The paper introduces MARI, a Memory-Augmented Recommendation system that uses a Decision Memory Bank to store Structured Decision Memories (SDMs) capturing users’ goals, constraints, and trade-offs. Unlike prior methods that flatten heterogeneous behaviors into token sequences, MARI retrieves relevant SDMs to enhance LLM reasoning, improving performance on next-item prediction and a new Difficult Choice Prediction task while keeping latency low. The approach also provides interpretable, human-readable insights into user decision-making.
By Leikun Liang, Guoshuai Wang, Xingsheng He, Yushan Han, Yunyi Xuan, Xiaoxiao Xu, Lin Qu
The paper introduces Residual Advantage (RA), a new guidance method for reinforcement learning with verifiable rewards (RLVR) that treats the teacher–student probability residual as a bounded one‑step reward, subtracts the student’s state value to form an advantage, and centers it within each response. RA preserves the verifier’s mean label while redistributing credit among response steps, and its extension CoRA updates the teacher model using verifier advantages to adapt guidance to the student’s attempts. Experiments with Qwen3 models show that RA combined with GRPO or REINFORCE++ consistently outperforms teacher‑only on‑policy distillation across three mathematical benchmarks, improving macro Avg@8 and Pass@8 scores, with CoRA providing additional gains.
By Xiaobing Chen, Zhiqi Pang
The paper introduces Propose-Then-Project Unlearning (PTP-U), a method for large language model unlearning that first weakens target knowledge associations and then projects non-target outputs back to the original distribution to recover capabilities. PTP-U aims to satisfy forgetting constraints while preserving fluent generation and performance on non-target tasks. Experiments on three benchmarks show PTP-U achieves a strong forgetting‑retention trade‑off, with 81.22%–91.03% forgetting and an average of 94.20% non‑target utility.
By Xunlei Chen, Qinghui Gong, Jingkun Xue, Qihe Liu, Shijie Zhou, Fei Ye
HarnessSQL introduces a harness‑native post‑training framework that keeps the full stateful, multi‑turn interaction structure of real‑world database agents during both supervised fine‑tuning and reinforcement learning. By building isolated executable database environments with hidden execution oracles and retaining only verified trajectories for full‑sequence SFT, the method enables execution‑reward RL that aligns training with inference. Experiments on Spider 2.0‑SQLite show dramatic gains in execution accuracy for compact models, and the approach transfers effectively to out‑of‑distribution interactive benchmarks such as BIRD‑Interact and LiveSQLBench.
By Haolin Yang, Jipeng Zhang, Jian Xie, Shuaishuai Gong, Sirui Han, Yike Guo
FastBench is a new benchmark that tests streaming Video‑Large‑Language Models (VLMs) on high‑dynamic real‑world video streams, where events occur rapidly and cannot be captured by sparse 1–2 FPS sampling. The benchmark uses a trajectory‑grounded pipeline that generates QA pairs from high‑FPS clips, filters questions answerable at 2 FPS, verifies answers with SAM3 and CoTracker3 trajectories, and incorporates three rounds of human inspection, yielding 306 QA pairs across eight domains and multiple temporal scopes. Experiments show that even the strongest model, Gemini‑3.5‑Flash, achieves only 50.7 % accuracy, and that denser sampling improves performance but saturates when history is compressed; a training‑free baseline, ProactiveFrame, outperforms uniform sampling but still falls short of oracle‑guided focusing.
By Yuxuan Hu, Weikang Shi, Yang Bo, Xudong Lu, Xintong Guo, Shuhan Li, Yuyang He, Huankang Guan, Peiwen Sun, Yunqiao Yang, Wenbo Li, Rui Liu, Hongsheng Li
The paper introduces MESSENGER, a memory‑enhanced sequential scene flow estimation framework that uses a memory buffer to store past flow estimates and latent states. By retrieving and correlating these stored flows for each new frame, the method performs next‑frame forecasting and incorporates an uncertainty‑aware reweighting module to reduce error accumulation. Experiments on nuScenes and Argoverse 2 show significant improvements, reducing EPE3D by 71.6% and 67.7% respectively in long‑horizon predictions.
By Jiuming Liu, Jianing Li, Mengmeng Liu, Hongyang He, Hesheng Wang, Per Ola Kristensson
The paper introduces SAVER, a dense-to-sparse post‑training framework that leverages dense video views as training‑time references to improve sparse‑frame inference. By optimizing paired dense and sparse views with grounding and reliability‑gated reference rewards, SAVER learns to preserve task‑relevant temporal evidence while using far fewer frames. Evaluated on three temporal grounding and six video question‑answering benchmarks, SAVER consistently outperforms dense‑frame baselines, matching or surpassing them with substantially reduced frame budgets.
By Wenfang Sun, Yingjun Du, Cees G. M. Snoek
LVSPM is a generalizable model that jointly estimates camera poses and synthesizes novel views from uncalibrated image collections. It is trained only with RGB images and pose supervision, avoiding dense 3D ground truth, and uses test‑time training layers to scale to hundreds of input views. On datasets such as RealEstate10k, Co3Dv2, and DL3DV, LVSPM outperforms VGGT in pose estimation across 16–256 views and achieves state‑of‑the‑art pose‑free novel view synthesis quality, maintaining high performance even as scene scale increases.
By Xi Chen, Yachi Zhang, Linghao Chen, Minghua Liu, Hao Su, Zexiang Xu, Xiaoshuai Zhang
The paper addresses the challenge of identity‑preserving video generation when a single reference portrait cannot capture the subject’s appearance across varied expressions. It introduces a compact, expression‑diverse reference gallery and a new evaluation metric that measures identity similarity for mild, intense, and extreme expressions. By extending the Stand‑In model to use these galleries and curating diverse face crops from training videos, the authors demonstrate improved identity preservation, especially for extreme expressions, on both real and synthesized reference sets.
By Tianwen Fu, Wenbin Teng, Gonglin Chen, Junyi Ouyang, Haolin Xiong, Yajie Zhao
The paper introduces TopoEnhance, a framework that improves the reliability of scene topology graphs for autonomous driving by treating topology enhancement as a denoising reconstruction task. It learns to recover structural consistency from corrupted ground‑truth graphs, thereby correcting logical inconsistencies and unreliable connections. Experiments show consistent gains in both continuous topology metrics (TOP score) and discrete connectivity (Topology Jaccard Similarity) across multiple baselines without retraining.
By Xiaoqi Wang, Dingyi Zhaung, David Paz, Wenbin He, Yucai Bai, Peng Zhou, Rui Zhang, Jinhua Zhao, Liu Ren
TKCAM is a new framework for generating controllable camera motion conditioned on text and sparse keyframes, using a 12‑dimensional kinematic representation and a two‑stage masked transformer with residual vector quantization. It allows users to provide free‑form text prompts and RGB observations at selected timestamps, enabling temporally localized visual guidance for coherent trajectory synthesis. The authors also introduce the RealEstate10K‑Cap dataset and a Universal CLaTr Evaluator, showing that TKCAM outperforms recent baselines on FID, text‑motion matching, and retrieval metrics.
By Haozhe Yang, Zhiyang Dou, Zekai Gu, Cheng Lin, Wenping Wang, Yuan Liu, Taku Komura
AutoAdapt is a framework for reliable few-shot adaptation of large pretrained clinical models to new hospitals. It introduces an Adapter that defines a flexible space of adaptation recipes and an Automator that combines these recipes based on evidence from limited adaptation patients. A reliability rule selects the most effective strategy across available patients, enabling state‑of‑the‑art performance on critical care, emergency care, and diagnostic datasets with only a few labeled patients.
By Song Wang, Jie Peng, Davis Hobley, Zachary Plotkin, Tianlong Chen
ChainDance is a scalable framework for generating coordinated group dance choreography from music, overcoming the fixed‑size limitation of prior transformer‑based methods. It reformulates the task as a Chain‑of‑Dancers, using a frozen single‑dancer diffusion backbone plus two lightweight modules—Role‑Aware Text Encoder (RATE) and Group‑Aware Motion Encoder (GAME)—to condition on per‑dancer semantics and aggregate previously generated dancers. Experiments on AIOZ‑GDance show that ChainDance delivers state‑of‑the‑art motion quality and group coordination while preserving individual identities, using 3–4× fewer parameters and 3–6× less training time than existing approaches.
By Jing Xu, Cunjian Chen, Qiuhong Ke
RT-DETRv4 introduces a cost‑effective distillation framework that leverages Vision Foundation Models to improve lightweight real‑time object detectors. The method uses a Deep Semantic Injector to embed high‑level VFM representations into deep detector layers and a Gradient‑guided Adaptive Modulation strategy to adjust semantic transfer intensity based on gradient norms. These techniques yield consistent performance gains across DETR‑based models without adding inference overhead, achieving state‑of‑the‑art COCO AP scores at high FPS.
By Zijun Liao, Yian Zhao, Xin Shan, Yu Yan, Chang Liu, Lei Lu, Xiangyang Ji, Jie Chen
PointLearner is a point cloud representation learning network inspired by biological vision. It uses a point-focused attention mechanism that mimics foveal vision by competitively weighting local neighbors against spatially downsampled features extracted via learnable inducing points. Additionally, a context-scan state space emulates eye saccades, inferring global scene structure through a Hilbert‑curve guided scan path, enabling robust performance on various point cloud tasks.
By Kanglin Qu, Pan Gao, Qun Dai, Yuanhao Sun
FastJEV addresses redundancy in JEV models by reusing context and intermediate states during candidate evaluation, eliminating repeated token matching and full backbone execution. It introduces shared context anchoring and candidate prefix sharing to preserve necessary states for branching, and applies decision‑guided pruning based on score changes from a small unlabeled set. Experiments on OmniJev models across five benchmarks show that FastJEV can reduce candidate depth by up to ~45% while maintaining over 94% of original task scores.
By Jie Ma, Jie Gao, Yihang Liu, Zhike Qiu, Junle Li, Chongyi Zhuang, Jiayi Ji, Xiaoshuai Sun
OX-NeRF is a new framework for 3D X‑ray tomography reconstruction from ultra‑sparse views, combining cross‑scene feature learning with scene‑specific optimisation. It uses a CNN to extract shared features, maintains multi‑resolution hash grids for each scene, and fuses these with an MLP that is jointly optimised end‑to‑end. Benchmarks on parallel‑beam and cone‑beam X‑ray datasets show that OX‑NeRF achieves significantly higher reconstruction accuracy than existing radiance field methods for scenes with ten or fewer views.
By Thomas Welsch, Min-Hsin Tu, David J. Chapman, Daniel E. Eakins
S$^3$Geo introduces a structure‑semantic synergistic learning framework for cross‑view geo‑localization, combining a Decoupled Query Pooling module to capture local structural patterns, an optimal‑transport based contrastive learning scheme for soft cross‑view correspondences, and a Semantic Knowledge Distillation strategy from a frozen CLIP teacher to embed semantic priors. The method jointly models fine‑grained structural and semantic information, enabling robust matching between images from vastly different viewpoints such as drone and satellite views. Experiments on University‑1652 and SUES‑200 datasets show that S$^3$Geo outperforms state‑of‑the‑art approaches without increasing inference complexity.
By Ziqian Mo, Hill Zhang, Haosheng Tan, Ling Li, Jiaheng Wei
VESSI is a Vision‑Language Model (VLM) framework that enhances automated video surveillance by converting salient visual features into textual descriptions for prompt‑driven interrogation. The authors evaluate four state‑of‑the‑art VLMs and introduce the Composite Model Utility Score (CMUS) to assess performance on largely unlabeled datasets. Results show that the best model flags relevant activity in over 66% of videos while cutting review time by more than 85%, and the CMUS ordering aligns with manual false‑positive rates, supporting its operational use.
By Saverio Cavasin, Pietro Tedeschi, Mattia Tamiazzo, Alessandro Brighente, Simone Milani, Mauro Conti
The paper introduces Vantage, a training‑free, model‑agnostic framework that pairs a vision‑language model with a 3D foundation model to perform Seek‑and‑View reasoning for multi‑view spatial understanding. Vantage first analyzes a question to plan a relevant viewpoint, then augments reasoning with geometry‑grounded visual evidence from that viewpoint. Experiments across six VLMs on five benchmarks show consistent improvements without fine‑tuning, demonstrating that revealing spatial evidence through view‑grounded reasoning reduces reliance on language‑based cross‑view alignment.
By Qixiang Chen, Cheng Zhang, Fucai Ke, Chi-Wing Fu, Jianfei Cai, Jingwen Ye