The paper proposes a cost‑efficient privacy enforcement method for email‑generating large language model (LLM) agents that intercepts drafts, creates an information‑release contract, rewrites the draft locally with a Gemma rewriter, and then optionally calls a final PrivacyChecker. Experiments on normal and strategically malicious prompts show that the Gemma + Checker approach reduces leakage from 11.41 % to 4.03 % and improves usable output rates, while also cutting the cost amplification of checker‑only pipelines from 1.32× to 1.02×. The results demonstrate that moving repair to a cheaper local layer can significantly lower both privacy leakage and operational cost, yet still rely on a checker for high‑risk sends.
By Qikun Cai, Tao Huang, Chen Hou
HaloBlock introduces a difficulty‑aware adaptive halo mechanism for diffusion large language models (dLLMs) to address two inference challenges: decoding difficulty dominated by right‑side tokens and boundary contextual discontinuity in block‑wise decoding. By using bidirectional cumulative entropy to assess block difficulty, HaloBlock allocates halo regions on both sides of each block, smoothing cross‑block context propagation and reducing difficulty for hard tokens. Experiments on LLaDA‑8B‑Instruct, Dream‑7B‑Instruct, and LLaDA‑1.5 show consistent accuracy gains and faster convergence on mathematical reasoning and code generation tasks.
By Zhaohui Zheng, Chenhang He, Jinrui Zhang, Ming-Ming Cheng, Lei Zhang
SWEPatchTest is a benchmark that evaluates software tests by measuring their ability to reproduce issues, accept correct patches, and reject incorrect ones across 619 tasks and 82,579 candidate patches. It introduces the Joint Validation Rate metric, revealing that even the best configurations only achieve 9.21% success when all three criteria are met. Using these validation scores to select demonstrations for supervised fine‑tuning can roughly triple the Joint Validation Rate and enhance downstream repair performance on multiple SWE‑bench datasets.
By Hua Lu, Pengcheng Wang, Han Li, Zichen Tian, Yanyu Chen, Huang He, Bingjin Chen
Virtual Head Attention (VHA) is a lightweight attention mechanism that improves long‑context large language models by adding two linear operations—Q Premix and Linear Postmix—to recover head diversity and fuse inter‑head features without nonlinearities. VHA reduces KV cache size, training and prefill FLOPs, and maintains or surpasses model quality compared to existing methods such as GQA, MLA, and MFA across multiple scales and benchmarks. The design allows VHA to be folded into standard GQA inference pipelines, enabling efficient deployment with only two KV heads and full exploitation of Tensor Core parallelism.
By Xibo Ding, Guoxia Wang, Jinle Zeng, Jiabin Yang, Sirui Tao, Chao Yang, Dianhai Yu
The paper introduces Predictability-Aware Canvas Selection and Supervision (PACS), a supervised fine‑tuning method for diffusion language models that aligns training with inference by using the model’s own predictability to determine which tokens are visible and when they are supervised. PACS selects the attention‑visible input length for each example and supervises reference tokens just before they become confident, addressing mismatches caused by random masking. Experiments on reasoning and code‑generation benchmarks show PACS consistently outperforms standard SFT, achieving up to 12.84 percentage‑point gains on individual tasks.
By Miao Li, Hanyang Jiang, Hengyu Fu, Baihe Huang, Yating Liu, Yuhang Cai, Changkun Guan, Zhongzhu Chen, Jiecheng Lu, Michael A. Boateng, Pascal Van Hentenryck
FETALFusion is a dual‑path CNN‑Mamba encoder‑decoder designed for fetal head segmentation in 2D ultrasound across transabdominal and transperineal acquisition types. It uses four‑directional state‑space scans at the highest‑resolution stages, domain‑specific channel gates, and a weak landmark shape prior, achieving a Dice score of 0.963 and an HD95 of 2.10 mm on a held‑out test set of 471 images. Ablation studies show that the landmark prior and domain gates drive most of the performance gains, while the state‑space branch has minimal impact.
whyItMatters":"The method provides highly accurate fetal head segmentation without test‑time augmentation or domain labels, enabling reliable head circumference measurement for monitoring fetal growth."
By Ammar Bhilwarawala, Mainak Bandyopadhyay
GIFT‑Eval‑CTX is a new benchmark for contextual forecasting that tests whether models can use verified future information to produce accurate time‑series predictions. It builds on GIFT‑Eval and contains 10,000 instances across 20 datasets, each annotated with macro‑level and micro‑level effects derived from actual future events. The benchmark evaluates traditional TSFMs, advanced LLMs, and agentic systems, showing that structured effects can substantially lower MASE for models like GPT‑5.6 Sol and Claude Opus 4.8, with MoiraiAgent achieving the best overall performance.
By Xu Liu, Hanshu Yan, Taha Aksu, Quang Pham, Zhongzheng Qiao, Jun Hao Liew, Silvio Savarese, Doyen Sahoo, Junnan Li
QK-Edit introduces a training‑free query‑key manipulation strategy for Multimodal Diffusion Transformers (MM‑DiTs), enabling precise text‑guided image and video editing. By isolating and adjusting critical attention components, it balances prompt fidelity with structural consistency, supporting tasks such as object addition, removal, replacement, background changes, material and color adjustments, and style transformation. The method works seamlessly with state‑of‑the‑art models like FLUX and HunyuanVideo, improving editing performance without requiring additional training.
By Tiancheng Shen, Zilong Huang, Xiangtai Li, Zhijie Lin, Jiyang Liu, Yitong Wang, Jiashi Feng, Ming-Hsuan Yang, Jun Hao Liew
Diff Diffusion is a diffusion-inspired large language model framework designed for efficient code editing. It predicts edit paths from an original snippet to its revised form by representing denoising as structured edit operations, and uses multi-stage training and efficient parallel decoding to focus computation on changed regions. The approach matches or outperforms autoregressive baselines while achieving up to 2.78× faster inference and a 39.0% reduction in median model latency in production.
By Xiaolong Xu, Hao Jiang, Bo Jiang, Zhiwen Deng, Ningyuan Sun, Bingzhou Chen, Jin Chen, Chenxing Wei
ProtoArena is a new black-box benchmark that tests AI agents on 80 cross‑component vulnerability‑exploitation environments across 11 protocol and application domains and 7 attack categories. It evaluates agents at three information levels—guided reproduction, exploitation with context, and effect‑directed black‑box exploration—using a hybrid verifier that combines rule‑based checks with LLM review of traffic and logs. In experiments, Claude Opus 4.8 achieved the highest success rates, reaching 48.8% under guided conditions and 12.5% in effect‑directed black‑box scenarios, while also demonstrating significant DNSBomb optimization performance.
By Fasheng Miao, Zhaoyu Hu, Xiang Li, Changqing An, Jiutian Zeng, Longtao Huang, Haixin Duan, Jilong Wang
The study evaluates whether signals from a frozen vision‑language model, InstructBLIP, can reliably indicate answerability in visual question answering. Using a 10K subset of UNK‑VQA, the authors test predictive uncertainty, cross‑modal alignment, visual grounding, visual dependence, internal representations, and learned probes, finding that all signals yield weak discrimination (AUROC 0.50‑0.54). Even interventions that remove question‑relevant visual evidence or increase probe capacity do not substantially improve answerability detection, underscoring that visual dependence and related metrics are not robust proxies for answerability.
By Minjun Kim, Hyunmin Song, Changwon Lim
The paper introduces FaithAlign, a theory‑scale benchmark for evaluating faithfulness in autoformalization, consisting of 460 items with 4,982 annotated phrase‑level links and a companion set FaithAlign‑Gap with 146 items lacking informal context. It frames faithfulness evaluation as a context‑aware align‑then‑label task, requiring identification of phrase‑level correspondences between informal and formal statements under specified contextual interpretations. Experiments on eight models show that GPT‑6 Astra achieves the best defect F1 of 47.8% but still struggles to recover context‑dependent correspondences, highlighting the challenge of fine‑grained, context‑aware faithfulness assessment.
By Yuming Feng, Tadd Mao, Dhruva Arekar, Tianjun Zhong, Jiani Huang, Xujie Si, Ziyang Li
The paper introduces Orchestra, a robotic agent that separates task planning from action generation to follow multimodal, long‑horizon instructions. Orchestra decomposes instructions into ordered subtasks, tracks progress, updates object locations, and uses a progress evaluator to decide when to advance. On the new RMBench‑Compose benchmark, Orchestra outperforms prior methods, achieving higher success rates and longer average subtask sequences, and shows improved performance on real‑world multi‑step tasks.
By Youpeng Wen, Gang Wang, Tianyu Huang, Weiwei Zhang, Zhijian Huang, Wei Chen, Xieyuanli Chen, Yunhui Liu
MemoryVLN introduces a memory-driven navigation system that compresses long navigation histories and reuses past spatial experience. It uses forgetting-curve-inspired temporal sampling for long sequence compression, an ego-aware memory representation that fuses 2D semantic, 3D geometric, and 6D pose tokens, and a hierarchical memory bank for in-episode and context-level memory management. The system retrieves top‑K relevant memories to augment current visual tokens, achieving significant improvements on R2R‑CE and RxR‑CE benchmarks.
By Xueyao Wei, Yifei Wang, Delun Li, Shaoan Wang, Junzhi Yu, Wenjie Song, Xingyu Chen
SolBehave is a benchmark that evaluates the behavioural correctness of smart contracts generated by large language models, going beyond simple compilation checks. It applies three tests—differential fuzzing, historical‑transaction replay, and DCR‑guided synthetic‑transaction replay—on 250 real Ethereum contracts to assess whether generated code matches the original’s behaviour. The study finds that while compilation success varies widely, only a minority of compiled contracts are behaviourally equivalent, highlighting that compilation alone is insufficient for correctness.
By Giovanni Pio Delvecchio, Stefano Fantazzini, Giacomo Frisoni, Gianluca Moro, Lorenzo Bacchiani, Mario Bravetti, Thomas Troels Hildebrandt
The paper investigates whether deeper predictor blocks in adaptive‑compute world models truly improve rollout accuracy and whether a router can select the appropriate depth per step. Experiments on nine DeepMind Control tasks show that deeper predictors do not always yield better accuracy, and that a greedy oracle selecting depth per step can reduce error by up to 37% compared to fixed‑depth mixtures, yet no trained router achieves this advantage. The study also finds that supervising early exits only at the first rollout step can mitigate depth‑inversion issues observed in some tasks.
By Achyuthan Sivasankar
The paper investigates how to best utilize demonstration quality during vision‑language‑action (VLA) pretraining to boost downstream task performance. It compares weighting, filtering, and conditioning strategies across different reward definitions and label granularities, using 5,000 suboptimal human demonstrations and 24 target tasks while keeping demonstration sets and training budgets matched. The study introduces AXIS‑Bench, a benchmark built on crowdsourced simulation demonstrations that allows recomputation of quality labels and replayable trajectories for both simulation and real‑world tasks, showing that simulation pretraining can raise real‑world success rates from 9.6% to 31.7%.
By Mingxuan Yan, Litian Gong, Dihong Huang, Zhixu Li, Tianyu Zhang, Wenqian Zhang, Hefeifei Jiang, Peihao Li, Jian Zhang, Mengfei Zhao, Yikai Tang, Hai Zhai, Zehao Wang, Ruijian Liang, Yanjia Huang, Lin Shao, Changliu Liu, Kaiyu Hang, Zhiwen Fan, Masayoshi Tomizuka, Jianfei Yang, Jiachen Li
The paper introduces GDEPO, a reinforcement learning framework designed to improve sample efficiency in automated theorem proving. GDEPO incorporates dynamic additional sampling, an equal‑right advantage function that separates correctness from reward magnitude, and dynamic additional iterations to refine learning from initially failed samples. Experiments on MinF2F‑test, MathOlympiadBench, and PutnamBench demonstrate that these mechanisms synergistically enhance data utilization and optimization efficiency.
By Zhengqing Yan, Fan Guo, Xinyang Liu, Yi Zhang, Yao Liu, Qi Liu, Jihao Huang, Kang Song
The Token Games (TTG) is an evaluation framework that pits language models against each other by having them create and solve programming puzzles. Using pairwise duels and Elo ratings, the authors benchmarked ten frontier models, achieving rankings comparable to established tests while spending under $200 and requiring no human puzzle creation. The study highlights that generating high-quality puzzles remains difficult for current models and proposes a new paradigm for reasoning evaluation that also tests creativity and task creation.
By Simon Henniger, Gabriel Poesia
ScoringBench is an open, extensible benchmark that evaluates tabular regression models using a comprehensive set of proper scoring rules—such as weighted CRPS, energy scores, and interval scores—alongside traditional point metrics like RMSE. It covers 101 datasets from diverse domains and offers two ranking protocols: an ordinal Demsar/autorank approach and a magnitude‑preserving z‑score ranking. Experiments show that model rankings vary with the chosen scoring rule, revealing that evaluation metric selection is a core modeling decision, especially when different error types have varying consequences.
By Jonas Landsgesell, Pascal Knoll, Tizian Wenzel