MedMemoryBench is a new benchmark designed to evaluate memory mechanisms in personalized healthcare agents, addressing the shortcomings of existing open‑domain conversation benchmarks. It uses a human‑agent collaborative pipeline to generate realistic, long‑horizon medical trajectories from synthetic patient archetypes, producing a dataset of roughly 2,000 sessions and 16,000 interaction turns. The benchmark introduces an "evaluate‑while‑constructing" streaming assessment protocol that mirrors dynamic memory accumulation in production, and it formally studies memory saturation—how continuous information influx can degrade retrieval and reasoning. The results highlight significant bottlenecks in current architectures, especially in complex medical reasoning and noise resilience, underscoring the need for more robust, production‑ready medical agents.
By Yihao Wang, Haoran Xu, Renjie Gu, Yixuan Ye, Xinyi Chen, Xinyu Mu, Yuan Gao, Chunxiao Guo, Peng Wei, Jinjie Gu, Huan Li, Ke Chen, Lidan Shou
Large Reasoning Models often overthink, producing redundant reasoning steps that consume many tokens. The authors observe that efficient reasoning steps cluster in a concentrated region of latent space, while verbose steps lie outside this region. By projecting deviating hidden states back into the region via a quadratic program, they propose a training‑free framework that cuts token generation by up to 52.8% while improving accuracy by up to 12.1% across several benchmarks.
By Zhinan Hou, Xingchen Li, Keyou You
The paper introduces Humanity's Sixth Sense (HSS), a benchmark designed to evaluate intuitive visual reasoning in multimodal large language models. HSS covers a wide range of image and video inputs, organizes them under a structured taxonomy, and pairs each with human‑written prompts that probe implicit temporal, spatial, social, and abstract cues that humans can infer at a glance. Current state‑of‑the‑art models, including GPT‑6‑astra, achieve far lower accuracy than humans (53.6 % vs. 93.1 %), highlighting a significant gap in intuitive visual understanding.
By Xingang Guo, Jing Gu, Brian Jang, Renxiong Wang, Utkarsh Tyagi, Daniel Quigley, Steven Li, David Yan, Daniel Yue Zhang, Darvin Yi, Forrest Huang, HiJae Kim, Tianyi Zhang, Jared Lichtarge, Jihua Huang, Le Xue, Manan Tomar, Qiuyi Richard Zhang, Ruofei Yu, Seth Neel, Yaning Hu, Marcella Valentine, Xinzhe Jiang, Daniel Evans, Chenguang Wang, Dustin Tran, Tong Zhao, Yinfei Yang, Yunzhong He
The paper introduces GAP, a methodology that automatically generates mathematically equivalent variants of existing math problems using surface renames and kernel rewrites. Applying GAP to all 1,051 William Lowell Putnam Competition problems creates PutnamGAP, a 6,306-item dataset with 5,255 unseen variants. Evaluation of 18 commercial and open-source LLMs shows accuracy drops across all models, especially under kernel rewrites, indicating a persistent weakness in transferring proof plans to new mathematical settings.
By Yuren Hao, Xiang Wan, ChengXiang Zhai
The study investigates how exposing large language models to Arabic translations of English benchmark tasks can hide data contamination from standard English-only detection methods. By testing four instruction-tuned models with increasing Arabic exposure, the authors find that traditional probes like TS-Guessing and Min-K%++ fail to detect contamination, yet English performance still improves. They propose a new Translation-Aware Contamination Detection (TACD) method that uses cross-lingual consistency and choice reordering to reveal contamination effects that are otherwise invisible to English-only diagnostics.
By Chaymaa Abbas, Nour Shammaa, Mariette Awad
The paper "ActionCodec: What Makes for Good Action Tokenizers" addresses the lack of design guidance for action tokenization in Vision‑Language‑Action (VLA) models. By deriving information‑theoretic principles—maximizing temporal token overlap, reducing vocabulary redundancy, boosting multimodal mutual information, and ensuring token independence—the authors propose ActionCodec, a tokenizer that improves training efficiency and VLA performance. Experiments show that a SmolVLM2‑2.2B fine‑tuned with ActionCodec reaches a 95.5 % success rate on LIBERO, rising to 97.4 % with architectural tweaks, setting a new state‑of‑the‑art for VLA models without robotics pre‑training.
By Zibin Dong, Yicheng Liu, Shiduo Zhang, Baijun Ye, Yifu Yuan, Fei Ni, Jingjing Gong, Xipeng Qiu, Hang Zhao, Yinchuan Li, Jianye Hao
The paper proposes a heterogeneous spatial‑temporal graph neural network (HSTGNN) to create virtual smart heat meters for district heating networks, addressing sparse instrumentation and sensor faults. It integrates network functional relationships and learns graph structures and temporal dynamics for flow, temperature, and pressure, enabling joint modeling of cross‑variable and spatial correlations. A new laboratory dataset from the Aalborg Smart Water Infrastructure Laboratory is introduced, and experiments show the HSTGNN outperforms existing baselines.
By Keivan Faghih Niresi, Christian M{\o}ller Jensen, Carsten Skovmose Kalles{\o}e, Rafael Wisniewski, Olga Fink
APEX is a new image quality assessment framework that eliminates assumptions by using the Sliced Wasserstein Distance as a similarity measure. It is embedding‑agnostic, scalable to high‑dimensional spaces, and employs open‑vocabulary models CLIP and DINOv2 for feature extraction. Benchmarks show that APEX is more robust to visual degradations and remains stable across different datasets, including out‑of‑domain ones.
By Caterina Gallegati, Monica Bianchini, Franco Scarselli, Vittorio Murino, Barbara Toniella Corradini
The paper introduces the World-Ego Model (WEM), which separates an embodied video into a background‐only world component and an ego component comprising the robot and manipulated objects. By using role‑conditioned attention and asymmetric query budgets for state prediction, coupled with a semantic‑routed mixture‑of‑experts diffusion generator, WEM disentangles the generation of scene evolution and instruction following. The authors also release HTEWorld, a large dataset of 125K training clips and 300 evaluation trajectories for long‑horizon navigation‑manipulation tasks, and demonstrate that WEM outperforms prior methods on this benchmark.
By Zuyao Lin, Jianhui Zhang, Peidong Jia, Xiaoguang Zhao, Shanghang Zhang, Jingdong Wang, Xingyu Chen
The paper introduces a minimax framework to analyze the trade-off between model utility and vulnerability to distillation attacks, featuring an adaptive evaluation rule that reweights high-value examples and a teacher-side defense that suppresses outputs most useful for imitation. It proposes a cheap, forward-pass-only defense called Product-of-Experts (PoE) that combines the teacher with a proxy student during generation. Empirical results show a significant passive–adaptive gap, with adaptive students recovering more capability than passive evaluation suggests, narrowing the robustness gap between expensive defenses and PoE while keeping PoE cheaper and preserving higher-quality reasoning traces.
By Youssef Allouah, Mahdi Haghifam, Sanmi Koyejo, Reza Shokri
The paper introduces a corrected protocol for measuring the hackability of speech quality predictors, addressing two key flaws in previous methods. Using this protocol, it finds that four published predictors vary widely in vulnerability, with NISQA being the most hackable. The study also audits an attack‑detect‑patch loop, showing it only hardens predictors within its own attack space and can even reduce performance.
By Ali Alavi, Donald S. Williamson
The paper introduces Dirichlet‑Sinkhorn Constant Averaging (DSCA), a calibration method for memetic symbolic regression that partitions data into subsets with varying covariate distributions, calibrates each candidate independently, and evaluates it using the mean of the resulting parameters. DSCA reduces the excess loss compared to centralized calibration for correctly specified models and improves functional recovery and the accuracy‑complexity trade‑off on synthetic and real datasets. The authors demonstrate that DSCA’s performance gains stem from a controlled heterogeneous calibration that adds a complementary source of selection pressure in symbolic‑regression search.
By Mattia Billa, Veronica Guidetti, Federica Mandreoli
F$^3$NO is a frequency‑decomposed finite‑time flow‑map neural operator that uses low‑frequency features to guide the refinement of high‑frequency information through cross‑scale conditioning within each layer. The model predicts future states directly for specified times and adjusts the contributions of its two branches based on the prediction interval. Experiments on five PDE benchmarks show that F$^3$NO outperforms autoregressive and direct‑prediction baselines, and ablation studies reveal that frequency‑decomposed refinement improves accuracy with fewer parameters while segmentation benefits vary with spatial resolution and dynamical regime.
By Fan Wu, Cheng Jing, Kookjin Lee
The paper introduces Ranking Prior Alignment, a model‑agnostic framework that injects external ranking priors—derived from domain experts, teacher models, or large language models—into credit‑risk scoring models via a temperature‑scaled KL divergence loss. It unifies neural (MIL attention) and tree‑based (XGBoost) architectures under a single loss formulation, requiring no external model at inference and tolerating significant annotation noise. Experiments on a large industrial dataset and public Amex data show consistent AUC gains across multiple model families, with the benefits increasing when data, model capacity, or feature quality are limited.
By Qiye Lu, Jiang Ji, Liang Zhang
SACQ introduces a structured prediction head for long‑horizon time‑series forecasting that replaces the traditional flatten readout. It uses a two‑stage decoding process: a coarse patch‑grid scaffold is first generated, then each future position is refined via cross‑attention over historical memory, with a learned gate merging the corrections. A batch‑adaptive scaled log‑cosh loss further stabilizes training under long horizons and noisy labels, leading to superior MSE/MAE performance across multiple backbones with minimal overhead.
By Guo Cheng, Zhengzhuo Xu, Chenchen Jing, Jingyi Hou
CARE is a lightweight plug‑in module that can be added to any deterministic time‑series forecaster. It resamples historical data to align with the forecast horizon, learns residual corrections, and applies scale‑aware bounded updates controlled by per‑coordinate sigmoid risk gates. The module is trained with a multi‑objective loss to improve accuracy, track residuals, align risk, and anchor the base model, and it demonstrates accuracy gains on eight benchmarks with minimal extra parameters or latency.
By Guo Cheng, Changlong Lv, Jingyi Hou
The paper introduces a framework that unifies diffusion and autoregressive (AR) generative models as paths on a corruption lattice, defining a cost for decoding schedules based on discarded parallel steps. It shows that the minimal number of steps for a zero‑cost schedule is determined by the data’s geometry—specifically, the graph’s treedepth for Markov data—providing a theoretical limit for both token and continuous field models. By estimating pairwise dependencies from pretrained weights, the authors predict the ranking of decoding schedules before execution, and validate these predictions across text, image, and video generation tasks.
By T. Y. Tsui, Jiatao Gu, Lingjie Liu
SteerCast is a retrieval‑based latent steering technique that enhances decoder‑only time‑series forecasters during inference without updating model parameters. It builds a database from the training set, storing each history window’s representation and a steering vector that captures the difference between the model’s prediction and the ground‑truth continuation in latent space. During testing, SteerCast retrieves nearest neighbors, aggregates their steering vectors, and injects the resulting signal into the forecaster’s hidden states at each autoregressive step, guiding predictions toward trajectories seen in similar training cases. Experiments on diverse multivariate benchmarks and multiple horizons demonstrate consistent accuracy improvements over fine‑tuned backbones and other retrieval‑based baselines, all while requiring no additional training beyond the original fine‑tuning and using only the training set as a retrieval corpus.
By Van Dai Do, Huu Hiep Nguyen, Minh Hoang Nguyen, Hung Le
SpatialOPSD is a self‑distillation framework that teaches a multimodal large language model (MLLM) to perform spatial reasoning without relying on external tools. By feeding the MLLM with verified execution traces from a spatial coding agent, the model learns an internal spatial chain‑of‑thought. The method includes Repetition‑Aware Distillation to prevent privileged‑information leakage, and experiments show it outperforms SFT and GRPO on spatial and out‑of‑distribution benchmarks.
By Rongxue Li, Meng Yang, Yiru Mao, Yongliang Tao, Lulu Hu, Bin Yang, Zhao Xu, Weihua Luo, Bowen Xu
Bernoulli Flow Models (BFM) introduce a continuous global Bernoulli probability flow path for binary data generation, replacing sequential one‑step Markov diffusion chains. By deriving analytical closed‑form posterior transitions over arbitrary time intervals, BFM eliminates the need for approximating multi‑step likelihoods and allows inference with far fewer function evaluations while maintaining sample quality. Experiments demonstrate that a BFM trained with 256 steps achieves an FID of 9.22 on LSUN Churches 256×256 using only 16 sampling steps, outperforming state‑of‑the‑art discrete baselines and remaining competitive under full‑step inference.
By Hao Mo, Liying Yang, Shumin Yao, Xinxing Yu, Ajian Liu, Xudong Mao, Yanyan Liang