ReuF2F is a Lean 4 benchmark comprising 244 formalized open problems sourced from Research Experiences for Undergraduates (REU). It aims to evaluate coding agents on less prominent, entry‑level research problems across diverse mathematical subfields, using a Docker‑based Lean project setup and a unified harness called Leanimum‑agent. The study finds that even the best model solves only about 40 % of the problems, indicating significant room for improvement.
By Zeyu Zheng, Yiyang Liu, Yuyang Zhao, Congmin Zheng, Jiachen Zhu, Jianian Chen, Kangning Zhang, Kayue Daniel Wong, Lingyu Yang, Rong Shan, Shihao Chen, Thomas Zhu, Wentao Zhang, Xiaopeng Li, Yan Hu, Yongxi Lin, Yuxiang Chen, Zhuoying Ou, Jianghao Lin, Jeremy Avigad, Prasad Tetali, Jingbang Chen, Sean Welleck
The paper investigates whether pretrained language models can act as lightweight auxiliary supervision for supervised vision training, without requiring paired image-text data or complex objectives. Two forms of language guidance—Explicit Language Guidance (ExLG) and Implicit Language Guidance (ImLG)—are evaluated across six generalization regimes, including in‑distribution, out‑of‑distribution, shortcut resistance, texture/shape bias, adversarial robustness, and continual learning. Results show that ExLG improves in‑distribution, low‑data performance, and continual learning retention, while ImLG is more effective in shortcut‑sensitive and adversarial settings, and both add minimal parameters and training overhead.
By Shruthi Gowda, Nikita Jain, Elahe Arani, Bahram Zonooz
arXiv:2610. 28058v1 Announce Type: new Abstract: We study universal online convex optimization with delayed feedback.
By Hao Qiu, Andrew Jacobsen, Emmanuel Esposito, Mengxiao Zhang
The paper introduces Planning Under Noise (PUN), a neural operator that learns noise‑aware value fields for planning in unseen geometries. By imposing reflecting Neumann conditions on obstacles and a Dirichlet condition at the goal, PUN allows sampled paths to explore free space until reaching the target. It extends the Monte Carlo solver Walk on Stars to higher dimensions, learns geometry‑to‑field mappings from noisy labels, and achieves high success rates and low runtimes on 4‑D replanning benchmarks across multiple noise levels.
By Michael Chen, Hong Chul Nam, Xi Deng, Anima Anandkumar
The paper introduces SiamSA-PPM, a self‑supervised framework that augments event logs for predictive process monitoring by generating realistic trace variants using control‑flow semantics and frequent behavioral patterns. It combines Siamese learning with statistically grounded data augmentation to learn generalizable prefix representations without requiring labeled data. Experiments on diverse real‑world logs show that SiamSA-PPM matches or outperforms state‑of‑the‑art baselines in next‑activity and final‑outcome prediction tasks.
By Sjoerd van Straten, Alessandro Padella, Marwan Hassani
PACE is a trajectory inference framework that reconstructs continuous transport dynamics from destructive single‑cell time‑course snapshots. It uses a state‑ and time‑dependent anisotropic Riemannian metric to guide transport along tangent directions, alternates between refining cross‑time couplings and fitting neural bridges, and distills the learned dynamics into a global velocity field. Across multiple datasets, PACE outperforms existing methods, reducing key distance metrics by an average of 23.7% and improving RNA‑velocity alignment by 15.4%.
By Chenglei Yu, Chuanrui Wang, Bangyan Liao, Tailin Wu
The paper "Fluent but Fallible: Benchmarking Hallucination and Factual Reliability in Large Language Models" evaluates ten advanced LLMs on a benchmark of 70 academic prompts across four complexity levels. Using five metrics—Accuracy, Precision, Recall, Response Deviation Index, and Hallucination Index—reviewers found that Perplexity and DeepSeek V3 achieved the highest overall accuracy, Gemini balanced recall and performance, LLaMA 3 resisted hallucinations, and Falcon had perfect precision but low recall and high deviation. The study highlights a trade‑off between response completeness and factual reliability, especially as prompt complexity increases.
By Osama Shabih, Dayam Nadeem, Shahab Saquib Sohail, Mohammad Nadeem
The paper introduces a drill‑down evaluation framework that breaks down Text‑to‑SQL queries into simpler executable sub‑queries and natural‑language questions, enabling analysis of hallucination patterns along the generation path. It distinguishes recurrent hallucinations—those appearing in both early and final steps—from emergent hallucinations that only surface at the end. Experiments on five large language models and an expanded BIRD‑mini benchmark reveal varying recurrence and emergence probabilities across categories, with conversational history generally increasing recurrence and decreasing emergence.
By Hayden Moore, Abhinav Verma, Anshuman Dutt
The paper reports UvA-MT’s submission to the WMT 2026 General Machine Translation shared task, competing in the unconstrained track for all 23 translation directions. The approach leverages test‑time methods of Large Language Models, using sequential sampling with a refinement prompt to generate multiple translation candidates and then employing LLMs as pairwise judges in a round‑robin voting scheme to select the best candidate. Offline evaluation on the WMT25 benchmark shows a clear improvement over traditional point‑wise best‑of‑N selection.
By Di Wu, Sergey Troshin, Wafaa Mohammed, Seth Aycock, Evgeniia Tokarchuk, Vlad Niculae, Christof Monz
WaveSpec is a multispectral iris presentation attack detection framework that adapts features across near‑infrared wavelengths using a Set Conditioned Spectral Coherence Adapter. Each band is encoded with a DINOv2 backbone, then updated based on consensus, deviation, and wavelength identity before being projected into structural, spectral, and anomaly representations for classification. On the MSIrPAD benchmark, WaveSpec significantly lowers error rates, achieving an average D‑EER of 3.33% and BPCER@APCER=1% of 12.07% under a strict cross‑artefact protocol.
By Raghavendra Ramachandra
PixelSR introduces a single‑stage Pixel Diffusion Transformer that operates directly in pixel space for real‑world image super‑resolution. It combines Pixel‑Space Conditional Diffusion—integrating low‑resolution conditioning, semantic guidance, and texture refinement—to avoid VAE compression, with Orthogonal Flow Decoupling that separates a low‑rank semantic component from a noise‑free texture component for specialized optimization. The approach yields perceptually realistic details while preserving structural fidelity, as shown by competitive results on multiple benchmarks.
By Xiang Lv, Kai Zhao, Qiang Wang, Qiao Mo, Chao Zhou, Mingwen Shao
Touch-JEPA is a framework that learns tactile representations from hand‑object videos using limited tactile supervision. It first trains a self‑supervised model that predicts masked visual features at hand‑surface locations, capturing latent contact dynamics. The model is then jointly fine‑tuned on a small labeled subset, improving contact discrimination and tactile magnitude estimation on human‑hand benchmarks and transferring to robotic hands without robot tactile data.
By Minseong Kweon, Qiao Feng, Yiming Huang, Lingjie Liu
SARES-DEIM is a SAR ship detection framework that builds on the DETR architecture. It introduces a sparse mixture‑of‑experts module, SARESMoE, which routes features to wavelet, spatial, and frequency experts, and a Space‑to‑Depth Enhancement Pyramid (SDEP) neck that preserves high‑resolution cues for small targets. The method achieves state‑of‑the‑art performance on HRSID and SAR‑Ship‑Dataset, delivering 76.4% mAP50:95 and 93.8% mAP50 while maintaining efficient inference at 136.5 FPS on an NVIDIA A40 GPU.
By Fenghao Song, Shaojing Yang, Xi Zhou
DeclineCraft is a dynamic abstention benchmark for retrieval‑augmented generation (RAG) that tests whether models can recognize when neither retrieved documents nor parametric knowledge can answer a question. The benchmark uses unanswerable Minecraft feature‑addition queries that appear after a corpus cut‑off, allowing evaluation of future models. Across 27 large language models, the study measures Correct Abstention Rate (CAR), Answerable Attempt Rate, and Answer Accuracy, finding that single‑shot retrieval boosts CAR, iterative RAG reduces CAR with minimal impact on accuracy, and larger models improve accuracy but not CAR.
By Atharva Thakur, Abulhair Saparov, Bruno Ribeiro
Object-Oriented Harness (OO-Harness) is a framework that optimizes agent harnesses by evolving a hierarchy of executable harness classes through object-oriented design. During training, an evolution agent uses execution records to create specialized harness subclasses, while at test time a harness router selects the appropriate harness for each task based on learned routing lessons. Experiments on three long-horizon benchmarks show OO-Harness outperforms other methods, achieving up to a 15% absolute improvement over the base harness.
By Gaotang Li, Bhavana Dalvi Mishra, Yifei Ming, Yanfei Chen, Zhongying CuiZhu, Lesly Miculicich, Long T. Le, Rui Meng, Chun-Liang Li, Hanghang Tong, Chen-Yu Lee, Tomas Pfister
The paper introduces cache engineering for large language model agents, enabling direct control over the active key-value cache without altering the trajectory or triggering costly re-prefilling. It defines cache operations such as drop and repos, formalizes their effects, and addresses the challenge of history-dependent cache states. The authors co-design a stateless AKV serving engine that manages stateful cache reuse, eviction, and restoration, achieving significant performance and cost improvements over traditional summarization methods.
By Tianyu Fu, Ruoxi Wang, Zihan Min, Zhongqi Ren, Zhihao Li, Tengxuan Liu, Shengxu Mu, Hanling Zhang, Yu Wang
SciFigBench is a new benchmark for evaluating Vision Language Models on scientific method figure understanding. It focuses on three fine‑grained capabilities—dense content perception, visual relation comprehension, and information‑flow tracing—using 4,352 curated figures and 40,670 QA pairs. The authors also introduce SciFigCue, a training‑free visual prompting module that improves performance on several tasks without updating model parameters.
By Lijia Wu, Zheng Li, Yuxuan Liu, Jingwen Zhang, Penghai Zhao, Xiang Li
The paper introduces a reinforcement learning framework that uses large language models to generate RTL circuits while optimizing for Pareto-optimality across power, performance, and area (PPA). It proposes a generation-based training pipeline and a Hypervolume Contribution reward to improve both the quality and diversity of the entire set of generated circuits. Experiments on the RTLLM v2 benchmark show significant gains in hypervolume and diversity compared to baseline and single-objective methods.
By Koki Takeshita, Kenji Uebayashi, Hikaru Tomonari, Masaya Tsunokake, Koichiro Yawata, Tatsuya Sasaki, Yusuke Ohtsubo, Kota Dohi, Yasuhiro Sogawa
EDMA: Entropy-Driven Multimodal Answering introduces a framework that formalises multimodal question answering as an iterative entropy‑reduction process. By quantifying each modality’s contribution through logical entropy, EDMA constructs measurable answering trajectories and employs entropy‑driven fusion and selection to minimise logical entropy. Across standard MMQA benchmarks, EDMA surpasses prompting baselines, achieving significant gains in F1 and LLM‑J scores, and demonstrates that higher entropy correlates with more false positives, validating entropy as a reliable performance proxy.
By Emanuele Mezzi, Gertjan Burghouts, Fabio Massacci, Mengyuan Zhang
TxnQA is a new analytical QA benchmark that tests reasoning and evidence grounding over transaction event sequences. It contains 27 task families with 81 templates and a 5‑scenario fraud‑risk reasoning suite built on synthetic, privacy‑mitigated transaction data. The benchmark evaluates models using format‑following rate, conditional content accuracy, and end‑to‑end correctness, revealing distinct bottlenecks in schema compliance, semantic calculation, and evidence attribution.
By Rabimba Karanjai, Yiming Bian, Yun-Shiuan Chuang, Jenna L. Ballard, Zijie Pan, Shivani Shekhar, Yixi Li, Yirou Ge, Avinash Thangali, Richard Williamson, Linsey Pang, Prakhar Mehrotra