin, Lu Bai
Robotics
PartiGen: Part-Aware Representations for Language-Conditioned Humanoid Motion Generation
Chuxuan Huang, Zhanxiang Cao, Tianyue Zhan et al. (+14 more)
Translating
language into humanoid motion requires coordinating body parts, determining motion duration, and producing references suitable for
robot execution. We introduce PartiGen, a part-aware latent
diffusion framework that incorporates anatomical structure into a shared whole-body motion representation. Part-specific
attention processes anatomically masked frame features, and its outputs are fused to combine local limb information with whole-body context. Conditioned on text and motion history, the model generates motion primitives autoregressively. Forward-kinematics supervision encourages kinematic consistency, while a learned end-of-sequence predictor selects a frame-level endpoint after generation. On HumanML3D retargeted to the Unitree G1, PartiGen achieves a segment-level FID of 3.165, a 16.8% reduction over TextOp, and improves R@1 by 1.4 percentage points. In the duration study with 16-frame primitives, endpoint prediction reduces mean absolute duration error from 12.471 to 3.080 frames compared with a fixed-length cutoff. We further evaluate
robot execution using a pretrained, frozen GRIT tracker: 409 of 500 generated references (81.8%) pass the posture-based stability screen in MuJoCo. Following simulation screening and manual selection, 26 of 27 attempted motions are judged successful on the real G1.
Read the full abstract ▾
All authors Chuxuan Huang, Zhanxiang Cao, Tianyue Zhan, Haoyang Li, Yang Zhang, Liyun Yan, Chenlin Ming, Keqi Zhu, Yizhi Chen, Shengcheng Fu, Zihao Zhou, Xinyi Peng, Kaixuan Zhang, Boyang Xing, Jiawei Bao, Yang Yang, Yue Gao
Cryptography and Security
Towards Safe World Models: Risks, Defenses, and Evaluation
Xiaoma Wang, Litian Zhang, Xinze Shi et al. (+11 more)
World models support planning, control, learning from imagined experience, and model-mediated evaluation by predicting action-conditioned consequences. Their safety cannot be inferred from predictive accuracy alone: errors in observations, state and
memory, dynamics, evaluation, execution, or feedback can propagate into unsafe decisions and persist through updates. This work organizes the literature on world-model safety around a six-stage lifecycle covering data and observations, state and
memory, dynamics and imagination, evaluation and planning, action and control, and feedback and update. We synthesize non-
adversarial failures, attacks, defenses, evaluation protocols,
benchmarks, and deployment conditions, and distinguish evidence directness from outcome level. The current evidence corpus contains 163 cited works. We use the framework to identify where claims are supported, where attribution remains unresolved, and which measurements are needed to connect model-level behavior to closed-loop and operational safety.
Read the full abstract ▾
All authors Xiaoma Wang, Litian Zhang, Xinze Shi, Zhengxuan Hu, Zihan Yang, Mingqi Ren, You Wu, Bingyu Yan, Zelong Yu, Zhenjie Su, Yuming Shang, Xi Zhang, Chaozhuo Li, Philip S. Yu
Computation and Language
Overcoming Data Imbalances with Expert Selection Biases
Ayush K. Tarun, Clara Meister and Antoine Bosselut
Mixture-of-Experts (MoE) models have become the standard architecture for frontier LLMs. The multiple parallel experts in each
transformer block theoretically enable expert specialization according to categories such as domains,
languages, or modalities during pretraining. However, when data is imbalanced across categories, minority categories can become concentrated in a small fraction of the expert pool or spread inefficiently across them, effectively limiting their available capacity. In this work, we demonstrate that category-conditioned expert-selection biases can preferentially route
tokens from different categories toward specialized expert pools. By applying biases based on input category, we provide a simple way to induce soft category-aligned specialization. This intervention encourages categories toward distinct expert subsets while retaining cross-category sharing when useful. More importantly, it enables the over-allocation of expert capacity to categories that would otherwise be disadvantaged by data scarcity. We observe performance improvements across modalities, low-resource
languages, and code in models up to 10.6B parameters. At our smallest and largest model scales, these gains are equivalent to allocating 18.5% and 2.7% additional baseline training
tokens, respectively. Overall, our results show that expert biases can serve as a lightweight control for category-level expert allocation, making room for low-resource categories while retaining standard MoE inference.
Read the full abstract ▾
Machine Learning
Opening the Black Box of Classifier-Free Guidance via Information Bottleneck
Jiayang Gao, Tianyi Zheng, Jiayang Zou et al. (+5 more)
Classifier-free guidance (CFG) is a standard technique for improving conditional generation in
diffusion models, yet its activation schedule is usually selected heuristically. In practice, most existing CFG-based methods rely on a pre-defined activation policy throughout inference, while recent dynamic variants often depend on hand-designed schedules without a principled theoretical explanation of when conditional information is most useful. In order to understand how the informational relevance between conditioning inputs and the score function evolves over the course of generation, we explore when guidance should be activated from an information-theoretic perspective. In this work, we formulate the
diffusion trajectory as a time-dependent information bottleneck and show that the effectiveness of conditional information is not uniform throughout sampling, but becomes significantly stronger after a critical stage characterized by a Fisher-information balance condition. Motivated by this insight, we propose a training-free dynamic CFG strategy for inference. Rather than prescribing a fixed schedule in advance, our method constructs a
stochastic process from conditional and unconditional score statistics and activates guidance when the estimated guidance gain crosses a prescribed threshold. This yields an adaptive activation rule that aligns condition injection with the underlying generation dynamics, while also recovering interval-based guidance as a limiting special case. Extensive experiments across diverse architectures and tasks, demonstrate that our method is effective, robust, and broadly applicable. Beyond empirical improvements, our framework provides an interpretation of dynamic guidance methods and a principled foundation for inference-time control in conditional
diffusion models.
Read the full abstract ▾
All authors Jiayang Gao, Tianyi Zheng, Jiayang Zou, Fengxiang Yang, Luyao Fan, Lv Tang, Bo Li, Jia Wang
Hardware Architecture
Anechoic: Onsager-Corrected Parallel Annealing for Sub-0.1 ms Dense Max-Cut on an FPGA
Seungki Hong, Kyeongwon Jeong and Taekwang Jang
Ising machines solve combinatorial optimization problems by annealing a network of
spins. On dense problems, the least expensive hardware step updates all
spins at once:
stochastic cellular automata (SCA) decide every
spin in parallel and read couplings only for the
spins that flip. Such synchronous updates create an echo: the previous state of each
spin returns to it through its neighbors. On K2000, a Max-Cut
benchmark on a 2,000-node complete graph, two-thirds of all flips undo the same
spin's flip one step earlier. This paper presents Anechoic, an FPGA Ising machine that subtracts this echo from every decision. Dynamical mean-field theory identifies the echo as an Onsager reaction; its one-step strength is the number of
spins whose flip probability is strictly between 0 and 1, divided by twice the temperature. The decision logic already flags these
spins, so the correction costs one population count per step; a temperature-scaled form needs none. An RTL engine executes a dense 2,000-
spin step in 18.1 cycles plus 0.14 per flip, so every saved step and flip saves time. Both forms find better K2000 cuts at every step budget than five published binary-
spin algorithms at their papers' settings. Twelve engines at 250 MHz on an AMD Alveo V80 reach a cut of at least 33,000 with 99% probability in 0.05 ms, 5.2 times faster than the fastest published hardware, at 5.4 mJ per solution; with 2-bit couplings, the temperature-scaled form is also faster than uncorrected SCA on 49 of 51 G-set graphs. On an NVIDIA GH200 GPU, the same engine is 2.3 times slower and needs 8.0 times more energy per solution, though its 240 concurrent anneals give 2.3 times the throughput. As a GlobalFoundries 22 nm ASIC at 1.25 GHz, twelve engines would reach the target in about 10 microseconds, 26 times faster than the fastest published hardware. The code is available at https://github.com/skethz/anechoic.
Read the full abstract ▾
Cryptography and Security
Secrets in Preamble: Practical and Robust PHY Information Hiding for Wi-Fi
Rui Pan, Guanxiong Shen, Linning Peng et al. (+2 more)
Physical layer (PHY) information hiding refers to the concealment of secrets within PHY waveforms without compromising the integrity of the primary communication link, thereby enabling critical applications such as covert communication, PHY steganography, and watermarking. The widespread deployment of Wi-Fi has motivated substantial research interest in hiding secrets within Wi-Fi waveforms. However, existing Wi-Fi information hiding approaches suffer from two limitations: restricted steganographic capacity and insufficient robustness against environmental variations, both of which severely constrain practical deployment. This work proposes to embed secrets within the Wi-Fi legacy short training field (L-STF). The designed method preserves the autocorrelation properties of Wi-Fi L-STF, ensuring that packet detection and coarse frequency offset compensation functionalities remain uncompromised. The secret
embedding and extraction processes employ a pair of neural networks designed to learn optimal
embedding rules and maximize steganographic capacity. We developed a prototype system and conducted comprehensive experiments across both indoor and outdoor environments. Experimental results show that our method can reliably embed 46 secret bits within Wi-Fi packets across diverse channel conditions, achieving significantly enhanced steganographic capacity and environmental robustness. The hardware prototype study validates the feasibility of the proposed Wi-Fi information hiding scheme.
Read the full abstract ▾
All authors Rui Pan, Guanxiong Shen, Linning Peng, Xinyu Qi, Junqing Zhang
Distributed Computing
Stacked Intelligent Metasurface-Based MultiModal Feature Fusion for Collaborative Inference in Edge Networks
Yushi Wang, Zhengxin Yu, Guhan Zheng et al. (+3 more)
In navigating the challenges of computational and communication overhead within the 6G-era Collaborative Inference (CI), this paper explores the feasibility of employing Stacked Intelligent Metasurface (SIM) to perform CI, termed SIM-CI. The SIM functions as a physical computing unit, receiving multi-user Over-The-Air (OTA) aggregated features within the same spectrum. It thus executes the necessary spatial-frequency combining and inference in a single physical wave propagation.SIM-CI also incorporates a tri-heterogeneity robustness adapter to address the network heterogeneity, such as multi-modal data and user inference models, as well as a quality-aware contrastive approach that align features from different modalities and extraction levels into a unified feature space. Considering the heterogeneous and dynamic channels, we further integrate Channel State Information (CSI) as an additional input into the adapter design, enabling users' early perception of potential channel states. Moreover, to maximally mitigate the impact of channel variations, we design a new hybrid beamforming network that combines the robustness of model-driven pre-equalization with the adaptability of data-driven learning. Experimental results demonstrate the effectiveness of our SIM-CI compared to digital-domain inference, achieving 98.6% accuracy to conversional digital-domain CI approaches under various heterogeneous conditions, while reducing computational energy by 25% and requiring 75% fewer transmission time slots.
Read the full abstract ▾
All authors Yushi Wang, Zhengxin Yu, Guhan Zheng, Haris Pervaiz, Zheyi Chen, Shunqing Zhang
Computers and Society
Adversarial Attacks Reveal Divergent Deception Competence in Humans and Generative Language Models
Lucca Pfründer, Riccardo Loconte, Bruno Verschuere et al. (+1 more)
Machine learning models typically outperform humans in deception detection. A threat to machine learning models is their vulnerability to
adversarial attacks: subtle yet targeted perturbations of input data that trick a model into misclassification. Yet, this also provides an opportunity: understanding how humans versus machines construct attacks offers new insights into second-order beliefs about deception. In a preregistered experiment (n = 501), we compare the effectiveness, efficiency, and nature of a paraphrasing attack by humans and generative
language models on an automated deception classifier. We find that humans were more effective attackers when rewriting truthful statements to be deceptive, while the generative
language model was more effective and efficient at making lies sound like truths. Computational text analysis suggests that this is because humans make targeted use of deception cues, while generative
language models, instead, apply generic changes that signal truth. The failure to mimic deception is two-sided: while it exposes a blind spot in the machines increasingly used to detect lies, it also means that
language models might facilitate successful lying.
Read the full abstract ▾
All authors Lucca Pfründer, Riccardo Loconte, Bruno Verschuere, Bennett Kleinberg
Software Engineering
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
Jiahao Wang, Hongjin Qian, Yuyang Hu et al. (+2 more)
As the production of software passes into the hands of software itself, coding
agents are increasingly commissioned to build
agents fitted to their users' own workflows. This delegation carries a quiet assumption, that the delivered
agent will faithfully serve the requirement it was built for, rather than merely appear to. No existing
benchmark tests that assumption. We introduce AgentSWE, a
benchmark for
agent software engineering spanning three stages of the
agent development lifecycle: Creation of a complete
agent from a natural-
language requirement, Editing of production
agent codebases, and Optimization toward a measurable target. Each of its 25 tasks is framed as a commission rather than an exam: the builder iterates against a handful of illustrative cases, while acceptance exercises the frozen submission on held-out cases designed to span the full requirement, grading behavior against requirement-derived rubrics under hard pass/fail gates or, for Optimization, against the source
benchmark's own objective. Evaluating a range of coding
agents across models and harnesses, we find that none comes close to delivering what a requirement asks for: the gap between what a builder demonstrates and what it delivers is the norm, not the corner case. An accompanying empirical analysis of the failure trajectories identifies recurring failure modes, among them overfitting to the visible cases, hollow runtimes behind code that looks complete, and unsupported claims of completion, each a signal by which coding
agents can learn to build the
agents they are asked for.
Read the full abstract ▾
All authors Jiahao Wang, Hongjin Qian, Yuyang Hu, Jiajun Zhang, Zheng Liu
Cryptography and Security · Cross-listed from eess.SP
ECGWmark: Deep Watermarking for Electrocardiogram Signals
Kuan-Yu Chen, Kuan-Chen Wang, Yu Tsao et al. (+1 more)
Secure transmission of electrocardiogram (ECG) signals remains a major challenge in telecardiology. While traditional mathematical watermarking can embed patient metadata directly into ECG signals, it is highly fragile against temporal desynchronization such as circular phase shifts. In contrast, existing deep watermarking approaches for temporal signals, such as audio, typically rely on lossy quantization, which compromises the high-frequency micro-morphological details critical for clinical diagnosis. To bridge this gap, we propose ECGWmark, a fully convolutional and quantization-free deep watermarking framework for medical time-series. ECGWmark operates in a continuous, full-resolution latent space and utilizes a localized frame-level mechanism to improve the robustness to severe desynchronization. Crucially, we introduce a differentiable Physiological Loss guided by a clinical classifier to preserve diagnostic semantics. Experiments on the
benchmark MIT-BIH database show that ECGWmark maintains channel-bit extraction accuracy above 90% under the evaluated circular temporal shifts. Furthermore, under a rigorous inter-patient diagnostic evaluation, Class-V sensitivity changes from 71.97% to 76.50%, whereas Macro-F1 changes from 35.25% to 34.71% and Class-S sensitivity from 8.89% to 1.39%; these results indicate mixed class-level effects.
Read the full abstract ▾
All authors Kuan-Yu Chen, Kuan-Chen Wang, Yu Tsao, Jian-Jiun Ding
Artificial Intelligence
SWE-Critix: Post-Training Generalist Verifiers for Assessing Coding-Agent Trajectories
Jiajun Hu and Chun Yong Chong
Coding
agents are increasingly deployed for end-to-end software engineering tasks, where a readily available oracle evaluator that defines the success criteria is often absent. This motivates verifications for coding
agents that can reliably judge and explain their behavior. However, verification is non-trivial as coding
agents are increasingly capable of solving long-horizon software engineering tasks. Their execution trajectories can be extremely long, containing hundreds of iterative interactions with code repositories and external tools. Existing verifiers either rely on general-purpose LLMs without specialized training or train dedicated critics that predominantly focus on producing scalar rewards, highlighting the relatively limited exploration of training-based generalist verification. We introduce SWE-Critix, a post-training pipeline for training generalist verifiers of coding-
agent trajectories. Given a task and its execution trajectory, the learned verifier predicts the completion label and generates a natural-
language analysis that explains its judgment. Our pipeline combines supervised fine-tuning (SFT) with reinforcement learning (RL) and trains a series of 30B models on trajectories from CodeForge-Preview. On a held-out test set, SWE-Critix improves classification accuracy from 57% for the base Qwen3-30B-A3B
reasoning model to 66% after SFT and 78% after RL. We release our models,
datasets, and training code to facilitate further research.
Read the full abstract ▾
Numerical Analysis
On the Convergence of the Mesh-Free Discretization of a Non-Linear Bond-Based Peridynamic Model with a Bond-Breakage Criterion
Lukas Pflug, Michael Stingl and Max Zetzmann
We show the convergence of a commonly used mesh-free discretization of a dynamic non-linear bond-based peridynamic model with irreversible damage implemented via a bond-breakage criterion for a fixed horizon $\delta >0$. The model chosen here is a mathematically well-posed regularization of a discontinuous peridynamic operator often found in the literature, which assumes that fracture results from strain in tension only. We explicitly include the one-point quadrature used in this scheme in the analysis. The convergence is with respect to the natural norm of the solution space given by $C^1([0,T],L^\infty(\Omega)^d)$, and therefore $L^\infty(\Omega)^d$ in the spatial coordinates. This result heavily constrains the spatial regularity of the continuous solution which we prove afterwards. More specifically, for locally $L$-Lipschitz continuous input data, local Lipschitz bounds are derived for the displacement, regardless of damage. This is consistent with interpreting the peridynamic model used here as an approximation to a truly discontinuous fracture model with a length-scale parameter given by the horizon $\delta>0$. We note that we do not deal with the limit of the solutions as $\delta \searrow 0$.
Read the full abstract ▾
Artificial Intelligence
AlphaNIST: Emergent Assembly Skills from Binary Task Rewards
Octi Zhang, Iretiayo Akinola, Rosario Scalise et al. (+12 more)
The NIST Assembly Task Board #1 remains a challenging
benchmark for learning-based methods in robotic manipulation. Over short horizons, it demands precise, contact-rich control of small, tight-fitting parts. Over long horizons, it requires recovery and coordination across extended action sequences. Demonstrations, task decomposition, and dense progress rewards have enabled research advances by providing guidance on how to assemble. But how much of the strategy must be specified? The progression from AlphaGo to AlphaGo Zero showed that RL at scale can discover superhuman strategies without prescribing how to win. AlphaNIST asks whether a sparse binary task reward and task diversity admitted by this task-agnostic reward can similarly let a
robot discover how to assemble. In simulation, we train a single policy from scratch across all 20 assembly tasks on the NIST Assembly Task Board #1, reaching 98.2% overall training success without demonstrations, progress shaping, scripted primitives, or per-task tuning. The framework also trains policies to complete ten assembly pairs in a single episode without a prescribed order, with task success rewarded only when all ten pairs are simultaneously complete. The learned policies discover dexterous and planning-like behaviors: choosing grasps for later assembly, using surrounding geometry to reorient parts, and preparing one part, assembling another, then returning to finish the first without an explicit planner. For
videos, see our website: https://alpha-nist.github.io/
Read the full abstract ▾
All authors Octi Zhang, Iretiayo Akinola, Rosario Scalise, Mateo Guaman Castro, Patrick Yin, Tobias Widmer, Miles Macklin, Chris Amevor, Dylan Turpin, Dirk Van Gelder, Jan Carius, Ankur Handa, Gavriel State, Yunrong Guo, Yashraj Narang
Multimedia
EmoGavel: Grounding-Aware Reinforcement Learning with Verifiable Rewards for Multimodal Emotion Preference Learning
Zilong Huang, Kong Aik Lee, Zezhong Jin et al. (+2 more)
Multimodal emotion preference learning uses pairwise comparisons to identify which open-ended emotion descriptions align more closely with human judgments. However, existing methods primarily focus on the final preference decision and provide little explicit supervision for localizing and correcting erroneous content. To address this limitation, we propose EmoGavel, a Grounding-Aware reinforcement learning framework with Verifiable rewards for
multimodal Emotion preference Learning. We first introduce grounded supervised fine-tuning (Grounded SFT), which exploits controlled-error construction information to jointly learn preference prediction, error localization, repair, and missing-evidence recovery. We then optimize these verifiable outcomes with grounding-aware group relative policy optimization (GA-GRPO), combining preference-gated grounding rewards with dynamic zero-variance rescue to prioritize preference correctness and recover learning signals from all-wrong groups. Experiments on EmoPrefer and controlled-error evaluation sets show that EmoGavel improves both natural preference prediction and controlled-error discrimination while maintaining strong error-grounding performance. Our code and model checkpoints will be released once the paper is accepted.
Read the full abstract ▾
All authors Zilong Huang, Kong Aik Lee, Zezhong Jin, Man-Wai Mak, Tatsuya Kawahara
Social and Information Networks
Polarization Dynamics under Algorithmic Recommendations in LLM-based Online Social Networks
Ali Safarpoor Dehkordi, Mohammad Shirzadi, Mehwish Nasim et al. (+1 more)
How do recommendation
algorithms contribute to the formation of polarized opinions in online social networks? This question has been studied through empirical analyses of real platforms and mathematical models of opinion dynamics. Empirical studies capture user behavior but are often limited by noisy, incomplete, or inaccessible platform data, while mathematical models offer control and interpretability but usually rely on simplified opinion update rules. More recently, large
language model (LLM)-based social simulations have provided an alternative for studying recommendation and opinion dynamics through natural-
language interactions. However, how different ways of applying recommendations shape opinion dynamics remains underexplored. In this work, we provide a controlled comparison of two different recommendation mechanisms in an LLM-based online social network simulation. Edge Recommendation (ER) introduces new connections in the network, while Content-Exposure Recommendation (CR) exposes additional users to individual posts without modifying the underlying graph. Our framework represents users as LLM-based
agents who interact in natural
language while maintaining evolving opinion states. We measure polarization as opinion dispersion, extremization as the average distance of opinions from the neutral midpoint, and reinforcement as interactions that amplify already aligned opinions. Our results show higher mean polarization and extremization across recommendation conditions compared with the no-recommendation baseline, with absolute differences of up to $0.12$ in our main experiments.
Read the full abstract ▾
All authors Ali Safarpoor Dehkordi, Mohammad Shirzadi, Mehwish Nasim, Ahad N. Zehmakan
Machine Learning
Hidden Data Challenges in Clinical Machine Learning
Carolin Cissée, Dennis Fast, Raneen Younis et al. (+3 more)
Widely reused preprocessing pipelines for electronic health records can encode assumptions about when information became available, what clinical episode a label represents, and what missingness means. These assumptions are rarely audited systematically and may propagate across subsequent studies when preprocessing pipelines are reused or adapted. Using MIMIC-IV as a case study, we identify and quantify violations of three complementary forms of preprocessing
alignment, i.e., temporal, semantic, and population
alignment, which manifest as temporal leakage, feature-task misalignment, and biased imputation. We show how ambiguities in timestamp semantics, episode definitions, endpoint construction, modality selection, and missingness handling can alter the information available to a model and affect the interpretation of its predictions. Through analyses of MIMIC-IV data and empirical case studies, we demonstrate that such choices may remain largely invisible in aggregate predictive performance while changing the signals on which models rely. We derive practical recommendations for temporally valid, clinically aligned, and reproducible preprocessing of MIMIC-IV and related electronic health record
datasets. The code for our experiments, including our proposed preprocessing, is available at: https://github.com/m2l-lab/mimic-iv-preprocess.
Read the full abstract ▾
All authors Carolin Cissée, Dennis Fast, Raneen Younis, Alexander Löser, Wolfgang Nejdl, Zahra Ahmadi
Computation and Language
Test-Time Memory Construction
Zijun Wang, Yifan Ji, Haidong Xin et al. (+6 more)
Large
Language Models (LLMs) struggle to answer questions over long interaction histories, where relevant evidence may be scattered across distant sessions. Existing
memory systems typically construct and maintain sophisticated
memory representations offline, before user queries are observed, incurring substantial overhead and limiting their ability to adapt
memory construction to query-specific information needs. In this paper, we propose T$^2$Mem, a Test-Time
Memory construction framework that defers
memory construction to inference and progressively builds
memory along the
reasoning process. Specifically, T$^2$Mem extracts lightweight episode-level metadata offline to capture key events and temporal associations as
retrieval cues, while retaining links to the original interactions to facilitate the reconstruction of dependencies across episodes. At test time, the LLM adaptively retrieves and organizes metadata and raw interactions, progressively constructing query-specific
memory as
reasoning unfolds. Experiments on LoCoMo and LongMemEval show the effectiveness of T$^2$Mem by consistently outperforming strong baselines across different backbone LLMs while substantially reducing the overhead of
memory construction. Further analysis demonstrates that deferring
memory construction to inference enables the
memory construction process to dynamically adapt to the information needs of individual queries. All code is available at https://github.com/OpenBMB/T2Mem.
Read the full abstract ▾
All authors Zijun Wang, Yifan Ji, Haidong Xin, Zhenghao Liu, Yukun Yan, Zulong Chen, Yu Gu, Ge Yu, Maosong Sun
Robotics
Evo-RL: Iterative Robot Reinforcement Learning from Selective Experience
Kai Ye, Yifan Jin, Qianhang Luo et al. (+7 more)
Existing real-world reinforcement learning methods improve
robot policies by combining human corrections with autonomous rollouts. However, accumulated experience varies in its value for policy improvement, and even successful trajectories can contain inefficient or redundant behavior, motivating more selective experience reuse. To this end, we propose Evo-RL, an iterative reinforcement learning method that combines human interventions with advantage-guided experience selection. Evo-RL uses human corrections and experience with positive estimated advantages from successful trajectories to iteratively refine the policy. To support learning from these different experience sources, we introduce RW-RL-
Dataset, which spans over 1,000 hours across four
robot embodiments. On a real-world high-precision manipulation task, Evo-RL improves attempt success rate over RECAP by 30.46%. Our code, model checkpoints, and
dataset are publicly available.
Read the full abstract ▾
All authors Kai Ye, Yifan Jin, Qianhang Luo, Hongru Zhang, Jiedong Zhang, Shuyuan Hu, Song-Yuan Wu, Jihai Xu, Jie Zhao, Bo Zhao
Computer Vision
Why DiT Models Underperform as Representation Learners without Long Skip Connections
Benyuan Meng, Yiliang Zhang, Jin-Wen Wu et al. (+3 more)
Diffusion feature, i.e., intermediate activations extracted from
diffusion model backbones, has emerged as a promising approach for unifying generation and
vision understanding (discrimination). Despite the advancement of
diffusion backbones into
Diffusion Transformers (DiT), it has been found in previous literature that DiT models tend to yield weaker features despite their stronger generation capability. This poses a threat to good generation-understanding unification. In this work, we show via controlled from-scratch training that long skip connections (LSCs) affect feature quality: a DiT with LSCs can produce better features. Furthermore, through detailed diagnosis, we reveal the mechanism causing this effect. Specifically, the input to a
diffusion backbone contains both the latent image and the
diffusion noise, where the
diffusion noise, though appearing in the form of noises, actually reflects the information of the
diffusion trajectory. To make the final prediction of a noisy variable ($v$ or $\epsilon$), the
diffusion backbone needs both signals. However, carrying both signals through the main backbone effectively "pollutes" the image signal with noises, leading to poor feature quality. In comparison, LSCs can shortcut
diffusion noises directly to where they are needed, leaving much weaker noise signals in the main backbone, thus improving feature quality. To name the broader impact, this reveals that LSCs are not a mere trick but closely related to the
diffusion objective. The code is at https://github.com/Darkbblue/dit-lsc.
Read the full abstract ▾
All authors Benyuan Meng, Yiliang Zhang, Jin-Wen Wu, Qianqian Xu, Qingming Huang, Longtao Huang
Machine Learning
Understanding Mamba-3 Through Oscillatory Dynamics
Shutian Liang and Yue Song
Mamba-3 introduces input-dependent rotations into recurrent state updates, but their contribution beyond state retention remains unclear. We interpret these rotations as underdamped second-order dynamics, providing a framework for investigating the role of oscillation in sequence modeling. Controlled comparisons show that underdamped dynamics outperform non-oscillatory damping regimes on state-tracking tasks, while
language-model pretraining demonstrates benefits for long-context prediction and length generalization. Interventions in pretrained models further establish the importance of learned phase, and state analysis suggests that rotation regulates how retained information combines with incoming updates. Taken together, these findings identify underdamped dynamics as a useful mechanism for recurrent computation and suggest that oscillation supports long-context processing by regulating how information accumulates.
Read the full abstract ▾
Computers and Society
Who Can Say No? Control, Accountability and the Pace of Frontier AI
Rikard Rosenbacke, Carl Rosenbacke, Victor Rosenbacke et al. (+1 more)
In September 2026, the leaders of the largest frontier AI laboratories said in public that the pace of capability improvement is too fast and that none of them can slow down alone, and
agents under evaluation at OpenAI broke out of their test environment and reached a third party's live systems, which a UN-mandated panel called an early warning of one route to loss of human control. This paper examines the institutional side of that risk: not the machine, but the person who must control it. We argue that control is not taken by AI but given away, one delegation at a time, by organisations arranged so that nobody is positioned to say no, and we define control as a person who can detect a serious problem, challenge the decision, stop the system and reverse it, and who has something to lose by not doing so. Three questions follow, which any builder, funder, regulator or deployer can answer in public: (1) who inside the organisation can say no, and who answers for saying yes; (2) who outside the organisation checks the decision, and what can they do when they disagree; (3) what happens when AI builds AI faster than anyone can check it. A checkable answer needs a named person, a date, a number, and a stop that has been used or tested. Three measures meet that test: a named person on every consequential decision, including the standing instructions AI
agents carry out; a supervisor with a stop authority; and two published clocks, time to build against time to check. We show that governance theory arrives at the same conditions, and code fifty older mechanisms against the questions, finding that almost every one began as one rule at one point, taken by one party without waiting for the rest. That is the paper's design: a start, not a solution. Publication is the first step and one laboratory can take it alone; the contract, the independent body and the law come after.
Read the full abstract ▾
All authors Rikard Rosenbacke, Carl Rosenbacke, Victor Rosenbacke, Martin McKee
Computer Vision
VideoAWR: Reinforcing Video Flow Models with Decoupled-Timestep Advantage-Weighted Rectification
Zheming Fu, Wenbin Ning, Ruizhe He et al. (+6 more)
Regression-based RL for flow models mandates sampling without classifier-free guidance (CFG) to align the rectification target with the reference marginal velocity. Extending Advantage-Weighted Rectification (AWR), which dynamically scales this regression target based on reward, to
video generation encounters a cold-start pathology: unguided rollouts collapse into low-fidelity regions where intra-group reward gaps remain wide but no longer track perceptual quality, rendering the optimization target arbitrary. Conversely, utilizing guided rollouts to restore fidelity, while keeping the regression target tied to the guided sampling distribution, provably triggers a
geometric guidance blow-up under exponential-moving-average (EMA) updates. We propose VideoAWR, which resolves this tension by untying sampling from the optimization anchor. An EMA reference policy generates guided rollouts to restore reward separability, while the trainable policy is rectified against the unguided conditional anchor. We prove this decoupled objective is mathematically exact: the centered advantage cancels the anchor in expectation, filtering out the reward-neutral guidance offset and retaining only the reward-correlated covariance increment. Consequently, inference runs single-branch and CFG-free, yet inference-time CFG preserves its original semantics. Furthermore, we introduce Decoupled Timestep Rectification, supervising each rollout at an independently sampled timestep without bias to improve sample efficiency. Experiments on Wan2.1-1.3B and 14B demonstrate stable optimization and superior visual and motion quality in both CFG-free and CFG-enabled regimes. Code is available at \pdfstartlink attr{/Border [0 0 0]} user{/Subtype /Link /A << /S /URI /URI (https://github.com/Fuzm7/FlowAWR) >>}% \textcolor{pearDark}{VideoAWR}\pdfendlink.
Read the full abstract ▾
All authors Zheming Fu, Wenbin Ning, Ruizhe He, Wei Shang, Xiaoxiao Ma, Lianrui Mu, Chang Liu, Lei Wang, Siming Fu
Computer Vision
Exploring Attention Sinks in Text-to-Image Transformers from Diffusion and Autoregressive Models
Binglei Li, Mengping Yang, Zhiyu Tan et al. (+3 more)
Attention sinks are
tokens or positions that consistently receive disproportionate
attention mass, often beyond what their semantic content would suggest. Despite extensive study in large
language models, their behavior and functional relevance in text-to-image (T2I)
Transformers remain largely unexplored. To fill this gap and better understand how text-to-image
Transformers allocate and route
attention, we explore this phenomenon of text-to-image
transformers across bidirectional
diffusion models and causal autoregressive models. Specifically, we explore whether
attention sinks exist, where they emerge, and how they functionally influence generation. Through systematic
attention analysis and targeted interventions, we identify that
attention sinks emerge in both bidirectional
diffusion and autoregressive models but exhibit distinct patterns. The autoregressive model concentrates
attention on the first and global-scale
tokens, whereas the bidirectional models develop sinks near modality boundaries, within text regions, and at spatial corners. Moreover, these patterns vary across generation stages and are localized to a sparse subset of layers and heads. To assess their functional relevance, we perturb sink-related heads and
token positions. It turns out that perturbing sink-ranked heads produces larger average changes than non-sink comparisons, while some sink-position perturbations have limited effects on bidirectional model scores. We further introduce register
tokens to redistribute
attention sinks and find that several configurations preserve or modestly improve generation performance. Overall,
attention sinks are architecture-, inference-stage-, and component-dependent phenomena, providing a basis for mechanistic analysis and principled
attention design in text-to-image
transformers.
Read the full abstract ▾
All authors Binglei Li, Mengping Yang, Zhiyu Tan, Wenxu Qian, Junping Zhang, Hao Li
Machine Learning
AgentTune: A Unified Training Lifecycle for Tool-Using LLM Agents
R E Zera Marveen Lyngkhoi, Abhivansh Gupta, Vidushee Vats et al. (+3 more)
Tool-using
agents increasingly require a common substrate for building, training, evaluating, distilling, and maintaining agentic systems, yet existing tooling typically addresses these stages in isolation. We present AgentTune, a source-available Python library that provides a unified framework for developing and optimizing tool-using
agents. At its core, AgentTune introduces a two-tier EventLog representation that preserves
agent trajectories while supporting consistent projections for training, evaluation, and auditing. Building on this substrate, the library provides a shared
agent runtime, tool-calling reinforcement learning, agentic
distillation,
retrieval-augmented generation, decision workflows, evaluation utilities, and a designed path toward self-healing
agents. These components are exposed through composable interfaces so that
agent strategies, training methods, evaluation procedures, and recovery mechanisms can be combined without requiring separate representations or pipelines. We demonstrate the framework across agentic tasks including legal contract search and domain-specific applications, illustrating its use for trajectory-based optimization, tool-use evaluation, and cross-task
agent development. AgentTune is released as a source-available library at https://github.com/Lexsi-Labs/AgentTune
Read the full abstract ▾
All authors R E Zera Marveen Lyngkhoi, Abhivansh Gupta, Vidushee Vats, Ram Mohan Rao Kadiyala, Vinay Kumar Sankarapu, Pratinav Seth
Machine Learning
PDE-PFN: Prior-Data Fitted Neural PDE Solver
Jaehyeon Park, Mingu Kang, Dongseok Lee et al. (+5 more)
Motivated by the success of large
language models (LLMs) with broad generalizability and robustness to noisy or unreliable pre-training data, we explore in-context prediction of PDE solution fields using approximate pre-training data. In addition, inspired by the posterior predictive mean-based inference mechanism of the in-context learning in prior-data fitted networks (PFNs), we propose PDE-PFN, a prior-data fitted neural solver that approximates the posterior predictive mean of PDE solutions via in-context learning. PDE-PFN builds on a PFN architecture with self- and cross-
attention mechanisms of
Transformer and is pre-trained on noisy approximate solutions generated by physics-informed neural networks, serving as diverse but not necessarily exact priors. Through experiments on a range of two-dimensional PDEs, we demonstrate that PDE-PFN achieves empirical generalization across heterogeneous equations, robustness under noisy priors, and zero-gradient-update in-context inference capability. PDE-PFN achieves competitive or improved predictive accuracy relative to the evaluated baselines while supporting flexible context and query representations.
Read the full abstract ▾
All authors Jaehyeon Park, Mingu Kang, Dongseok Lee, Woojin Cho, Kookjin Lee, Anthony Gruber, Youngjoon Hong, Noseong Park
Computer Vision
LETHE: Defending against Unauthorized Image-to-Video Generation via Value Diversity Collapse
Boheng Li, Yueming Xing, Yiming Li et al. (+2 more)
Image-to-
video (I2V)
diffusion models can animate a single reference image into high-quality, subject-consistent
videos, posing threats of misuse as anyone can turn personal photos into fabricated
videos without consent. A promising solution is to proactively add an
adversarial perturbation to the image before release, such that
videos generated from it would be unfaithful or visually corrupted. However, due to an insufficient understanding of what governs how subsequent frames read the reference image in modern I2V models, existing protections struggle to offer strong protection under common perturbation budgets. In this paper, we approach the problem by exploring this underlying reading mechanism. Through a convex
geometric analysis, we theoretically identify the spread of the reference value vectors as a key factor governing this reading process, and empirically find that even a mild contraction of this spread can severely degrade the generated
video, making it an efficient lever for protection. Guided by this insight, we propose LETHE, a novel protection method that directly disrupts this reading process. LETHE pairs two complementary objectives, where a value collapse objective drives the reference value vectors toward uniformity, while an
attention attraction objective draws subsequent frames toward these collapsed
tokens to amplify corruption. We further construct two human-centric I2V protection
benchmarks and conduct extensive experiments with multiple modern I2V models, which demonstrate the state-of-the-art protection performance and versatility of LETHE.
Read the full abstract ▾
All authors Boheng Li, Yueming Xing, Yiming Li, Jie Zhang, Tianwei Zhang
Artificial Intelligence
Mengdi Liu, Chenghao Jia, Hong Chang et al. (+1 more)
Designing tool-integrated
reasoning agent systems for complex chemical tasks remains a fundamental challenge. Although recent chemical
agents have demonstrated promising performance through the integration of external tools, they are still built on the autoregressive paradigm, which is inherently constrained by causal
attention and left-to-right generation. As a result, they often struggle with bidirectional molecular understanding, global coordinated tool planning and multi-source evidence integration. To address this challenge, we introduce ChemDiffAgent, the first masked
diffusion language agent for tool-integrated chemical
reasoning. It reformulates chemical
reasoning as iterative denoising over interleaved
reasoning and tool-use trajectories, enabling tool-use decisions to be jointly determined under full-trajectory context and thereby better aligning the
agent's decision boundary with its knowledge boundary. We further analyze how this paradigm benefits agentic
reasoning through harder infilling subproblems during training and flexible arbitrary-order decoding during inference. Then we develop a two-stage post-training framework, comprising agentic supervised fine-tuning and variance-reduced preference optimization, to enhance tool-use and
reasoning capabilities. We also construct a comprehensive
benchmark covering six chemistry tasks across both single-turn and multi-turn tool-use scenarios. Under matched budgets and evaluation, ChemDiffAgent consistently outperforms autoregressive
agents, achieving an overall improvement of approximately 20%. Notably, our 8B model achieves performance comparable to, in some cases surpassing, significantly larger frontier models, such as GPT 5.5 and Claude Opus 4.7. Code is available at https://github.com/bilemond/ChemDiffAgent.
Read the full abstract ▾
All authors Mengdi Liu, Chenghao Jia, Hong Chang, Shiguang Shan
Machine Learning
DuoRate: Two-Rate Per-Expert Precision Allocation for Mixture-of-Experts Serving
Mengqi Guo, Zhe Wei, Yuan Yuan et al. (+3 more)
Serving a Mixture-of-Experts (MoE) model pays for expert precision twice: the resident footprint pays for every expert the router can reach, while each batch-one decode step pays only for the experts it selects. These are two different rate functionals over the same per-expert widths, yet published per-expert MoE allocators constrain only the first. We call the difference the store-read gap; in the released BitsMoE checkpoint, 2.54 bits per weight (bpw) of storage become 7.50 bpw of decode traffic. We propose DuoRate, which stores once and reads per workload. Offline, it stores one uniform tier of a prefix-readable weight stream and bakes a distortion tensor. Online, it profiles a workload's own prefill without labels and solves a Lagrangian rate-distortion allocation under the decode functional; routing frequency provably cancels from every per-expert decision, and the solve costs one dot product per expert and a bisection. A deployment therefore keeps a single artifact that prefill reads at the stored rate and decode reads at the workload's own rate. On Qwen3-30B-A3B, DuoRate reading 2.3 bpw exceeds every 2-bit-class baseline by 6.3 to 47.0 points of average accuracy, and at 2.9 bpw it matches 4-bit formats that read 1.4-1.6 bpw more. The advantage is largest under long decoding: on GSM8K, DuoRate at 2.9 bpw matches or exceeds every 4-bit format and trails BF16 by only 0.25 points, and at 2.6 bpw it still matches or exceeds the 4.5-bit NF4 and FP4. Against controls in its own format, the allocation alone is worth 0.07-0.11 bpw of read rate. Code will be released upon publication.
Read the full abstract ▾
All authors Mengqi Guo, Zhe Wei, Yuan Yuan, Jiunn Bin Lim, Boyi Pan, Michael Bi Mi
Distributed Computing
Taking Tokenization off the Host for Agentic LLM Serving
Mohammad Siavashi, Pascal Schwinte, Ali Banaei et al. (+3 more)
Agentic workloads resend long contexts that barely change between calls. Prefix caching now skips most of their prefill, yet every request must still be tokenized in full before the cache can even be consulted. Tokenization, a CPU stage long treated as negligible, becomes the largest part of time-to-first-
token (TTFT), and a frontend that tokenizes for a whole GPU fleet limits how many GPUs it can feed. We present GraphTok, a GPU tokenizer that runs the entire encoding pipeline, from pre-tokenization through byte-pair encoding (BPE) merging to
token-ID assembly, as one replayable CUDA graph. GraphTok reproduces the reference
token IDs bit-for-bit on GPT-2, Llama 3, Qwen 2.5, DeepSeek-V3, and Gemma 3, hands short inputs to a CPU engine where the GPU cannot win, and plugs into unmodified vLLM and NVIDIA Dynamo. GraphTok encodes one million GPT-2
tokens in 1.7-3.2 ms including data transfers, and on Llama 3 it is 4.7-6.0x faster than gigatoken, the strongest CPU tokenizer. On TokTier's own single-request
benchmark, it is 2.6-3.9x faster than gigatoken and up to 18.1x faster than TokTier. In vLLM with a cached prefix, it lowers mean TTFT by 2.1-2.4x. The largest gain is capacity. Within a 2 s P99 target, one Dynamo frontend with GraphTok sustains 518 requests/s of 121k-
token traffic, the busiest-hour load of 56 GPUs, against 296 with gigatoken and 18.5 with the default tokenizer. Across 121k to one million
tokens, this is 1.75-2.5x the load gigatoken sustains, with a P99 up to 64x lower where gigatoken's tail breaks.
Read the full abstract ▾
All authors Mohammad Siavashi, Pascal Schwinte, Ali Banaei, Gerald Q. Maguire Jr., Marco Chiesa, Dejan Kostić