The paper investigates whether progress on spatial reasoning benchmarks translates into improved navigation performance. It finds a gap between benchmark-oriented spatial specialization and real navigation, and proposes aligning spatial supervision with navigation goals, phases, and decision learning. The authors introduce Spatial-Nav-100K, a two-stage fine-tuning approach, and Spatial-NPD, a teacher–student framework that eliminates explicit spatial reasoning at inference, achieving state‑of‑the‑art success rates on HM3D and MP3D with efficient GPU usage.
By Xun Huang, Shijia Zhao, Rongsheng Qu, Jiayuan Li, Xin Lu, Weixin Li, Chenglu Wen, Cheng Wang
WanPE is a 397‑B parameter prompt‑enhancement model that learns director‑level cinematic planning from 1.05 M real‑world videos. It generates shot‑level cinematic plans through video‑grounded reverse construction and uses Semantic‑Consistency GRPO (SC‑GRPO) to maintain user intent across shots and time. In evaluations, WanPE improves human preference over raw prompts by up to 50.86 points for 30‑second videos and outperforms commercial offerings for shorter durations.
By Yubo Zhu, Yawen Shao, Ziyun Dai, Zixun Fang, Kai Zhu, Siyang Sun, Haolan Xue, Chuxin Wang, Tingyu Weng, Jingming Luo, Chen Shi, Lianghua Huang, Yufeng Ai, Yuzheng Wang, Wenyuan Zhang, Yu Shang, Yuxiang Bao, Zoubin Bi, Jie Xiao, Jinbo Xing, Jiaxing Zhao, Chongyang Zhong, Hengjian Chen, Chenwei Xie, Akide Liu, Zhehan Kan, Yu Liu, Wei Zhai, Sheng Zhong, Wei Tong
The Representation World Model (RWM) learns states, transitions, and executable plans directly within a representation space, bypassing traditional explicit dynamics models and action-space search. It uses inverse-dynamics supervision along latent paths to shape the representation geometry, enabling direct planning by constructing a latent path between current and goal states and recovering actions via inverse dynamics. Experiments on continuous-control benchmarks and robotic manipulation tasks demonstrate RWM’s effectiveness and potential for complex embodied control.
By Yijun Yuan, Weicheng Zheng, Weibang Wang, Minghui Qin, Chang Sun, Junhao Huang, Kenan Li, Anmin Liu, Yicheng Yao, Hang Zhao
ToCo-Mesh is a dynamic mesh reconstruction framework that preserves topology consistency while achieving high‑fidelity geometry from multi‑view temporal images. It uses a dual‑mesh representation, coupling a canonical template mesh to time‑varying coarse guide meshes via barycentric parameterization, and applies error‑driven split‑and‑merge operations on the template to refine detail. A Surface‑Aligned 2DGS module anchors flattened Gaussians to mesh faces, using rendered normals to guide fine‑tuning and suppress surface irregularities, resulting in photorealistic rendering.
By Chuanjin Fan, Wenjie Chang, Aibing Li, Bingzhou Wang, Wenfei Yang, Tianzhu Zhang
LeafTrackNet is a deep learning framework that combines a YOLOv10-based leaf detector with a MobileNetV3-based embedding network to track individual leaves over time. The authors introduce CanolaTrack, a large benchmark dataset of 5,704 RGB images with 31,840 annotated leaf instances from 184 canola plants. When evaluated without prior fine‑tuning, LeafTrackNet outperforms existing methods on CanolaTrack, KOMATSUNA, and MSU‑PID datasets, achieving HOTA scores of 88.03, 87.33, and 74.20 respectively.
By Shanghua Liu, Majharulislam Babor, Christoph Verduyn, Breght Vandenberghe, Bruno Betoni Parodi, Cornelia Weltzien, Marina M. -C. H\"ohne
WaterClear-GS introduces a physics-informed Gaussian splatting method tailored for underwater 3D reconstruction and appearance restoration. It models underwater degradation as intrinsic Gaussian attributes and employs a dual-branch optimization that separates clean appearance from degradation while preserving photometric consistency. The approach incorporates depth-guided geometry regularization, perception-driven supervision, exposure constraints, adaptive regularization, and spectral regularization, achieving strong novel view synthesis and image restoration performance at over 160 FPS.
By Xinrui Zhang, Yufeng Wang, Zesheng Wang, Dacheng Qi, Wenrui Ding, Shuangkang Fang
The paper introduces AnchorCache, a parameter‑free token‑layout and attention‑mask design that decouples reference tokens from the target in in‑context diffusion transformers. By inserting static text anchors, the method conditions reference representations on the instruction during cache construction, enabling exact key‑value reuse across denoising steps. To restore quality lost by this structural change, the authors employ teacher‑forced velocity distillation followed by a brief on‑policy stage, achieving full‑attention quality while delivering up to 6.40× speedup in diffusion transformer inference across image, speech, and video benchmarks.
By Yangshuai Liu, Zheming Li, Jiaao Li, Kang He, Ziliang Lai, Zhitai Liu, Chengru Song
The paper extends the neural network verification framework to graph neural networks by introducing GraphStar sets, which model uncertainty over both node and edge features. This allows sound propagation of linear message‑passing operations and ReLU nonlinearities for GCN and GINE layers. Experiments on power system tasks (PF, OPF, CFA) and graph classification benchmarks (ENZYMES, PROTEINS) show that the method, called GNNV, yields tighter robustness guarantees than CORA and provides, for the first time, edge‑aware guarantees for GINE‑based models under joint node and edge perturbations.
By Anne M. Tumlin, Ben Wooding, Zhenxuan Shao, Diego Manzanas Lopez, Tyler Derr, Taylor T. Johnson
COILD is an Indic‑centric parallel corpus that contains over 1.16 million human‑translated and verified sentence pairs across 20 Indian language pairs from four language families. The corpus is sourced from original Indian language materials in eight domains, and a 2,000‑sentence domain‑centric benchmark is provided for consistent multilingual evaluation. Experiments with IndicTrans2‑Distilled and NLLB‑200 show consistent improvements in automatic metrics and human judgments, underscoring the value of high‑quality Indic‑centric data.
By Kshetrimayum Boynao Singh, Nitin Kumar Mishra, Palash Pratim Dutta, Atai Waris Khan, Aparna Kaushik, Avinash Kumar, Deeksha, Deepak Kumar, Saroj Kumar Jha, Saloka Sengupta, Anansa Roy, Umalatha Kannoth, Saifulla Samar, Meena Sharma, Manpreet Kaur, Jyoti Sharma, Ashwini Vaidya, Muralikrishna SN, Md Shad Akhtar, Poonam Bansal, Amita Dev, Sanasam Ranbir Singh, Samit Bhattacharya, Tanmoy Chakraborty, Asif Ekbal
The paper introduces MORE-PLR, a method that tackles the partial label ranking problem by employing multi-output regression. It uses an encoder to transform incomplete rankings with ties into regression targets during training, and applies post‑hoc layers during inference to convert regression outputs into bucket orders. Experiments show that this framework competes with state‑of‑the‑art partial label ranking methods.
By Santo M. A. R. Thies, Juan C. Alfaro, Viktor Bengs
The paper introduces BaCVA, a Bayesian Context-aware personalized Value Alignment method for large language models. It treats personal values as priors and context-dependent preferences as posteriors, estimating contextual value salience from normative responses and using a dual-view personalization module to infer posterior preferences. Experiments show BaCVA outperforms strong baselines, offering more accurate and data‑efficient personalized value alignment.
By Hanze Guo, Aixuan Song, Jing Yao, Xiangxu Zhang, Xiaoyuan Yi, Xing Xie, Xiao Zhou
RECLAIM is a benchmark that tests whether AI agents can reproduce results from 100 NeurIPS 2025 papers, with each paper pre‑defined with a target result, success criteria, and GPU‑hour budget. Papers are categorized into three difficulty tiers—Run, Retrain, and Reimplement—based on the availability of code, data, and weights. In experiments, the best agent succeeded on only 41% of Run‑tier papers, 27% of Retrain‑tier, and 15% of Reimplement‑tier, often stopping early with unused budget and frequently making errors such as writing code without verifying against the paper’s numbers.
By Mithil Salunkhe, Haochen Ding, Samridhi Verma, Volodymyr Kindratenko
The paper introduces TRACER, a multi‑turn user simulator that models evolving user intent and aligns simulated behavior with real interaction trajectories. TRACER is trained first with supervised fine‑tuning on real dialogues and then with reinforcement learning that uses hierarchical outcome‑ and trajectory‑level rewards to address reward sparsity and credit assignment. In real customer‑service sessions, TRACER‑7B outperforms the best baseline by 11.4 conversion F1, achieves the lowest group‑level conversion‑rate error and semantic trajectory distance, and generalizes to out‑of‑distribution scenarios, while human Turing tests show its conversations appear natural. The authors also present the Dynamic Marketing Benchmark, which evaluates both persuasion effectiveness and response quality of large language models through simulated interactions, demonstrating that higher response quality does not always lead to higher conversion rates.
By Geng Chen, Ruotong Pan, Zhirui Yang, Qiqi He, Jiawei Chen, Zhang Yunfei, Chongyuan Chen, Minxuan Lv, Zheng Yang, Win-Bin Huang, Xiangyu Wu, Wenwu Ou
IndicBankBench is a 799‑case benchmark designed to evaluate the safety and reliability of language model assistants in Indian retail banking. It covers five operational domains, a capability/refusal domain, and twenty primary axes, assessing each case at four stages: safety, action and tool use, response adequacy, and advisory quality. The benchmark uses deterministic safety checks, a narrow resolver for ambiguous confirmation‑before‑write scenarios, and an LLM judge for semantic response adequacy, reporting strict pass rates that reveal a gap between strict reliability (43.7%–58.2%) and at‑least‑once success (60%–74%).
By Suvradip Paul, Chandra Bhushan, Harsh Sharma, Nitin Kukreja, Yatharth Dedhia, Keyur Doshi, Prashant Devadiga
SkinAgent AI is a multimodal, safety‑grounded framework designed for non‑diagnostic skincare support. It routes visual concerns (Acne, Pores, Wrinkles), estimates skin type from photographs, and provides count‑informed acne severity, all while grounding recommendations in a database and enforcing deterministic safety, privacy, and evidence checks. The system achieved high accuracy in routing and skin‑type estimation, and demonstrated no safety or privacy violations in controlled tests, though tool‑selection errors and incomplete product grounding remain.
By Muhammad Muhtasim Shahriar, Abdullah Mohammad Sayem, Tze Hui Liew, M. F. Mridha, Md. Mahiuddin
The paper introduces a new multi‑agent micro‑benchmark called Delay‑of‑Gratification, modeled after the Stanford marshmallow experiment, to evaluate large language models (LLMs) in long‑horizon, multi‑turn interactions. In the benchmark, ReAct agents use a per‑step “raise a question” tool under various constraints—social context (broadcast vs. isolated), persona traits (age, hedonic drive), and tool‑use policy (mandatory vs. optional). Across 19,200 trajectories, the study finds that most agents exhibit an early impulse to “eat,” only 75.9% persist to the end, and factors such as isolation and hedonic drive significantly influence survival and questioning behavior, with ablations showing that removing hedonic drive and age can improve completion rates.
By Olga Manakina, Igor Bogdanov, Chung-Horng Lung
The paper introduces CARE, a contrastive accuracy reward estimation method that adaptively adjusts reasoning length for large language models. By comparing beneficial length adjustments from online sampled responses, CARE applies adaptive length rewards within Group Relative Policy Optimization without extra hyperparameters or inference cost. Experiments on multiple reasoning benchmarks show that CARE improves Pass@1 by up to 4% while reducing reasoning length by 37%, achieving higher token efficiency.
By Zhengdong He, Yunfan Zhou, Jianguo Yao, Haibing Guan, Xijun Li
ENDOPROMPT is a white‑box attack that learns prefixes to degrade the utility of language model responses without producing harmful content. It uses clean victim continuations as pseudo‑references, performing local search and preference fitting on unlabeled instructions to distill a generator that outputs a single prefix per request. In experiments across four instruction‑tuned models and seven benign benchmarks, ENDOPROMPT reduces utility by an average of 26.8 percentage points, with 27 of 28 test cells showing negative change.
By Qingyu Wu, Zeyu Feng, Yongda Yu, Yuzhe Luo, Hua Cheng
PrivDrift is a benchmark that tests whether user‑disclosed secrets can still be recovered by large language models after the conversation shifts to unrelated topics. It includes 1,000 controlled multi‑turn dialogues with seeded secrets, topic‑drift turns, and standardized extraction probes. Experiments on three LLMs with extended context windows show that dialogue‑level leakage remains substantial—between 38.7% and 54.6%—and is influenced by model, secret type, and persuasion intensity, while additional topic drift does not reliably reduce leakage.
By Luciano Maldonado
SciWalker is a framework that automatically synthesizes scientific coding problems by sampling operator chains from scientific library interfaces and using execution feedback to refine generated problem statements, solutions, and tests. It produces 8,178 high‑quality problems across five scientific domains and 32 subdomains, and training a large language model with these problems improves its scientific coding accuracy by nearly 10 percentage points. The approach combines structured workflow composition with verification and quality review to enable scalable, scientifically grounded task generation.
By Chenxi Li, Wenxuan Zeng, Yun Luo, Fangchen Yu, Peng Ye, Yu Cheng, Jun Zhang