Skills let language agents reuse experience across tasks as natural-language documents. Recent work evolves skills from interaction, but existing methods have limitations when skills evolve online. Prior methods update a skill offline from a batch of trajectories collected over many tasks. Some then adopt the new skill without any check, which can degrade the skill over time. Others keep it only if it improves performance on a held-out validation set, which is too slow and costly to run after every update. We introduce EvoGuard, a pairwise skill selector that mitigates skill degradation at low latency and cost. We first run the agent iteratively to build pools of skills and score each skill on a private test set to obtain utility labels. We pair the skills and use rejection sampling to keep only pairs with a clear utility gap. The selector is trained on the logits of two label tokens to pick the better skill. At deployment it decides in a single forward pass, which is significantly faster than evaluation on the private test set. Compared with adopting every new skill, EvoGuard improves success by 11.0% on ALFWorld, 5.9% on ScienceWorld, and 2.7% on LiveCodeBench.
In recent years, AIGC technologies are capable of generating hyper-realistic forged facial images, posing severe threats to facial security safety. To tackle these challenges, early Deepfake research primarily centered on real-fake classification. Recently, growing efforts have been devoted to the Deepfake Attribution (DFA) of generated content. Nevertheless, existing research on Deepfake detection and attribution faces saturated performance in binary classification, limited diversity in datasets and algorithms, and imperfect evaluation protocols, which severely impede practical application. To address these limitations, we propose a comprehensive deepfake detection and attribution benchmark named DeepfakeGenome (DFG). It contains 100 facial forgery algorithms and 2M images in total, achieving 4x to 100x larger than prior DFA benchmarks. We further design 4 protocols for practical evaluation, including a novel retrieval-based attribution paradigm. Unlike previous open-set evaluation metrics, the proposed retrieval metrics are more aligned with the real-world active defense situation of blacklist registration mechanisms. Based on these elaborate designs, we investigate the performance ceiling of deepfake attribution task. Over 2k experimental evaluations are conducted, and 10 insightful findings are derived. We believe this work can provide new insights into the DFA research field. The Code and Benchmark of DFG are avaliable on https://huggingface.co/spaces/DeepfakeGenome/DFG_Page.
Constructive neural routing solvers usually score the next action by matching a decoder context against candidate embeddings computed once per instance from static node features such as coordinates, demands, and time windows; deterministic one-step consequences such as travel, waiting, slack, and capacity changes are not provided as inputs and must be inferred implicitly by the model. We propose LINC, a decoder-side candidate decision architecture that decouples computing these consequences from learning how much they matter. The environment computes the consequences explicitly at each step; candidate-level consequences are scored by a state-conditioned shared linear comparator, while feasible-set summaries modulate the decoder context. This keeps the cached embeddings and standard global matching while giving the scoring function direct access to deterministic transition quantities. The Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) serves as the main constrained-routing testbed, and the same interface extends to the Capacitated Vehicle Routing Problem (CVRP) and Traveling Salesman Problem (TSP). Across external benchmarks and no-retraining scale-transfer settings, LINC consistently improves strong neural baselines, with the advantage becoming more pronounced as test size moves further beyond the training scale, especially on constrained routing problems.
Peng Kuang, Haibo Jin, Xiaoyu Han, Yanli Wang, Xiaopeng Yuan, Ye Yu, Kaidi Xu, Haohan Wang
Process Reward Models (PRMs) have been proven to be highly effective in guiding test-time scaling (TTS) methods, which significantly boost the capabilities of LLM-based multi-agent systems. However, existing PRMs are text-based: they re-encode the entire trajectory text from scratch. In long multi-agent rollouts, the scoring cost, growing quadratically with respect to sequence length L, creates a severe computational bottleneck, severely limiting PRMs' application in long-context scenarios. To resolve this, we introduce KV-PRM, a highly efficient process reward model that eliminates the heavy text re-encoding by directly reading the KV cache produced naturally during the LLM's generation phase. By processing a single "verify token" against the pre-existing KV cache, KV-PRM reduces the scoring cost from O(L^2) to O(L). We formally prove that the KV cache contains strictly greater information capacity than text, and is more efficient for downstream reward modeling. Empirically, across the MATH, GSM8K, and AIME benchmarks, KV-PRM matches or strictly outperforms text-PRMs under various TTS methods such as Beam Search, MCTS, and Weighted Voting, with up to a 5,000x reduction in scoring FLOPs, a 37x reduction in latency, and a 34x reduction in per-sequence memory footprint compared to text-based PRMs.
Speculative decoding accelerates rollout generation, which dominates the cost of reinforcement learning (RL) post-training. Online co-training keeps the draft aligned with the evolving policy, further increasing acceptance length and speedup. However, training large models on long contexts requires context parallelism (CP) and pipeline parallelism (PP), and co-training drafts under this layout poses two obstacles: (1) the draft's branch attention is unsupported by causal CP, and (2) the target features that the draft consumes are spread across PP stages. In this paper, we present the design and implementation of online draft co-training in NeMo-RL. For CP, we extend packed, load-balanced zigzag ring attention by merging rank-local branch attention with causal main-sequence attention. For PP, TapChannel transports intermediate target features across stages via a separate path, leaving the pipeline schedule unaffected. Experiments show that co-training keeps the policy's reward and accuracy close to the no-draft baseline while delivering $1.48$--$2.25\times$ rollout speedups and $1.19$--$1.88\times$ end-to-end training speedups across model scales up to 122B. Our CP design achieves strong scaling at 256K sequence length with significant memory savings over prior work, and our PP transport incurs modest overhead. Code can be found at https://github.com/NVIDIA-NeMo/RL/issues/3698.
Hongcheng Gao, Jianshu Zeng, Alex Zhou, Shijia Ge, Xuan Shangguan, Yazhe Wang, Yexing Wan, Zelin Zheng 等 16 位作者
Agents built on vision language models (VLMs) now control robots by writing code, but the strongest of them refine task-specific code over many rounds of trial and revision. Such code tends to overfit the benchmark on which it was tuned, can exploit the benchmark's shortcuts, and is costly to carry over to a new benchmark or robot. We present HexaShell, a thin, task-agnostic harness designed to test how far general tools can go instead. We design a simple and complete set of 14 basic tools, 4 for motion, 5 for perception and 5 for state, which the agent calls in one Python cell per turn. Perception pairs SAM 3, which segments an object from a few words or a click, with NumPy and OpenCV geometry that turns its outlines in several cameras, or a depth image, into its position, size and grasp angle. The API, agent loop and sandbox are shared across all tasks; a new simulator or robot needs only an adapter of 350 to 1,450 lines and a prompt describing its world, and the only task-specific input is an optional skill of about a page of guidance. The same code runs on four simulated benchmarks, built on three simulators with two kinds of arm, and on real robots. With GPT-6-Astra and no task-specific input, HexaShell reaches success rates of 69.9% on RoboDojo, 59.3% on RoboCasa365, 78.3% on LIBERO-PRO and 82.2% on RoboTwin 2.0. With task skills, it reaches 85.0% (score 88.8), 85.6%, 96.7% and 96.8%, respectively, above the best published result we are aware of on each benchmark. The RoboCasa365 result is listed first on its leaderboard. On real dual-arm robots, HexaShell succeeds in 76 of 90 episodes across nine tasks.
Adapting step-distilled text-to-image (T2I) models through post-training incurs additional computational costs and affects native few-step generation behavior. This motivates a complementary route beyond style-specific adaptation: drawing on the visual knowledge already encoded in step-distilled T2I models to elicit stylistic capabilities through language. Pursuing this direction requires textual guidance that captures how visual attributes jointly define a style and remain applicable as the depicted content changes. To explore this approach, we introduce StyleForge, a fully automatic, training-free framework that expresses reference styles as reusable rendering instructions. By integrating overall rendering characteristics with local color and lighting behavior, StyleForge organizes visual evidence from reference images into a coherent specification of how the target style should be expressed. The specification is then compiled into textual guidance that can be reused across content prompts, enabling frozen step-distilled T2I models to render different subjects and scenes in the reference style while retaining native few-step generation. Extensive experiments show relative gains of up to 29.47\% in generation quality scores over the strongest baseline, while Pareto analysis indicates that improved stylization is accompanied by strong adherence to the requested content. Code is available at https://github.com/sunshy-1/StyleForge.
Visualizing multivariate scientific simulation data is essential for understanding complex physical phenomena, yet the massive scale of such data imposes a severe storage bottleneck in high-performance computing environments. Implicit neural representations (INRs) have emerged as a promising continuous alternative to traditional lossy compressors, but existing methods struggle to explicitly capture cross-variable interactions and to adapt to the diverse spectral characteristics of physical variables. To address these limitations, we propose SWAP-INR, an architecture for high-fidelity compression of multivariate, time-varying simulation data. SWAP-INR introduces two key components: (1) The switchable attention path (SWAP) module explicitly models cross-variable interactions through a cross-variable attention layer, then routes each variable to an attention or non-attention path by a deterministic, loss-driven criterion, and prunes the inactive layers for structural compression. (2) The adaptability amplification SIREN (A2-SIREN) replaces rigid frequency scaling with channel-wise learnable modulation, adapting to heterogeneous physical features with negligible overhead. Combined with post-training 16-bit quantization, SWAP-INR achieves state-of-the-art performance on small modular reactor (SMR) datasets and remains competitive on public physics simulation datasets, while maintaining a constrained memory footprint. Our code is https://github.com/hvcl/SWAP-INR.
Long-context inference is bound by the KV cache, which 2-bit codes could shrink about sixfold if their error stays in check. The usual remedy rotates keys before quantizing, assuming rotation always helps. However, we observe that it does not: rotation flattens a spiky piece of a key but can give a flat one a new peak, and both coexist within one key. We therefore propose FlagRot, which stores each 32-channel piece rotated or as is, whichever has the smaller calibrated attention error, and rotates the query instead when reading. We further show that how often to decide follows the quantization axis: per-channel caches such as KIVI need one decision per head, at no write cost, while per-token caches decide per key, with the choice stored in existing sign bits. At 2.5 bits per value, FlagRot lowers the twelve-model KL divergence of KIVI from 18.7 to 12.4 x 10^-3, below RotateKV on 9 of 12 models. Served as a vLLM plugin, it raises RULER at 128k tokens by +2.9 points and the LongBench v1 mean by +0.4 to +1.9 points on the three of four models where 2-bit KIVI loses accuracy. It also lowers the error of 4-bit integer and FP4 caches.
Zhaoyang Wang, Qianhui Wu, Boyu Gou, Demi Wang, Jiacheng Zhang, Peichun Hua, Qize Li, Shi Qiu 等 57 位作者
GUI web agents require realistic and interactive environments that remain stable during both evaluation and training. Live website evaluation suffers from changing content, access restrictions, and limited coverage of authenticated workflows, while existing synthetic environments for training often lack functional depth and visual fidelity. We propose WebHarbor, a suite of 153 executable and dockerized website environments that faithfully reconstruct the visual interfaces, content, and functional workflows of real-world web services under deterministic control. We build WebHarbor through task-driven human-agent collaboration, where environments evolve as new tasks expand their functional scope, coding agents implement the required changes, and human contributors audit visual fidelity, functional correctness, and task validity. The resulting suite comprises 3,518 tasks spanning multi-step and post-authentication workflows, where existing works often fail to cover. Beyond reliable evaluation, WebHarbor provides high-fidelity synthetic web environments for cost-efficient reinforcement learning (RL). We demonstrate that a Qwen 4B-level GUI agent trained with RL in WebHarbor generalizes to unseen tasks and performs competitively with leading open-source GUI web agents across four benchmarks. The code and data are publicly available at https://github.com/aiming-lab/WebHarbor.
Fuyuan Ma, Yili Wang, Xin He, Xin Zhang, Ying Wang, Xin Wang
Dynamic graphs capture evolving interactions in Web systems, including online communication, user engagement, and recommendation. Graph foundation models aim to reuse predictive capabilities across networks, but extending this capability to dynamic graphs faces two challenges. First, node identities are graph-specific, and similar interaction patterns can appear at different temporal and activity scales, making histories difficult to represent consistently across networks. Second, the predictive roles of historical patterns vary across interaction dynamics, limiting transfer from a narrow range of training processes. In this paper, we show that a common interaction language, combined with learning across diverse interaction processes, offers a basis for dynamic graph foundation modeling. We propose CL-DGFM, a Common Language Dynamic Graph Foundation Model for temporal link prediction. Its Dynamic Interaction Tokenizer encodes interaction memory, historical connectivity, and ordered destination transitions into calibrated tabular inputs, making dynamic histories accessible to existing predictors without an additional trainable graph encoder. Its Dynamic Interaction Prior trains a pretrained tabular foundation predictor across diverse synthetic dynamics. Labeled historical context guides prediction, with a Contextual Value Adapter modulating attention values. The model transfers to real graphs without target-graph parameter updates. Experiments on TGB and TGB-Seq demonstrate competitive performance. Further comparisons show that the common language supports basic machine learning and neural models, and enables direct use of frozen tabular foundation models on dynamic graphs.
Asynchronous reinforcement learning (RL) is becoming a mainstream paradigm for post-training large language models (LLMs), as it improves training efficiency by updating the model as rollouts arrive. Recent work argues that group-wise sampling in GRPO does not naturally fit asynchronous RL training because rollout lengths follow a long-tailed distribution, and proposes Single-rollout Asynchronous Optimization (SAO) to improve stability and off-policy behavior, especially on agentic tasks. However, SAO relies on a closed-source value model obtained from scaling value pretraining. In this paper, we analyze GRPO and LiSAO, our lightweight reproduction of SAO without the pretrained value model, on single-turn reasoning and multi-turn agentic tasks. We find that GRPO outperforms LiSAO on reasoning tasks, while the reverse holds on agentic tasks. Since the value model of LiSAO is not yet well trained , LiSAO lags behind GRPO substantially in early training steps. To address this issue, we propose Li$^{2}$SAO, which Lifts Light Single-rollout Asynchronous Optimization with group-relative actor-critic bootstrapping: while the critic is cold, a group-relative advantage trains the actor and a group-centered return trains the critic, after which training switches to single-rollout GAE. Experiments show that Li$^{2}$SAO catches up with or surpasses GRPO in early training and further raises the later performance ceiling. Code is available at https://github.com/jd-cc-coder/Li2SAO.
Xiaopeng Yuan, Suijin Wang, Yanli Wang, Haibo Jin, Peng Kuang, Jerry Wang, Lijun Yu, Haohan Wang
LLM-based formal provers can retrieve relevant lemmas and prior proofs, but relevance alone does not say whether a mathematical method can be used on the current theorem. A method has prerequisites, a target, an intended action, and obligations that its use leaves to prove. Methods that look equally related to a theorem may therefore differ substantially in whether they offer a plausible next step. We formulate this as an applicability-aware method-selection problem and introduce self-advertisement: before candidates are ranked, a model generates a problem-specific proposal for each one, stating what part of the goal it targets, what action it would take, and what conditions that action requires. We organize 82 reusable methods from Putnam 2000-2014 as Method Contracts, which pair applicability descriptions with Mathlib anchors, a checked example or scaffold, and expected proof obligations. A single batched call elicits proposals across the library; vague or unsupported proposals are demoted, yielding a ranked shortlist accompanied by inspectable claims about each candidate's use. We analyze when similarity-based representations cannot distinguish methods with different applicability, how errors in applicability estimates affect shortlist quality, and what a checked scaffold guarantees under its stated assumptions. Against lexical, embedding, and embedding-plus-LLM reranking baselines, self-advertisement achieves 95.0% hit@5 on Putnam 2015-2025, compared with 84.2% for the strongest reranker. On IMO ProofBench, it achieves 91.7% compared with 88.3%. These results indicate improved coverage of annotated methods in the retrieved shortlists, particularly on Putnam.
Localized false-merge correction in connectomics requires converting an identified error into an executable native 3D partition. Still, neither intermediate correction predictions nor seeded graph solvers alone determine which native source-sink constraints should define that partition. We formulate this problem as a 3D-native partition-aware contact selection problem: selecting candidate contacts based on the complete partitions they induce. Our solution comprises two components: ContactCut (CC)-Auto, which uses learned single-contact selection, and ContactCut (CC)-Search, which enables bounded paired-contact search, generates compact alternatives, and applies learned panel ranking. On historical two-way false-merge corrections from MICrONS and FlyWire, CC-Auto improves failure-aware 3D partition (FA-3D Part.) mIoU by 5.9 and 3.4 points over the strongest graph-based controls and by 12.5 and 10.0 points over the best learned baselines. A MICrONS-trained ranker transfers to FlyWire with negligible degradation. CC-Search expands oracle reachability and prioritizes near-oracle alternatives. The code is available at https://github.com/bentake/ContactCut.
Localization is an essential capability of intelligent robots. However, in real-world constrained environments, robots often cannot rely on external localization systems. Therefore, this paper proposes an online particle-filter localization method based on Monte Carlo sampling. The method is applied to the localization of a mobile robot in a maze environment and requires only four range sensors mounted on the front, rear, left, and right sides of the vehicle, without the need for any heading information. Simulation experiments demonstrate that even in a maze where local regions are highly similar, the vehicle can still achieve relatively accurate localization, thereby providing a fast and effective approach for robot localization in complex environments.
Ali Al-Lawati, Nafis Tripto, Abolfazl Ansari, Fali Wang, Jason Lucas, Suhang Wang, Dongwon Lee
The emergence of multi-agent systems introduces novel moderation challenges that extend beyond content filtering. Agents with $\textit{malicious intent}$ may contribute content that appears benign to evade content-based moderation, while serving a hidden objective. To address this challenge, we introduce Bot-Mod a moderation framework that grounds detection in agent intent rather than traditional content-level signals. Bot-Mod interrogates a target agent using a fixed set of diagnostic probes and infers its intent through Bayesian posterior inference over candidate intents. It starts with a calibrated content-only prior and updates the posterior using the $\textit{likelihood}$ of each answer under every candidate intent, as assigned by the moderator, until a stopping criterion is met. To evaluate our approach, we construct Molt-Intent, a dataset derived from Moltbook, capturing diverse benign and malicious behaviors within real community structures through posts and comments. Across Molt-Intent, an external social engineering dataset, and three agent LLMs, Bot-Mod outperforms content-only moderators, a fine-tuned classifier, and judge-based interrogators both in- and out-of-distribution. On Molt-Intent, it reaches a composite F1 of 0.736 in distribution and 0.684 out of distribution, 22% and 14% above the best baseline.
Hongcheng Gao, Zelin Zheng, Alex Zhou, Jay Zhu, Shijia Ge, Yazhe Wang, Xiping Li, Jianshu Zeng 等 21 位作者
Embodied agents are expected to accomplish real-world goals by deciding for themselves how and when to interact with their surroundings. For many such goals, changing the physical state of the world is the means rather than the end: the agent changes the state of a system through interaction, observes the feedback, and explores until it either achieves the goal itself (e.g., discovering physical laws) or learns what the goal requires; it acts to know. Manipulation benchmarks stop at acting to change, with the target state given; scientific discovery benchmarks evaluate knowing but replace physical experimentation with symbolic actions or numerical queries. We introduce EmbodiedLab, a benchmark in which a robot must act to know. Quantitative science experiments make this setting controllable: the parameters of the system are resampled for every instance, so the result cannot be known in advance, yet it can be scored exactly. Across 30 experiments in five disciplines on 14 apparatus platforms, the agent receives an objective and the apparatus, chooses its own procedure, and operates a physics-based simulated laboratory, reading the instruments through its cameras. Evaluation checks what the agent came to know, grounded in readings recorded during the episode, as well as the physical state the task requires; every experiment is verified to be solvable. We evaluate six frontier models in their developers' coding agents with native observation and control, with added perception and motion planning, and with manipulation skills distilled from the strongest agents. With native control, even the strongest agent succeeds less than half the time; for every agent, most failures arise in interacting with the apparatus, and manipulation skills help weaker agents more than perception and planning. The benchmark and its evaluation code are available at https://github.com/HexaFuture/EmbodiedLab.
Yunqi Liu, Peiyu Wang, Bingli Wang, Leo Luo, Xuanwu Song, Xudong Ma, Hongxin Li, Yaxin Du 等 24 位作者
Command-line interface (CLI) agents increasingly execute longer-horizon tasks, yet existing benchmarks either assume that user instructions remain fixed throughout execution or evaluate task outcomes without capturing runtime collaboration quality. In this paper, we present REALCLI, a paired benchmark that unifies the evaluation of task execution and human-agent collaboration quality on realistic CLI tasks. To comprehensively measure collaboration, REALCLI builds a five-dimensional framework, operationalized into 22 sub-dimensions and 51 interaction scenario templates. REALCLI provides two modes for paired comparison: REALCLI-EXEC supplies complete specifications upfront, whereas REALCLI-COLLAB introduces a scripted user that delivers progressive information, corrections, constraint revisions, and authorization decisions. Evaluations reveal profound capability gaps in frontier agents: the best configuration attains an average score of only 49.2%, and when evaluated by the joint collaboration score ($J_{\mathrm{collab}}$), the peak performance reaches merely 37.6%. These results show that task effectiveness diverges sharply from collaboration quality, with models consistently struggling in steerability, trust calibration, and safety boundary management. REALCLI provides a rigorous, reproducible foundation for measuring and advancing collaborative CLI agents.
Yanli Wang, Yanlin Wang Sun Yat-Sen University, Bowen Zhang, Yiwei Zhang, Daya Guo, Jiachi Chen, Hongyu Zhang Chongqing University, Zibin Zheng Sun Yat-sen University
Code Large Language Models face critical Time-To-First-Token (TTFT) latency challenges when handling long code completion due to the quadratic complexity (O(n^2)) of attention mechanisms. While existing sparse attention methods attempt to address this issue, they suffer from three key limitations: (1) general sparse patterns cause excessive accuracy degradation without considering code structure, (2) code-specific methods achieve only logical sparsity without actual computational speedup, and (3) limited adaptation to complex scenarios such as repository-level completion. We propose SabreCoder, a training-free Structure-aware block-sparse attention mechanism that bridges the gap between logical and computational sparsity. SabreCoder parses code into semantic chunks, constructs chunk-level sparse patterns through dependency analysis and similarity matching, and maps them to GPU-friendly block-sparse formats. Extensive experiments on LCC and CrossCodeEval benchmarks demonstrate that SabreCoder reduces TTFT by 45-55% while maintaining accuracy within 3% of dense attention.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Qianli Wang, Van Bach Nguyen Phillips-Universität Marburg, Yihong Liu Ludwig-Maximilians-Universität München, Fedor Splitt Technische Universität Berlin, Nils Feldhus, Christin Seifert Phillips-Universität Marburg, Hinrich Schuetze, Sebastian Möller 等 9 位作者
Counterfactuals refer to minimally edited inputs that cause a model’s prediction to change, serving as a promising approach to explaining the model’s behavior. Large language models (LLMs) excel at generating English counterfactuals and demonstrate multilingual proficiency. However, their effectiveness in generating multilingual counterfactuals remains unclear. To this end, we conduct a comprehensive study on multilingual counterfactuals. We first conduct automatic evaluations on both directly generated counterfactuals in the target languages and those derived via English translation across six languages. Although translation-based counterfactuals offer higher validity than their directly generated counterparts, they demand substantially more modifications and still fall short of matching the quality of the original English counterfactuals. Second, we find the patterns of edits applied to high-resource European-language counterfactuals to be remarkably similar, suggesting that cross-lingual perturbations follow common strategic principles. Third, we reveal that multilingual counterfactual data augmentation (CDA) yields larger model performance improvements than cross-lingual CDA, especially for lower-resource languages. Yet, the imperfections of the generated counterfactuals limit gains in model performance and robustness. Finally, we identify and categorize four main types of errors that consistently appear in the generated counterfactuals across languages.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
3D Visual Grounding (3DVG) locates objects in 3D scenes based on natural language descriptions. However, existing methods are primarily confined to small-scale indoor data or rely on heavy supervision, failing to generalize to the complexity of large-scale urban environments. To address this limitation, we present CityVG, the first city-scale zero-shot 3D visual grounding framework capable of localizing urban objects without manual annotations. Our approach adopts a retrieval-and-reasoning paradigm comprising two key components. Specifically, we propose a contrastive fine-tuning strategy to align textual queries with urban scene graphs. By leveraging an LLM-driven graph clustering mechanism, we automatically construct high-quality positive and negative training pairs and fine-tune the text encoder via contrastive learning, resulting in a scene-adaptive text encoder that enables efficient alignment without grounding supervision. Complementing this, we introduce a multi-trajectory reward-based Chain-of-Thought (CoT) reasoning strategy for inference. This mechanism iteratively evaluates candidate objects by aggregating reward scores across diverse reasoning trajectories, selecting the target that is most consistent with both appearance and spatial constraints. Extensive experiments on city-scale 3D grounding benchmarks demonstrate that CityVG achieves strong zero-shot localization performance and generalizes effectively to unseen urban environments.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Rui Miao, Yixin Liu Griffith University, Yili Wang, Xu Shen, Yue Tan Griffith University, Yiwei Dai, Shirui Pan Griffith University, Xin Wang
The security of LLM-based multi-agent systems (MAS) is critically threatened by propagation vulnerability, where malicious agents can distort collective decision-making through inter-agent interactions. While existing supervised defense methods demonstrate promising performance, they may be impractical in real-world scenarios due to their heavy reliance on labeled malicious agents to train a supervised malicious detection model. To enable practical and generalizable MAS defenses, in this paper, we propose BlindGuard, an unsupervised defense method that learns without requiring any attack-specific labels or prior knowledge of malicious behaviors. To this end, we establish a hierarchical agent encoder to capture individual, neighborhood, and global interaction patterns of each agent, providing a comprehensive understanding for malicious agent detection. Meanwhile, we design a corruption-guided detector that consists of directional noise injection and contrastive learning, allowing effective detection model training solely on normal agent behaviors. Extensive experiments show that BlindGuard effectively detects diverse attack types across MAS with various communication patterns while maintaining superior generalizability compared to supervised baselines.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jinrui Wang Beijing University of Posts and Telecommunications, Zhenfeng Gao, Wendan Wang, Huili Wang, Zichen Qin, Linjie Zhu, Hongke Fu Sangfor Technologies Inc., Shangguang Wang Xiong'an Aerospace Information Research Institute and Beijing University of Posts and Telecommunications 等 9 位作者
The increasing use of video data in training multimodal large language models (MLLMs) raises significant concerns on privacy leakage and copyright violations, highlighting the need for detecting improperly used training videos through membership inference attacks (MIAs). Most existing video MIA methods assess model memorization of key semantic concepts within a video (e.g., the name of a well-known movie character). However, such concepts usually appear repeatedly throughout the training corpus, and memorization of them does not constitute reliable evidence that a specific video was used during training. Besides, while some methods mitigate this limitation by capturing relationships between frames, they require a model logit-accessible setting and are impractical in realistic black-box scenarios. To address these challenges, we propose a black-box MIA framework, named VideoMIA, that can provide reliable evidence of specific video data usage for training MLLMs. The key of our method is to leverage temporal dependencies across video frames to evaluate the model’s memorization of sequential dynamics within the video data, which cannot be inferred solely from general world knowledge or individual image data. The results across ten MLLMs and four benchmarks demonstrate that our method consistently achieves superior performance over all baselines in black-box evaluation settings. Code is available in https://github.com/jinruiwang258/VideoMIA.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yuanhong Huang Beijing University of Posts and Telecommunications, Huili Wang, Xueying Bai, Jinrui Wang Beijing University of Posts and Telecommunications, Jiajun Liu, Ziqin Wang, Wanchun Ni ETHZ - ETH Zurich, Shangguang Wang Xiong'an Aerospace Information Research Institute and Beijing University of Posts and Telecommunications 等 9 位作者
Membership inference attack (MIA) has emerged as a promising tool for auditing the training data of LLMs, supporting data privacy and copyright protection. Most existing MIA methods rely on the assumption that LLMs assign higher confidence scores to training samples than to non-training ones.However, since LLMs generate text by sampling high-confidence tokens, they naturally produce AI-generated texts (AIGTs) that also satisfy this assumption.In this work, we empirically confirm that such AIGTs, regardless of whether they are generated by the target LLM, can lead existing MIAs to assign even higher membership likelihoods than those of true training samples, thereby significantly undermining their reliability.To address this challenge, we propose a robust membership inference framework for reliably identifying training data.Our method adopts a mixture-of-experts formulation to jointly model interactions across complementary features derived from multiple MIA methods and AIGT detectors, which can remain robust against adversarially generated samples.Furthermore, by leveraging expert components, our method provides explainable insights into the characteristics of member data.Experiments on various datasets and LLMs show that adversarial samples substantially degrade the performance of baselines, whereas our method preserves performance close to that of the unattacked setting.Codes and datasets are released at https://github.com/kong-hyh/MoMIA.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表