Chehre is an emoji‑prompted video dataset designed to study perceptual flexibility in video language models. It contains 2,111 videos of 203 participants expressing 40 facial emojis, with each video annotated by about 30 perceivers, yielding 1,242 annotators in total. The dataset introduces a new task—distributional expression recognition—that evaluates a model’s ability to reproduce the variation seen in human annotations, and shows that persona prompting can shift model perception to better match human variability.
By Bita Azari, Zoe Stanley, Avneet Batra, Poorvi Bhatia, Hali Kil, Manolis Savva, Angelica Lim
The paper introduces EASE, a method that enhances multimodal reinforcement learning with verifiable rewards (RLVR) by adding visual‑evidence process supervision. EASE transforms annotated evidence regions into smoothed visual‑token targets and uses them to guide attention during RL training, but only on high‑reward trajectories. Experiments on Qwen2.5‑VL‑7B, Qwen3‑VL‑4B, and Qwen3‑VL‑8B show that EASE improves average scores over DAPO by 2.5 to 3.1 points across perception, hallucination, visual math, and multimodal reasoning benchmarks, and diagnostics confirm better alignment of visual attention with annotated evidence.
By Ruina Hu, Chen Wang, Lai Wei, Jionghao Bai, Bin Yu, Weiran Huang, Kai Wang, Yue Wang
SignSeek is a new method for learning transferable sign representations that enables efficient retrieval of signs from dictionaries using only a query video. It employs contrastive learning with saliency‑guided articulator masking, aligning same‑gloss signs across signers while focusing on the single most critical articulator per sign. Trained on 266K samples from multiple sign languages, SignSeek achieves state‑of‑the‑art cross‑corpus retrieval performance and zero‑shot generalisation to unseen British Sign Language, also improving isolated sign recognition and subtitle alignment.
By Sobhan Asasi, Ozge Mercanoglu Sincan, Richard Bowden
The paper introduces DSAQuant, a quantization‑aware training framework tailored for video diffusion models (VDMs). It aligns quantization with the denoising stages of VDMs, using denoising‑stage‑oriented supervision during training and denoising‑stage gated guidance during inference to preserve structure while improving detail reconstruction. Experiments on Wan and CogVideoX models under aggressive W3A3 and W4A4 quantization settings show that DSAQuant outperforms state‑of‑the‑art QAT baselines, boosting VBench scores by up to 6.60 while maintaining strong text‑video alignment.
By Shuaiting Li, Zelin Gao, Haibin Shen, Yujun Shen, Haotong Qin, Yinghao Xu
The study investigates whether large language models can extract Architectural Design Decisions (ADDs) from source code commits. Using four LLMs (Gemini 3 Pro, DeepSeek R1, Kimi K2, Qwen3) with zero‑shot and few‑shot prompting on 30 developer‑written ADDs, the authors evaluate outputs with ROUGE‑L, BLEU, METEOR, and BERTScore. Results show all models achieve a BERT‑F1 above 0.81, with few‑shot prompting slightly improving alignment, but the generated ADDs tend to be overly long, implementation‑focused, and lack the rationale behind the decisions.
By Amey Karan, Rudra Dhar, Mohamed Soliman, Karthik Vaidhyanathan
ENEAS is a unified, text‑promptable method that simultaneously provides precise instance tracking and high‑quality segmentation, and enables open‑concept discovery of any instance named by a text query. It extends the SeC architecture with a text‑prompting adapter and temporal memory to maintain targets through disappearance and avoid drifting, while a semantic verification layer combines visual embedding matching with conditional VLM refinement to filter ontological errors. Designed for 3D reconstruction, ENEAS delivers robust semantic tracking and segmentation across videos, libraries, and unordered collections, distinguishing true instances from look‑alike doppelgangers.
By Javier del Pino (SperidLabs), Salvador Rodr\'iguez (SperidLabs), Alejandro Garabito (SperidLabs), Javier \'Alvarez (SperidLabs), Chema Garabito (SperidLabs)
The paper audits zero‑shot vision‑language models—CLIP, BioCLIP, BioCLIP2, and a multilingual Jina CLIP v2—on 10,321 images of seven Bangladeshi freshwater fish species. BioCLIP2 achieves the highest accuracy (72.36% on BFF‑15 with English common names and 68.91% on SylFishBD with scientific names), while generic CLIP lags far behind (25.15% and 14.40%). The study also shows that Bengali prompts perform poorly, that Jina CLIP partially improves Bengali discrimination, and that image context (blur, masking) affects performance differently across species.
By Nazim-E-Alam, Tarek Rahman, Md Kishor Morol
The paper introduces MINT‑Safe, a new open‑source dataset of 11,270 multi‑image dialogues and 500 refusal VQA pairs designed to expose safety risks in multi‑modal large language models during open‑ended conversations. It also proposes TAD‑Align, a turn‑aware dual‑objective reward framework that dynamically up‑weights dialogue turns with inconsistent safety behavior, improving safety metrics on Qwen2.5‑VL‑7B‑Instruct and LLaVA‑Next‑7B. The results show over 10% reduction in attack success rate and notable gains in harmlessness and helpfulness while maintaining overall model performance.
By Han Zhu, Jiale Chen, Chengkun Cai, Shengjie Sun, Haoran Li, Yujin Zhou, Chi-Min Chan, Pengcheng Wen, Lei Li, Yike Guo, Sirui Han
The paper introduces FGLGuard, a privacy‑preserving federated graph learning framework that trains a graph attention detector on each operator’s own multi‑agent system (MAS) episode graphs, sharing only model updates. By combining a proximal local objective, domain‑balanced aggregation, threshold calibration, and guarded rewrite mechanisms, FGLGuard adapts to non‑IID data across organizations and outperforms centralized and local‑only baselines on Agent‑SafetyBench, R‑Judge, and AgentDojo. The method achieves significant reductions in attack success rates—up to 43% on AgentDojo—without compromising utility, API cost, or model capability.
By Jinxi Yu, Eric Hanchen Jiang, Levina Li, Dong Liu, Zhi Zhang, Wenxiao Zhao, Yanxuan Yu, Kai-Wei Chang, Ying Nian Wu
The paper introduces CreaEval, an automated creativity evaluator designed for complex multi-step tasks (CGPST). It separates the evaluation process into two phases: a memory‑augmented analysis that transforms responses into structured evidence, and an evidence‑based judging step that scores without seeing raw outputs. Experiments show CreaEval outperforms existing baselines by an average of 22.74% across CGPST and two simpler creativity tasks.
By Xiangyu Wang, Jin Wu, Xiaoyu Li, Chanjin Zheng, Yifeng Zhou
The paper introduces CUAHarm, a benchmark comprising 104 expert‑written realistic misuse scenarios for computer‑using agents (CUAs), such as disabling firewalls or leaking data. Using a sandbox with verifiable rewards, the authors evaluate frontier language models—including GPT‑5, Claude 4 Sonnet, Gemini 2.5 Pro, Llama‑3.3‑70B, and Mistral Large 2—and find that even without jailbreak prompts, these models can successfully execute many malicious tasks at high rates (e.g., 90% for Gemini 2.5 Pro). The study also shows that newer models, while safer in traditional safety benchmarks, exhibit higher misuse risks as CUAs, and that monitoring CUAs’ actions remains challenging, with current methods achieving only about 77% accuracy.
By Aaron Xuxiang Tian, Ruofan Zhang, Janet Tang, Ji Wang, Tianyu Shi, Jiaxin Wen
The paper introduces Alignment-Free Text‑Audiobox (Text‑AB), a unified diffusion‑based framework that performs high‑quality voice dubbing and full‑duplex dialogue synthesis without requiring forced alignment. Text‑AB uses a latent diffusion model with DAC‑VAE features, achieving over 10× compression compared to prior EnCodec representations, and learns text‑speech alignment via cross‑attention. The authors pretrain a 3B‑parameter model on 480k hours of monolingual speech and fine‑tune it for cross‑lingual dubbing, full‑duplex dialogue, and emotional dialogue, reporting significant improvements in prosody, voice similarity, naturalness, and emotional expressivity over existing internal systems.
By Sanyuan Chen, Min-Jae Hwang, Sho Inoue, Anna Sun, Bokai Yu, David Kant, Dongmin Hyun, Dorian Desblancs, Gregory Antonovsky, Oleg Repin, Peng-Jen Chen, Xutai Ma, Zehai Tu, Juan Pino, Wei-Ning Hsu
The article discusses how current world models, while achieving high predictive likelihood and visual fidelity, often fail to preserve the evidence needed for safe decision-making in embodied systems. It identifies three structural mismatches—likelihood versus risk, prediction versus intervention, and finite-horizon prediction versus accumulated consequences—and proposes the Risk‑Informed World Model (RIWM) as a decision‑centric framework. RIWM emphasizes consequences, intervention, epistemic uncertainty, and recoverability, integrating decision‑relevant representation, counterfactual reasoning, safety‑critical episodic memory, and runtime safety assurance to better support safety‑critical embodied systems.
By Kailang Ma, Heye Huang, Inhi Kim, Kitae Jang
The paper argues that federated learning, while touted as privacy‑preserving, still concentrates control over the resulting model with the entity that orchestrates training. It identifies three layers—storage, circulation, and learning—where creative communities can exert governance, noting that current practices allow consent for training but not for model ownership or federation. The authors propose four design principles for a creative data commons that extends governance to models, ensures legibility of terms at contribution, incorporates refusal as a first‑class state, and makes stewardship transparent and accountable.
By Phoenix Perry, George Simms, Elizabeth Wilson, Yasmine Boudiaf, Nick Bryan-Kinns, Tega Brain, R. Luke DuBois, Alix Rule, Rachel Meade Smith, Kelani Nichole, Atharva Pravin Pawar, Rebecca Fiebrink
FailBench is a new benchmark for robot failure detection, containing 2,197 manipulation attempts from 14 public sources, with 75% of failures occurring naturally. The study evaluates 13 vision‑language model (VLM) detectors, finding the best model achieves only 0.77 mean balanced accuracy, and that fine‑tuned failure detectors often underperform general‑purpose VLMs. Performance varies with visual evidence, excelling when object motion is observable but dropping to near chance on contact‑intensive assembly tasks, and input‑level cropping of outcome‑relevant regions improves the top detector by 2.4 percentage points.
By Zaruhi Navasardyan, Tatul Danielyan, Hrant Davtyan
EraseSAE is a framework for surgical concept erasure in text-to-video diffusion models. It uses a Partitioned Convolutional Sparse Autoencoder to decompose activations into interpretable sparse features, a contrastive attribution mechanism to isolate concept-specific kernels, and timestep‑resolved masks to confine erasure to active regions. Experiments show precise removal with minimal quality loss, outperforming existing methods.
By Xinghao Wang, Dong Li, Wei Yu, Yingwei Pan, Tao Gong, Qi Chu, Nenghai Yu, Ting Yao
GazeFS is a model that predicts and stabilizes target‑centered gaze trajectories using a variable‑length gaze‑head history, without requiring target information during inference. It maps this history to the next target‑center direction and a short‑horizon Search/Focus estimate, improving focus target centering and reducing residual gaze error. Across 7,960 acquisition episodes from 30 participants, GazeFS reduces Focus episode bias, dispersion, and P90 target error by 0.182°, 0.257°, and 0.400°, respectively, while maintaining high phase‑balanced accuracy and AUPRC.
By Yaozheng Xia, Zaiping Zhu, Bo Pang, Minghao Xie, Hui Li, Shaorong Wang, Sheng Li
IndicSafeEval is a new evaluation framework that tests the safety robustness of large language models against persuasion-based jailbreak attacks in Indian languages. The benchmark includes 7,200 adversarial prompts covering ten safety-critical content categories, six persuasive strategies, and four languages (Hindi, Bengali, Marathi, Punjabi). Experiments show that model safety varies significantly across languages, prompt styles, and risk categories, revealing that current English-centric safety evaluations miss important multilingual vulnerabilities.
By Saikat Mondal, Mamta, Deeksha Varshney, Oana Cocarascu, Asif Ekbal
The paper explores whether a language model’s refusal behavior can be predicted from its intermediate activations before decoding. Using linear probes on residual stream activations at each transformer block, the authors show that refusal signals are linearly decodable well before the final layer, indicating that safety-relevant behavior is encoded early in the network. They then introduce Mechanistic AutoDAN, a probe‑guided variant of AutoDAN that replaces full-model fitness evaluation with partial forward passes and probe-based scoring within a genetic prompt search loop, achieving competitive attack success rates while reducing per‑iteration search time by up to 72% and improving cross‑model transfer in several configurations, especially for larger models.
By Matteo Gioele Collu, Riccardo Conte, Alberto Giaretta, Denis Kleyko, Mauro Conti, Matteo Zavatteri, Roberto Confalonieri
The paper introduces a new measure of generative‑process diversity for language models, using Normalised Compression Distance on raw outputs after controlling for permutation effects. Across 38 models, this metric uncovers population structure that semantic similarity misses and predicts lower correlated failures across ten benchmark families, independent of semantic similarity or model capability. The authors argue that higher generative‑process diversity reduces correlated failures in multi‑model systems, offering a practical tool for safety‑relevant applications.
By Ross Tieman, Evan Markou