WeVisOne‑VL is a new vision‑language model that unifies object‑level and pixel‑level perception through structured next‑token prediction. It supports a wide range of visual tasks—object detection, referring localization, counting, OCR, segmentation, and matting—allowing localization at any user‑specified granularity from points and boxes to masks and alpha mattes. The architecture incorporates a lightweight pixel decoder via a special token, and the model is trained with a three‑stage curriculum followed by mixed‑task reinforcement learning, achieving leading performance on over 30 benchmarks while robustly rejecting object‑absent queries.
By Tianming Liang, Junyi Pan, Yuxi Li, Canmiao Fu, Zhipeng Huang, Chong Sun, Chen Li, Jing Lyu
UniMem is a framework that unifies visual and textual memory with low‑level control in a single Vision‑Language‑Action model, using an event classifier, keyframe encoder, and caching to manage memory efficiently. It addresses non‑Markovian tasks that require remembering past scenes, avoiding the memory bottleneck of separate VLMs and the performance drop from fixed‑interval frame sampling. Across five simulation and four hardware tasks, UniMem outperforms fixed‑interval sampling (93.4% vs. 68.2%) and hierarchical VLM‑VLA baselines (80.0% vs. 43.5%) while offering faster inference and a simpler training pipeline.
By Lars Osterberg, Maggie Wang, Mac Schwager
The paper evaluates recent large language models (LLMs) for knowledge graph‑to‑text generation against traditional fine‑tuned encoder‑decoder models. Using both reference‑based metrics and graph‑grounded, reference‑free LLM‑based evaluation, the study finds that LLMs largely close the performance gap on reference metrics and outperform fine‑tuned models on graph‑grounded metrics. The results also show that LLM‑based evaluation correlates weakly with reference metrics, indicating that each captures distinct aspects of generation quality, and that prompted LLMs produce fewer contradictions and omissions, especially as graph size grows.
The paper investigates how performance predictors can improve model routing for large language models (LLMs). It shows that a predictor only outperforms a model’s self‑confidence when the candidate model is suboptimal or the predictor adds significant capacity. The authors provide theoretical bounds, extend the analysis to two‑model cascades, and confirm that predictors help most with suboptimal models, with gains diminishing after fine‑tuning.
During the deep learning era, innovations in remote sensing (RS) vision models primarily focused on optimizing network architectures for specific tasks and conducting end-to-end training. However, wit...
IndoNutriQA is a curated benchmark of 911 Indonesian clinical nutrition question‑answer pairs derived from real reports. Each pair includes a clinically relevant context, structured metadata, and answers in free‑form prose, with expert validation indicating 6‑10% of answers are contestable. The dataset supports generative, extractive, and retrieval‑augmented QA, and is used to evaluate fine‑tuned models and large language models in zero‑ and few‑shot settings.
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 remains largely unsolved, with the best model achieving only a 23.7% macro F1 score. To improve performance, the authors fine‑tune an 8B‑parameter MLLM, RetGrounder, using an OCT‑specific strategy that raises the grounding F1 to 33.9%, a 43% relative gain.
whyItMatters":"The benchmark establishes a concrete reference for measuring progress in OCT diagnosis with multimodal large language models, highlighting the current limitations and guiding future improvements in visual grounding for clinical applications."
The paper investigates natural language autoencoders (NLAs) that convert large language model activations into readable text. It examines whether NLAs capture compositional structure, generalize to fine‑tuned models, recover fine‑tuning objectives, and can be specialized for model diffing. Experiments across multiple models show positive results: NLAs preserve compositionality, transfer to lightweight fine‑tunes, recover hidden objectives, and outperform a black‑box baseline on diffing tasks, though they lag behind state‑of‑the‑art methods for misalignment detection.
HimmBO is a two‑stage method for preference‑driven merging of large collections of lightweight adapters used to fine‑tune text‑to‑image models. The first stage screens adapters via level‑set estimation on pre‑generated images, avoiding costly image generation, while the second stage optimizes merge coefficients in the reduced subspace using a style‑aware kernel. Experiments show HimmBO attains higher similarity with far fewer image generations than prior methods and scales robustly to collections of up to 1,080 adapters.
Lean Finder V2 is a semantic search framework for Mathlib that tackles the problem of short, ambiguous user queries by using adversarial training and query rewriting to disambiguate them. It achieves a new state‑of‑the‑art performance on the MathlibQR benchmark, improving retrieval by at least 11% over previous engines, and further boosts proof success when integrated with Claude Code.
ProQuant is a weight‑only post‑training quantization framework designed for Mixture‑of‑Experts large language models. It introduces PoP, which jointly prunes experts and assigns mixed‑precision bit‑widths to better use a tight bit budget, and RoCo, a routing‑consistent RMSNorm fine‑tune that mitigates routing drift caused by quantization errors. Experiments show ProQuant outperforms existing methods on a range of generative tasks.
ULME introduces a post‑training framework that unifies text embedding and generation within a single causal decoder‑only LLM by inserting learnable slots for semantic and reasoning embeddings. It trains three objectives—contrastive, language‑modeling, and stop‑gradient predictive—so that generation actively shapes the embedding space. Across Qwen2.5 and Qwen3 models (0.5B–8B), ULME improves MTEB scores by 3.3–4.1 points, tightens BRIGHT and AdvBench‑IR, and matches generation‑only fine‑tunes on standard accuracy benchmarks.
Doob-Based Generative Policy Optimization (DoobPO) introduces a likelihood‑free framework for fine‑tuning generative policies such as diffusion and flow models. It first learns a lightweight ratio model to represent the desired policy update and then distills this ratio‑tilted target policy into the generative policy via conditional drift matching. The method supports standard policy improvement rules like PPO and GRPO, proves gradient equivalence to an ideal Doob drift matching objective, and demonstrates strong performance on MuJoCo continuous‑control benchmarks.
By Maojiang Su, Qijie Zhu, Yu Xiao, Shuyang Yu, Minshuo Chen, Zhaoran Wang, Han Liu
The paper introduces a method to transform opaque fine‑tuned neural AI‑text detectors into interpretable classifiers by training sparse autoencoders on the detector’s residual stream. Each autoencoder feature is labeled with a human‑readable concept, and only features meeting a fidelity threshold are kept. A lightweight classifier built on these interpretable features outperforms both neural and stylometric detectors, especially on out‑of‑distribution texts.
By Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen, Chris Callison-Burch
The study fine‑tuned DINOv2‑base on 2,253 weighed Nutrition5k dishes to estimate food mass from a single RGB photo and compared it with the zero‑shot vision‑language model Qwen3‑VL‑32B across Nutrition5k and three external datasets. While the fine‑tuned model achieved a mean absolute error of 24.8 g on its own domain—better than Qwen—it failed to outperform Qwen on external datasets and its gram scale varied with scene coverage and differed per food type. Adding a single labeled photo of the same food reduced external error significantly, indicating that single‑image mass estimators require labeled photographs from the target setting.
By Yongsheng Lian
AgentPolicyForge is a framework that leverages executable environments to build policy benchmarks and iteratively train policy guards for task agents. It introduces PolicyGym, a dataset covering ten policy families with 698 training scenarios and 254 human‑reviewed test scenarios, where tool calls are executed and outcomes recorded to assess compliance. The framework supports an attacker‑generated request loop that fine‑tunes guards on balanced, environment‑verified examples, leading to a 40‑44% reduction in attack success rates compared to policy prompting alone.
By Chao Shang, Wanyu Du, Yi Zhang, Sandesh Swamy
The paper investigates the use of classically simulated quantum layers in vision‑language models, specifically adding or replacing components in the LLaVA‑OneVision backbone. Experiments with 4‑ and 8‑qubit circuits show that inserting a quantum layer after the visual projector can raise development accuracy by 2.28 percentage points, while replacing the final decoder feed‑forward sublayer reduces trainable parameters by 13.07 million with only a modest drop in accuracy. The study also evaluates robustness against six attack scenarios, mapping where quantum layers improve or degrade performance and security in multimodal models.
By Md Khalid Syfullah, Alvi Ataur Khalil
The paper introduces two theories—Fiber Regulation Theory (FRT) and Fiber Acquisition Theory (FAT)—to analyze how a learner’s current predictions influence future learning dynamics. FRT focuses on protecting and repairing existing functions, while FAT examines how new trainable structures are installed and renewed. The authors demonstrate that combining these operations can yield learning responses unattainable by either alone, and they validate this by implementing coordinated repair and renewal strategies in the Qwen2.5-32B model, achieving superior teacher‑forced accuracy and lower macro NLL compared to separate policies.
By Qinyou Wang
ClusterFusion is a hybrid text‑clustering framework that uses a large language model (LLM) as the core clustering engine, guided by lightweight embedding methods. The process involves embedding‑guided subset partitioning, LLM‑driven topic summarization, and LLM‑based topic assignment, allowing direct incorporation of domain knowledge and user preferences. Experiments on three public benchmarks and two new domain‑specific datasets show that ClusterFusion achieves state‑of‑the‑art performance on standard tasks and significant gains in specialized domains, with the authors releasing a new dataset and benchmark results.
By Yiming Xu, Yuan Yuan, Vijay Viswanathan, Graham Neubig
MARS is a scalable knowledge‑graph question‑answering system that integrates large language models with knowledge graphs without requiring model fine‑tuning. It uses a structured retrieval process that links question entities to the graph, iteratively fetching next‑hop information and deciding when to stop and generate a SPARQL query. Evaluated on three KGQA benchmarks—including multilingual settings—MARS matches state‑of‑the‑art performance while remaining efficient and scalable.
By Nikit Srivastava, Daniel Vollmers, Ren\'e Speck, Nikolaos Karalis, Hamada M. Zahera, Axel-Cyrille Ngonga Ngomo