Autonomous web agents powered by large language models (LLMs) remain brittle on long-horizon browser workflows. A key bottleneck is a grounding gap in existing skill formulations: textual workflow skills provide natural language guidance but cannot be directly executed, while code-based skills execute without giving the agent step-level guidance for adaptation or recovery. We introduce WebXSkill, a framework that bridges this gap with executable skills, each pairing a parameterized action program with step-level natural-language guidance. WebXSkill operates in three stages: skill extraction mines reusable action subsequences from readily available synthetic agent trajectories and abstracts them into parameterized skills, skill organization indexes them into a URL-based graph for context-aware retrieval, and skill deployment exposes two complementary modes, grounded mode for fully automated execution and guided mode where skills serve as step-by-step instructions the agent follows with its native planning. WebXSkill demonstrates consistent improvements on WebArena, WebVoyager, and Online-Mind2Web. We further find that better skill deployment mode depends on a model’s plan and execution capability. The code is available at https://github.com/aiming-lab/WebXSkill.

We study Latent Recurrent Transformer (LRT), a lightweight augmentation of autoregressive transformers that reuses a high-level source-layer hidden state from the previous token as recurrent memory for the next token. Because this state is already computed during ordinary decoding, LRT introduces a cross-token, cross-layer latent pathway while preserving the standard attention mechanism, KV-cache interface, and one model forward per generated token. To pretrain this recurrence without sequentially unrolling the full sequence, we introduce interleaved parallel training: one full-sequence initialization forward constructs a shared buffer, followed by sequential refinement of disjoint position subsets with parallel computation within each subset. This provides every token with recurrent-memory-aware supervision at approximately 2x ideal token compute. Across 1.3B- and 2.1B-parameter nanochat-style backbones and a wide range of training budgets, LRT improves both BPB and CORE under matched effective compute. Additionally, LRT outperforms two-forward PonderLM-2 and matches a three-loop Transformer in BPB, while retaining one-forward-per-token decoding with 9% latency overhead over the standard Transformer.

Recent studies suggest that video generative models can exhibit certain forms of zero-shot visual reasoning through generated frames. Yet reliable evaluation remains challenging: benchmarks should adopt inputs aligned with the visual priors of current video models, require valid evolving processes rather than only plausible final states, and calibrate task difficulty to remain challenging yet partly feasible. To this end, we introduce VGI-Bench, containing 27 tasks and 810 instances, organized by a two-level taxonomy of domains and skill tags for evaluating the visual reasoning capabilities of video generative models. Our evaluations show that current generative systems can solve a subset of visually grounded reasoning tasks, but remain far from reliable, with even the strongest model, Seedance 2.0, achieving only 51.0% under our evaluation criteria. Our findings further reveal the input sensitivity, skill-bounded transfer from synthetic fine-tuning, and limited denoising-time self-correction, where later steps mainly refine early hypotheses rather than correct reasoning errors. We hope VGI-Bench will help stimulate the development of next-generation video generative models. Data and code are available at [our project page](https://hexuan21.github.io/VGI-Bench/).

The widely adopted Top-K router in Mixture-of-Experts (MoE) models is mathematically non-differentiable, making them harder to train than dense counterparts. We propose DenseMixer, a simple and effective MoE post-training technique that provides a more precise router gradient estimation by trading in some extra compute. It is universally applicable to TopK routed MoE, operates in a plug-and-play manner, preserves inference-time efficiency, and remains fully compatible with popular training libraries as well as parameter-efficient methods such as LoRA. Extensive experiments demonstrate that DenseMixer consistently outperforms conventional approaches when fine-tuning MoE models across scales (7B–30B), architectures (with and without shared experts), pre-training regimes of the underlying MoE models (from scratch or via up-cycling), and post-training data types (instruction tuning and long chain-of-thought). Furthermore, our analysis on efficiency shows that the additional computation introduced by DenseMixer remains modest, providing substantial improvements at an acceptable cost.

Web agents struggle to adapt to new websites due to the scarcity of environment specific tasks and demonstrations. Recent works have explored synthetic data generation to address this challenge, however, they suffer from data quality issues where synthesized tasks contain hallucinations that cannot be executed, and collected trajectories are noisy with redundant or misaligned actions. In this paper, we propose SynthAgent, a fully synthetic supervision framework that aims at improving synthetic data quality via dual refinement of both tasks and trajectories. Our approach begins by synthesizing diverse tasks through categorized exploration of web elements, ensuring efficient coverage of the target environment. During trajectory collection, tasks are refined only when conflicts with observations are detected, which mitigates hallucinations while preserving task consistency. After collection, we conduct trajectory refinement with global context to mitigate potential noise or misalignments. Finally, we fine-tune open-source web agents on the refined synthetic data to adapt them to the target environment. Experimental results demonstrate that SynthAgent outperforms existing synthetic data methods, validating the importance of high-quality synthetic supervision. The code is publicly available at https://github.com/aiming-lab/SynthAgent.

Recent research has shown that Large Language Models (LLMs) are vulnerable to automated jailbreak attacks, where adversarial suffixes crafted by algorithms appended to harmful queries bypass safety alignment and trigger unintended responses. Current methods for generating these suffixes are computationally expensive and have low Attack Success Rates (ASR), especially against well-aligned models like Llama2 and Llama3. To overcome these limitations, we introduce ADV-LLM, an iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability. Our framework significantly reduces the computational cost of generating adversarial suffixes while achieving nearly 100% ASR on various open-source LLMs. Moreover, it exhibits strong attack transferability to closed-source models, achieving 99% ASR on GPT-3.5 and 49% ASR on GPT-4, despite being optimized solely on Llama3. Beyond improving jailbreak ability, ADV-LLM provides valuable insights for future safety alignment research through its ability to generate large datasets for studying LLM safety.

In this work, we address the memory overhead of deploying Mixture-of-Experts (MoE) architectures in Large Language Models (LLMs). While MoE layers improve LLM performance without increasing inference costs, the ever-growing number of experts inflates memory requirements, hindering practical deployment. Our empirical study reveals that some experts encode redundant knowledge during pre-training. We thus propose a method of grouping and pruning similar experts to improve the model’s parameter efficiency. We validate the effectiveness of our method by pruning three state-of-the-art MoE architectures, including Mixtral, Deepseek-MoE, and Qwen. The evaluation shows that our method outperforms other model pruning methods on a range of natural language tasks.

Large language models (LLMs) are increasingly used in interactive applications, and human evaluation remains the gold standard for assessing their performance in multi-turn conversations. Since human studies are costly, time-consuming, and hard to reproduce, recent work explores using LLMs to simulate users for automatic assistant evaluation. However, there is no benchmark or systematic study to evaluate whether these simulated users are reliable stand-ins for real users. To address this, we introduce SimulatorArena, a benchmark of 909 annotated human–LLM conversations on two interactive tasks—math tutoring and document creation. SimulatorArena evaluates simulators based on how closely their messages match human behavior and how well their assistant ratings align with human judgments. Experiments on various simulator methods show that simulators conditioned on user profiles, capturing traits like background and message styles, align closely with human judgments. They reach Spearman’s 𝜌 of 0.7 on both tasks, providing a practical, scalable alternative to human evaluation. Using the best simulator for each task, we benchmark 18 assistants, including the latest LLMs such as GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro.

Compositionality is a core property of natural language, and compositional behavior (CB) is a crucial goal for modern NLP systems. The research literature, however, includes conflicting perspectives on how CB should be defined, evaluated, and achieved. We propose a conceptual framework to address these questions and survey researchers active in this area.We find consensus on several key points. Researchers broadly accept our proposed definition of CB, agree that it is not solved by current models, and doubt that scale alone will achieve the target behavior. In other areas, we find the field is split on how to move forward, identifying diverse opportunities for future research.

Inductive reasoning is a core component of human intelligence. In the past research of inductive reasoning within computer science, formal language is used as representations of knowledge (facts and rules, more specifically). However, formal language can cause systematic problems for inductive reasoning such as disability of handling raw input such as natural language, sensitiveness to mislabeled data, and incapacity to handle ambiguous input. To this end, we propose a new paradigm (task) for inductive reasoning, which is to induce natural language rules from natural language facts, and create a dataset termed DEER containing 1.2k rule-fact pairs for the task, where rules and facts are written in natural language. New automatic metrics are also proposed and analysed for the evaluation of this task. With DEER, we investigate a modern approach for inductive reasoning where we use natural language as representation for knowledge instead of formal language and use pretrained language models as ”reasoners”. Moreover, we provide the first and comprehensive analysis of how well pretrained language models can induce natural language rules from natural language facts. We also propose a new framework drawing insights from philosophy literature for this task, which we show in the experiment section that surpasses baselines in both automatic and human evaluations. We discuss about our future perspectives for inductive reasoning in Section 7. Dataset and code are available at https://github.com/ZonglinY/Inductive_Reasoning.

Knowledge-grounded dialogue systems are challenging to build due to the lack of training data and heterogeneous knowledge sources. Existing systems perform poorly on unseen topics due to limited topics covered in the training data. In addition, it is challenging to generalize to the domains that require different types of knowledge sources. To address the above challenges, we present PLUG, a language model that homogenizes different knowledge sources to a unified knowledge representation for knowledge-grounded dialogue generation tasks. We first retrieve relevant information from heterogeneous knowledge sources (e.g., wiki, dictionary, or knowledge graph); Then the retrieved knowledge is transformed into text and concatenated with dialogue history to feed into the language model for generating responses. PLUG is pre-trained on a large-scale knowledge-grounded dialogue corpus. The empirical evaluation on two benchmarks shows that PLUG generalizes well across different knowledge-grounded dialogue tasks. It achieves comparable performance with state-of-the-art methods in the fully-supervised setting and significantly outperforms other approaches in zero-shot and few-shot settings.

The retriever-reader framework is popular for open-domain question answering (ODQA) due to its ability to use explicit knowledge. Although prior work has sought to increase the knowledge coverage by incorporating structured knowledge beyond text, accessing heterogeneous knowledge sources through a unified interface remains an open question. While data-to-text generation has the potential to serve as a universal interface for data and text, its feasibility for downstream tasks remains largely unknown. In this work, we bridge this gap and use the data-to-text method as a means for encoding structured knowledge for open-domain question answering. Specifically, we propose a verbalizer-retriever-reader framework for ODQA over data and text where verbalized tables from Wikipedia and graphs from Wikidata are used as augmented knowledge sources. We show that our Unified Data and Text QA, UDT-QA, can effectively benefit from the expanded knowledge index, leading to large gains over text-only baselines. Notably, our approach sets the single-model state-of-the-art on Natural Questions. Furthermore, our analyses indicate that verbalized knowledge is preferred for answer reasoning for both adapted and hot-swap settings.