The paper introduces BrailleBench, a benchmark of 259 real-world Grade‑1 and Grade‑2 Braille images, and evaluates various vision‑language models (VLMs) on the task of transcribing Braille from images. It finds that general‑purpose VLMs perform poorly, while tool‑assisted systems do better. By training a 2B VLM on a synthetic Braille generation pipeline, the authors achieve 87% exact match on Grade‑1 and 61% on Grade‑2 BrailleBench, surpassing other evaluated systems, but also demonstrate that high synthetic validation accuracy does not translate to real‑world performance.
By Kartik Jangra, Atharva Gawane, Siddharth Pulipaka, Hemachandran Karah, Raj Dabre
PRISM-BENCH is a new benchmark comprising 1,000 expert‑curated vision‑language questions across five under‑represented reasoning categories, each paired with an idealized reasoning trace validated by a second annotator. It introduces diagnostics beyond final‑answer accuracy: a perception probe that attributes about 85% of frontier‑model failures to misperception, and trace alignment that predicts answer correctness with a macro AUROC of 0.66. The benchmark also presents CORE, an efficiency‑aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length, revealing that recent gains are driven by recognition rather than spatial grounding.
By Vansh Bhatia, Mehdi Rezagholizadeh, Parsa Ashrafi Fashi, Igor Gilitschenski, Vikram Appia, Emad Barsoum
The paper introduces Multi-Level Consistency and Conflict Decoupling (MLCCD), a method for detecting fake news in multimodal content by leveraging the layered knowledge of multimodal large language models (MLLMs). MLCCD first selects relevant visual and textual tokens using the MLLM’s final decision representation, then identifies tokens whose representations change sharply across adjacent layers to capture layer‑specific knowledge. Finally, it decouples this knowledge into consistency and conflict evidence, aggregating each across layers with dedicated routers that adapt to varying conflict levels, achieving superior performance on three benchmarks.
By Kai Qian, Dongjie Xu, Weijie Shi, Jiaqi Wang, Haolun Ma, Jia Zhu, Jiajie Xu
The paper introduces Watermark Basin Alignment (WM‑Basin), a black‑box method for detecting distilled large language models (LLMs). It exploits a persistent watermark basin—a key‑specific geometric signal embedded in the teacher’s parameter space during fine‑tuning—that remains detectable even when token‑level signals vanish. WM‑Basin maps suspect model outputs to gradient representations, removes ordinary fine‑tuning directions, and tests alignment with the watermark basin, achieving significantly higher detection rates than token‑level baselines across a range of teacher and student models.
By Hongyan Chang, Hady Elsahar, Tom Sander, Pierre Fernandez, Saeed Mahloujifar, Sylvestre-Alvise Rebuffi, Tom\'a\v{s} Sou\v{c}ek, Tuan Tran, Valeriu Lacatusu, Surya Parimi, Rashel Moritz, Vanessa Stark, Alexandre Mourachko
The paper presents an agent designed for the GLEE Competition that engages in bargaining, negotiation, and persuasion games against human and AI opponents. The agent combines a classical game‑theoretic solver, an online Bayesian layer for estimating opponent archetypes, and a language model for generating messages. Experiments include 13,200 Monte Carlo games against 11 synthetic archetypes, an ablation study of the opponent‑modeling layer, and an analysis of 9,337 live‑game messages showing a strong correlation between moral rhetoric framing and buyer purchase decisions.
By Hayden Arnold
FullFront is a new benchmark that evaluates Multimodal Large Language Models (MLLMs) across the entire front‑end engineering workflow, covering webpage design, perception QA, and code generation. It uses a two‑stage process to convert real‑world webpages into clean, standardized HTML, avoiding copyright issues and providing diverse visual designs. Experiments show that current MLLMs struggle with page perception, image handling, layout, and interaction implementation, revealing a large gap compared to human experts.
By Haoyu Sun, Huichen Will Wang, Jiawei Gu, Linjie Li, Yu Cheng
HEART-Bench is a benchmark designed to assess whether large language models (LLMs) can make decisions that are consistent with a given persona and past experiences. It features 11 synthetic characters anchored in the Big Five personality traits, each with 1,000 autobiographical memories, and 64 decision scenarios guided by the DIAMONDS framework, resulting in 673 multiple‑choice questions. Models are evaluated with and without memory retrieval, showing that memory‑augmented configurations improve accuracy to 64.9%, while human annotators achieve 80.8% on a matched subset.
By Weihan Peng, Chenxu Zhang, Qianao Wang, Yuling Shi, Heng Lian, Qihong Mao, Jiahao Pang, Chunliang Feng, Bowen Li, Xiaodong Gu
The paper introduces Causal Attribution Pruning (CAP), a training‑free technique that identifies critical attention heads in large language models by measuring their causal impact on reasoning tasks. CAP estimates performance loss when each head is masked on a small calibration set, then translates these causal scores into weight‑level importance for pruning. Compared to magnitude‑only or activation‑based methods, CAP achieves up to 61% relative accuracy gains over Wanda on ARC‑Challenge at 20% sparsity and improves performance on GSM8K, StrategyQA, and ARC‑Challenge across Llama‑3‑8B‑Instruct and Mistral‑7B‑Instruct at 10%, 20%, and 50% sparsity.
By Amogh Sheth, Andrew Lin, Biruk Assefa, Yi Wen Huang, Yuhao Ge
ReMCTS is a Monte Carlo Tree Search–style framework that enhances code generation by large language models (LLMs). It organizes program candidates as tree states, retains branch‑local debugging context, and retrieves failure experience across branches, distinguishing failed checks from unavailable evidence. Experiments on HumanEval and MBPP‑Sanitized show that visible‑test ReMCTS improves performance over direct generation in 8 of 10 model‑dataset pairs, while a pilot on HumanEval‑X C++ demonstrates compatibility with compiler‑backed execution.
By Huifei Wang, Xinying Huang, Yiheng Sun, Yifan Yuan
The paper introduces DYNASHIELD, a black-box moving target defense for large language models that dynamically customizes decoding hyperparameters and system prompts during inference. It first identifies decoding configurations that lower the probability of successful jailbreak attacks, then samples from a weighted pool of these configurations to add controlled variability to the model’s behavior. Experiments on seven open-source LLMs and four state-of-the-art jailbreak attacks show that DYNASHIELD significantly reduces attack success rates while preserving response quality and adding minimal inference overhead.
By Xiaoqun Liu, Weiming Qi, Qiben Yan
ShredBench is a new benchmark for evaluating multimodal large language models (MLLMs) on the task of reconstructing documents from shredded fragments. It uses an automated pipeline that generates fragmented documents from Markdown, covering English, Chinese, code, and table scenarios with varying fragmentation levels. Experiments show that while MLLMs perform well on intact documents, their performance drops sharply when documents are shredded, revealing a gap in fine‑grained cross‑modal reasoning.
By Zichun Guo, Yuling Shi, Wenhao Zeng, Chao Hu, Haotian Lin, Terry Yue Zhuo, Jiawei Chen, Xiaodong Gu, Wenping Ma
The paper analyzes 2,487 tasks from five public coding‑agent benchmarks and 14,922 runs of six agent configurations, introducing the Spread‑Novelty‑Centrality (SNC) profile to describe how reference solutions are distributed, added, and the code’s dependencies, history, and complexity. It finds that every pair of benchmarks differs on at least two SNC axes, even when sharing the same task label, and that resolved runs tend to involve less distributed changes and lower centrality values, with patch sizes varying between Claude and Qwen configurations. The authors recommend that practitioners choose evaluation tasks based on change characteristics, inspect task descriptions, and assess patch scope alongside test outcomes to better align benchmarks with real‑world workloads.
By Radin Shayanfar, Keheliya Gallaba, Ahmed E. Hassan
Any3R is a feed‑forward 3D reconstruction model that works with any central camera type, regardless of lens distortion or field‑of‑view limits. It conditions on a dense ray field—unit direction vectors through every pixel—so its positional encoding and loss function are lens‑agnostic. Evaluated on thirteen camera types across three datasets, Any3R consistently outperforms other methods, enabling scalable reconstruction from heterogeneous, crowdsourced imagery.
By Jingong Chen, Qingwen Zhang, Youming Deng, Chulwoo Pack, Olaf Wysocki, Boris Jutzi, Kwanghee Won, Kyle Gao
UT‑HMR is a one‑stage monocular 3D human mesh recovery framework that introduces uncertainty‑guided temporal modeling to address ambiguities caused by weak observations such as distance, occlusion, and motion blur. It predicts a Gaussian distribution over SMPL pose with a low‑rank-plus‑diagonal covariance, converting this into joint‑wise precision weights for heteroscedastic supervision while keeping inference deterministic. Temporal early fusion incorporates neighboring‑frame evidence before pose decoding, enabling the model to resolve pose ambiguities and achieve state‑of‑the‑art results on VIP‑HockeyPose‑v2, WorldPose, and competitive performance on 3DPW using a lightweight ViT‑B backbone.
By Zhibo Wang, Sirisha Rambhatla, Yuhao Chen, Maria Koshkina, David Clausi
BoxWorld introduces a novel approach to real‑time interactive video worlds by decoupling interaction reasoning from generative rendering. A coding agent builds a Box World Engine that applies user‑defined rules to determine interaction outcomes and exposes object states via a lightweight interface of layout boxes, persistent attributes, and dynamic status variables. A separate Box‑State Generative Renderer then autoregressively renders these evolving states into video, enabling online control over object motion, state switches, and conditional events. The system employs a three‑stage training strategy—learning object grounding from open‑domain videos, state transitions from gameplay videos, and rollout distillation on self‑generated histories—to handle real‑time state updates while maintaining visual coherence. BoxWorld outperforms end‑to‑end world models and remains competitive across multiple public benchmarks, with a new PairedStateBench benchmark for fine‑grained interaction control.
By Zike Wu, Ahmed Hatem, Chintan Trivedi, Wallace Michel Pinto Lira, Xuebin Qin, Lele Wang, Renjie Liao
The paper presents a lightweight extension to the Monst3R 4D reconstruction model by adding an event-camera input branch while keeping the rest of the architecture unchanged. Across three benchmark datasets, this RGB‑event fusion reduces absolute relative error by an average of 3.4% compared to the RGB‑only baseline, with the largest gains occurring under high object motion where RGB‑only methods falter. The study demonstrates that incorporating event data can meaningfully improve 4D reconstruction robustness without a full architectural redesign.
By Dulanga Weerakoon, Darshana Rathnayake, Van Minh Do, Hemanth Sabbella, Kaichen Zhou, Fangneng Zhan, Alok Prakash, Sanjay E. Sarma, Archan Misra, Paul Pu Liang
SAGE (Saliency‑Annealed Gated Encoder) is a Vision Transformer‑based backbone that explicitly leverages saliency, global context, and high‑frequency details through a saliency‑guided architecture, an annealing training paradigm, and a two‑stage loss transition. It achieves state‑of‑the‑art or competitive results on a wide range of dense prediction tasks, including COCO object detection and segmentation, ADE20K segmentation, TinyImageNet classification, and DIOR‑R oriented detection, all using the same backbone with task‑specific decoders. The framework is notable for being the first to deliver multi‑task SOTA performance on Huawei Ascend NPUs.
By Le Yang, Yucheng Huang, Junyu Chen, Qi Dang, Xin Kong, Haifeng Huo, Qiang Zhao
Efficient Remote Sensing Image-Text Retrieval via Interaction-Free Recall and Text-Guided Reranking proposes a fast-then-fine (FTF) retrieval framework that first performs interaction-free multi-granular recall to quickly match query embeddings with pre-computed gallery embeddings, then refines the top‑K candidates using a parameter‑free text‑guided cross‑modal reranking module called the Balanced Text‑guided Interaction Block (BTIB). The method introduces an inter‑ and intra‑modal loss to jointly optimize cross‑modal alignment across multi‑granular representations. Experiments on public benchmarks show that FTF achieves competitive retrieval accuracy while significantly improving retrieval efficiency compared with existing methods.
By Xi Chen, Xu Chen, Xiamgyang Jia, Wei Wang, Xu Zhang, Zhenyuan Shun
The paper critiques current GUI grounding benchmarks that count a click as correct only when it falls inside a single annotated box, arguing that many interfaces expose the same action through multiple elements. The authors propose redefining correctness as achieving the same outcome as clicking the annotated target, and introduce a probe that executes and fingerprints thousands of clicks per page to create outcome maps and an execution judge. Applying this method to several large‑scale grounders on real web applications shows that a significant fraction of clicks rejected by box‑based scoring actually produce the intended outcome, indicating that box scoring under‑counts correct clicks and that evaluation metrics need adjustment.
By Xiaokang Yuan, Binhui Yao, Luwei Yang, Ruimin Chen, Zeng Wang
FeatureArc is a new benchmark that evaluates coding agents on the evolution of software features, requiring agents to extend a feature across multiple stages while preserving earlier functionality. The benchmark uses a pipeline that constructs tasks from real repository histories, validates fail‑to‑pass and pass‑to‑pass tests, and covers four programming languages. In tests on 73 tasks, agents performed well on the first stage (62.7% success) but dropped to 28.5% when the entire feature was evaluated, with inherited defects causing most failures.
By Yunlong Wang, Alyssa Wu, Cuiyun Gao, Jia Feng, Songyuanyi Lu, Zhichao Hu