Recent studies suggest that video generation models can exhibit certain forms of zero-shot visual reasoning through generated frames. Yet reliable evaluation remains challenging: benchmarks should adopt inputs aligned with the visual priors of current video models, require valid evolving processes rather than only plausible final states, and calibrate task difficulty to remain challenging yet partly feasible.
Training fast generators on new domains with limited data remains challenging for two reasons. First, adapting a pretrained diffusion or flow model to a new domain leaves its costly multi-step sampling unaddressed, and existing acceleration methods are tied to the source parameterization--$ε$, $x$, $v$, or $u$--leaving heterogeneous pretrained models with no common acceleration target.
State-of-the-art vision-language-action (VLA) models such as $π_{0.5}$ exhibit strong semantic understanding, instruction following and task behavior. However, when deployed on new robots, even minor...
The paper introduces TextQ‑German, a dataset suite for evaluating German natural language generation (NLG) from a Quality of Experience (QoE) perspective, covering tasks such as summarization and machine translation. Human ratings collected via crowdsourcing identify perceptual quality dimensions, and the authors develop automatic QoE prediction models—including transformer‑based, linguistic feature‑based, and hybrid approaches—showing that hybrid models outperform pure transformers and that linguistic features alone can rival fine‑tuned language models. The dataset is further enriched with large language model (LLM) outputs annotated with overall QoE scores, and validation on held‑out data demonstrates generalization to unseen data.
"whyItMatters":"The resource provides a publicly accessible benchmark and baseline models for human‑centered NLG evaluation, enabling the development of systems that better align with human quality perception."
Falcon Perception-HD applies reinforcement learning (GRPO) to autoregressive perception models, aligning them directly with precision and recall metrics rather than relying on maximum‑likelihood fine‑tuning. The RL framework introduces reward design for set‑structured outputs and multi‑head sampling control, enabling state‑of‑the‑art performance in very dense scenes (up to 500 objects) and eliminating common issues such as mask repetitions, NMS, and coordinate deduplication. Hybrid self‑annotation pipelines tailored for difficult referring expressions and dense scenes further boost RL training, with improvements observed across all difficulty levels on PBench and SACO‑Gold, and the model preserves object existence knowledge without negative samples.
The paper examines how multilingual medical adaptation affects the internal representations of Whisper ASR models by performing layer‑wise encoder analysis. It compares several adaptation strategies—zero‑shot decoding, English‑only fine‑tuning, German‑only diagnostic fine‑tuning, two‑stage EN→EN+DE continuation, and direct EN+DE fine‑tuning—across different Whisper sizes, finding that fine‑tuning improves performance but the best model varies by setting. Layer‑wise results show that English medical fine‑tuning drives the main encoder shift, while multilingual continuation largely preserves the adapted representation space, with domain and language information remaining recoverable across layers.
"whyItMatters":"The study provides insight into how multilingual medical adaptation reshapes Whisper’s internal representations, guiding the selection of model sizes and fine‑tuning strategies for improved MedASR performance."
The paper presents the first systematic evaluation of uncertainty quantification (UQ) methods applied to a foundation model for semantic segmentation. By fine‑tuning a lightweight DPT decoder on the pretrained SAM2 encoder, the authors benchmark four UQ approaches—Monte Carlo Dropout, Deep Sub‑Ensemble, Test‑Time Augmentation, and Evidential Deep Learning—across Cityscapes, NYUv2, and two out‑of‑domain settings. The study compares segmentation accuracy, calibration, uncertainty quality, and inference time, revealing trade‑offs between predictive performance, reliability, and computational cost.
The paper presents a local traditional OCR pipeline that can be iteratively fine‑tuned on both layout and appearance levels of complex historical Sanskrit manuscripts. By adapting to the specific manuscript distribution, the pipeline improves transcription accuracy across subsequent pages, reducing the need for costly human annotation. The authors apply the method to three manuscripts, release a dataset with detailed layout and Unicode annotations in PAGE‑XML format, and benchmark the results against leading multimodal large language models.
TranslatePsy-AfriSLM is an open‑source machine‑translation resource set for 19 Sub‑Saharan African languages, comprising curated parallel data, African‑specialized synthetic data, and a suite of fine‑tuned small language models (SLMs). The authors demonstrate that a unified quality‑estimation filtering can discard up to 96% of training tokens while preserving quality, and that filtered synthetic data dominates the quality‑efficiency Pareto frontier. Models trained on this curated mixture outperform larger systems such as TranslateGemma‑27B and Qwen3.5‑122B‑A10B, achieving superior performance with as few as 0.8 B parameters.
The paper investigates how guided protein language models can collapse onto off‑manifold representations when heavily steered to optimize a property. This collapse causes generated sequences to become low‑complexity and statistically similar to random amino‑acid input, yet the property oracle may still rate them highly. The authors propose a cheap, training‑free Mahalanobis filtering step that removes such off‑manifold candidates, improving both property scores and structural plausibility without altering the generator.
Data-DPO is a target model‑oriented supervised fine‑tuning data selection method that uses one‑step probing of the target model to generate pairwise data preferences, trains a lightweight reward model to capture these preferences, and then selects a training subset by combining target‑model preference, external quality scores, and marginal diversity. Experiments on Vision‑Flan and LLaVA‑CoT demonstrate that Data‑DPO consistently outperforms existing data selection baselines across multiple data budgets and even surpasses full data training performance.
By Peng Sun, Yi Yang, Antong Zhang, Chunxiao Li, Yanbo Wang, Dianbo Liu, xin chen, Kai Yu, Lu Chen, Tianfan Fu
The paper introduces LSem2Vec, a two‑stage method that first uses a large language model to extract source code semantics and then applies a sentence embedding model to produce vector representations. This approach removes the need for task‑specific training or fine‑tuning, addressing errors in LLM outputs. Experiments on three datasets across multiple programming languages show that LSem2Vec outperforms five state‑of‑the‑art unsupervised methods.
By Zixiang Xian, Chenhui Cui, Rubing Huang, Chunrong Fang, Zhenyu Chen
PTXBench is a benchmark designed to evaluate and adapt large language models (LLMs) for GPU kernel optimization using architecture‑specific PTX code. It assesses functional correctness, runtime execution of target instructions, and speedup over state‑of‑the‑art libraries on GEMM and attention workloads on H100 and B200 GPUs. The study finds uneven success rates, especially on complex attention backward tasks, and shows that executing target instructions does not guarantee competitive performance, with no model consistently outperforming frontier libraries. The authors also fine‑tune Qwen3.6‑27B, noting that repair‑conditioned training improves some tasks but generalization remains inconsistent, highlighting the importance of data coverage, balance, and teacher quality.
By Genghan Zhang, Yixin Dong, Chengze Fan, Zhichen Zeng, Yueming Yuan, Shaowei Zhu, Kunle Olukotun
The paper introduces MASS, a hierarchical data selection method that first groups data using low‑dimensional manifold coordinates learned by a dense autoencoder, then applies a TopK sparse autoencoder for quality‑aware feature coverage within each group. This approach addresses the shortcomings of traditional diversity metrics that mix semantic, supervisory, and noise signals. Experiments on Vision Flan and LLaVA‑CoT demonstrate that MASS outperforms existing baselines across various budgets and can match or exceed full‑data training with only a small subset.
By Peng Sun, Yi Yang, Antong Zhang, Chunxiao Li, Yanbo Wang, Dianbo Liu, xin chen, Kai Yu, Lu Chen, Tianfan Fu
OraclePhys is a fine‑tuning framework for large language models on structural mechanics, comprising a graded benchmark (OraclePhys‑Bench), a 30K supervision dataset (OraclePhys‑30K), and a controlled training study. The study shows that the form of the label’s answer, rather than its length, determines what the model learns, and that certain training objectives can produce models that match or exceed existing LLMs on spatial structural response tasks. The trained 8B model reaches the data‑precision frontier, outperforming zero‑shot and 32‑shot baselines at a specialist level.
By Mingyu Li, Guorui Song, Jing Lin, Haoqian Wang
The paper introduces a small language model (SLM)-based framework that detects and classifies GNSS spoofing attacks on autonomous vehicles by converting vehicle states from GNSS and other sensors into structured semantic narratives. The SLM achieves performance comparable to large language models, with an average accuracy of 96.99%, while offering lower inference latency and reduced GPU memory usage. Field tests in Clemson, South Carolina, confirm the framework’s real‑time detection capabilities on resource‑constrained vehicular platforms.
By Abyad Enan, Sagar Dasgupta, Mizanur Rahman, Mashrur Chowdhury
The paper introduces QWM, a framework that integrates world models with standard Q‑learning to perform test‑time search over imagined trajectories. By training the policy and value function solely on real transitions, QWM avoids compounding model bias while still benefiting from predictive search. Experiments on the Robomimic and LIBERO manipulation benchmarks show that QWM outperforms strong prior state‑of‑the‑art methods in both sample efficiency and performance.
By Perry Dong, Yueru Jia, Chelsea Finn, Dorsa Sadigh
The paper introduces OVRSISBench, a unified benchmark for open‑vocabulary remote sensing image segmentation, and evaluates existing OVS/OVRSIS models, uncovering their shortcomings in remote sensing contexts. Leveraging insights from this evaluation, the authors propose RSKT‑Seg, a new framework featuring a Multi‑Directional Cost Map Aggregation module, an Efficient Cost Map Fusion transformer, and a Remote Sensing Knowledge Transfer module. Experiments on the benchmark demonstrate that RSKT‑Seg outperforms strong baselines by +3.8 mIoU and +5.9 mACC while achieving twice the inference speed.
By Bingyu Li, Haocheng Dong, Da Zhang, Zhiyuan Zhao, Junyu Gao, Xuelong Li
The paper proposes a Learnware-based framework for deploying scene‑specific CSI feedback models in 6G systems. A centralized AI data center maintains a catalog of pre‑trained models, each tagged with semantic and statistical specifications. Base stations retrieve the most relevant model using only statistical fingerprints, which reduces data privacy risks, lowers retrieval latency, and cuts fine‑tuning effort, achieving up to 57.7% performance gains over a general model.
By Xiangyi Li, Jiajia Guo, Chao-Kai Wen, Xin Geng, Shi Jin, Zhi-Hua Zhou
The paper introduces MCTH (Monte Carlo Tree Hallucination), an inference-only framework that performs all‑atom biomolecular sequence‑structure co‑design by treating pretrained folding and inverse‑folding models as black‑box operators. MCTH uses Monte Carlo Tree Search to allocate a fixed inference budget across competing design trajectories, incorporating model confidence, uncertainty, and cross‑expert consensus. Experiments across protein‑RNA, protein‑DNA, protein‑protein, and protein‑ligand design show that adaptive search outperforms simpler sampling strategies, and evaluations with AlphaFold3 and Chai‑1 demonstrate transferability beyond the search‑time oracle.
By Xuefeng Liu, Mingxuan Cao, Xiao Luo, Songhao Jiang, Tobin Sosnick, Jinbo Xu, Louis Maher, Rick Stevens