PROBE is a benchmark that compiles 2,457 bugs fixed over 18 months in 20 GitHub repositories and provides a method to match candidate findings to this reference set. Evaluation of frontier coding agents shows they achieve less than 21% recall and under 9% precision, with stronger models improving recall from 8% to 17% but not precision. Despite low precision, 85% of findings are judged valid by LLM and human evaluators, and the distribution of bugs found by agents differs substantially from those fixed by maintainers.
By Marco Paina, Mark Niklas M\"uller, Niels M\"undler-Sasahara, Martin Vechev
The paper introduces Dirichlet Process Mixtures (DPMs) as a new non‑parametric conditional probability distribution (CPD) model for Semiparametric Bayesian Networks (SPBNs). By replacing kernel density estimators with DPMs, the authors achieve a more compact representation that reduces inference costs and improves density estimation accuracy. Experiments on synthetic and real‑world data show that DPM‑based SPBNs outperform existing SPBNs and other continuous‑variable Bayesian network models in both density estimation and structure learning tasks.
By Victor Alejandre, Pedro Larra\~naga, Concha Bielza
The paper introduces Contrastive Evidence Paths (CEP), a method that scores search candidates by their strongest individual paths on a query‑conditioned document graph and calibrates these scores against a background visitation distribution to reduce structural bias. Building on CEP, Structured Corpus Interaction (SCI) organizes candidates and paths into a structured map, allowing agents to verify sources and expand from anchor documents. Experiments on BrowseComp‑Plus and TREC‑COVID demonstrate that SCI achieves state‑of‑the‑art answer accuracy and nDCG@10 while cutting tool calls and generation costs by over 50% and 40%, respectively.
By Junchen Li, Liang Xu, Rongzheng Wang, Qizhi Chen, Junhao Wen, Shuang Liang
Concorde is an energy‑based framework for link prediction that tackles sparse graphs by decomposing link likelihood into three decoupled branches: a parameter‑free spectral anchor for topological stability, a mixed‑curvature residual on a learnable hyperbolic‑spherical product manifold regularized by a curvature proxy, and a continuous vector field on the hyperbolic tangent bundle for non‑local alignment. These branches are fused via learned gating and trained with an N‑pair ranking loss. On five diverse benchmarks, Concorde achieves state‑of‑the‑art early retrieval precision, nearly doubling P@10% over leading baselines in sparse settings.
By Ege Demirci, Ambuj Singh
The paper introduces RePPO, a method that adapts Proximal Policy Optimization to jointly optimize a compact population of diverse, high‑quality policies using a population‑dependent, dynamically shaped reward. RePPO achieves a strong quality‑diversity tradeoff on standard continuous‑control tasks and scales to larger policy and behavior spaces, maintaining quality while significantly increasing diversity in pixel‑based tasks. The authors also demonstrate RePPO’s applicability to fine‑tuning large language models, showing for the first time that quality‑diversity reinforcement learning can scale to this domain.
By Saeed Hedayatian, Harsh Shah, Celine Mneimneh, Stefanos Nikolaidis
The paper proposes Mixture Proposal On-Policy Distillation (MPOPD) to address a mismatch issue in on‑policy distillation (OPD) where the rollout policy is fixed to the current student. By treating the rollout policy as an optimization target, MPOPD derives a geometric mixture of student and teacher policies and introduces a proposal policy for efficient rollout generation. Experiments on 16 reasoning benchmarks show consistent gains over OPD, notably a 6.91% accuracy improvement on math reasoning with Qwen3-4B-Base and a 41% reduction in generation length.
By Sirui Chen, Yunzhe Qi, Haobo Xu, Jiaru Zou, Mengting Ai, Zihao Li, Jingrui He
NextEval introduces an agent-based derivative-free optimization solver that leverages a large language model to guide the sequential selection of evaluation points. The system couples an LLM agent with a numerical environment, allowing the agent to choose tools, configure parameters, and allocate a limited evaluation budget while the environment maintains shared history and state across method switches. A benchmark protocol, NextEval Bench, is also presented to compare different agent configurations on shared DFO tasks under matched budgets and information constraints, and numerical experiments show that NextEval outperforms state-of-the-art baselines in function-evaluation efficiency.
By Shuo Chen, Li Cui, Cunxin Huang, Ya-Feng Liu, Minghang Xie, Meng Xu, Zaikun Zhang
The paper introduces Req2Test-Bench, a benchmark of 5,086 instances linking natural‑language Gherkin requirements to developer‑written test implementations across five languages and four BDD frameworks. It evaluates eight agent–model–reasoning setups, finding that while the best configuration (Codex with DeepSeek‑V4.1‑Flash) produces passing tests for 82.8% of cases, only 33.8% are behaviorally adequate, as measured by kill‑completeness. The study highlights diminishing returns from scaling agents, the impact of under‑specified requirements, and the potential of requirement‑grounded harnesses to improve test generation.
By Binhang Qi, Yiming Liu, Yun Lin, Yangjie Zhou, Hailong Sun, Jin Song Dong
The paper introduces Motion-Guided Asymmetric Prediction (MoGAP), a method that enhances World Action Models (WAMs) for robot control by focusing visual supervision on dynamic regions of video data. MoGAP asymmetrically corrupts static and dynamic areas, training the model to reconstruct heavily corrupted dynamic regions using less corrupted static context, all within the shared video diffusion backbone. Experiments on robot manipulation benchmarks show that MoGAP consistently improves policy performance, raising success rates on the GR-1 tabletop benchmark from 38.8% to 47.5% for Fast‑WAM and from 52.3% to 62.3% for Cosmos3‑Nano.
By Taeyoung Kim, Junyoung Sung, Heeseung Kwon, Minseop Kim, Jinwoo Shin
The paper introduces LOWN, a decoding strategy for diffusion vision‑language models that selectively processes visual tokens only when necessary. By predicting masked tokens without the image and only invoking image processing when predictions disagree, LOWN achieves 99.6% of LLaDA‑V’s performance while reducing decoding FLOPs to 35.3% on average. The method adapts to varying image‑dependency across benchmarks, generalizes to other LLaDA-based models, and can be combined with visual token pruning and key‑value caching for further efficiency.
By Donghyeon Cho, Joowon Kim, Seungho Shin, Eunho Yang
The study trains small GPT‑2 models on child‑scale language data from dense home recordings (BabyView) and synthetic child‑directed dialogue (TinyDialogues), varying data amount, source, and diversity. Models trained on a single child’s environment predict that child’s language far better than others, a pattern largely driven by lexical and distributional factors as shown by a 5‑gram baseline. While training on multiple children improves predictions on pooled held‑out language, it does not yield a clear advantage on grammatical benchmarks, and naturalistic versus synthetic inputs produce distinct grammatical and lexical‑semantic patterns.
By Steven Y. Feng, Michael C. Frank
Select‑TTT introduces a query‑conditioned rule for In‑Place Test‑Time Training (TTT) that scales the write strength of each context chunk based on its relevance to the user query. By using the model’s frozen output head to compute relevance scores and applying a temperature‑scaled softmax as learning‑rate multipliers, the method preserves the efficient, chunk‑wise computation of In‑Place TTT while preventing irrelevant content from diluting fast‑weight memory. Experiments on two benchmarks and two backbones show consistent improvements in long‑context question answering.
By Yue Wu, Yang Zhang, Weinan Gan, Chongming Gao, Xinsen Guo, Xiaoyuan Li, Fuli Feng
TAPESTRY repurposes a pretrained video diffusion transformer into a robotic policy by treating robot actions as single-patch tokens in the model’s latent space, eliminating the need for video generation at inference. The lightweight linear projection maps each action timestep to the VAE patch dimensions, and the shared backbone is trained with three diffusion objectives—video diffusion, joint video+action diffusion, and action-only diffusion—to learn dynamics priors and action denoising. At test time, the model produces smooth action chunks in a single DDIM step, achieving strong performance on simulation and real‑robot benchmarks with significantly lower inference cost than comparable joint video–action policies.
By Jeremy A. Collins, Seungjae Lee, Rachanon Wachakorn, Krishnan Srinivasan, Animesh Garg, Vitor Campagnolo Guizilini, Paarth Shah
AdaGround is a training‑free framework for GUI grounding that addresses shortcomings of existing zooming methods. It introduces Token‑Normalized Adaptive Zooming to match magnification to target size, a Verifier‑in‑the‑Loop Refinement that uses a simple Yes/No click to accept or reject zooms and decide when to abstain, and Compositional Structure Grounding for spreadsheet cells. Experiments on Qwen3‑VL models show AdaGround outperforms state‑of‑the‑art training‑free zooming techniques and many trained GUI models, boosting accuracy by 7.4 to 20.6 points on OSWorld‑G and ScreenSpot‑Pro and maintaining gains on unseen UI‑Vision data.
By Bingchao Wang, Yongcheng Jing, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
The paper introduces Energy-Based Generative Replay (EB-Gen) for Equilibrium Propagation (EqProp) trained Hopfield Networks, enabling continual learning without external data buffers or architectural overhead. EB-Gen leverages the network’s intrinsic generative dynamics to produce pseudo‑samples from a frozen copy of the model, allowing internal rehearsal. Experiments on split MNIST and split CIFAR‑10 show that EB-Gen improves stability and plasticity, achieving 75 % and 64 % average accuracy respectively, outperforming several baseline methods.
By Victor Cook (LIRMM | ADAC), Gilles Sassatelli (LIRMM | ADAC), Thierry Gil (LIRMM), Mohamed Watfa (LIRMM | ADAC), Marina Reyboz (CEA-LETI)
GCSR‑OCC introduces geometry calibration and semantic routing to improve open‑vocabulary 3D occupancy prediction from a single RGB image. It uses an Evidence‑Conserving Gaussian‑to‑Occupancy operator to mitigate scale inflation and a Multi‑Mode Query‑First Gaussian Semantic Routing module to align multiple semantic modes with text queries, refining voxel‑level predictions. The method achieves state‑of‑the‑art results on Occ‑ScanNet, improving geometric IoU to 63.31% and semantic mIoU to 27.44%, while reducing aggregation latency from 19.76 ms to 12.88 ms.
By Bohua Wang, Congyang Zhao, Tianxiong Lv, Chenxu Zhao, He Zheng, Hangqi Fan, Leigang Luo, Zhihao Liao, Yihong Guo, Yang Cai
Retrieve to Act (R2A) is a test‑time scaling framework that improves the robustness of Vision‑Language‑Action (VLA) policies to out‑of‑distribution visual shifts. It transforms OOD robustness into a geometric motion‑primitive retrieval problem by building an offline spatiotemporal motion‑primitive graph from successful training demonstrations. During inference, a graph neural network localizes the robot’s current state, retrieves compatible future motion primitives, and blends them with the frozen VLA’s actions via a learned gating mechanism, thereby correcting erroneous spatial trajectories while preserving local reactivity.
By Pham Tri Quang, Bao-Ngoc Dao, Minh Duc Nguyen, Duc-Duy Nguyen, Hung T. Ho, Linh Dang Le, Trong-Bao Ho, Thien-Loc Ha, Tan Q. Nguyen, Tuan Dam, Vu N. Duong, Daniel Sonntag, Ngan Le, Khoa D. Doan, Jan Peters, An Thai Le, Duy M. H. Nguyen, Vien Anh Ngo
The paper demonstrates that Mixture-of-Experts (MoE) models can avoid the need for calibration data when selecting which expert channels to load during decoding. Instead of using a calibrated threshold, a top‑k selection strategy—choosing the k highest‑scoring channels per token—achieves comparable or better performance without calibration. Additionally, the authors propose a reader that uses a ternary sketch and low‑rank mapping to recover skipped channels, achieving 98.1% of full‑model accuracy while reading only 17.8% of expert weights and speeding up inference by up to 2.69× compared to optimized vLLM.
By Geunhyoung Kim, Sung Ju Hwang
The paper introduces ImagiRep, a benchmark that tests spatial reasoning using seven question types across two settings, offering RGB and six alternative representations of imagined states. It shows that accurate RGB generation does not guarantee correct spatial reasoning and that different questions benefit from different representations such as depth, panoramic views, or top‑down geometry. Building on these insights, the authors present GEARS, a lightweight strategy that adaptively selects representations at test time, outperforming existing methods and demonstrating that reasoning about unseen states requires more than just RGB generation.
By Zengqi Zhao, Zun Wang, Shoubin Yu, Mohit Bansal, Yue Zhang
The study assesses how reliably individualized functional‑network attribution can be derived from resting‑state fMRI data for predicting cognitive scores in the ADNI cohort. Using a frozen functional‑connectivity predictor trained on 875 subjects, the authors find that split‑half reliability is near‑ceiling, but scan‑rescan reliability drops significantly. Even under a mismatched‑subject null, correlations remain substantial, indicating a strong population‑shared pattern, yet matched pairs show a significant person‑specific component. The authors recommend reporting scan‑rescan reliability alongside a permutation‑null control to avoid overstating individual specificity.
By Sheethal Bhat, Mahfuzur Rahman Chowdhury, Imam Ahmed Shaik, Rose Dawn Bharath, Siming Bayer, Andreas Maier