GUI agents that interact with graphical interfaces on behalf of users are a promising direction for practical AI assistants, yet training them is hindered by scarce suitable environments. We present InfiniteWeb, a system that automatically generates functional web environments at scale for GUI agent training. While LLMs perform well on generating a single webpage, building a realistic and functional website with many interconnected pages faces challenges. We address these challenges through unified specification, task-centric test-driven development, and combining website seed variation with reference design images. Our system also generates verifiable task evaluators enabling dense reward signals for reinforcement learning. Experiments show that our system surpasses commercial coding agents at realistic website construction, and GUI agents trained on our generated environments achieve significant performance improvements on OSWorld and Online-Mind2Web, demonstrating the effectiveness of the proposed system.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Anomaly detection aims at distinguishing between in-distribution samples, which belong to the same distribution as the training set, and out-of-distribution samples, which lie outside of it. In textual anomaly detection, recent approaches routinely apply anomaly detection algorithms directly to embeddings extracted from pre-trained embedding models (two-stage approaches). However, the geometric properties of pre-trained embeddings can hinder the effectiveness of detection algorithms, which often rely on distance-based measures. In this work, we first highlight the relevance of similarity-trained models for textual anomaly detection. Beyond being trained to capture semantic similarities, these models also exhibit geometric properties that appear better suited to detection algorithms. We further demonstrate that, besides model choice, a simple post-processing step can significantly improve anomaly detection by adapting embeddings to the assumptions made by classical detection algorithms. The bulk of our experiments is done on a reformulation of the classification tasks from the MTEB benchmark into anomaly detection tasks.
Large language models (LLMs) excel at general programming but struggle with domain-specific software development. This gap motivates research into domain specialization methods that enable LLMs to learn and utilize domain knowledge and data. However, existing domain-specific code benchmarks focus on assessing what knowledge LLMs possess rather than how they acquire and apply new knowledge, lacking explicit knowledge corpora for developing domain specialization methods. To this end, we present KOCO-bench, a novel benchmark designed for evaluating domain specialization methods in real-world software development. KOCO-bench contains 6 emerging domains with 11 software frameworks and 25 projects, featuring curated knowledge corpora alongside multi-granularity evaluation tasks including domain code generation (from function-level to project-level with rigorous test suites) and domain knowledge understanding (via multiple-choice Q A). Unlike previous benchmarks that only provide test sets for direct evaluation, KOCO-bench requires acquiring and applying diverse domain knowledge (APIs, rules, constraints, etc.) from the corpora to solve evaluation tasks. Our evaluations reveal that KOCO-bench poses significant challenges to state-of-the-art LLMs. Even with domain specialization methods (e.g., SFT, RAG, kNN-LM) applied, improvements remain marginal. Best-performing coding agent, Claude Code, achieves only 34.2%, highlighting the urgent need for more effective domain specialization methods. We release KOCO-bench, evaluation code, and baselines to advance further research at https://github.com/jiangxxxue/KOCO-bench.
OASIS: Mitigating Harmful Fine-tuning Attacks on LLMs via Orthogonal and Adaptive Safety Alignment Strategy
PDF ↗The "Fine-Tuning-as-a-Service" paradigm exposes large language models to catastrophic safety degradation from less harmful samples. Alignment-stage defenses address this by proactively injecting adversarial perturbations to bolster the model’s inherent robustness against harmful drift. However, existing methods rely on perturbation directions that often conflict with harmful gradients, inadvertently facilitating the acquisition of malicious features rather than suppressing them. To address this issue, we propose Orthogonal and Adaptive Safety Alignment Strategy (OASIS) to mathematically decouple safety enforcement from harmful feature acquisition. By projecting perturbations orthogonal to harmful gradients and concentrating optimization on adaptively selected safety-critical layers, OASIS effectively resolves directional conflicts while maximizing parameter efficiency. Extensive experiments on four LLMs across three datasets (SST2, GSM8K, and AGNews) demonstrate that OASIS reduces the Harmful Score by approximately 60% compared to competitive baselines, while maintaining stable downstream task utility.
Large Language Models (LLMs) demonstrate exceptional capability across diverse tasks. However, their deployment in long-context scenarios is hindered by two challenges: computational inefficiency and redundant information. We propose RAM (Read As HuMan), a context compression framework that adopts an adaptive hybrid reading strategy, to address these challenges. Inspired by human reading behavior (i.e., close reading important content while skimming less relevant content), RAM partitions the context into segments and encodes them with the input query in parallel. High-relevance segments are fully retained (close reading), while low-relevance ones are query-guided compressed into compact summary vectors (skimming). Both explicit textual segments and implicit summary vectors are concatenated and fed into decoder to achieve both superior performance and natural language format interpretability. To refine the decision boundary between close reading and skimming, we further introduce a contrastive learning objective based on positive and negative query–segment pairs. Experiments demonstrate that RAM outperforms existing baselines on multiple question answering and summarization benchmarks across two backbones, while delivering up to a 12x end-to-end speedup on long inputs (average length 16K; maximum length 32K).
Reinforcement Learning (RL) is the dominant paradigm for training Large Language Model (LLM) agents on long-horizon tasks. However, sparse and delayed rewards often lead to trajectory neglect, in which agents lose focus on the task goal and interaction history at intermediate steps. Prior work has explored step-level supervision using Shannon-entropy–based uncertainty signals, which conflate inherent state complexity with agent confidence and therefore provide unreliable estimates of decision reliability. To address this issue, we propose normalized entropy, which measures confidence deviations relative to an agent’s average behavior under a given state, thereby strengthening the association between low-quality actions and trajectory neglect. Building on this insight, we introduce Selective Trajectory-Aware Policy Optimization (STAPO), a hierarchical group-based RL framework. STAPO leverages normalized entropy to locate outlier steps associated with trajectory neglect and optimizes them via a joint mechanism of trajectory-aware reward and trajectory-independent penalty, enhancing trajectory awareness while preserving training stability. Extensive experiments on ALFWorld, WebShop, and Search-Augmented QA demonstrate that STAPO achieves state-of-the-art performance while substantially alleviating trajectory neglect, validating its effectiveness and robustness for agentic tasks.
Large Vision–Language Models (LVLMs) increasingly rely on retrieval to answer knowledge-intensive multimodal questions. Existing benchmarks overlook conflicts between visual and textual evidence and the importance of generating deflections (e.g., "Sorry, I cannot answer...") when retrieved knowledge is incomplete. These benchmarks also suffer from rapid obsolescence, as growing LVLM training sets allow models to answer many questions without retrieval. We address these gaps with three contributions. First, we propose a dynamic data curation pipeline that preserves benchmark difficulty over time by filtering for genuinely retrieval-dependent samples. Second, we introduce VLM-DeflectionBench, a benchmark of 2,775 samples spanning diverse multimodal retrieval settings, designed to probe model behaviour under conflicting or insufficient evidence. Third, we define a fine-grained evaluation protocol with four scenarios that disentangle parametric memorization from retrieval robustness. Experiments across 20 state-of-the-art LVLMs indicate that models usually fail to deflect in the presence of noisy or misleading evidence. Our results highlight the need to evaluate not only what models know, but how they behave when they do not, and serve as a reusable and extensible benchmark for reliable KB-VQA evaluation. All resources will be publicly available upon publication.
Embodied agents have successfully leveraged large language models (LLMs) to better transform human instructions and images into executable task plans. Furthermore, memories of agents can be leveraged to achieve continual self-learning and optimization. However, vector data quality problems emerge in memories when they are projected into vector space, especially in discerning contextually similar but semantically conflicting sentences and highly similar images. This is particularly detrimental to embodied AI as it potentially distorts the robot’s actions. To address this challenge, we propose Conflict Detection Rules (CDRs) to identify and manage data quality issues in vector knowledge bases, which assist in correcting the index structure and further improving the answer quality. Experimental results show that planners with CDRs exceed the basic LLM planner by 15.25% and 14.25% in grammatical accuracy (GA) and interpretation accuracy (IA) on average, respectively. Moreover, the entire workflow has been successfully integrated into various scenarios, demonstrating its practical applicability and robustness in the real world.
Fine-tuning large language models (LLMs) is intended to improve their reasoning capabilities, yet we uncover a counterintuitive effect: models often forget how to solve problems they previously answered correctly during training. We term this phenomenon Temporal Forgetting and show that it is widespread across model sizes, fine-tuning methods (both Reinforcement Learning and Supervised Fine-Tuning), and multiple reasoning benchmarks. Our analysis reveals on average more than 20% of final errors were once solved correctly at an earlier checkpoint. Inspired by the phenomenon of Temporal Forgetting, we proposed Temporal Sampling, a simple decoding strategy that draws outputs from multiple checkpoints along the training trajectory. This approach recovers forgotten solutions and leads to significant improvements in reasoning performance than final-ckpt-sampling only, gains from 4 to 19 points in Pass@k and consistent gains for majority-voting and Best-of-N across several benchmarks. Temporal sampling also outperforms strong baselines such as model merging. By leveraging the temporal diversity inherent in training, Temporal Sampling offers a practical, compute-efficient way to surface hidden reasoning ability and rethink how we evaluate LLMs.
LLMs often fail in hardware vulnerability detection due to the intrinsic semantic concurrency of HDLs (Hardware Description Language), where vulnerabilities arise from the interaction of multiple concurrent execution statements rather than a single sequential execution path. To address the problem, we propose VerilogLAVD, a LLM-Aided Vulnerability Detection framework by generating executable Traversal Detection Patterns (TDPs), i.e. the rules describing how to find the evidence of vulnerabilities in Verilog HDL. We first introduce a Unified Verilog Property Graph (VeriPG) that explicitly models parallel semantics by combining AST, CFG, and DDG. Furthermore, a semantic validation mechanism is designed to constrain and filter the LLM-generated TDPs. By executing these validated TDPs on VeriPG, our method produces stable and deterministic detection results. Experiments demonstrate that VerilogLAVD improves the F1 score by 133% compared to LLM-based methods. Furthermore, the framework successfully identifies real-world hardware vulnerabilities in open-source hardware design repositories.
Learning speech representations that are useful for a variety of downstream tasks has received considerable attention, due to the outstanding properties of Self-Supervised Learning (SSL) trained models. Despite advancements in modelling methods, understanding the difference in task performance on representations is limited. Mainly motivated by the no-free-lunch theorem and speech production, this work investigates changes in task performance in sparse speech representations, providing interpretability analysis under the Information Bottleneck (IB) framework. Autoencoders with varying sparsity levels were trained using three SSL features, and evaluated on six tasks of SUPERB: Speech Enhancement (SE), Speaker Identification (SID), Speech Emotion Recognition (SER), Phone Recognition (PR), Automatic Speech Recognition (ASR) and Slot Filling (SF). Experiments show that: 1) different tasks manifest different degrees of sensitivity to the sparsity levels; 2) the optimal sparsity level for task performance varies; 3) the choice of SSL features has a limited impact on most tasks but with an exception of PR; 4) overall PR and ASR require more preservation of relevant information about the labels, while SID and SER demand more compression of irrelevant information, where the input quality can shift this trade-off to some degree. These findings can contribute to the design of a universal sparse speech representation learner.
Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval
PDF ↗Sign Language Retrieval (SLRet) enables efficient access to sign language content but remains fragile in fine-grained scenarios where visually similar signs must be distinguished. We show that this limitation does not stem from model capacity, but from ineffective hard negative supervision. Specifically, we formulate fine-grained retrieval failures as a negative distribution mismatch: semantically distinct yet visually confusable signs are rarely treated as hard negatives, while existing text-based mining strategies fail to capture such visual ambiguity. To address this issue, we propose Sign-Aware Hard Negative Mining (SAN), which constructs hard negatives based on visual confusability in the sign embedding space rather than linguistic similarity. Experiments on PHOENIX-2014T demonstrate that SAN substantially improves fine-grained retrieval performance while preserving coarse-grained accuracy, highlighting the importance of aligning negative supervision with visual ambiguity in sign language retrieval.
Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning (Conditioned Stimulus–Unconditioned Stimulus (CS–US) associations shaping first decisions). Our 300-item suite employs a unified Learning/Priming-Interfere-Test protocol with first-attempt scoring. Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall, with top performers DeepSeek-R1 (65.3%), Qwen3-32B (64.1%), and GPT-5 (63.0%) far below human baselines. Analysis uncovers dramatic asymmetries (inhibition 17.6% vs. preference 75.0%) and universal bottlenecks requiring architectural innovations beyond parameter scaling. ImplicitMemBench reframes evaluation from "what agents recall" to "what they automatically enact".
Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities across a wide range of vision-language tasks. However, their performance as embodied agents, which requires multi-round interaction with spatial reasoning and sequential action prediction, needs further exploration. Our work investigates this potential in the context of Vision-and-Language Navigation (VLN) by introducing a unified and extensible simulation-free evaluation framework to probe MLLMs as zero-shot agents, named VLN-MME. Simplifying the evaluation with a highly modular and accessible design streamlines experiments, enabling structured comparisons and component-level ablations across diverse MLLM architectures, agent designs, and navigation tasks. Crucially, enabled by VLN-MME, we observe that enhancing prevalent agents with Chain-of-Thought (CoT) reasoning and self-reflection leads to an unexpected performance decrease. This suggests MLLMs exhibit poor context awareness in embodied navigation tasks; although they can follow instructions and structure their output, their 3D spatial reasoning fidelity is low. Furthermore, we demonstrate that agent performance could be largely improved with simple failure cases in context learning. VLN-MME lays the groundwork for systematic evaluation of general-purpose MLLMs in embodied navigation settings and reveals limitations in their sequential decision-making capabilities. We believe these findings offer crucial guidance for MLLM post-training as embodied agents.
Mobile GUI agents powered by large foundation models enable autonomous task execution in applications, but frequent updates that alter UI appearance and reorganize workflows cause agents trained on historical data to fail. Despite these surface changes, we observe that functional semantics and task intents remain fundamentally stable. Building on this insight, we introduce MAGNET, a memory-driven adaptive agent framework with dual-level memory: stationary memory that links diverse visual features to stable functional semantics for robust action grounding and procedural memory that captures stable task intents across varying workflows. Furthermore, we propose a dynamic memory evolution mechanism that continuously refines both memories by prioritizing frequently accessed knowledge. Evaluations on the online benchmark AndroidWorld demonstrate substantial improvements over memory-augmented baselines, while offline benchmarks confirm consistent gains under distribution shifts. These results validate that leveraging stable structures across interface changes improves agent performance and generalization in evolving software environments.
Supervised fine-tuning (SFT) on long chain-of-thought (CoT) trajectories has emerged as a crucial technique for enhancing the reasoning abilities of large language models (LLMs). However, the standard cross-entropy loss treats all tokens equally, ignoring their heterogeneous contributions across a reasoning trajectory. This uniform treatment leads to misallocated supervision and weak generalization, especially in complex, long-form reasoning tasks. To address this, we introduce **V**ariance-**C**ontrolled **O**ptimization-based **RE**weighting (VCORE), a principled framework that reformulates CoT supervision as a constrained optimization problem. By adopting an optimization-theoretic perspective, VCORE enables a principled and adaptive allocation of supervision across tokens, thereby aligning the training objective more closely with the goal of robust reasoning generalization. Empirical evaluations demonstrate that VCORE achieves the strongest overall average performance, with especially clear gains on lower-capacity models. Across both in-domain and out-of-domain settings, VCORE achieves substantial performance gains on mathematical and coding benchmarks, using models from the Qwen3 series (4B, 8B, 32B) and LLaMA-3.1-8B-Instruct. Moreover, we show that VCORE serves as a more effective initialization for subsequent reinforcement learning, establishing a stronger foundation for advancing the reasoning capabilities of LLMs.
You Can Have a Second Chance: Unbiased and Multi-bit Watermarking for Diffusion Language Models with Regret-based Remasking
PDF ↗The rapid development of Diffusion Language Models (DLMs) raises concerns about watermarking for DLM-generated detection. However, existing sequential LLM watermarking cannot be directly applied to DLMs, as DLMs’ generation order is arbitrary. While emerging studies adapt biased LLM watermarking to DLMs by temporarily predicting the watermark prefix, they suffer from degraded quality and unstable watermarking due to bias accumulation and prediction errors. Besides, they cannot carry multi-bit watermarks. In this paper, we propose unbiased multi-bit watermarking for DLMs. We introduce a stability-aware constraint that allows watermarking only in stable contexts and a bit-controlled, unbiased modulation to preserve the original DLM output distribution, achieving stable watermarking with minimal quality impact. To enhance detection robustness, we design a Regret-based Remasking, which grants a “second chance” for unwatermarked tokens to be regenerated. It can seamlessly integrate into DLM inference with no added diffusion steps and latency. Experiments across DLMs and various tasks show that our scheme is effective, achieving superior generation quality compared to baselines while maintaining high detection accuracy and multi-bit capacity. Our code is available here https://github.com/iieSKLCSDsg/UMR.
Recent Large Reasoning Models (LRMs) have achieved remarkable progress, yet their evaluation still relies on a narrow paradigm: evaluating one question at a time. This single-question setup suffers from two major limitations: (1) vulnerability to data contamination and diminishing difficulty, forcing costly creation of new questions with significant human effort, (2) failure to evaluate models under multi-context pressure, a key requirement for real-world deployment. To bridge this gap, we present **REST** (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that exposes LRMs to multiple problems simultaneously. Beyond basic reasoning, REST evaluates two under-tested capabilities: *contextual priority allocation* and *robustness against contextual interference*. Our evaluation of more than **30** advanced reasoning models on **9** reasoning benchmarks reveals several striking findings: Even state-of-the-art (SOTA) models such as ***DeepSeek-R1 exhibit substantial performance degradation under stress testing***, challenging the prevailing assumption that "LLMs are multi-problem solvers". Crucially, ***REST demonstrates stronger discriminative power*** than existing benchmarks, revealing performance gaps among models that exhibit similar, near-ceiling performance under traditional evaluation. Some key insights emerge from our analysis: (1) the ***"overthinking trap"*** is a critical factor contributing to the performance degradation; (2) models trained with the ***"Long2Short" technique preserve more of their single-problem accuracy*** under REST, outperforming their standard-trained counterparts. These results establish REST as a cost-efficient, future-proof evaluation paradigm while reducing reliance on continuous human annotation. Code is available at https://github.com/opendatalab/REST.
Multi-Hop Question Answering (MHQA) is crucial for evaluating the model’s capability to integrate information from diverse sources. However, creating extensive and high-quality MHQA datasets is challenging: (i) manual annotation is expensive, and (ii) current synthesis methods often produce simplistic questions or require extensive manual guidance. This paper introduces HopWeaver, the first cross-document framework synthesizing authentic multi-hop questions without human intervention. HopWeaver synthesizes bridge and comparison questions through an innovative pipeline that identifies complementary documents and constructs authentic reasoning paths to ensure true multi-hop reasoning. We further present a comprehensive system for evaluating the synthesized multi-hop questions. Empirical evaluations demonstrate that the synthesized questions achieve comparable or superior quality to human-annotated datasets at a lower cost. Our framework provides a valuable tool for the research community: it can automatically generate challenging benchmarks from any raw corpus, which opens new avenues for both evaluation and targeted training to improve the reasoning capabilities of advanced question answering models, especially in domains with scarce resources.
Multi-agent systems (MAS) built on large language models promise improved problem-solving through collaboration, yet they often fail to consistently outperform strong single-agent baselines due to error propagation at inter-agent message handoffs. In this work, we conduct a systematic empirical analysis of such failures and introduce an edge-level error taxonomy that identifies four dominant error types: Data Gap, Signal Corruption, Referential Drift, and Capability Gap, as primary sources of failure in multi-agent interactions. Building on this taxonomy, we propose AgentAsk, a lightweight clarification module designed to intervene at the edge level in MAS to prevent cascading errors. The module operates by strategically applying minimal clarifications at critical points within the system, improving the accuracy and efficiency of the overall task. AgentAsk is trained to balance the trade-offs between clarification cost, latency, and accuracy, while it is also architecture-agnostic and can be easily integrated into existing systems. Evaluated across five benchmarks, AgentAsk consistently improves accuracy by up to 4.69%, while keeping latency and extra costs below 10% compared to baseline MAS, showcasing its high efficiency and minimal overhead. The code is available at https://anonymous.4open.science/r/AgentAsk-3432.