The paper investigates how to estimate confidence for activation oracles—language models that translate internal activations of other models into natural language. Five confidence estimation methods are compared across four oracles from Qwen and Gemma families (8B–27B parameters) on a secret‑word recovery task. Results show that when a hypothesis space is available, scoring candidate answers directly yields the best accuracy and confidence separation; when free‑text answers are required, sampling agreement over twenty outputs provides the most calibrated confidence, while a rescaled probability from a calibration set performs comparably with fewer generations. Asking the oracle to output a numeric confidence score proves ineffective.
By Federico Torrielli, Peter Schneider-Kamp, Lukas Galke Poech
Foundation4D is a unified feed‑forward model that performs geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi‑camera driving video. It uses a shared backbone with factorized spatiotemporal attention to merge observations across cameras and time, and a video inference pipeline that refines camera trajectories via pose graph optimization and assembles object tracks. The authors also provide a data annotation pipeline and release the PAI‑Foundation4D benchmark, achieving state‑of‑the‑art results on depth, detection, and occupancy metrics and demonstrating that co‑training detection with geometry and occupancy improves performance.
By NVIDIA, :, S\'ergio Agostinho, Maximilian Apodaca, Emmanuel Attia, Guillem Braso, George Dong, Ilya Fedorov, Michele Fenzi, Gang Huang, Jiahui Huang, Shengyu Huang, Bingxin Ke, Laura Leal-Taix\'e, Kuan-Ting Lee, Mingxuan Liu, Tian Liu, Tim Meinhardt, Aljo\v{s}a O\v{s}ep, Naveen Kumar Rai, Kefei Ren, Cristiano Saltori, Keqiang Wu, Ning Xu, Fei Xue, Jin-Bey Yu, Lei Zhang, Qunjie Zhou
The paper introduces SKD-ETR, a one‑shot federated unsupervised domain adaptation framework that uses Smoothed Knowledge Distillation and Exponential Moving Average Teacher Refinement. It trains a student model on an unlabeled target domain with smoothed pseudo labels from an ensemble of source models, then refines each source model by moving its parameters toward the student. This bidirectional refinement improves pseudo‑label quality and student generalization while eliminating complex aggregation and multiple communication rounds, achieving competitive performance on OfficeHome, Office‑Caltech, and DomainNet with low communication and computation overhead.
By Ali Abedi, Q. M. Jonathan Wu, Ning Zhang, Farhad Pourpanah
The paper introduces MIME, a task‑agnostic visual in‑context learning (vICL) dataset that removes textual shortcuts by omitting explicit rules linking demonstration images to answers, forcing models to infer the rule visually. Fine‑tuning vision‑language models on MIME boosts vICL accuracy by 8.1–18.0 points across three VLMs while preserving single‑image performance and improving multi‑image and text‑only in‑context learning. The authors also propose DTAD, a technique that randomly drops demonstration labels during training to encourage visual inference, achieving comparable accuracy with more consistent attention to demonstration images and outperforming frontier models on the TrueMICL benchmark.
By Nimrod Shabtay, Asaf Yehudai, Avihu Dekel, Eli Schwartz, Raja Giryes
The paper introduces RetinaGroundBench, a new benchmark for visual grounding of retinal OCT images, featuring 815 B-scans annotated with 4,537 bounding boxes for 13 biomarkers across 8 diseases. It evaluates 27 multimodal large language models (MLLMs) and finds that visual grounding is largely unsolved, with the best model achieving only a macro F1 of 23.7%. The authors also present RetGrounder, an 8‑B parameter MLLM fine‑tuned for OCT, which raises grounding F1 to 33.9%, a 43% relative improvement.
By Ronald Fecso, Jos\'e Morano, Botond Fazekas, Paul Widmann-Sedlnitzky, Dorota Himel, Laetita Hinterhuber, Virginia Mares, Andreas Pollreisz, Hrvoje Bogunovi\'c
UniVOE is a removal‑driven framework that trains both video object removal and injection models. It first builds an effect‑aware remover by leveraging video inpainting priors and fine‑tuning on object‑present/object‑absent pairs, then uses the remover to generate large‑scale paired data for training an injection model that restores objects with their dynamics and interaction effects. A reward‑based post‑training step aligns injection results with human preferences, and experiments show UniVOE outperforms existing open‑source and commercial systems, especially in challenging interaction‑effect scenarios.
By Shuzhou Yang, Lanqing Guo, Lingen Li, Weiqi Li, Zhao Zhong, Liefeng Bo, Jian Zhang
The paper investigates how to best integrate large language model (LLM) guidance into visual question answering (VQA) systems. By prompting the 70‑billion‑parameter Llama 3.3 model to identify visual concepts, distractors, and external facts for each question, the authors inject this guidance into a frozen Vision Transformer (ViT) and a T5 text‑to‑text model via two methods: decision‑level late fusion and attention‑matrix injection. On the OK‑VQA benchmark, the shallow late‑fusion approach outperforms the deeper attention injection, achieving 39.4% exact match versus 31.8%, and a paraphrasing pipeline further expands the training set tenfold, indicating that the depth of integration should match the data regime.
By Aneerban Saha, Vijay Kumar, Sameer Madhav D., Priyam Srivastava, Dhruv Bansal
The paper presents an evidence-driven, multi-agent AI system for head and neck proton radiotherapy planning. It combines a DRL-based planning agent that generates multiple treatment plans with an LLM-based conversational agent that refines the chosen plan through natural language interactions. The system, trained on 21 patients, improves instruction-following and domain-specific reasoning, yielding more clinically appropriate trade-offs in patient-specific plan refinement.
By Qingqing Wang, Liqiang Xiao, Chang Chang
SLoSa is a neuro‑symbolic framework that compresses and fine‑tunes foundation models to create small, low‑power, and safety‑aware robotic task planners. It builds a safety‑aware dataset using classical planners, a teacher LLM, and formal verifiers, then applies supervised and reinforcement‑learning fine‑tuning guided by a verification‑based reward machine. Evaluations on PDDL domains, household simulators, and a desktop robotic arm show that verifier‑guided post‑training boosts success rates to 70.5–76.0% and that INT4 deployment on a Jetson Orin consumes only 17.85–26.41 W of power.
By Weizhe Xu, Jialiang Fan, Essien Taylor, Zijun Wang, Russell Joseph, Fanxin Kong
SyRA-CS is a synthetic-to-real adaptation framework for code‑switched ASR that uses Reptile meta‑learning to create pair‑specific initializations from synthetic CS speech. For each language pair and pretrained model, the system meta‑trains on a synthetic CS‑FLEURS subset and then adapts with just one hour of labeled real target speech. Across four architectural settings and four language pairs, SyRA-CS consistently outperforms continual fine‑tuning, achieving up to 17.7% relative WER reduction for XLS‑R and 5.3–5.8% for other models.
By Fadel Thior, Ahmer Tabassum, Haania Siddiqui, Hanan Aldarmaki, Thamar Solorio
The paper introduces a training‑free method called “Track, Then Bind” for video‑language models to reason about object identity over time. It separates semantic identity binding from temporal identity maintenance by using a promptable video tracker to follow candidate objects and a frozen vision‑language model to bind an object to a track based on visible evidence. The approach achieves state‑of‑the‑art performance on several benchmarks, outperforming fine‑tuned trajectory models and other training‑free agents on tasks such as VET‑Bench, Perception Test cup games, and TOC‑Bench.
By Di Yang, Xiaochang Li, Muyang Liu, Yufei Wang, Yanhai Xiong
The paper introduces the Pitfall Anticipation World Model (PAWM), a selective guidance system for black-box large language model agents. PAWM predicts imminent failures in a bounded multi-step future and intervenes only when a potential pitfall is detected, reducing costly and often uninformative step-by-step predictions. Trained first with supervised fine-tuning on synthetic pitfall-labeled data and then co-trained with a black-box LLM, PAWM boosts Claude Sonnet 4.5’s performance by 24.3% on ScienceWorld and 7.7% on WebArena-Lite, outperforming several baselines.
By Jiongxiao Wang, Han Ding, Youzhi Luo, Yawei Wang, Yun Zhou, Linbo Liu, Panpan Xu
WindINR is a latent‑state implicit neural representation that enables fast, continuous high‑resolution local wind queries and corrections in complex terrain using sparse observations. It decouples reusable representation learning from sample‑specific latent correction, training a privileged encoder to infer latent states from high‑resolution data and a deployable predictor to estimate initial states from inference‑time inputs. During inference, only the low‑dimensional latent state is updated via a regularized objective, yielding a 14‑fold speedup over full‑network fine‑tuning while maintaining continuous spatial queries.
By Yi Xiao, Qilong Jia, Hang Fan, Pascal Fua, Robert Jenssen, Xiaosong Ma, Wei Xue
Lean Finder V2 is a semantic search framework for Mathlib that resolves ambiguous user queries by combining adversarial training at the model level with input-level query rewriting. It addresses the common scenario where mathematicians and coding agents issue short, context‑poor queries, which existing search engines struggle to interpret. On the MathlibQR benchmark, Lean Finder V2 achieves at least an 11% relative improvement over prior engines and further boosts retrieval performance through optimized query rewriting, while also enhancing downstream theorem‑proving success when integrated with Claude Code.
By Jialin Lu, Xiaoyu Liu, Hengrui Luo, Rodrigo Stehling, Linxi Liu, Weiran Sun, Yuepeng Wang, Wuyang Chen
The paper introduces OCFormer, a transformer-based neural network for real‑time multi‑agent path planning. It maps a simple latent space where agents evolve independently to the physical space of interacting agents, using a physics‑informed loss that incorporates the Hamiltonian ODE residual and boundary matching. Trained on a distribution of boundary conditions, OCFormer produces near‑optimal trajectories for unseen configurations in about 220 ms and can be adapted to out‑of‑distribution scenarios via lightweight transfer learning.
By Qian Zhang, Shanqing Liu, Jerome Darbon, George Karniadakis
The paper investigates token sparsification for audio transformers, evaluating various pruning and merging methods across multiple datasets and model sizes. It introduces a new approach that blends attention scores with spectral cues from token embeddings, showing that frequency-aware techniques outperform attention-only methods. The study finds that spectrogram patch intensity-based sparsification is particularly effective for longer audio inputs, underscoring the benefit of incorporating frequency-domain information into sparsification strategies.
By Shreya Sajid, Fabrizio Julian Piva, Hiram Rayo Torres Rodriguez, Elif Ceren Gok Yildirim, Murat Onur Yildirim, Joaquin Vanschoren
The study evaluates four open language models (8–27 billion parameters) on the Spanish MIR residency exam, covering 1,552 questions across eight editions. The largest models could rank within the top 600–3,000 candidates and had 41–43 specialties still open, yet abstaining never improved performance because even low‑confidence answers were more often correct than the 25% break‑even threshold. Medical fine‑tuning lowered overconfidence but also reduced accuracy, did not improve the correlation between confidence and correctness, and removing or swapping the provided images did not affect accuracy.
By Miguel Fernandez-de-Retana, Unai Zulaika, Ruben Sanchez-Corcuera, Aritz Bilbao-Jayo, Aitor Almeida
MemPlan is a memory‑conditioned PDDL planning framework designed for partially observable interactive text environments. It keeps the PDDL domain fixed and reconstructs the problem at each step using parsed observations, candidate hypotheses, and interaction memory, employing positive graph memory to assign relevance‑based action costs and negative memory to record refuted assumptions as exclusion constraints. Experiments on TextWorld, ALFWorld, and Robotouille demonstrate that MemPlan improves task completion and step efficiency over language‑only and PDDL‑aware baselines while significantly reducing online token cost compared with LLM‑planning baselines such as ReAct, and its fine‑tuned interface generalizes across task types without task‑specific fine‑tuning.
By Tengwei Song, Robert Hoehndorf
IQuest‑Q1 is a 320 B sparse Mixture‑of‑Experts foundation model that activates 15 B parameters, designed to enhance agentic command‑line interface systems. Its development follows a three‑stage pipeline: stable pre‑training with advanced normalization and curriculum techniques, fine‑grained mid‑training that expands data domains and context length, and multi‑scaffold post‑training that blends supervised fine‑tuning, domain‑specialized reinforcement learning, and multi‑objective policy distillation. The model merges experts from different stages to reduce parameter conflicts, achieving strong coding and general task performance, and is released as open‑source to aid community research.
By A. Jian Yang, An Huang, Bo Dan, Chen Mu, Chengran Yang, Chuan Shi, Feng Chang, Hongxin Li, Hongyuan Tao, Jinxing Zhang, Jujie He, Kun Wang, Lin Jing, Ming Yang, Qicheng Hu, Renyuan Li, Shark Liu, Shawn Guo, Shi Qing, Shi Zhong, Tianyu Zheng, Wei Liu, Wei Zhang, Wenrui Cai, Wenxiao Chen, Xuxing Chen, Yan Xing, Yi Xing, Yifan Luo, Yiran Xing, Yizhi Li, Yongsheng Kang, Yucheng Ding, Yue Song, Yun Qing, Yunqi Liu, Zelong Huang, Zhengmao Ye, Zhentao Fan, Zhiteng Li, Ziling Huang, Bihong Tang, Bingli Wang, Ce Jun, Changhao Chen, Chen Hu, Chen Xiao, Chi Liu, Chunjiang Ge, Danyang Li, Deyang Kong, Dian Qi, Ding Tan, Evan Yang, Fanglin Xu, Fei Bai, Gang Liu, Guochang Li, Hao Tang, Haoran Feng, Haowen Wang, Huahao Hu, Huanjin Yao, Huatong Song, Jia Deng, Jiajun Wu, Jialiang Cheng, Jiayi Liang, Jiayi Wu, Jie Deng, Jiejun Tan, Jing Yang, Jingda Xu, Junhang Cheng, Junxiao Liu, Kai Lv, Kaixiang Pan, Kecheng Lu, Kun Zeng, Kun Zheng, Kunbo Ding, Liang Qiu, Lingyi Sun, Luan Zhang, Mengxiao Lv, Ning Dong, Pei Wang, Peng Zhang, Qingsong Cai, Qixiang Yin, Qiyao Wang, Renbiao Liu, Renfei Dang, Ruihang Li, Ruirui Huang, Shanhui Yin, Shen Jian, Shengjie Fang, Shikai Li, Shinan Zhang, Shiqi Chen, Shuai Liu, Shuangqi Li, Shuo Tang, Shuqing Deng, Sijin Hao, Siwei Wu, Siyue Wang, Tao Hai, Tao Ren, TingYang Chen, Wei Shao, Weicheng Gu, Weihan Peng, Wenlong Miao, Xavier Yu, Xiangyuan Ji, Xiaocheng Li, Xiaodong Huang, Xiaohan Ge, Xiaoxiang Shi, Xiaoyang Zhang, Xiaoyu Yang, Xidong Wang, Xin Li, Xinyu Tang, Xiyuan Yang, Xudong Ma, Ya Tong, Yachao Wang, Yang Tian, Yaxin Du, Yeli Ruan, Yi Wang, Yi Yu, Yibin Feng, Yibo Tan, Yihang Wang, Yiming Ju, Yinda Chen, Yinda Chen, Yingjie Du, Yiwei Liu, Yizhi Zhou, Yuan Wei, Yuli Chen, Yuqi Zhu, Yuwen Li, Yuxi Dong, Yuxuan Zhang, Yuyang Song, Zhange Zhang, Zhaohan Zhang, Zhaoyi Li, Zhen Zhang, Zhengbo Zhang, Zhihang Hu, Zhiwei Yu, Zhixin Xie, Zhiyu Chen, Zhizhao Luo, Zhongyuan Wu, Zhou Fang, Zhu Lin, Zhuocheng Gong, Zixi Zhao, Ziyang Zeng, Ziyi Wang, Ziyu Zhao, Ran Tao, Mingjie Tang, Chuan Hao, Peihao Wu, Jianzhou Wang, Nan Jiang, Yuan Zhi, Bryan Dai
HimmBO introduces a two‑stage Bayesian optimization framework for merging large collections of lightweight adapters in text‑to‑image models. The screening stage identifies an active subset of adapters via level‑set estimation on pre‑generated images, avoiding costly image generation, while the refinement stage optimizes merge coefficients within this reduced subspace using a style‑aware kernel. Experiments show HimmBO outperforms existing methods in similarity and efficiency, maintaining performance even as the adapter collection scales to 1,080 items.
By Jinyuan Mao, Zhihao Shen, Chenxi Liu