The paper investigates how privileged information—such as a teacher’s full solution or reasoning trace—affects on‑policy self‑distillation (OPSD) in language models. Using the AMPLE‑Math benchmark, the authors compare distillation with and without extra teacher views, finding that reference‑free distillation explains most gains for Qwen3‑1.7B, while additional references provide modest benefits, especially for polished solutions. The study also shows that the impact of privileged data depends on the student’s training regime and that altering token‑level supervision can leave student behavior largely unchanged.
By XiuYu Zhang, Wei Chow, Junfeng Fang, Zhenkai Liang, Tat-Seng Chua
GeLaCo is an evolutionary method for compressing large language models by collapsing layers through parametrized weight merging. It uses population-based search with a fitness function that balances similarity of residual updates and language modeling KL divergence, enabling both single and multi-objective compression. The approach yields Pareto-optimal trade-offs between compression and quality, outperforming existing methods in perplexity and generative evaluations.
By David Ponce, Thierry Etchegoyhen, Javier Del Ser
The paper introduces a three‑layer checklist-and-judge framework to evaluate interpreter agents that mediate live conversation across languages. It assesses semantic, pragmatic, and cultural‑social dimensions—naturalness, intent, and social appropriateness—rather than just fidelity, in both single‑turn and multi‑turn settings. Extensive validation shows that conventional MT metrics miss failures in stronger interpreters, and that context, structured instructions, and cultural cues influence communicative success.
By Faiz Ghifari Haznitrama, Alice Oh
The paper introduces Redemption Score (RS), a multi‑modal evaluation framework for image captioning that combines three complementary signals: Mutual Information Divergence for global image‑text alignment, DINO‑based perceptual similarity of cycle‑generated images for visual grounding, and LLM text embeddings for contextual similarity to human references. RS fuses these signals to provide a more holistic assessment, achieving a Kendall‑τ of 58.42 on Flickr8k and outperforming most prior methods. The framework demonstrates consistent performance across Conceptual Captions and MS COCO, offering a robust evaluation that captures both visual accuracy and text quality.
By Ashim Dahal, Ankit Ghimire, Saydul Akbar Murad, Nick Rahimi
AMB3R‑SLAM is a real‑time monocular SLAM system that can reconstruct kilometer‑scale trajectories over 10,000 frames on a single consumer‑grade GPU. It combines a lightweight front‑end for low‑latency tracking with a hierarchical backend that enforces local, mid‑level, and global consistency, avoiding bundle adjustment and thus handling dynamic scenes naturally. The system also supports stereo, RGB‑D, and LiDAR inputs, achieving strong camera tracking performance and reducing absolute trajectory error by over 70% on several datasets, with sub‑meter accuracy when LiDAR is added.
By Hengyi Wang, Lourdes Agapito
The paper introduces Active Cross-View Object Geo-Localization (ActiveGeo), enabling mobile agents to actively select new viewpoints and decide when to stop to improve localization with fewer observations. It proposes the ActiveMoPT framework, which uses a three-stage training process: Multi-View Prompt-Preserving Adaptation, Trajectory-Guided Policy Initialization, and Cost-Aware Policy Refinement with GRPO. The authors also create a zero-shot test set, ActiveGeo-858, and demonstrate that ActiveMoPT outperforms prior methods on MoP-UAV and ActiveGeo-858.
By Shunyu Yao, Xiaohan Zhang, Zhuoran Yang, Haoqi Lai, Qi Ming, Xiaoxi Hu, Hui-Liang Shen, Si-Yuan Cao
GAPrompt++ is a multi-granular geometry-aware prompting method designed to adapt pre-trained 3D vision models to downstream tasks efficiently. It introduces a Point Shift Prompter for multi-scale geometric feature extraction, a Keypoint Prompter for local geometric saliency, and a Prompt Propagation mechanism to embed these cues throughout the model hierarchy. Experiments demonstrate that GAPrompt++ outperforms other prompting-based PEFT methods and even surpasses full fine-tuning while using less than 2% trainable parameters, and the authors provide two new challenging benchmarks for future research.
By Zixiang Ai, Zhenyu Cui, Yufei Guo, Wenwen Qiang, Lei Chen, Jiwen Lu, Jiahuan Zhou
The paper introduces Recency Forcing, a technique that addresses the long‑horizon degradation in autoregressive video generation caused by KV eviction mismatch. By applying a timestep‑dependent bias—Temporal Response Bias—derived from a positional response measure, the method reduces the influence of distant frames during inference without altering context length or training objectives. An exact reformulation, Biased Attention Reparameterization, enables this bias to be applied as a standard FlashAttention call with zero overhead, achieving state‑of‑the‑art long‑horizon generation quality on VBench datasets.
By Tri Cao, Hung Nguyen, Phong Nguyen, Khoi Nguyen
KoUniTalk is a lightweight, articulation‑centered benchmark that unifies Korean and English 3D talking‑face datasets onto a single mesh topology. By retargeting VOCASET and Korean speech‑based 3D data to a shared 1,176‑vertex template, it reduces output dimensionality from tens of thousands to 3,528 dimensions, focusing on the mouth and adjacent lower‑face regions. The benchmark includes 22 speakers, 4,978 sequences, and 642,781 frames, enabling controlled speech‑driven facial articulation training and cross‑dataset evaluation in a compact, identity‑neutral space.
By Hyunjung Chung, Unsang Park
CitySTAR introduces a training‑free framework that transforms billion‑scale urban point clouds into a query‑ready scene graph of open‑vocabulary 3D instances, using CodeLLM‑driven tools to supply multimodal evidence for node attributes and spatial relations. It models target‑context topology with paired hypergraphs and performs bidirectional topology verification for structural disambiguation, followed by a Reflective Cross‑modal Grounding module that integrates topology consistency and 2D visual evidence to decide over a metric‑aware 3D context graph. The authors also present CitySTAR‑3D, a benchmark that enhances semantic coverage, instance completeness, bounding‑box fidelity, and spatial‑relation complexity for city‑scale 3D grounding, and report extensive experiments showing consistent improvements in open‑world urban 3D grounding with strong interpretability and generalization.
By Shuai Zhang, Hongye Hou, Qinghe Liu, Zhuoxiao Li, Dongli Wu, Jing Ou, Yuan Liu, Wufan Zhao
Paint-Anything introduces a unified hex-prompt interface that allows users to specify any 24‑bit hex color for both image generation and editing. The method trains on a new Paint‑500K dataset created from real images with object grounding, perceptual color labeling, and editing‑pair synthesis, and supplements this with pure‑color anchors to address shadow‑induced color inaccuracies. Evaluated on the newly proposed Any Color Benchmark (ACBench), Paint‑Anything achieves significant improvements over the base FLUX.2‑4B model, boosting T2I and editing scores by 85.3 % and 28.3 % respectively, and outperforms competing methods on the CompColor metric.
By Ji Xie, Dewei Zhou, Xinyu Huang, Zhennan Chen, Xun Wang
The paper introduces Search-to-World, an end‑to‑end evaluation task that measures how well agentic systems can turn user requests into usable 3D worlds by combining request understanding, web visual‑content retrieval, and world delivery. It defines two metrics—Observed Retrieval Rate (ORR) and World Delivery Rate (WDR)—to separate the act of finding relevant content from successfully delivering a perceptually acceptable 3D world. The authors also present WorldSearcher, a reusable‑then‑reconstruction harness that connects existing search agents to world delivery, and a structured recovery controller that can revise queries or source videos when failures occur. Benchmarking with WorldSearcher shows that delivery success varies with the underlying agentic model and that observing relevant content alone does not guarantee world delivery; jointly training recovery agents improves both success and efficiency.
By Zixiao Gu, Yabo Chen, Xunzhi Xiang, Yu He, Haibin Huang, Chi Zhang, Yunbo Wang, Xuelong Li
The paper "What Do We Expect from LLMs? Mapping the Design of LLM Benchmarks" analyzes 14,767 arXiv submissions from 2022 to 2026 that introduce or update evaluation resources for large language models. It systematically maps changes in target systems, domains, evaluation materials, conditions, and scoring mechanisms, revealing a growing emphasis on action, interaction, and professional applications. The study also notes uneven development in model participation, with LLM-based scoring increasing in both agent and non-agent groups, while model-generated materials do not show a comparable rise.
By Chao Wang (Independent Researcher)
CliniCIRCA is a modular large‑language‑model framework that reconstructs longitudinal mental‑health patient journeys from raw electronic health record narratives. It temporally classifies clinical events in unstructured discharge summaries without explicit timestamps, producing 15,891 tagged events from 52 summaries and correcting 629 errors to create verified gold‑standard timelines. The framework then generates temporally grounded summaries, compressing each source by 1.52×, and scales to produce 1,000 silver‑standard timelines for training, showing that instruction tuning improves event extraction, temporal tagging, and summarization across models.
By Aiwei Ivy Zhang, Nimra Ishfaq, Mohit Chandra, Santiago Alvarez Lesmes, Adam Coscia, Khatiya Chelidze Moon, Xiaohan Ding, Munmun De Choudhury
The paper introduces Oph‑Guid‑RAG, a multimodal retrieval‑augmented generation system tailored for ophthalmology clinical question answering. It treats each guideline page as an independent evidence unit, retrieving page images to preserve tables, flowcharts, and layout, and employs a controllable retrieval framework with routing and filtering to reduce noise. Evaluated on HealthBench, the system outperforms GPT‑5.2 and GPT‑5.4 on hard cases, achieving significant gains in overall score and accuracy, and ablation studies confirm the importance of reranking, routing, and retrieval design.
By Shuying Chen, Sen Cui, Zhong Cao
The paper introduces a digital‑twin framework that simulates opinion dynamics in real Twitter networks by assigning agents attributes such as persona, emotions, centrality, stubbornness, and influence, and using Mistral‑7B to update opinions based on memory and social exposure. Validation on COVID‑19 and U.S. election 2020 datasets shows the framework reproduces opinion trajectories, reducing prediction error by over 50% compared to classical baselines, and improves structural alignment and polarization dynamics. Ablation studies reveal that agent attributes, memory, and social exposure all contribute to predictive fidelity, with agent attributes being the most critical.
By Omran Berjawi, Giuseppe Fenza, Rida Khatoun, Sherali Zeadally
arXiv:2609. 20353v1 Announce Type: new Abstract: We study repeated contract design when a principal observes outcomes but not the actions that generate them.
By Rui Ai, David Simchi-Levi, Han Zhong
PosteriorBench is a new benchmark that evaluates how well generative inverse solvers recover full posterior distributions rather than just a single reconstruction. It tests four physics-based inverse problems—Darcy flow inversion, Poisson source recovery, carbon capture and storage, and light transport material inference—using high-fidelity reference posteriors generated by rejection sampling and MCMC. The benchmark employs five metrics (posterior-mean error, posterior-standard-deviation error, maximum mean discrepancy, sliced Wasserstein distance, and radially averaged power-spectrum error) to assess pointwise accuracy, uncertainty, distributional alignment, and global frequency fidelity, revealing significant distribution-matching gaps in current solvers and highlighting the importance of neural operators, guidance weights, and generation noise for posterior-variance calibration.
By Jiachen Yao, Zi-Siang Hsu, Xi Deng, Aditi Gupta, Xin Ju, Sally M Benson, Gege Wen, Anima Anandkumar
Elastic Spectral State Space Models (ES-SSM) are a train‑once, export‑many sequence modeling framework that achieves elasticity by spectrally approximating the state‑space operator. The method builds on Hankel spectral filtering, using fixed spectral channels to represent long‑range token mixing and combining input‑adaptive gates with budget dropout to enable reliable deployment across different resource budgets. ES‑SSM is evaluated on byte‑level language modeling, Long Range Arena, Speech Commands V2, and offline reinforcement learning, showing that a single trained model can be truncated to competitive compact models while maintaining smooth quality‑cost curves across a wide range of truncation levels.
By Dachuan Song, Junyu Yin, Zechen Hu, Xuan Wang
The article surveys model fusion for large language models, defining it as the integration of capabilities from source models into a single target model. It categorizes prior work into three levels—parameter-level, representation-level, and behavior-level fusion—and reviews related metrics, benchmarks, and applications. The survey also outlines current challenges and future research directions, providing a comprehensive map of the field.
By Shuo Cai, Yanggan Gu, Zihao Wang, Yuanyi Wang, Yibo Yan, Wenjun Wang, Yuhang Liu, Guanghao Zhu, Sirui Huang, Ming Li, Hongxia Yang