On-policy exploration is a crucial component for training robust Vision-Language Navigation agents, as it exposes the policy to a broader state distribution. However, such exploration inevitably leads to trajectories that deviate from expert demonstrations, resulting in a semantic mismatch between the executed visual stream and the original language instruction. In this work, we address this challenge by introducing Phi-Nav, a unified on-policy framework that leverages hindsight reasoning to align instructions with the agent's actual exploratory journey. Specifically, Phi-Nav operates through a three-stage dual-supervision cycle: 1) the agent performs oracle-guided on-policy exploration, sampling a trajectory while learning from expert action feedback, 2) a hindsight speaker synthesizes a path-level hindsight instruction grounded in the collected visual observations, and 3) the agent conducts a second imitation pass, treating the synthesized trajectory-instruction pair as an additional expert demonstration. Through this process, Phi-Nav bridges the critical semantic supervision gap inherent in on-policy methods, transforming semantically unlabeled movement into dense training signals. Evaluations on the R2R-CE and RxR-CE benchmarks show that Phi-Nav yields competitive performance while requiring only a fraction of the expert demonstrations used by current baselines. These results underscore the necessity of semantic exploration in VLN, positioning Phi-Nav as an effective solution for training embodied agents with limited data.

Training vision models with language supervision enables general and transferable representations. However, many visual domains, especially non-object-centric domains such as medical imaging and remote sensing, contain itemized text annotations: multiple text items describing distinct and semantically independent findings within a single image. Such supervision differs from standard multi-caption supervision, where captions are redundant or highly overlapping. Here, we introduce ItemizedCLIP, a framework for learning complete and explainable visual representations from itemized text supervision. ItemizedCLIP employs a cross-attention module to produce text item-conditioned visual embeddings and a set of tailored objectives that jointly enforce item independence (distinct regions for distinct items) and representation completeness (coverage of all items). Across four domains with naturally itemized text supervision (brain MRI, head CT, chest CT, remote sensing) and one additional synthetically itemized dataset, ItemizedCLIP achieves substantial improvements in zero-shot performance and fine-grained interpretability over baselines. The resulting ItemizedCLIP representations are semantically grounded, item-differentiable, complete, and visually interpretable.

This technical report presents K-EXAONE 2.0, an open-weight multilingual foundation model developed by LG AI Research as a step in our effort toward global frontier-scale foundation models. Rather than training from scratch, we upcycle K-EXAONE and expand its architecture, yielding a Mixture-of-Experts (MoE) model with 750B total parameters and approximately 37B activated per token—more than three times the capacity of its predecessor. K-EXAONE 2.0 supports context lengths of up to 256K tokens and expands multilingual coverage from six to ten languages. Its training pipeline combines continual pre-training, difficulty-focused mid-training, and post-training to strengthen reasoning, agentic coding, multilingual capability, and safety grounded in Korean sociocultural contexts. Across nine evaluation categories selected to reflect the conditions of practical use, K-EXAONE 2.0 improves over K-EXAONE and remains competitive with open-weight models, showing its largest gains in agentic coding and long-context understanding and its clearest strengths in long-context retrieval and safety. Released under the Apache 2.0 license, K-EXAONE 2.0 enables the wider AI ecosystem to evaluate, deploy, adapt, and build upon it, while marking the beginning—rather than the endpoint—of our challenge toward the global frontier.

We present a scalable pipeline for automatically generating high-quality training data for web agents. In particular, a major challenge in identifying high-quality training instances is trajectory evaluation - quantifying how much progress was made towards task completion. We introduce a novel constraint-based evaluation framework that provides fine-grained assessment of progress towards task completion. This enables us to leverage partially successful trajectories, which significantly expands the amount of usable training data. We evaluate our method on a new benchmark we propose called BookingArena, which consists of complex booking tasks across 20 popular websites, and demonstrate that our distilled student model outperforms open-source approaches and matches or exceeds commercial systems, while being a significantly smaller model. Our work addresses the challenge of efficiently creating diverse, realistic web interaction datasets and provides a systematic evaluation methodology for complex structured web tasks.

The scalability of current language-image pre-training for 3D medical imaging, such as CT and MRI, is constrained by the need for radiologists to manually curate raw clinical studies. In this work, we pioneer pre-training directly on uncurated studies, which both aligns more closely with the radiologist's workflow and provides a natural path to scalability. However, the unique structure of such data presents new challenges for existing model architectures, which were originally designed for 2D slices or single 3D scans. To address this, we introduce a novel hierarchical attention mechanism inspired by the intrinsic hierarchy of radiology data: slice, scan, and study. We denote our framework as Hierarchical attention for Language-Image Pre-training (HLIP). Trained on 220K studies with 3.13 million scans for brain MRI and 240K studies with 1.44 million scans for head CT, HLIP achieves state-of-the-art performance, e.g., +10.5

Step-by-step verifiers—also known as process reward models (PRMs)—are a key ingredient for test-time scaling, but training them requires expensive step-level supervision. This work aims to build data-efficient PRMs as verbalized step-wise reward models that verify every step in the solution by generating a verification chain-of-thought (CoT). We propose ThinkPRM, a long CoT verifier fine-tuned on orders of magnitude fewer process labels than those required by discriminative PRMs. Our approach capitalizes on the inherent reasoning abilities of long CoT models, and outperforms LLM-as-a-Judge and discriminative verifiers—using only 1% of the process labels in PRM800K—across several challenging benchmarks. Specifically, ThinkPRM beats the baselines on ProcessBench, MATH-500, and AIME ’24 under best-of-N selection and reward-guided search. In an out-of-domain evaluation over subsets of GPQA-Diamond and LiveCodeBench, our PRM surpasses discriminative verifiers trained with the full PRM800K by 8% and 4.5%, respectively. Lastly, under the same token budget, ThinkPRM scales up verification compute more effectively compared to LLM-as-a-Judge, outperforming it by 7.2% on a subset of ProcessBench. This work highlights the value of generative, long CoT PRMs that can scale test-time compute for verification while requiring minimal supervision for training.

As Large Language Models (LLMs) are increasingly deployed in real-world applications, balancing both helpfulness and safety has become a central challenge. A natural approach is to incorporate safety constraints into Reinforcement Learning from Human Feedback (RLHF), where recent studies have shown promising progress. However, these methods often rely on auxiliary networks or multi-stage pipelines, thereby increasing complexity. In this work, we revisit the safety alignment objective itself and demonstrate that it admits a closed-form solution, yielding a theoretically grounded and provably equivalent reformulation that enables a direct and tractable optimization procedure. Building on this insight, we propose SafeDPO, a lightweight method derived from this formulation, which preserves the optimality of the underlying safety-constrained objective while requiring only one additional hyperparameter and minimal modifications to existing preference-based training methods. At the same time, it eliminates the need for reward models, cost models, and online sampling. Despite its simplicity, SafeDPO achieves comparable or superior results to state-of-the-art safety alignment methods in both theoretical soundness and empirical performance. Experiments on the PKU-SafeRLHF-30K benchmark show that SafeDPO consistently improves safety while maintaining competitive helpfulness. Ablation studies further show that the additional hyperparameter provides a flexible mechanism to enhance safety without altering the theoretical optimum, and confirm that SafeDPO scales reliably to LLMs with up to 13B parameters. Overall, our results highlight that a simple, theory-driven objective can provide a lightweight yet effective solution for safety alignment in practice.

Large language models (LLMs) are increasingly used as judges to evaluate agent performance, particularly in non-verifiable settings where judgments rely on agent trajectories including chain-of-thought (CoT) reasoning. This paradigm implicitly assumes that the agent's CoT faithfully reflects both its internal reasoning and the underlying environment state. We show this assumption is brittle: LLM judges are highly susceptible to manipulation of agent reasoning traces. By systematically rewriting agent CoTs while holding actions and observations fixed, we demonstrate that manipulated reasoning alone can inflate false positive rates of state-of-the-art VLM judges by up to 90

Frontier artificial intelligence (AI) models have advanced rapidly through training on internet-scale public data, yet such systems lack access to private clinical data. Neuroimaging is underrepresented in the public domain due to identifiable facial features within magnetic resonance imaging (MRI) and computed tomography (CT) scans, restricting model performance in clinical medicine. Here we show that frontier models underperform on neuroimaging tasks and that learning directly from uncurated data generated during routine clinical care at health systems, a paradigm we call 'health system learning', yields high-performance, generalist neuroimaging models. We introduce NeuroVFM, a visual foundation model trained on 5.24 million clinical MRI and CT volumes using a scalable volumetric predictive architecture. NeuroVFM learns comprehensive representations of brain anatomy and pathology, achieving state-of-the-art performance across multiple clinical tasks, including radiologic diagnosis and report generation. The model embeds MRI and CT scans into a shared neuroanatomic latent space and grounds diagnostic findings. When paired with open-source language models, NeuroVFM generates radiology reports that surpass frontier models in accuracy, clinical triage and expert preference. NeuroVFM reduces hallucinated findings and critical errors, offering safer clinical decision support. These results establish health system learning as a paradigm for building generalist medical AI and provide a scalable framework for clinical foundation models.

Recent advances in large language models have enabled deep research systems that generate expert-level reports through multi-step reasoning and evidence-based synthesis. However, evaluating such reports remains challenging: report quality is multifaceted, making it difficult to determine what to assess and by what criteria; LLM-based judges may miss errors that require domain expertise to identify; and because deep research relies on retrieved evidence, report-wide claim verification is also necessary. To address these issues, we propose DEER, a benchmark for evaluating expert-level deep research reports. DEER systematizes evaluation criteria with an expert-developed taxonomy (7 dimensions, 25 subdimensions) operationalized as 101 fine-grained rubric items. We also provide task-specific Expert Evaluation Guidance to support LLM-based judging. Alongside rubric-based assessment, we propose a claim verification architecture that verifies both cited and uncited claims and quantifies evidence quality. Experiments show that while current deep research systems can produce structurally plausible reports that cite external evidence, there is room for improvement in fulfilling expert-level user requests and achieving logical completeness. Beyond simple performance comparisons, DEER makes system strengths and limitations interpretable and provides diagnostic signals for improvement.

Achieving reliable robotic manipulation, such as dexterous grasping, requires a synergy between physically stable interactions and semantic task guidance, yet these objectives are often treated as separate, disjoint goals. In this paper, we investigate how to integrate dexterous grasping techniques, i.e., physically stable grasps for object lifting and language-guided grasp generation, to achieve both physical stability and semantic understanding. To this end, we propose SECOND-Grasp (SEmantic CONtact-guided Dexterous Grasping), a unified framework that enables robotic hands to dynamically adjust grasping strategies based on semantic reasoning while ensuring physical feasibility. We begin by obtaining coarse contact proposals through vision-language reasoning to infer where contacts should occur based on object properties, followed by segmentation to localize these regions across views. To further ensure consistency across multiple viewpoints, we introduce Semantic-Geometric Consistency Refinement (SGCR), which refines initial contact predictions by enforcing semantic consistency across views and removing geometrically invalid regions, yielding reliable 3D contact maps. Then, we derive a feasible hand pose for each contact map via inverse kinematics, generating a supervision signal for policy learning. Our approach, trained on DexGraspNet, consistently outperforms baselines in lifting success rate on both seen and unseen categories, achieving 98.2

Neuroimaging is a ubiquitous tool for evaluating patients with neurological diseases. The global demand for magnetic resonance imaging (MRI) studies has risen steadily, placing substantial strain on health systems, prolonging turnaround times and intensifying physician burnout. These challenges disproportionately impact patients in low-resource and rural settings. Here we utilize data from a large academic health system to develop Prima, an AI foundation model for neuroimaging that supports real-world, clinical MRI studies as input. Trained on over 220,000 MRI studies, Prima uses a hierarchical vision architecture that provides general and transferable MRI features. Prima was tested in a 1-year health system-wide study that included 29,431 MRI studies. Across 52 radiologic diagnoses from major neurologic disorders, Prima achieved a mean diagnostic area under the curve (AUC) of 92.0%, outperforming other state-of-the-art general and medical AI models. Prima offers explainable differential diagnoses, worklist priority for radiologists and clinical referral recommendations. Prima demonstrates algorithmic fairness across sensitive groups. These findings highlight the transformative potential of health system-scale AI training and Prima's role in advancing AI-driven healthcare.

Sparse upcycling provides an efficient way to initialize a Mixture-of-Experts (MoE) model from a pretrained dense checkpoint instead of training from scratch.However, since all experts start from identical weights and the router is randomly initialized, the model suffers from expert symmetry and limited early specialization.We propose Cluster-aware Upcycling, a strategy that embeds semantic structure into MoE initialization.The method clusters the dense model’s input activations to identify latent subspaces, initializes each expert using a data-aware truncated SVD of the dense weights within its cluster, and initializes the router with the corresponding cluster centroids.This cluster-aware initialization breaks expert symmetry and encourages early specialization aligned with the data structure.In addition, we introduce an Expert-Ensemble Self-Distillation loss that regularizes training by guiding uncertain routing with stable predictions from an ensemble teacher.Applied to CLIP ViT-B/16 and ViT-B/32 models, Cluster-aware Upcycling achieves consistent improvements over standard upcycling across zero-shot and few-shot benchmarks, and produces more diverse and disentangled expert representations.

Natural-language critiques provide supervision beyond scalar rewards for non-verifiable generation, which lacks deterministic verifiers. In critique-guided refinement, a critic gives feedback on an initial response and an actor revises it. However, final revision quality does not reveal whether the critique was actually useful: a capable actor may improve without following the feedback, while valid feedback may fail if the actor cannot execute it. We frame critique as actor-conditioned revision guidance, where usefulness depends on whether the feedback helps the target actor address the intended weakness. We introduce TAIScore (Targeted Actionable Improvement Score), a reward that evaluates the instruction, initial response, critique, and revision together, assessing whether the critique targets a real weakness, whether the actor follows it, and whether the intended aspect improves. We use this reward to train an actor-tailored critic with GRPO, and use critique-guided refinements to construct DPO preference pairs for the actor, forming a co-evolving critic-actor loop where the critic adapts to the actor's changing capability. Experiments show that an 8B critic trained with TAIScore outperforms both a zero-shot 120B critic and critics trained with outcome-only or critique-only reward signals. Co-evolving the critic and actor further improves performance, suggesting that effective critique supervision should adapt as the actor changes.

This technical report presents K-EXAONE, a large-scale multilingual language model developed by LG AI Research. K-EXAONE is built on a Mixture-of-Experts architecture with 236B total parameters, activating 23B parameters during inference. It supports a 256K-token context window and covers six languages: Korean, English, Spanish, German, Japanese, and Vietnamese. We evaluate K-EXAONE on a comprehensive benchmark suite spanning reasoning, agentic, general, Korean, and multilingual abilities. Across these evaluations, K-EXAONE demonstrates performance comparable to open-weight models of similar size. K-EXAONE, designed to advance AI for a better life, is positioned as a powerful proprietary AI foundation model for a wide range of industrial and research applications.

This technical report introduces EXAONE 4.5, the first open-weight vision language model released by LG AI Research. EXAONE 4.5 is architected by integrating a dedicated visual encoder into the existing EXAONE 4.0 framework, enabling native multimodal pretraining over both visual and textual modalities. The model is trained on large-scale data with careful curation, particularly emphasizing document-centric corpora that align with LG's strategic application domains. This targeted data design enables substantial performance gains in document understanding and related tasks, while also delivering broad improvements across general language capabilities. EXAONE 4.5 extends context length up to 256K tokens, facilitating long-context reasoning and enterprise-scale use cases. Comparative evaluations demonstrate that EXAONE 4.5 achieves competitive performance in general benchmarks while outperforming state-of-the-art models of similar scale in document understanding and Korean contextual reasoning. As part of LG's ongoing effort toward practical industrial deployment, EXAONE 4.5 is designed to be continuously extended with additional domains and application scenarios to advance AI for a better life.

We present Spanning Tree Autoregressive (STAR) modeling, which can incorporate prior knowledge of images, such as center bias and locality, to maintain sampling performance while also providing sufficiently flexible sequence orders to accommodate image editing at inference. Approaches that expose randomly permuted sequence orders to conventional autoregressive (AR) models in visual generation for bidirectional context either suffer from a decline in performance or compromise the flexibility in sequence order choice at inference. Instead, STAR utilizes traversal orders of uniform spanning trees sampled in a lattice defined by the positions of image patches. Traversal orders are obtained through breadth-first search, allowing us to efficiently construct a spanning tree whose traversal order ensures that the connected partial observation of the image appears as a prefix in the sequence through rejection sampling. Through the tailored yet structured randomized strategy compared to random permutation, STAR preserves the capability of postfix completion while maintaining sampling performance without any significant changes to the model architecture widely adopted in the language AR modeling.