Retrieval-augmented generation (RAG) improves factuality but adds latency and engineering overhead at serving time. We propose RING (Retrieval-Internalized Generation), a holistic paradigm spanning both architecture and training that injects large-scale external knowledge into a Mixture-of-Memory Experts and learns parametric search over this internal memory via reinforcement learning, removing the external retriever entirely. Training proceeds in three stages: continued pre-training injects new corpora into a Knowledge Expert via our novel Dual Causal Attention; supervised fine-tuning teaches a “search-then-answer” pattern; and reinforcement learning with hierarchical rewards optimizes the routing-and-search policy over the parametric memory. Unlike prior parametric injection methods that pair internal memory with a fixed or rule-based retriever, RING learns its retrieval policy directly from task signals. We further frame RING theoretically as a search-free approximation to the classical RAG objective. To evaluate large-scale injection of genuinely new knowledge without test-time leakage, we further construct News-2025, a benchmark built from news strictly post-dating the base LLM’s pretraining cutoff. RING matches or surpasses both search-based RAG and parametric injection baselines in accuracy and efficiency.

Latent reasoning offers a computation-efficient alternative to Chain-of-Thought but often suffers from performance degradation due to distributional misalignment and ambiguous chain definitions. Ideally, latent reasoning should function as a superposition of multiple reasoning paths. To realize this, we introduce Latent-SFT, a unified framework addressing challenges at three levels: token, chain, and learning. First, we define the Latent-Vocab to constrain hidden states within the pre-trained vocab-space. Second, we construct the Latent-Chain via Induction-Supervision Masking to ensure semantic compactness and sufficiency. Third, we employ Latent-Optim with stochastic Gumbel-Softmax to guide the model toward generalizable solutions. Empirical results demonstrate that Latent-SFT consistently outperforms explicit SFT across six mathematical benchmarks (e.g., GSM8k, AIME24) while achieving a 2.7x to 5.5x reduction in reasoning length. Analysis confirms that our method effectively captures a superposition of diverse reasoning trajectories rather than merely compressing a single path.

Latent reasoning offers a more efficient alternative to explicit reasoning by compressing intermediate reasoning into continuous representations and substantially shortening reasoning chains. However, existing latent reasoning methods mainly focus on supervised learning, and reinforcement learning in latent space remains highly unstable. We study this problem through the lens of Group Relative Policy Optimization (GRPO), and show that directly adapting GRPO to latent reasoning is fundamentally non-trivial: latent reasoning changes both the probability density and the sampling mechanism, causing three coupled bottlenecks: absence of intrinsic latent manifolds, where unconstrained exploration pushes rollouts off the valid latent manifold; exploration-optimization misalignment, where trajectory-level rewards can induce incorrect token-level updates; and latent mixture non-closure, where jointly reinforcing multiple correct latent paths can produce an invalid averaged state. To address them, we propose Latent-GRPO, which combines invalid-sample advantage masking, one-sided noise sampling, and optimal correct-path first-token selection. Across four low-difficulty benchmarks (e.g., GSM8K-Aug) and four high-difficulty benchmarks (e.g., AIME), Latent-GRPO improves over its latent initialization by 7.86 Pass@1 points on low-difficulty tasks and surpasses explicit GRPO by 4.27 points on high-difficulty tasks while using 3–4 shorter reasoning chains. It also achieves stronger pass@k performance under Gumbel sampling. These results establish Latent-GRPO as an effective approach for stable and efficient latent reasoning.

Long-form chain-of-thought reasoning can improve LLM performance on complex tasks, but models often continue generating unnecessary reasoning after a correct answer has emerged. We call this behavior overthinking. We study this phenomenon from the perspective of GRPO-style reinforcement learning (RL) post-training, framing it as a training-time credit-assignment problem rather than merely a decoding-time stopping problem. In rollouts sampled at the onset of GRPO training, we observe that successful trajectories can exhibit a slightly higher degree of overthinking than unsuccessful trajectories for the same prompts. This early imbalance provides a starting point for an undesirable feedback loop: because GRPO assigns sequence-level credit, it cannot distinguish the solution-reaching prefix from the unnecessary continuation that lengthens a successful trajectory. Both receive positive update signal, allowing the initial imbalance to grow into more severe overthinking during training. To address this issue, we introduce Dynamic Rollout Editing (DRE), a training-time intervention for successful trajectories that continue thinking after answer emergence. DRE preserves the accepted verified prefix, edits the remaining thinking, and prefers the edited trajectory within the same RL group, weakening the preference signal for unnecessary thinking without penalizing the reasoning needed to reach the answer. Experiments across diverse tasks show the effectiveness of DRE.

LLM-based agent systems increasingly rely on open-registry skills, which are difficult to vet and expose agents to greater attack risks. These attacks mainly take two forms. Malicious injection attacks embed explicit harmful instructions into skills and are readily caught by auditing. Direct vulnerability attacks target latent flaws in non-malicious skills; however, directly crafted attack prompts often lead to unstable or unverified outcomes. We introduce **SkillAttack**, an automated red-teaming framework that systematically explores executable attack paths from latent vulnerabilities without modifying skill files. Evaluated across 11 target LLMs, SkillAttack substantially outperforms baselines, showing that well-intended skills can pose serious risks in realistic agent interactions. Our evaluation also yields trajectory-level attack traces with verified risk labels, providing a foundation for defense. We also release **SkillRiskBench**, a benchmark for skill risk analysis, and train a lightweight pre-execution risk guard that predicts risky invocations from the user request and skill file. The guard outperforms frontier LLM judges and existing safety guard models, highlighting the value of trajectory-grounded supervision for proactive skill defense. Code is available at https://github.com/Zhow01/SkillAttack.

Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content, restricting their safe deployment. While traditional methods (e.g., alignment) adjust output preferences, they fail to eliminate underlying toxic regions in parameters, leaving models vulnerable to adversarial attacks. Prior mechanistic studies characterize toxic regions as “toxic vectors” or “layer-wise subspaces”, yet our analysis identifies critical limitations: i) Removed toxic vectors can be reconstructed via linear combinations of non-toxic vectors, demanding targeting of entire toxic subspace; ii) Contrastive objective over limited samples inject noise into layer-wise subspaces, hindering stable extraction. These highlight the challenge of identifying robust toxic subspace and removing them. Therefore, we propose GLOSS (GLobal tOxic Subspace Suppression), a lightweight method that mitigates toxicity by identifying and eliminating this global subspace from FFN parameters. Experiments on LLMs (e.g., Qwen3) show GLOSS achieves SOTA detoxification while preserving general capabilities without requiring large-scale retraining.

Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., lt;/think gt;). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.

A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs’ pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.

Unmanned Aerial Vehicle (UAV) Vision-and-Language Navigation (VLN) is vital for applications such as disaster response, logistics delivery, and urban inspection. However, existing methods often struggle with insufficient multimodal fusion, weak generalization, and poor interpretability. To address these challenges, we propose FlightGPT, a novel UAV VLN framework built upon Vision-Language Models (VLMs) with powerful multimodal perception capabilities. We design a two-stage training pipeline: first, Supervised Fine-Tuning (SFT) using high-quality demonstrations to improve initialization and structured reasoning; then, Group Relative Policy Optimization (GRPO) algorithm, guided by a composite reward that considers goal accuracy, reasoning quality, and format compliance, to enhance generalization and adaptability. Furthermore, FlightGPT introduces a Chain-of-Thought (CoT)-based reasoning mechanism to improve decision interpretability. Extensive experiments on the city-scale dataset CityNav demonstrate that FlightGPT achieves state-of-the-art performance across all scenarios, with a 9.22% higher success rate than the strongest baseline in unseen environments. Our implementation is publicly available.

The extensive utilization of large language models (LLMs) underscores the crucial necessity for precise and contemporary knowledge embedded within their intrinsic parameters. Existing research on knowledge editing primarily concentrates on monolingual scenarios, neglecting the complexities presented by multilingual contexts and multi-hop reasoning. To address these challenges, our study introduces MLaKE (Multilingual Language Knowledge Editing), a novel benchmark comprising 4072 multi-hop and 5360 single-hop questions designed to evaluate the adaptability of knowledge editing methods across five languages: English, Chinese, Japanese, French, and German. MLaKE aggregates fact chains from Wikipedia across languages and utilizes LLMs to generate questions and answer. We assessed the effectiveness of current multilingual knowledge editing methods using the MLaKE dataset. Our results show that due to considerable inconsistencies in both multilingual performance and encoding efficiency, these methods struggle to generalize effectively across languages. The accuracy of these methods when editing English is notably higher than for other languages. The experimental results further demonstrate that models encode knowledge and generation capabilities for different languages using distinct parameters, leading to poor cross-lingual transfer performance in current methods. Transfer performance is notably better within the same language family compared to across different families. These findings emphasize the urgent need to improve multilingual knowledge editing methods.

“以大语言模型为代表的生成式人工智能迅猛发展,标志着人工智能从判别时代向生成时代的转变。这一进步极大地推动了信息检索技术的发展,本文对大语言模型对信息检索领域的影响进行了深入的综述。从性能改进到模式颠覆,逐步展开论述大语言模型对信息检索领域的影响。针对传统信息检索流程,大语言模型凭借强大的语义理解和建模能力,显著增强索引、检索和排序等信息检索模块的性能。同时,文章也探讨了大语言模型可能取代传统信息检索的趋势,并催生了新的信息获取方式,或将是新一次信息时代的寒武纪。此外,大语言模型对内容生态的深远影响也值得关注。”

Retrieval-augmented language models show promise in addressing issues like outdated information and hallucinations in language models (LMs). However, current research faces two main problems: 1) determining what information to retrieve, and 2) effectively combining retrieved information during generation. We argue that valuable retrieved information should not only be related to the current source text but also consider the future target text, given the nature of LMs that model future tokens. Moreover, we propose that aggregation using latent variables derived from a compact latent space is more efficient than utilizing explicit raw text, which is limited by context length and susceptible to noise. Therefore, we introduce RegaVAE, a retrieval-augmented language model built upon the variational auto-encoder (VAE). It encodes the text corpus into a latent space, capturing current and future information from both source and target text. Additionally, we leverage the VAE to initialize the latent space and adopt the probabilistic form of the retrieval generation paradigm by expanding the Gaussian prior distribution into a Gaussian mixture distribution. Theoretical analysis provides an optimizable upper bound for RegaVAE. Experimental results on various datasets demonstrate significant improvements in text generation quality and hallucination removal.

The task of response selection in multi-turn dialogue is to find the best option from all candidates. In order to improve the reasoning ability of the model, previous studies pay more attention to using explicit algorithms to model the dependencies between utterances, which are deterministic, limited and inflexible. In addition, few studies consider differences between the options before and after reasoning. In this paper, we propose an Implicit Relational Reasoning Graph Network to address these issues, which consists of the Utterance Relational Reasoner (URR) and the Option Dual Comparator (ODC). URR aims to implicitly extract dependencies between utterances, as well as utterances and options, and make reasoning with relational graph convolutional networks. ODC focuses on perceiving the difference between the options through dual comparison, which can eliminate the interference of the noise options. Experimental results on two multi-turn dialogue reasoning benchmark datasets MuTual and MuTualplus show that our method significantly improves the baseline of four pre-trained language models and achieves state-of-the-art performance. The model surpasses human performance for the first time on the MuTual dataset.