Hallucination detection remains a fundamental challenge for the safe and reliable deployment of large language models (LLMs), especially in applications requiring factual accuracy. Existing hallucination benchmarks often operate at the sequence level and are limited to English, lacking the fine-grained, multilingual supervision needed for a comprehensive evaluation. In this work, we introduce PsiloQA, a large-scale, multilingual dataset annotated with span-level hallucinations across 14 languages. PsiloQA is constructed through an automated three-stage pipeline: generating question-answer pairs from Wikipedia using GPT-4o, eliciting potentially hallucinated answers from diverse LLMs in a no-context setting, and automatically annotating hallucinated spans using GPT-4o by comparing against golden answers and retrieved context. We evaluate a wide range of hallucination detection methods -- including uncertainty quantification, LLM-based tagging, and fine-tuned encoder models -- and show that encoder-based models achieve the strongest performance across languages. Furthermore, PsiloQA demonstrates effective cross-lingual generalization and supports robust knowledge transfer to other benchmarks, all while being significantly more cost-efficient than human-annotated datasets. Our dataset and results advance the development of scalable, fine-grained hallucination detection in multilingual settings.

Despite interpretability work analyzing VIT encoders and transformer activations, we don't yet understand why Multimodal Language Models (MLMs) struggle on perception-heavy tasks. We offer an under-studied perspective by examining how popular MLMs (LLaVA-OneVision, Qwen2.5-VL, and Llama-3-LLaVA-NeXT) process their visual key-value tokens. We first study the flow of visual information through the language model, finding that image value tokens encode sufficient information to perform several perception-heavy tasks zero-shot: segmentation, semantic correspondence, temporal correspondence, and referring expression detection. We find that while the language model does augment the visual information received from the projection of input visual encodings-which we reveal correlates with overall MLM perception capability-it contains less visual information on several tasks than the equivalent visual encoder (SigLIP) that has not undergone MLM finetuning. Further, we find that the visual information corresponding to input-agnostic image key tokens in later layers of language models contains artifacts which reduce perception capability of the overall MLM. Next, we discuss controlling visual information in the language model, showing that adding a text prefix to the image input improves perception capabilities of visual representations. Finally, we reveal that if language models were able to better control their visual information, their perception would significantly improve; e.g., in 33.3% of Art Style questions in the BLINK benchmark, perception information present in the language model is not surfaced to the output! Our findings reveal insights into the role of key-value tokens in multimodal systems, paving the way for deeper mechanistic interpretability of MLMs and suggesting new directions for training their visual encoder and language model components.

Controllers for structured LM reasoning (e.g., Chain-of-Thought, self-consistency, and Tree-of-Thoughts) often entangle what to try next with how to execute it, exposing only coarse global knobs and yielding brittle, compute-inefficient, and hard-to-audit behavior. We introduce Natural Language Edge Labelling (NLEL), a labeller-tuner overlay that attaches a free-form natural-language directive to each search edge and translates it into a schema-bounded control vector for decoding, search (branch quotas, exploration $β$), generation bundle size, retrieval mixtures, and verification passes. A labeller $Λ$ emits labels from the parent state and a compact context; a tuner $Ψ$ maps $(P, L, C)\to Π$, with strict schema validation and trust-region projection around safe defaults. Downstream selection remains ToT-style with score $S=μ+βσ$ and depth-annealed $β$. We show NLEL strictly generalizes CoT/ToT, prove an anytime-monotonicity property for top-$k$ selection under label-conditioned bundles, and bound selector shortfall by control-vector distortion, providing decision-relevant justification for guards like trust regions and verification passes. We instantiate $Ψ$ as a prompt-only JSON Parameter Emitter and preregister an evaluation on GSM8K, MATH (subset), StrategyQA, and ARC-Challenge with compute-aware reporting (success@compute, tokens-per-success) and ablations over $Λ$, $Ψ$, trust-region radius, and control quantization; preregistered forecasts anticipate accuracy gains at comparable token budgets and improved success@compute under constraints. NLEL offers an interpretable, model-agnostic interface that separates intent from execution for controllable, auditable LM inference.

Recent progress in large language models demonstrates that hybrid architectures--combining self-attention mechanisms with structured state space models like Mamba--can achieve a compelling balance between modeling quality and computational efficiency, particularly for long-context tasks. While these hybrid models show promising performance, systematic comparisons of hybridization strategies and analyses on the key factors behind their effectiveness have not been clearly shared to the community. In this work, we present a holistic evaluation of hybrid architectures based on inter-layer (sequential) or intra-layer (parallel) fusion. We evaluate these designs from a variety of perspectives: language modeling performance, long-context capabilities, scaling analysis, and training and inference efficiency. By investigating the core characteristics of their computational primitive, we identify the most critical elements for each hybridization strategy and further propose optimal design recipes for both hybrid models. Our comprehensive analysis provides practical guidance and valuable insights for developing hybrid language models, facilitating the optimization of architectural configurations.

Recent advancements in large language models (LLMs) and agentic systems have shown exceptional decision-making capabilities, revealing significant potential for autonomic finance. Current financial trading agents predominantly simulate anthropomorphic roles that inadvertently introduce emotional biases and rely on peripheral information, while being constrained by the necessity for continuous inference during deployment. In this paper, we pioneer the harmonization of strategic depth in agents with the mechanical rationality essential for quantitative trading. Consequently, we present TiMi (Trade in Minutes), a rationality-driven multi-agent system that architecturally decouples strategy development from minute-level deployment. TiMi leverages specialized LLM capabilities of semantic analysis, code programming, and mathematical reasoning within a comprehensive policy-optimization-deployment chain. Specifically, we propose a two-tier analytical paradigm from macro patterns to micro customization, layered programming design for trading bot implementation, and closed-loop optimization driven by mathematical reflection. Extensive evaluations across 200+ trading pairs in stock and cryptocurrency markets empirically validate the efficacy of TiMi in stable profitability, action efficiency, and risk control under volatile market dynamics.

Humans are good at learning on the job: We learn how to solve the tasks we face as we go along. Can a model do the same? We propose an agent that assembles a task-specific curriculum, called test-time curriculum (TTC-RL), and applies reinforcement learning to continue training the model for its target task. The test-time curriculum avoids time-consuming human curation of datasets by automatically selecting the most task-relevant data from a large pool of available training data. Our experiments demonstrate that reinforcement learning on a test-time curriculum consistently improves the model on its target tasks, across a variety of evaluations and models. Notably, on challenging math and coding benchmarks, TTC-RL improves the pass@1 of Qwen3-8B by approximately 1.8x on AIME25 and 2.1x on CodeElo. Moreover, we find that TTC-RL significantly raises the performance ceiling compared to the initial model, increasing pass@8 on AIME25 from 40% to 62% and on CodeElo from 28% to 43%. Our findings show the potential of test-time curricula in extending the test-time scaling paradigm to continual training on thousands of task-relevant experiences during test-time.

As Large Language Models (LLMs) demonstrate remarkable capabilities learned from vast corpora, concerns regarding data privacy and safety are receiving increasing attention. LLM unlearning, which aims to remove the influence of specific data while preserving overall model utility, is becoming an important research area. One of the mainstream unlearning classes is optimization-based methods, which achieve forgetting directly through fine-tuning, exemplified by Negative Preference Optimization (NPO). However, NPO's effectiveness is limited by its inherent lack of explicit positive preference signals. Attempts to introduce such signals by constructing preferred responses often necessitate domain-specific knowledge or well-designed prompts, fundamentally restricting their generalizability. In this paper, we shift the focus to the distribution-level, directly targeting the next-token probability distribution instead of entire responses, and derive a novel unlearning algorithm termed \textbf{Di}stribution \textbf{P}reference \textbf{O}ptimization (DiPO). We show that the requisite preference distribution pairs for DiPO, which are distributions over the model's output tokens, can be constructed by selectively amplifying or suppressing the model's high-confidence output logits, thereby effectively overcoming NPO's limitations. We theoretically prove the consistency of DiPO's loss function with the desired unlearning direction. Extensive experiments demonstrate that DiPO achieves a strong trade-off between model utility and forget quality. Notably, DiPO attains the highest forget quality on the TOFU benchmark, and maintains leading scalability and sustainability in utility preservation on the MUSE benchmark.

Open-vocabulary learning requires modeling the data distribution in open environments, which consists of both seen-class and unseen-class data. Existing methods estimate the distribution in open environments using seen-class data, where the absence of unseen classes makes the estimation error inherently unidentifiable. Intuitively, learning beyond the seen classes is crucial for distribution estimation to bound the estimation error. We theoretically demonstrate that the distribution can be effectively estimated by generating unseen-class data, through which the estimation error is upper-bounded. Building on this theoretical insight, we propose a novel open-vocabulary learning method, which generates unseen-class data for estimating the distribution in open environments. The method consists of a class-domain-wise data generation pipeline and a distribution alignment algorithm. The data generation pipeline generates unseen-class data under the guidance of a hierarchical semantic tree and domain information inferred from the seen-class data, facilitating accurate distribution estimation. With the generated data, the distribution alignment algorithm estimates and maximizes the posterior probability to enhance generalization in open-vocabulary learning. Extensive experiments on $11$ datasets demonstrate that our method outperforms baseline approaches by up to $14\%$, highlighting its effectiveness and superiority.

Web 3.0 represents the next generation of the Internet, which is widely recognized as a decentralized ecosystem that focuses on value expression and data ownership. By leveraging blockchain and artificial intelligence technologies, Web 3.0 offers unprecedented opportunities for users to create, own, and monetize their content, thereby enabling User-Generated Content (UGC) to an entirely new level. However, some self-interested users may exploit the limitations of content curation mechanisms and generate low-quality content with less effort, obtaining platform rewards under information asymmetry. Such behavior can undermine Web 3.0 performance. To this end, we propose \textit{LMM-Incentive}, a novel Large Multimodal Model (LMM)-based incentive mechanism for UGC in Web 3.0. Specifically, we propose an LMM-based contract-theoretic model to motivate users to generate high-quality UGC, thereby mitigating the adverse selection problem from information asymmetry. To alleviate potential moral hazards after contract selection, we leverage LMM agents to evaluate UGC quality, which is the primary component of the contract, utilizing prompt engineering techniques to improve the evaluation performance of LMM agents. Recognizing that traditional contract design methods cannot effectively adapt to the dynamic environment of Web 3.0, we develop an improved Mixture of Experts (MoE)-based Proximal Policy Optimization (PPO) algorithm for optimal contract design. Simulation results demonstrate the superiority of the proposed MoE-based PPO algorithm over representative benchmarks in the context of contract design. Finally, we deploy the designed contract within an Ethereum smart contract framework, further validating the effectiveness of the proposed scheme.

Retrieval-Augmented Generation (RAG) is a powerful technique for enriching Large Language Models (LLMs) with external knowledge, allowing for factually grounded responses, a critical requirement in high-stakes domains such as healthcare. However, the efficacy of RAG systems is fundamentally restricted by the performance of their retrieval module, since irrelevant or semantically misaligned documents directly compromise the accuracy of the final generated response. General-purpose dense retrievers can struggle with the nuanced language of specialised domains, while the high accuracy of in-domain models is often achieved at prohibitive computational costs. In this work, we aim to address this trade-off by developing and evaluating a two-stage retrieval architecture that combines a lightweight ModernBERT bidirectional encoder for efficient initial candidate retrieval with a ColBERTv2 late-interaction model for fine-grained re-ranking. We conduct comprehensive evaluations of our retriever module performance and RAG system performance in the biomedical context, fine-tuning the IR module using 10k question-passage pairs from PubMedQA. Our analysis of the retriever module confirmed the positive impact of the ColBERT re-ranker, which improved Recall@3 by up to 4.2 percentage points compared to its retrieve-only counterpart. When integrated into the biomedical RAG, our IR module leads to a state-of-the-art average accuracy of 0.4448 on the five tasks of the MIRAGE question-answering benchmark, outperforming strong baselines such as MedCPT (0.4436). Our ablation studies reveal that this performance is critically dependent on a joint fine-tuning process that aligns the retriever and re-ranker; otherwise, the re-ranker might degrade the performance.