GeoFlip is a new benchmark for video generation models that tests their ability to produce physically correct outcomes when the initial geometry of a scene is altered. The benchmark presents 482 pairs of scenes with controlled changes in extent, placement, and mechanism configuration, and evaluates models on event correctness, structural integrity, and similarity to reference videos. Across 16 evaluated models, the best event correctness averaged only 27.98%, highlighting significant gaps in geometry‑sensitive physical reasoning.
By Yuqi Li, Junming Liu, Bowen Hu, Ruize Niu, Huiran Duan, Yingli Tian, Hao Wu
TalkFlow introduces a structured latent flow framework for generating holistic co-speech motion, learning separate continuous latent representations for upper-body, lower-body, and facial expressions. The model incorporates lightweight cross-stream interactions to preserve part-specific dynamics and maintain coordination across motion components. Experiments on the BEAT2 dataset show improved facial expression accuracy, body gesture quality, and speech-motion synchronization compared to state-of-the-art methods.
By Xinmu Wang, Fumi Honda, Xiang Gao, Xianfeng Gu, Xiyun Song, Heather Yu
The paper introduces Semap, an open benchmark dataset of 1,439 manually annotated map patches that captures the diversity of historical map collections. It proposes a generalizable semantic segmentation workflow that combines procedural data synthesis with multiscale integration, achieving the best performance among evaluated methods on heterogeneous map data. The study demonstrates that segmentation accuracy remains stable across different map styles, scales, geographic regions, and publication contexts, thereby facilitating the integration of diverse cartographic archives into historical geographic research.
By Remi Petitpierre
The paper investigates how implicit assumptions about tasks in continual learning affect reported performance metrics. It distinguishes between the underlying data, the partitioning of that data into tasks, and the learner’s experience, showing that different granularities, orders, and similarity structures can lead to markedly different outcomes. The authors use controlled teacher–student experiments to illustrate these effects and review existing benchmarks to assess which properties are explicitly stated.
By Kathrin Korte, Christian Metzner, Akshay Bedhotiya, Viet Anh Khoa Tran, Eleni Nisioti, Sebastian Risi
PPS: Plug-and-Play Saccadic Vision for Fine-Grained Classification proposes a lightweight extension that transforms any pretrained backbone with spatial feature maps into a saccadic classifier. The method adds only a small priority head and a single cross-attention fusion block, using a weight-tied encoder to process a downsampled peripheral glance and high-resolution glimpse crops selected by a priority distribution. A sequential sampler with non-maximum suppression draws a configurable number of glimpses, and a residual multi-head cross-attention block fuses the glance with the glimpses; the priority head is trained via a per-glimpse information-gain target from a lightweight memory bank, avoiding reinforcement learning. The approach achieves linear complexity in the number of glimpses and decouples peak memory from input resolution, consistently improving performance across nine fine-grained classification benchmarks.
By Johann Schmidt, Sebastian Stober, Joachim Denzler, Paul Bodesheim
The paper introduces Proof‑Carrying Materials (PCM), a three‑stage framework that improves the reliability of machine‑learned interatomic potentials (MLIPs) used for high‑throughput materials screening. PCM combines adversarial falsification, bootstrap envelope refinement, and Lean 4 formal certification to close the large gap between MLIP predictions and density functional theory (DFT) stability, achieving a 93% reduction in missed stable materials on a 25,000‑material benchmark. The method uncovers architecture‑specific blind spots, provides a risk model with high predictive performance, and boosts discovery yield by 25% in a thermoelectric screening case study.
By Abhinaba Basu, Pavan Chakraborty
SAT-ACT introduces an outcome‑derived learning framework for improving branching decisions in conflict‑driven clause learning (CDCL) SAT solvers. It extracts action preferences from independent branching rollouts using three outcome comparison criteria, allowing the model to learn branching priorities without forcing a predefined ordering. The approach also incorporates non‑dominated heuristic action supervision, and experimental results show consistent reductions in CDCL search costs and improved PAR‑2 performance on SAT Competition benchmarks.
By Zhenchao Sun, Chongyang Tao, Ping Lu, Tongkai Lu, Jianling Gao, Shuai Ma
CryoAtlas is a large, rigorously curated dataset and unified benchmark for density‑guided atomic model building of protein chains from cryo‑EM maps. It compiles 17,809 map‑model‑sequence entries from EMDB and PDB (up to March 2026), with a 16,657‑entry training split, and defines a 618‑target benchmark that evaluates structural accuracy, map fit, and stereochemical validity. The resource enables systematic comparison of learning‑based methods with traditional tools like Phenix and AlphaFold3, and demonstrates that retraining a CNN component on CryoAtlas improves performance.
By Mingrui Li, Minzhang Li, Weichen Qin, Yufan Xie, Zhangzhi Xiong, Sixian Shen, Jiakai Zhang, Yuan Pei, Jingyi Yu
CAMEL (Cross-Architecture Masked Embedding Learning) is a new distillation framework that transfers knowledge from large foundation models to lightweight student models using an asymmetric masking strategy during embedding learning. The method is designed for retinal image analysis, where existing foundation models like RETFound excel but are computationally heavy. Experiments on four datasets demonstrate that CAMEL maintains accuracy while dramatically reducing computational complexity, making it suitable for resource‑limited clinical settings.
By Andre Guarnier De Mitri, Giovani Decico Lucafo, Miller Matheus Lima Anacleto Rocha, Diego Furtado Silva
The paper introduces EDB-Bio, a benchmark of 96,000 episodes drawn from real biological studies—drug‑perturbation transcriptomics, CRISPR screens, and GWAS—to test whether language models can weigh empirical evidence and predict hypothesis truth. Initial results show that both open‑weight and frontier models are systematically underconfident, matching a trivial base‑rate forecaster in Brier score, but post‑training with proper scoring rules improves calibration and discrimination. In a downstream experiment‑selection task, a calibrated model (Ibis0) outperforms the best frontier model by 2.5×, and Bayes‑factor analysis reveals how models update beliefs, align with expert relevance judgments, and expose biases.
By Conor Igoe, Gabriel Grand, Zane Koch, Albert Bou, Eli Moss, Michaela M. Hinks, Samuel G. Rodriques, Andrew White, Siddharth Narayanan
The paper introduces a degree‑resolved framework for nonlinear interactions between irreducible representations in graph neural networks, enabling mixing of non‑isomorphic representations that linear equivariant maps cannot achieve. Using this framework, the authors build Specht Triple, a residual module for unordered graph triples, and demonstrate that it reduces aggregate standardised test MSE by 11.05% on a synthetic structural benchmark compared to a parameter‑matched linear baseline. The results highlight irreducible decomposition as a useful inductive bias for nonlinear higher‑order equivariant processing.
By Mani Shayestehfar
TalentRL is a strategy‑guided reinforcement learning framework that routes natural‑language skills hierarchically, using coarse‑grained talents to guide fine‑grained skill retrieval. The method trains talent routing via reward contrasts and optimizes execution with group‑relative advantages, achieving significant performance gains on open‑domain QA benchmarks and agentic tasks such as ALFWorld and WebShop.
By Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang, Michelle Chang, Jingwen Lu
The paper introduces a controlled evaluation framework to determine when encoder adaptation is beneficial after correcting for class imbalance in long‑tailed recognition tasks. By separating unique support from representation allocation and comparing classifier‑only versus full‑network continuations, the study shows that encoder updates become worthwhile only when representation allocation is highly imbalanced, a transition that persists across various corrections and scales to ImageNet. The framework uses stage‑specific controls, multiple objectives, whole‑distribution metrics, and hierarchical inference to robustly estimate incremental encoder value.
By Zexing Zhang, Jichao Li, Tianyang Lei, Yiyin Zhang, Kewei Yang
NormDetect is a new detection method that anchors a learned per-section head to a normalized zero-shot score to improve low‑false‑positive detection of machine‑generated web text. By blending the two scores with a weight chosen on training folds under a human false‑positive rate cap, it raises recall by 5.0 points over normalized Binoculars at a calibrated 0.1% FPR on the RAID benchmark. The method maintains a human FPR bound of at most 1% in every section, outperforming other detectors within this bound.
By Hanshuai Cui, Fanshuai Meng, Zhiqing Tang
The paper introduces a finite‑dimensional orbit framework that builds data‑adapted dictionaries, sparse realizations, and predictive models using wrapped weighted shifts to create orbit dictionaries and local interpolation maps. It presents CCDS, a deviation‑controlled sparse realization procedure that achieves MP/OMP‑like behavior at minimal deviation while enabling feasible interior paths with specified residual geometry. The authors prove generic orbit independence, exact block interpolation with norm control, bounds on sparse approximation distortion, guarantees on selected supports, and stability under perturbations, and demonstrate that orbit scaffolds outperform random or unstructured alternatives in synthetic and benchmark experiments.
By Yunied Puig
The paper examines the bias in group-relative advantage estimation used in reinforcement learning from verifier rewards (RLVR) for large language models. It shows that for binary rewards, hard prompts suffer from negative estimation errors while easy prompts have positive errors, leading to underweighting correct hard answers and insufficient penalties for incorrect easy answers. The authors propose History‑Aware Adaptive Difficulty Weighting (HA‑DW), an adaptive reweighting scheme that corrects these biases, and demonstrate through theory and experiments on five mathematical reasoning benchmarks that HA‑DW improves performance when integrated into GRPO and its variants.
By Fengkai Yang, Zherui Chen, Xiaohan Wang, Xiaodong Lu, Jiajun Chai, Guojun Yin, Wei Lin, Shuai Ma, Fuzhen Zhuang, Deqing Wang, Yaodong Yang, Jianxin Li, Yikun Ban
The paper evaluates three backpropagation‑free learning rules—Direct Feedback Alignment, Direct Random Target Projection, and NoProp—on 112 univariate time‑series classification datasets using 1D convolutional networks. Each method achieves mean F1 scores between 0.76 and 0.79, slightly below state‑of‑the‑art backpropagation baselines. The study also analyzes how dataset properties, especially intra‑class temporal variability, affect performance gaps and examines computational trade‑offs such as FLOPs, training speed, and memory usage.
By Miha Kap\v{s}, Bla\v{z} Bertalani\v{c}, Carolina Fortuna
GENSTRAT is a system that evaluates strategic reasoning in large language models (LLMs) by procedurally generating two-player zero-sum imperfect-information betting games on demand. Each game is scored across five complexity axes—state space, temporal contingency, opponent dependence, private information, and precision—and assessed via Monte Carlo simulation. Using this generator, the authors created a 50-game benchmark, ran a tournament of twelve frontier and open-weight models with over 40,000 LLM vs. LLM matches, and found that the performance gap on the hardest games is 1.7 times larger than across all games, with higher thinking budgets, opponent strategy disclosure, and tailored reasoning improving outcomes.
By Vartan Shadarevian, Kia Ghods, Alex Kenich, Anany Kotawala
The paper introduces a species‑level Acoustic Biodiversity Index (ABI) that normalises Shannon diversity by a regional reference species pool (ABI_E) and reports temporal stability separately. Benchmarking on a tropical dataset showed BirdNET embeddings with logistic regression achieved the best classifier performance, while stricter post‑hoc filtering removed implausible detections. In field tests across a Zimbabwean savanna and an Abu Dhabi zoo, ABI_E correlated strongly with observed species richness (Spearman ρ = 0.93), outperforming six conventional acoustic indices.
By Muhammad Saeed, Dmitry Mikhaylov, Mohamed H. Alhosani, Yasser F. Al Wahedi, Rory Pilossof, Ahmed Arabi, Walid Shaaba
The paper investigates whether adding a loop—re‑applying a block of layers—to a trained language model is worth the extra inference cost. Experiments on Qwen3 and Qwen3.5 models from 0.8B to 9B parameters show that only the 0.8B looped model outperforms both a non‑looped model with the same parameters and a model with the same FLOPs. For larger models (4B and 9B), the looped version does not beat the baseline trained without loops, indicating that loops only pay off for small models with a fixed parameter budget.
By Rhea Sanjay Sukthanker, Pashmina Cameron