Harness Scaling introduces a new scaling dimension for agentic reinforcement learning by training models across diverse harness configurations instead of a single fixed harness. The approach constructs a factorized harness space, selects a compact training subset with balanced factorial coverage, and organizes cross-harness training to combine global coverage with controlled variation for the same task. Experiments across multiple benchmarks, model scales, and RL algorithms show that Harness Scaling consistently improves generalization to unseen white-box configurations and enables effective transfer to unseen external black-box harnesses.
By Fei Bai, Huatong Song, Chuan Hao, Ran Tao, Jian Yang, Bryan Dai, Wayne Xin Zhao, Ji-Rong Wen
The paper introduces methods to improve training of stochastic flow maps for stochastic differential equations. It proposes representing Brownian paths with a shifted Legendre polynomial basis and using a Wong‑Zakai styled objective to enable derivative‑based losses. The authors also provide a fine‑tuning recipe that turns large pre‑trained flow models into strong stochastic flow maps, achieving a three‑fold improvement on ImageNet compared to SSFM.
By Zander W. Blasingame, Antonio Franca, Raul Minan, Christian Koke, Niklas Rindtorff, Alexander Tong
The paper introduces two novel techniques for enhancing graph neural networks on Petri nets: a positional encoding derived from directed place‑transition walk return probabilities (PNRWPE) and a diffusion‑kernel‑based message gating using the incidence matrix (DA‑PNGNN). These methods explicitly capture directed cycles and invariants that were previously hidden, leading to significant performance gains. On the Process Discovery Contest 2023 dataset, PNRWPE raises macro balanced accuracy from 0.751 to 0.796, nearly doubling loop detection accuracy, while DA‑PNGNN further improves it to 0.830 and boosts net‑to‑tree retrieval on ProcRosetta by 4.1–5.9 points.
By Geri Skenderi
FrontierCS 2 is a new benchmark comprising 48 open‑ended computer‑science research tasks that place language‑model agents inside a research environment, allowing them to submit incremental patches to a working repository. Each patch is scored by a judge on a development suite, while the final evaluation uses a held‑out suite and a private reference derived from the original authors’ code. The benchmark covers 13 domains, includes a private set of over 100 future tasks, and evaluates 12 agents, with Claude Opus 5.5 achieving the highest pass rate of 58%.
By Qiuyang Mang, Wenhao Chai, Runyuan He, Dongruixuan Li, Zihan Zheng, Yujun He, Xuanyi Zhou, Hao Wang, Siddhant Ray, Bo Peng, Ruyi Ji, Shuo Yang, Haocheng Xi, Boru Chen, Xinyang David Han, Melissa Pan, Mayank Mishra, Boxi Yu, Sijia Liu, Kerem Akillioglu, Yi Pan, Zirui Wang, Mert Cemri, Diogo Antunes, Yichuan Wang, Aiwei Ivy Zhang, Zhenyang Chen, Qizheng Zhang, Kaiyuan Liu, Wenyuan Huang, Sherry Ruan, Shu Liu, Dacheng Li, Zerui Cheng, Zeyu Shen, Shang Zhou, Hanfei Chen, Yuzhong Zhang, Bohan Lyu, Qu Ao, Han Zheng, Chonghe Jiang, Jingzhuo Hu, Xiaokun Chen, Kunjun Li, Zhiqing Zhong, Haoye Tian, Tong Zhao, Yunzhong He, Hanchen Li, Ofir Press, Jiantao Jiao, Kunle Olukotun, Karthik R. Narasimhan, Jingbo Shang, Koushik Sen, Matei Zaharia, Alex Dimakis, Ion Stoica, Joseph E. Gonzalez, Alvin Cheung
HyperDrone is a high‑performance quadrotor simulator that runs more than 20× faster than current batch simulators by combining optimized multirotor dynamics with the HyperLux ray‑tracing engine. HyperLux supports multiple graphics backends (Nvidia OptiX, Vulkan, Apple Metal, WebGPU) and can render RGB, depth, segmentation, normals, and optical flow at configurable fidelity—including reflections, realistic lighting, motion blur, anti‑aliasing, and PBR—while achieving millions of frames per second on consumer GPUs. The system was used to train a vision‑based yaw‑control policy in simulation, which was successfully transferred to a real Crazyflie quadrotor with an onboard camera.
By Jonas Eschmann, Giuseppe Loianno
arXiv:2610. 28994v1 Announce Type: cross Abstract: We address the Diverse Traveling Salesman Problem (D-TSP) that seeks a set of $k$ distinct TSP tours.
By Ssu-Yuan Lo, Ching-Kai Wang, Kuan-Lun Chen, Hao-Tsung Yang
arXiv:2610. 29052v1 Announce Type: cross Abstract: A typed decision maps evidence, a question and a fixed set of options to a probability for every option, computed in one forward pass, by reading a language model's logits or a small head over its hidden states, instead of generating and parsing text.
By Yuchuan Weng, Chi Liu, Xin Chen, Srinivasan Manoharan
SAFTER is a safety-aware filtering method that embeds a differentiable backup control barrier function (CBF) into Soft Actor-Critic training, unifying safety, finite-horizon backup safety, terminal recoverability, and input admissibility into a single differentiable certificate. The method achieves an exact minimum-intervention projection without iterative online optimization, leading to higher training success and fewer transitions compared to vanilla SAC and other CBF-filtered baselines. In quadruped navigation experiments, SAFTER outperforms state-of-the-art methods on unseen simulated environments and real-world maps, with ablation studies highlighting the importance of accurate differentiation through the safety filter for robustness.
By Benedictus C. G. Cinun, Bin Hu
The paper introduces QSelect, a method for compressing large language models for space robots by composing low‑bit quantized modules from AWQ, GPTQ, and AutoRound checkpoints. In fidelity mode, QSelect selects substitutions that reduce forward KL divergence from a BF16 reference, achieving lower perplexity and KL than any single‑family model across several 3‑ and 4‑bit settings. An optional target‑adapted mode further improves accuracy on labeled data, demonstrating the approach’s potential for meeting strict memory and behavioral‑fidelity constraints in space‑borne deployments.
By Jia Deng, Zizhe Wang, Xin Lu, Yi Huang, Zehao Jia
The paper introduces OnQ, a low‑bit recovery pipeline for quantized large language models that uses on‑policy distillation and candidate‑wise Dynamic KL to address trajectory divergence and distributional misalignment. OnQ starts with block‑wise calibration, then lets the quantized student generate responses and receive full‑precision teacher supervision on its own prefixes, dynamically weighting forward and reverse KL losses per candidate. Experiments on Qwen3 models show that OnQ outperforms baselines on nine benchmarks, especially long‑horizon reasoning, and produces shorter, less truncated responses.
By Boyuan Zheng, Jing Xiong, Chenwei Wu, Hui Shen, Zesen Zhao, Hei Ting Una Chan, Xueshen Liu, Zhen Zhang, Liyue Shen, Zechun Liu, Z. Morley Mao
VX‑One is a small multi‑head neural network that outputs typed, calibrated decision primitives— a calibrated gate, a closed‑set choice with an explicit unknown option, and an ordinal score—trained in a multi‑task, temperature‑calibrated manner. In experiments on UNSW‑NB15, CIC‑IDS2017, and EMBER2024, VX‑One outperforms traditional models on detection metrics and achieves high recall when combined with a signature engine in a two‑phase cascade, though its calibration degrades under temporal drift and it struggles with multi‑class attribution. The study provides a detailed empirical assessment of where such typed, calibrated decisions aid perimeter triage and where they fall short.
By Juan S. Santillana
NodeSynth is a method that generates socially aligned synthetic data by combining a fine‑tuned taxonomy generator with credible sources and extracted sensitive attributes to produce queries that reflect real societal contexts and nuanced harms. In evaluations across ten open‑weight and proprietary large language models on Medical Advice and Self‑Harm tasks, NodeSynth outperformed generic synthetic baselines, structured synthetic benchmarks, and authentic human‑authored queries, inducing higher failure rates that highlight model vulnerabilities. Ablation and variance attribution studies show that the granular Level 3 taxonomic expansion and sociotechnical grounding of queries drive these failure rates, underscoring the importance of nuanced data for high‑stakes evaluation.
By Qazi Mamunur Rashid, Xuan Yang, Zhengzhe Yang, Yanzhou Pan, Erin van Liemt, Darlene Neal, Kshitij Pancholi, Jamila Smith-Loud
The paper introduces a new approach to learning disentangled representations by directly sculpting the aggregate posterior using the non‑parametric Maximum Mean Discrepancy (MMD) and programmable priors. It demonstrates that traditional Variational Autoencoders and Adversarial Training consistently fail to enforce a factorized latent space, as quantified by the Latent Predictability Score (LPS). The proposed method achieves state‑of‑the‑art unsupervised statistical independence, a 70% reduction in entanglement, optional semi‑supervised semantic alignment, and precise shaping of the posterior without compromising reconstruction quality.
By Quentin Fruytier, Akshay Malhotra, Shahab Hamidi-Rad, Aditya Sant, Aryan Mokhtari, Sujay Sanghavi
The paper presents a method for synthesizing and verifying Deterministic Register Automata (DRAs) over ordered data sequences such as time series and sensor streams. It introduces a local‑search synthesis procedure and an exact robustness checker that runs in polynomial time for a fixed number of registers, reducing robustness to a shortest‑path problem. The framework, implemented in RAExc, is applied to learn interpretable surrogate models for neural sequence classifiers, providing statistical guarantees of fidelity and local stability while verifying robustness on benchmark languages and sequentialized MNIST.
By Chih-Duo Hong, Hongjian Jiang, Anthony W. Lin, Oliver Markgraf, Julian Parsert, Tony Tan
LIFT introduces Lift-Bench, a new instruction‑following benchmark that asks language models to draft a document based on a topic and a structured outline, with constraints scoped to specific sections for finer control. The benchmark includes both verifiable and judged constraints and a training counterpart, Lift‑Train, designed to reduce reward hacking by providing both main‑task and instruction‑following constraints. Experiments show that state‑of‑the‑art models struggle with the harder subsets, and training on Lift‑Train is less prone to reward hacking than traditional IF training.
By Eran Goldstein, Yuval Belfer, Yonatan Belinkov, Daniel Gissin
Omnilingual-MsASR is a new benchmark designed to evaluate omni‑modal large language models on long‑form, multi‑speaker automatic speech recognition across six languages, with recordings up to 30 minutes and up to 20 speakers. The benchmark provides parallel audio and video inputs to test cross‑modal consistency and the models’ ability to use complementary cues from both modalities. Evaluations of leading omni‑modal LLMs show significant performance gaps, with many models achieving word error rates above 100%, underscoring challenges in temporal localization, speaker disambiguation, and cross‑modal synergy.
By Yuan Ge, Pei Zhang, Xize Cheng, Lingzhou Mu, Mingyang Wang, Jin Xu, Baosong Yang, Shengxiang Gao, Jingbo Zhu, Zhengtao Yu, Tong Xiao
RCSD (Reinforcement and Continual Learning for Speculative Decoding) is a new training framework that improves speculative decoding for large language models. It aligns the draft model’s optimization with the goal of maximizing acceptance length by using accepted and rejected sequences from inference to create implicit preference pairs, while masking loss to avoid penalizing target‑aligned tokens. Across multiple benchmarks, RCSD boosts average acceptance length by 35% over HASS and delivers 1.35× faster end‑to‑end inference without sacrificing generation quality.
By Mingkang Liu, Weiming Meng, Peipei Jiao, Jiawei Zhang, Tianyuan Liu, Junyan Li, Lichen Pang
The paper introduces a framework that links the foundational capabilities of large language models (LLMs) to their agentic performance by measuring operation-level gaps. It constructs foundational profiles from benchmark data and operation profiles from real interaction trajectories, then quantifies how much observed operation performance deviates from predictions based on foundational capabilities. Experiments on fifteen model services in the τ²-Bench demonstrate that foundational profiles alone only partially predict agentic outcomes, while incorporating operation profiles significantly improves task-score predictions and reveals consistent execution gaps across task subsets.
By Wei Guo, Yuanjian Xu, Jianing Hao, Yongyao Duan, Ran Chen, Zhong Li, Guang Zhang
The paper introduces MAPS, a Macro-Adaptive Parallel Reasoning framework designed to regulate search momentum in Retrieval-Augmented Generation (RAG). MAPS retains retrieved knowledge and a refined reasoning state to reduce historical influence on sub-query generation, then expands the sequential reasoning process into a parallel paradigm that identifies complementary retrieval directions and consolidates evidence. Experiments on six question-answering benchmarks show that MAPS consistently outperforms representative RAG methods across different backbone models.
By Xinze Li, Dingwen Zhang, Zhenghao Liu, Yukun Yan, Shuo Wang, Yu Gu, Ge Yu, Maosong Sun
The paper introduces CS‑SeLeRoSa, a synthetic dataset of over 600k code‑mixed Romanian sentences with English, German, French, and Italian, generated via four methods (random replacement, phrase translation, dependency parsing, and large language models). It evaluates multilingual large language models on satire detection in this dataset and a human‑annotated gold benchmark, CS‑SeLeRoSa‑Gold, showing that newer models like Gemma 4 31B outperform older open models and achieve competitive zero‑shot performance, with fine‑tuning further improving results.
By R\u{a}zvan-Alexandru Sm\u{a}du, Ioana Buhnila, Daniela Tanase, Dumitru-Clementin Cercel, Florin Pop, Mihaela-Claudia Cercel