Daily Papers

Computer use agents (CUAs) have advanced rapidly in desktop automation, and a growing number of users deploy CUAs such as OpenClaw on Mac Mini for always-on automation. However, existing benchmarks, including those for macOS, evaluate agents without framework augmentation and rely on binary evaluation. As a result, they fail to capture both the framework capabilities leveraged by modern CUAs and the partial progress on long-horizon, multi-application tasks. We present MacAgentBench, a comprehensive macOS agent benchmark comprising 676 tasks across 25 applications, with nearly 60% involving both GUI and CLI interaction. The benchmark adopts deterministic rule-based evaluation and introduces fine-grained multi-checkpoint scoring with capability annotations for multi-application tasks. Experiments across three frameworks and 16 models show that the best configuration, Claude Opus 4.6 on OpenClaw, attains 73.7% Pass@1, while this advantage is primarily driven by the skill library rather than by framework design. Fine-grained metrics further reveal that models with similar Pass@1 can differ substantially in sub-goal completion. Our code and data are publicly available at https://github.com/JetAstra/MacAgentBench.

We introduce eCLIP, an enhanced version of the CLIP model that integrates expert annotations in the form of radiologist eye-gaze heatmaps. It tackles key challenges in contrastive multi-modal medical imaging analysis, notably data scarcity and the "modality gap" -- a significant disparity between image and text embeddings that diminishes the quality of representations and hampers cross-modal interoperability. eCLIP integrates a heatmap processor and leverages mixup augmentation to efficiently utilize the scarce expert annotations, thus boosting the model's learning effectiveness. eCLIP is designed to be generally applicable to any variant of CLIP without requiring any modifications of the core architecture. Through detailed evaluations across several tasks, including zero-shot inference, linear probing, cross-modal retrieval, and Retrieval Augmented Generation (RAG) of radiology reports using a frozen Large Language Model, eCLIP showcases consistent improvements in embedding quality. The outcomes reveal enhanced alignment and uniformity, affirming eCLIP's capability to harness high-quality annotations for enriched multi-modal analysis in the medical imaging domain.

Recent advancements in multimodal reasoning have largely overlooked the audio modality. We introduce Audio-Reasoner, a large-scale audio language model for deep reasoning in audio tasks. We meticulously curated a large-scale and diverse multi-task audio dataset with simple annotations. Then, we leverage closed-source models to conduct secondary labeling, QA generation, along with structured COT process. These datasets together form a high-quality reasoning dataset with 1.2 million reasoning-rich samples, which we name CoTA. Following inference scaling principles, we train Audio-Reasoner on CoTA, enabling it to achieve great logical capabilities in audio reasoning. Experiments show state-of-the-art performance across key benchmarks, including MMAU-mini (+25.42%), AIR-Bench chat/foundation(+14.57%/+10.13%), and MELD (+8.01%). Our findings stress the core of structured CoT training in advancing audio reasoning.

DeepSeek-R1-Zero has successfully demonstrated the emergence of reasoning capabilities in LLMs purely through Reinforcement Learning (RL). Inspired by this breakthrough, we explore how RL can be utilized to enhance the reasoning capability of MLLMs. However, direct training with RL struggles to activate complex reasoning capabilities such as questioning and reflection in MLLMs, due to the absence of substantial high-quality multimodal reasoning data. To address this issue, we propose the reasoning MLLM, Vision-R1, to improve multimodal reasoning capability. Specifically, we first construct a high-quality multimodal CoT dataset without human annotations by leveraging an existing MLLM and DeepSeek-R1 through modality bridging and data filtering to obtain a 200K multimodal CoT dataset, Vision-R1-cold dataset. It serves as cold-start initialization data for Vision-R1. To mitigate the optimization challenges caused by overthinking after cold start, we propose Progressive Thinking Suppression Training (PTST) strategy and employ Group Relative Policy Optimization (GRPO) with the hard formatting result reward function to gradually refine the model's ability to learn correct and complex reasoning processes on a 10K multimodal math dataset. Comprehensive experiments show our model achieves an average improvement of sim6% across various multimodal math reasoning benchmarks. Vision-R1-7B achieves a 73.5% accuracy on the widely used MathVista benchmark, which is only 0.4% lower than the leading reasoning model, OpenAI O1. The datasets and code will be released in: https://github.com/Osilly/Vision-R1 .

Instructional videos are an excellent source for learning multimodal representations by leveraging video-subtitle pairs extracted with automatic speech recognition systems (ASR) from the audio signal in the videos. However, in contrast to human-annotated captions, both speech and subtitles naturally differ from the visual content of the videos and thus provide only noisy supervision for multimodal learning. As a result, large-scale annotation-free web video training data remains sub-optimal for training text-video models. In this work, we propose to leverage the capability of large language models (LLMs) to obtain fine-grained video descriptions aligned with videos. Specifically, we prompt an LLM to create plausible video descriptions based on ASR narrations of the video for a large-scale instructional video dataset. To this end, we introduce a prompting method that is able to take into account a longer text of subtitles, allowing us to capture context beyond a single sentence. To align the captions to the video temporally, we prompt the LLM to generate timestamps for each produced caption based on the subtitles. In this way, we obtain human-style video captions at scale without human supervision. We apply our method to the subtitles of the HowTo100M dataset, creating a new large-scale dataset, HowToCaption. Our evaluation shows that the resulting captions not only significantly improve the performance over many different benchmark datasets for text-video retrieval but also lead to a disentangling of textual narration from the audio, boosting performance in text-video-audio tasks.

Reinforcement fine-tuning (RFT) has shown great promise in achieving humanlevel reasoning capabilities of Large Language Models (LLMs), and has recently been extended to MLLMs. Nevertheless, reasoning about videos, which is a fundamental aspect of human intelligence, remains a persistent challenge due to the complex logic, temporal and causal structures inherent in video data. To fill this gap, we propose VideoRFT, a novel approach that extends the RFT paradigm to cultivate human-like video reasoning capabilities in MLLMs. VideoRFT follows the standard two-stage scheme in RFT: supervised fine-tuning (SFT) with chain-of-thought (CoT) annotations, followed by reinforcement learning (RL) to improve generalization. A central challenge to achieve this in the video domain lies in the scarcity of large-scale, high-quality video CoT datasets. We address this by building a multi-expert-driven, cognition-inspired CoT curation pipeline. First, we devise a cognition-inspired prompting strategy to elicit a reasoning LLM to generate preliminary CoTs based solely on rich, structured, and literal representations of video content. Subsequently, these CoTs are revised by a MLLM conditioned on the actual video, ensuring visual consistency and reducing visual hallucinations. This pipeline results in two new datasets, i.e.VideoRFT-CoT-102K for SFT and VideoRFT-RL-310K for RL. To further strengthen the RL phase, we introduce a novel semantic-consistency reward that explicitly promotes the alignment between textual reasoning and visual evidence. This reward encourages the model to produce coherent, context-aware reasoning outputs grounded in visual input. Extensive experiments show that VideoRFT achieves state-of-the-art performance on six video reasoning benchmarks.

Bimanual manipulation is essential for achieving human-like dexterity in robots, but the large-scale and diverse bimanual robot datasets remain scarce due to hardware heterogeneity across robotic platforms. To address the challenge, we present RoboCOIN, a comprehensive multi-embodiment bimanual manipulation dataset with over 180,000 demonstrations collected from 15 distinct robotic platforms. The dataset covers 16 scenarios, including residential, commercial, and working environments, with 421 tasks systematically organized by bimanual coordination patterns and object properties. Our key innovation is a hierarchical capability pyramid that provides multi-level annotations, spanning trajectory-level concepts, segment-level subtasks, and frame-level kinematics. We further develop CoRobot, a comprehensive processing framework featuring Robot Trajectory Markup Language (RTML) for quality assessment, automated annotation generation, and unified multi-embodiment management. Extensive experiments demonstrate the reliability and effectiveness of RoboCOIN in multi-embodiment bimanual learning, with significant performance improvements across various model architectures and robotic platforms. The complete dataset and framework are open-sourced and publicly available for further research purposes. Project website: https://FlagOpen.github.io/RoboCOIN/.

Visual Text Rendering (VTR) remains a critical challenge in text-to-image generation, where even advanced models frequently produce text with structural anomalies such as distortion, blurriness, and misalignment. However, we find that leading MLLMs and specialist OCR models largely fail to perceive these structural anomalies, creating a critical bottleneck for both VTR evaluation and RL-based optimization. As a result, even state-of-the-art generators (e.g., SeedDream4.0, Qwen-Image) still struggle to render structurally faithful text. To address this, we propose TextPecker, a plug-and-play structural anomaly perceptive RL strategy that mitigates noisy reward signals and works with any textto-image generator. To enable this capability, we construct a recognition dataset with character-level structural-anomaly annotations and develop a stroke-editing synthesis engine to expand structural-error coverage. Experiments show that TextPecker consistently improves diverse text-to-image models; even on the well-optimized Qwen-Image, it significantly yields average gains of 4% in structural fidelity and 8.7% in semantic alignment for Chinese text rendering, establishing a new state-of-the-art in high-fidelity VTR. Our work fills a gap in VTR optimization, providing a foundational step towards reliable and structural faithful visual text generation.

Generalizing video matting models to real-world videos remains a significant challenge due to the scarcity of labeled data. To address this, we present Video Mask-to-Matte Model (VideoMaMa) that converts coarse segmentation masks into pixel accurate alpha mattes, by leveraging pretrained video diffusion models. VideoMaMa demonstrates strong zero-shot generalization to real-world footage, even though it is trained solely on synthetic data. Building on this capability, we develop a scalable pseudo-labeling pipeline for large-scale video matting and construct the Matting Anything in Video (MA-V) dataset, which offers high-quality matting annotations for more than 50K real-world videos spanning diverse scenes and motions. To validate the effectiveness of this dataset, we fine-tune the SAM2 model on MA-V to obtain SAM2-Matte, which outperforms the same model trained on existing matting datasets in terms of robustness on in-the-wild videos. These findings emphasize the importance of large-scale pseudo-labeled video matting and showcase how generative priors and accessible segmentation cues can drive scalable progress in video matting research.

In this paper, we introduce an underexplored problem in facial analysis: generating and recognizing multi-attribute natural language descriptions, containing facial action units (AUs), emotional states, and age estimation, for arbitrarily selected face regions (termed FaceFocalDesc). We argue that the system's ability to focus on individual facial areas leads to better understanding and control. To achieve this capability, we construct a new multi-attribute description dataset for arbitrarily selected face regions, providing rich region-level annotations and natural language descriptions. Further, we propose a fine-tuned vision-language model based on Qwen2.5-VL, called Focal-RegionFace for facial state analysis, which incrementally refines its focus on localized facial features through multiple progressively fine-tuning stages, resulting in interpretable age estimation, FAU and emotion detection. Experimental results show that Focal-RegionFace achieves the best performance on the new benchmark in terms of traditional and widely used metrics, as well as new proposed metrics. This fully verifies its effectiveness and versatility in fine-grained multi-attribute face region-focal analysis scenarios.

We present a validation dataset of newly-collected kitchen-based egocentric videos, manually annotated with highly detailed and interconnected ground-truth labels covering: recipe steps, fine-grained actions, ingredients with nutritional values, moving objects, and audio annotations. Importantly, all annotations are grounded in 3D through digital twinning of the scene, fixtures, object locations, and primed with gaze. Footage is collected from unscripted recordings in diverse home environments, making HDEPIC the first dataset collected in-the-wild but with detailed annotations matching those in controlled lab environments. We show the potential of our highly-detailed annotations through a challenging VQA benchmark of 26K questions assessing the capability to recognise recipes, ingredients, nutrition, fine-grained actions, 3D perception, object motion, and gaze direction. The powerful long-context Gemini Pro only achieves 38.5% on this benchmark, showcasing its difficulty and highlighting shortcomings in current VLMs. We additionally assess action recognition, sound recognition, and long-term video-object segmentation on HD-EPIC. HD-EPIC is 41 hours of video in 9 kitchens with digital twins of 413 kitchen fixtures, capturing 69 recipes, 59K fine-grained actions, 51K audio events, 20K object movements and 37K object masks lifted to 3D. On average, we have 263 annotations per minute of our unscripted videos.

We present v1, a lightweight extension to Multimodal Large Language Models (MLLMs) that enables selective visual revisitation during inference. While current MLLMs typically consume visual input only once and reason purely over internal memory, v1 introduces a simple point-and-copy mechanism that allows the model to dynamically retrieve relevant image regions throughout the reasoning process. This mechanism augments existing architectures with minimal modifications, enabling contextual access to visual tokens based on the model's evolving hypotheses. To train this capability, we construct v1g, a dataset of 300K multimodal reasoning traces with interleaved visual grounding annotations. Experiments on three multimodal mathematical reasoning benchmarks -- MathVista, MathVision, and MathVerse -- demonstrate that v1 consistently improves performance over comparable baselines, particularly on tasks requiring fine-grained visual reference and multi-step reasoning. Our results suggest that dynamic visual access is a promising direction for enhancing grounded multimodal reasoning. Code, models, and data will be released to support future research.

3D Gaussian Splatting (3D GS) has gained popularity due to its faster rendering speed and high-quality novel view synthesis. Some researchers have explored using 3D GS for reconstructing driving scenes. However, these methods often rely on various data types, such as depth maps, 3D boxes, and trajectories of moving objects. Additionally, the lack of annotations for synthesized images limits their direct application in downstream tasks. To address these issues, we propose EGSRAL, a 3D GS-based method that relies solely on training images without extra annotations. EGSRAL enhances 3D GS's capability to model both dynamic objects and static backgrounds and introduces a novel adaptor for auto labeling, generating corresponding annotations based on existing annotations. We also propose a grouping strategy for vanilla 3D GS to address perspective issues in rendering large-scale, complex scenes. Our method achieves state-of-the-art performance on multiple datasets without any extra annotation. For example, the PSNR metric reaches 29.04 on the nuScenes dataset. Moreover, our automated labeling can significantly improve the performance of 2D/3D detection tasks. Code is available at https://github.com/jiangxb98/EGSRAL.

One way to enhance the reasoning capability of Large Language Models (LLMs) is to conduct Supervised Fine-Tuning (SFT) using Chain-of-Thought (CoT) annotations. This approach does not show sufficiently strong generalization ability, however, because the training only relies on the given CoT data. In math problem-solving, for example, there is usually only one annotated reasoning path for each question in the training data. Intuitively, it would be better for the algorithm to learn from multiple annotated reasoning paths given a question. To address this issue, we propose a simple yet effective approach called Reinforced Fine-Tuning (ReFT) to enhance the generalizability of learning LLMs for reasoning, with math problem-solving as an example. ReFT first warmups the model with SFT, and then employs on-line reinforcement learning, specifically the PPO algorithm in this paper, to further fine-tune the model, where an abundance of reasoning paths are automatically sampled given the question and the rewards are naturally derived from the ground-truth answers. Extensive experiments on GSM8K, MathQA, and SVAMP datasets show that ReFT significantly outperforms SFT, and the performance can be potentially further boosted by combining inference-time strategies such as majority voting and re-ranking. Note that ReFT obtains the improvement by learning from the same training questions as SFT, without relying on extra or augmented training questions. This indicates a superior generalization ability for ReFT.

Existing Multivariate Time Series Anomaly Detection (MTSAD) frameworks increasingly rely on integrating Graph Neural Networks (GNNs) with sequence models to capture complex spatio-temporal dependencies. However, less attention is paid to the spatial over-generalization problem, where unconstrained structural modeling indiscriminately reconstructs anomalies, inevitably degrading detection recall. To tackle this problem, we propose a novel framework that unifies spatio-temporal modeling through a joint prior-observation adversarial learning paradigm. In the spatial dimension, the model alternately learns adjacency matrices as structural prior and models the association discrepancy between prior and data-driven observation in a minimax manner during training. Such adversarial optimization not only improves the model sensitivity for time-wise detection, but also enables the model to localize anomalies to specific channels. To systematically evaluate this anomaly localization capability, we further construct a synthetic benchmark equipped with precise channel-wise annotations. Extensive experiments across public datasets and our dedicated benchmark demonstrate that the proposed framework establishes a new state-of-the-art in both time-wise detection and spatial localization tasks. Our code, pre-trained models, and benchmark are publicly available at https://github.com/anocodetest1/POST.

LLM-as-a-Judge refers to the automatic modeling of preferences for responses generated by Large Language Models (LLMs), which is of significant importance for both LLM evaluation and reward modeling. Although generative LLMs have made substantial progress in various tasks, their performance as LLM-Judge still falls short of expectations. In this work, we propose Think-J, which improves generative LLM-as-a-Judge by learning how to think. We first utilized a small amount of curated data to develop the model with initial judgment thinking capabilities. Subsequently, we optimize the judgment thinking traces based on reinforcement learning (RL). We propose two methods for judgment thinking optimization, based on offline and online RL, respectively. The offline RL requires training a critic model to construct positive and negative examples for learning. The online method defines rule-based reward as feedback for optimization. Experimental results showed that our approach can significantly enhance the evaluation capability of generative LLM-Judge, surpassing both generative and classifier-based LLM-Judge without requiring extra human annotations.

This paper proposes PyFi, a novel framework for pyramid-like financial image understanding that enables vision language models (VLMs) to reason through question chains in a progressive, simple-to-complex manner. At the core of PyFi is PyFi-600K, a dataset comprising 600K financial question-answer pairs organized into a reasoning pyramid: questions at the base require only basic perception, while those toward the apex demand increasing levels of capability in financial visual understanding and expertise. This data is scalable because it is synthesized without human annotations, using PyFi-adv, a multi-agent adversarial mechanism under the Monte Carlo Tree Search (MCTS) paradigm, in which, for each image, a challenger agent competes with a solver agent by generating question chains that progressively probe deeper capability levels in financial visual reasoning. Leveraging this dataset, we present fine-grained, hierarchical, and comprehensive evaluations of advanced VLMs in the financial domain. Moreover, fine-tuning Qwen2.5-VL-3B and Qwen2.5-VL-7B on the pyramid-structured question chains enables these models to answer complex financial questions by decomposing them into sub-questions with gradually increasing reasoning demands, yielding average accuracy improvements of 19.52% and 8.06%, respectively, on the dataset. All resources of code, dataset and models are available at: https://github.com/AgenticFinLab/PyFi .

We propose a simple yet effective strategy to incorporate event knowledge extracted from event trigger annotations via posterior regularization to improve the event reasoning capability of mainstream question-answering (QA) models for event-centric QA. In particular, we define event-related knowledge constraints based on the event trigger annotations in the QA datasets, and subsequently use them to regularize the posterior answer output probabilities from the backbone pre-trained language models used in the QA setting. We explore two different posterior regularization strategies for extractive and generative QA separately. For extractive QA, the sentence-level event knowledge constraint is defined by assessing if a sentence contains an answer event or not, which is later used to modify the answer span extraction probability. For generative QA, the token-level event knowledge constraint is defined by comparing the generated token from the backbone language model with the answer event in order to introduce a reward or penalty term, which essentially adjusts the answer generative probability indirectly. We conduct experiments on two event-centric QA datasets, TORQUE and ESTER. The results show that our proposed approach can effectively inject event knowledge into existing pre-trained language models and achieves strong performance compared to existing QA models in answer evaluation. Code and models can be found: https://github.com/LuJunru/EventQAviaPR.

Large language models (LLMs) have achieved significant progress in solving complex reasoning tasks by Reinforcement Learning with Verifiable Rewards (RLVR). This advancement is also inseparable from the oversight automated by reliable verifiers. However, current outcome-based verifiers (OVs) are unable to inspect the unreliable intermediate steps in the long reasoning chains of thought (CoTs). Meanwhile, current process-based verifiers (PVs) have difficulties in reliably detecting errors in the complex long CoTs, limited by the scarcity of high-quality annotations due to the prohibitive costs of human annotations. Therefore, we propose the Outcome-based Process Verifier (OPV), which verifies the rationale process of summarized outcomes from long CoTs to achieve both accurate and efficient verification and enable large-scale annotation. To empower the proposed verifier, we adopt an iterative active learning framework with expert annotations to progressively improve the verification capability of OPV with fewer annotation costs. Specifically, in each iteration, the most uncertain cases of the current best OPV are annotated and then subsequently used to train a new OPV through Rejection Fine-Tuning (RFT) and RLVR for the next round. Extensive experiments demonstrate OPV's superior performance and broad applicability. It achieves new state-of-the-art results on our held-out \thisbench, outperforming much larger open-source models such as Qwen3-Max-Preview with an F1 score of 83.1 compared to 76.3. Furthermore, OPV effectively detects false positives within synthetic dataset, closely align with expert assessment. When collaborating with policy models, OPV consistently yields performance gains, e.g., raising the accuracy of DeepSeek-R1-Distill-Qwen-32B from 55.2\% to 73.3\% on AIME2025 as the compute budget scales.

Large language models (LLMs) have achieved significant progress in solving complex reasoning tasks by Reinforcement Learning with Verifiable Rewards (RLVR). This advancement is also inseparable from the oversight automated by reliable verifiers. However, current outcome-based verifiers (OVs) are unable to inspect the unreliable intermediate steps in the long reasoning chains of thought (CoTs). Meanwhile, current process-based verifiers (PVs) have difficulties in reliably detecting errors in the complex long CoTs, limited by the scarcity of high-quality annotations due to the prohibitive costs of human annotations. Therefore, we propose the Outcome-based Process Verifier (OPV), which verifies the rationale process of summarized outcomes from long CoTs to achieve both accurate and efficient verification and enable large-scale annotation. To empower the proposed verifier, we adopt an iterative active learning framework with expert annotations to progressively improve the verification capability of OPV with fewer annotation costs. Specifically, in each iteration, the most uncertain cases of the current best OPV are annotated and then subsequently used to train a new OPV through Rejection Fine-Tuning (RFT) and RLVR for the next round. Extensive experiments demonstrate OPV's superior performance and broad applicability. It achieves new state-of-the-art results on our held-out OPV-Bench, outperforming much larger open-source models such as Qwen3-Max-Preview with an F1 score of 83.1 compared to 76.3. Furthermore, OPV effectively detects false positives within synthetic dataset, closely align with expert assessment. When collaborating with policy models, OPV consistently yields performance gains, e.g., raising the accuracy of DeepSeek-R1-Distill-Qwen-32B from 55.2% to 73.3% on AIME2025 as the compute budget scales.

Amodal perception, the ability to comprehend complete object structures from partial visibility, is a fundamental skill, even for infants. Its significance extends to applications like autonomous driving, where a clear understanding of heavily occluded objects is essential. However, modern detection and tracking algorithms often overlook this critical capability, perhaps due to the prevalence of modal annotations in most datasets. To address the scarcity of amodal data, we introduce the TAO-Amodal benchmark, featuring 880 diverse categories in thousands of video sequences. Our dataset includes amodal and modal bounding boxes for visible and occluded objects, including objects that are partially out-of-frame. To enhance amodal tracking with object permanence, we leverage a lightweight plug-in module, the amodal expander, to transform standard, modal trackers into amodal ones through fine-tuning on a few hundred video sequences with data augmentation. We achieve a 3.3\% and 1.6\% improvement on the detection and tracking of occluded objects on TAO-Amodal. When evaluated on people, our method produces dramatic improvements of 2x compared to state-of-the-art modal baselines.

Multimodal large language models (MLLMs) demonstrate significant potential in the field of medical diagnosis. However, they face critical challenges in specialized domains such as ophthalmology, particularly the fragmentation of annotation granularity and inconsistencies in clinical reasoning logic, which hinder precise cross-modal understanding. This paper introduces FundusExpert, an ophthalmology-specific MLLM with integrated positioning-diagnosis reasoning capabilities, along with FundusGen, a dataset constructed through the intelligent Fundus-Engine system. Fundus-Engine automates localization and leverages MLLM-based semantic expansion to integrate global disease classification, local object detection, and fine-grained feature analysis within a single fundus image. Additionally, by constructing a clinically aligned cognitive chain, it guides the model to generate interpretable reasoning paths. FundusExpert, fine-tuned with instruction data from FundusGen, achieves the best performance in ophthalmic question-answering tasks, surpassing the average accuracy of the 40B MedRegA by 26.6%. It also excels in zero-shot report generation tasks, achieving a clinical consistency of 77.0%, significantly outperforming GPT-4o's 47.6%. Furthermore, we reveal a scaling law between data quality and model capability (L propto N^{0.068}), demonstrating that the cognitive alignment annotations in FundusGen enhance data utilization efficiency. By integrating region-level localization with diagnostic reasoning chains, our work develops a scalable, clinically-aligned MLLM and explores a pathway toward bridging the visual-language gap in specific MLLMs. Our project can be found at https://github.com/MeteorElf/FundusExpert.

Reconstructing 3D hand-face interactions with deformations from a single image is a challenging yet crucial task with broad applications in AR, VR, and gaming. The challenges stem from self-occlusions during single-view hand-face interactions, diverse spatial relationships between hands and face, complex deformations, and the ambiguity of the single-view setting. The first and only method for hand-face interaction recovery, Decaf, introduces a global fitting optimization guided by contact and deformation estimation networks trained on studio-collected data with 3D annotations. However, Decaf suffers from a time-consuming optimization process and limited generalization capability due to its reliance on 3D annotations of hand-face interaction data. To address these issues, we present DICE, the first end-to-end method for Deformation-aware hand-face Interaction reCovEry from a single image. DICE estimates the poses of hands and faces, contacts, and deformations simultaneously using a Transformer-based architecture. It features disentangling the regression of local deformation fields and global mesh vertex locations into two network branches, enhancing deformation and contact estimation for precise and robust hand-face mesh recovery. To improve generalizability, we propose a weakly-supervised training approach that augments the training set using in-the-wild images without 3D ground-truth annotations, employing the depths of 2D keypoints estimated by off-the-shelf models and adversarial priors of poses for supervision. Our experiments demonstrate that DICE achieves state-of-the-art performance on a standard benchmark and in-the-wild data in terms of accuracy and physical plausibility. Additionally, our method operates at an interactive rate (20 fps) on an Nvidia 4090 GPU, whereas Decaf requires more than 15 seconds for a single image. Our code will be publicly available upon publication.

In production, multi-modal large language models (MLLMs) are expected to support multi-turn queries of interchanging image and text modalities. However, the current MLLMs trained with visual-question-answering (VQA) datasets could suffer from degradation, as VQA datasets lack the diversity and complexity of the original text instruction datasets which the underlying language model had been trained with. To address this challenging degradation, we first collect a lightweight (6k entries) VQA preference dataset where answers were annotated by Gemini for 5 quality metrics in a granular fashion, and investigate standard Supervised Fine-tuning, rejection sampling, Direct Preference Optimization (DPO), and SteerLM. Our findings indicate that the with DPO we are able to surpass instruction-following capabilities of the language model, achieving a 6.73 score on MT-Bench, compared to Vicuna's 6.57 and LLaVA's 5.99 despite small data scale. This enhancement in textual instruction proficiency correlates with boosted visual instruction performance (+4.9\% on MM-Vet, +6\% on LLaVA-Bench), with minimal alignment tax on visual knowledge benchmarks compared to previous RLHF approach. In conclusion, we propose a distillation-based multi-modal alignment model with fine-grained annotations on a small dataset that reconciles the textual and visual performance of MLLMs, restoring and boosting language capability after visual instruction tuning.

The goal of dialogue relation extraction (DRE) is to identify the relation between two entities in a given dialogue. During conversations, speakers may expose their relations to certain entities by explicit or implicit clues, such evidences called "triggers". However, trigger annotations may not be always available for the target data, so it is challenging to leverage such information for enhancing the performance. Therefore, this paper proposes to learn how to identify triggers from the data with trigger annotations and then transfers the trigger-finding capability to other datasets for better performance. The experiments show that the proposed approach is capable of improving relation extraction performance of unseen relations and also demonstrate the transferability of our proposed trigger-finding model across different domains and datasets.

Teaching large language models (LLMs) to generate text with citations to evidence sources can mitigate hallucinations and enhance verifiability in information-seeking systems. However, improving this capability requires high-quality attribution data, which is costly and labor-intensive. Inspired by recent advances in self-improvement that enhance LLMs without manual annotation, we present START, a Self-Taught AttRibuTion framework for iteratively improving the attribution capability of LLMs. First, to prevent models from stagnating due to initially insufficient supervision signals, START leverages the model to self-construct synthetic training data for warming up. To further self-improve the model's attribution ability, START iteratively utilizes fine-grained preference supervision signals constructed from its sampled responses to encourage robust, comprehensive, and attributable generation. Experiments on three open-domain question-answering datasets, covering long-form QA and multi-step reasoning, demonstrate significant performance gains of 25.13% on average without relying on human annotations and more advanced models. Further analysis reveals that START excels in aggregating information across multiple sources.

The National Comprehensive Cancer Network (NCCN) provides evidence-based guidelines for cancer treatment. Translating complex patient presentations into guideline-compliant treatment recommendations is time-intensive, requires specialized expertise, and is prone to error. Advances in large language model (LLM) capabilities promise to reduce the time required to generate treatment recommendations and improve accuracy. We present an LLM agent-based approach to automatically generate guideline-concordant treatment trajectories for patients with non-small cell lung cancer (NSCLC). Our contributions are threefold. First, we construct a novel longitudinal dataset of 121 cases of NSCLC patients that includes clinical encounters, diagnostic results, and medical histories, each expertly annotated with the corresponding NCCN guideline trajectories by board-certified oncologists. Second, we demonstrate that existing LLMs possess domain-specific knowledge that enables high-quality proxy benchmark generation for both model development and evaluation, achieving strong correlation (Spearman coefficient r=0.88, RMSE = 0.08) with expert-annotated benchmarks. Third, we develop a hybrid approach combining expensive human annotations with model consistency information to create both the agent framework that predicts the relevant guidelines for a patient, as well as a meta-classifier that verifies prediction accuracy with calibrated confidence scores for treatment recommendations (AUROC=0.800), a critical capability for communicating the accuracy of outputs, custom-tailoring tradeoffs in performance, and supporting regulatory compliance. This work establishes a framework for clinically viable LLM-based guideline adherence systems that balance accuracy, interpretability, and regulatory requirements while reducing annotation costs, providing a scalable pathway toward automated clinical decision support.

In this work, we focus on semi-supervised learning for video action detection which utilizes both labeled as well as unlabeled data. We propose a simple end-to-end consistency based approach which effectively utilizes the unlabeled data. Video action detection requires both, action class prediction as well as a spatio-temporal localization of actions. Therefore, we investigate two types of constraints, classification consistency, and spatio-temporal consistency. The presence of predominant background and static regions in a video makes it challenging to utilize spatio-temporal consistency for action detection. To address this, we propose two novel regularization constraints for spatio-temporal consistency; 1) temporal coherency, and 2) gradient smoothness. Both these aspects exploit the temporal continuity of action in videos and are found to be effective for utilizing unlabeled videos for action detection. We demonstrate the effectiveness of the proposed approach on two different action detection benchmark datasets, UCF101-24 and JHMDB-21. In addition, we also show the effectiveness of the proposed approach for video object segmentation on the Youtube-VOS which demonstrates its generalization capability The proposed approach achieves competitive performance by using merely 20% of annotations on UCF101-24 when compared with recent fully supervised methods. On UCF101-24, it improves the score by +8.9% and +11% at 0.5 f-mAP and v-mAP respectively, compared to supervised approach.

Recent advancements in reasoning capability of Multimodal Large Language Models (MLLMs) demonstrate its effectiveness in tackling complex visual tasks. However, existing MLLM-based Video Anomaly Detection (VAD) methods remain limited to shallow anomaly descriptions without deep reasoning. In this paper, we propose a new task named Video Anomaly Reasoning (VAR), which aims to enable deep analysis and understanding of anomalies in the video by requiring MLLMs to think explicitly before answering. To this end, we propose Vad-R1, an end-to-end MLLM-based framework for VAR. Specifically, we design a Perception-to-Cognition Chain-of-Thought (P2C-CoT) that simulates the human process of recognizing anomalies, guiding the MLLM to reason anomaly step-by-step. Based on the structured P2C-CoT, we construct Vad-Reasoning, a dedicated dataset for VAR. Furthermore, we propose an improved reinforcement learning algorithm AVA-GRPO, which explicitly incentivizes the anomaly reasoning capability of MLLMs through a self-verification mechanism with limited annotations. Experimental results demonstrate that Vad-R1 achieves superior performance, outperforming both open-source and proprietary models on VAD and VAR tasks. Codes and datasets will be released at https://github.com/wbfwonderful/Vad-R1.

Large recommender models have extended LLMs as powerful recommenders via encoding or item generation, and recent breakthroughs in LLM reasoning synchronously motivate the exploration of reasoning in recommendation. Current studies usually position LLMs as external reasoning modules to yield auxiliary thought for augmenting conventional recommendation pipelines. However, such decoupled designs are limited in significant resource cost and suboptimal joint optimization. To address these issues, we propose \name, a unified large recommender model with intrinsic reasoning capabilities. Initially, we reconceptualize the model architecture to facilitate interleaved reasoning and recommendation in the autoregressive process. Subsequently, we propose RecPO, a corresponding reinforcement learning framework that optimizes \name\ both the reasoning and recommendation capabilities simultaneously in a single policy update; RecPO introduces a fused reward scheme that solely leverages recommendation labels to simulate the reasoning capability, eliminating dependency on specialized reasoning annotations. Experiments on three datasets with various baselines verify the effectiveness of \name, showing relative improvements of 68.67\% in Hit@5 and 45.21\% in NDCG@20. Code available at https://github.com/YRYangang/RRec.

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term cross capabilities. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the "Law of the Weakest Link," where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios.

Large Language Models (LLMs) have demonstrated human-like instruction-following abilities, particularly those exceeding 100 billion parameters. The combined capability of some smaller, resource-friendly LLMs can address most of the instructions that larger LLMs excel at. In this work, we explore how to route the best-performing LLM for each instruction to achieve better overall performance. We develop a new paradigm, constructing capability instructions with model capability representation, user instruction, and performance inquiry prompts to assess the performance. To learn from capability instructions, we introduce a new end-to-end framework called Model Selection with Aptitude Test (Model-SAT), which generates positive and negative samples based on what different models perform well or struggle with. Model-SAT uses a model capability encoder that extends its model representation to a lightweight LLM. Our experiments show that Model-SAT understands the performance dimensions of candidate models and provides the probabilities of their capability to handle various instructions. Additionally, during deployment, a new model can quickly infer its aptitude test results across 50 tasks, each with 20 shots. Model-SAT performs state-of-the-art model routing without candidate inference and in real-world new model-released scenarios. The code is available at https://github.com/Now-Join-Us/CIT-LLM-Routing

Visual captioning benchmarks have become outdated with the emergence of modern multimodal large language models (MLLMs), as the brief ground-truth sentences and traditional metrics fail to assess detailed captions effectively. While recent benchmarks attempt to address this by focusing on keyword extraction or object-centric evaluation, they remain limited to vague-view or object-view analyses and incomplete visual element coverage. In this paper, we introduce CAPability, a comprehensive multi-view benchmark for evaluating visual captioning across 12 dimensions spanning six critical views. We curate nearly 11K human-annotated images and videos with visual element annotations to evaluate the generated captions. CAPability stably assesses both the correctness and thoroughness of captions using F1-score. By converting annotations to QA pairs, we further introduce a heuristic metric, know but cannot tell (KT), indicating a significant performance gap between QA and caption capabilities. Our work provides the first holistic analysis of MLLMs' captioning abilities, as we identify their strengths and weaknesses across various dimensions, guiding future research to enhance specific aspects of capabilities.

Recent advances in Large Language Models (LLMs) have significantly enhanced their capabilities, highlighting the need for comprehensive evaluation frameworks that extend beyond task-specific benchmarks. However, existing benchmarks often focus on isolated abilities, lacking a holistic framework for assessing LLM capabilities. To address this gap, we propose the Cognition-Domain-Task (CDT) framework, which comprehensively measures a model's capabilities across three dimensions. We expand the scope of model capability definitions at the cognitive level by incorporating the Cattell-Horn-Carroll cognitive theory, refining the categorization of model capabilities. We apply CDT in two directions: dataset capability evaluation and data selection. Experiments show that our capability metrics correlate well with downstream performance and can support effective dataset analysis and construction. The experiments on data selection also show significant improvements in both general and specific benchmarks, achieving scores of 44.3 and 45.4, with an increase of 1.6 and 2.2 points over the baselines, respectively. These results validate the effectiveness and practicality of CDT. Source code and models are available at https://github.com/Alessa-mo/CDT.

Existing online benchmarks for mobile GUI agents remain largely app-centric and task-homogeneous, failing to reflect the diversity and instability of real-world mobile usage. To this end, we introduce VenusBench-Mobile, a challenging online benchmark for evaluating general-purpose mobile GUI agents under realistic, user-centric conditions. VenusBench-Mobile builds two core evaluation pillars: defining what to evaluate via user-intent-driven task design that reflects real mobile usage, and how to evaluate through a capability-oriented annotation scheme for fine-grained agent behavior analysis. Extensive evaluation of state-of-the-art mobile GUI agents reveals large performance gaps relative to prior benchmarks, indicating that VenusBench-Mobile poses substantially more challenging and realistic tasks and that current agents remain far from reliable real-world deployment. Diagnostic analysis further shows that failures are dominated by deficiencies in perception and memory, which are largely obscured by coarse-grained evaluations. Moreover, even the strongest agents exhibit near-zero success under environment variations, highlighting their brittleness in realistic settings. Based on these insights, we believe VenusBench-Mobile provides an important stepping stone toward robust real-world deployment of mobile GUI agents. Code and data are available at https://github.com/inclusionAI/UI-Venus/tree/VenusBench-Mobile.

AI agents that interact with the real world through tool calls pose fundamental safety challenges: agents might leak private information, cause unintended side effects, or be manipulated through prompt injection. To address these challenges, we propose to put the agent in a programming-language-based "safety harness": instead of calling tools directly, agents express their intentions as code in a capability-safe language: Scala 3 with capture checking. Capabilities are program variables that regulate access to effects and resources of interest. Scala's type system tracks capabilities statically, providing fine-grained control over what an agent can do. In particular, it enables local purity, the ability to enforce that sub-computations are side-effect-free, preventing information leakage when agents process classified data. We demonstrate that extensible agent safety harnesses can be built by leveraging a strong type system with tracked capabilities. Our experiments show that agents can generate capability-safe code with no significant loss in task performance, while the type system reliably prevents unsafe behaviors such as information leakage and malicious side effects.

Rigorous and reproducible evaluation is critical for assessing the state of the art and for guiding scientific advances in Artificial Intelligence. Evaluation is challenging in practice due to several reasons, including benchmark saturation, lack of transparency in methods used for measurement, development challenges in extracting measurements for generative tasks, and, more generally, the extensive number of capabilities required for a well-rounded comparison across models. We make three contributions to alleviate the above challenges. First, we present Eureka, an open-source framework for standardizing evaluations of large foundation models beyond single-score reporting and rankings. Second, we introduce Eureka-Bench as an extensible collection of benchmarks testing capabilities that (i) are still challenging for state-of-the-art models and (ii) represent fundamental but overlooked language and multimodal capabilities. The inherent space for improvement in non-saturated benchmarks enables us to discover meaningful differences between models at a capability level. Third, using Eureka, we conduct an analysis of 12 state-of-the-art models, providing in-depth insights into failure understanding and model comparison, which can be leveraged to plan targeted improvements. In contrast to recent trends in reports and leaderboards showing absolute rankings and claims for one model or another to be the best, our analysis shows that there is no such best model. Different models have different strengths, but there are models that appear more often than others as best performers for some capabilities. Despite the recent improvements, current models still struggle with several fundamental capabilities including detailed image understanding, benefiting from multimodal input when available rather than fully relying on language, factuality and grounding for information retrieval, and over refusals.

Trustworthy capability evaluations are crucial for ensuring the safety of AI systems, and are becoming a key component of AI regulation. However, the developers of an AI system, or the AI system itself, may have incentives for evaluations to understate the AI's actual capability. These conflicting interests lead to the problem of sandbagging, which we define as strategic underperformance on an evaluation. In this paper we assess sandbagging capabilities in contemporary language models (LMs). We prompt frontier LMs, like GPT-4 and Claude 3 Opus, to selectively underperform on dangerous capability evaluations, while maintaining performance on general (harmless) capability evaluations. Moreover, we find that models can be fine-tuned, on a synthetic dataset, to hide specific capabilities unless given a password. This behaviour generalizes to high-quality, held-out benchmarks such as WMDP. In addition, we show that both frontier and smaller models can be prompted or password-locked to target specific scores on a capability evaluation. We have mediocre success in password-locking a model to mimic the answers a weaker model would give. Overall, our results suggest that capability evaluations are vulnerable to sandbagging. This vulnerability decreases the trustworthiness of evaluations, and thereby undermines important safety decisions regarding the development and deployment of advanced AI systems.

The ability to recognize one's own limitations and decide whether to solve a problem or delegate is fundamental for reliable intelligent systems. Yet we show that modern large language models systematically lack this ability: across diverse model families and scales, they overestimate their competence and attempt queries they cannot solve. We refer to this ability as Capability Self-Assessment (CSA) and formulate it as a policy-learning problem, aiming to improve self-assessment while preserving the model's original capabilities. Our results show that reinforcement learning teaches CSA effectively, significantly outperforming supervised fine-tuning while preserving original capabilities. In contrast, supervised fine-tuning severely degrades the capabilities the model is meant to assess. Moreover, learned self-assessment behavior generalizes well out of distribution, suggesting that CSA is a transferable model trait. Finally, CSA is practically useful: it improves local-cloud decision making at inference time and provides a signal for targeted data selection during training.

We present a scientific reasoning foundation model that aligns natural language with heterogeneous scientific representations. The model is pretrained on a 206B-token corpus spanning scientific text, pure sequences, and sequence-text pairs, then aligned via SFT on 40M instructions, annealed cold-start bootstrapping to elicit long-form chain-of-thought, and reinforcement learning with task-specific reward shaping, which instills deliberate scientific reasoning. It supports four capability families, covering up to 103 tasks across workflows: (i) faithful translation between text and scientific formats, (ii) text/knowledge extraction, (iii) property prediction, (iv) property classification, (v) unconditional and conditional sequence generation and design. Compared with specialist systems, our approach broadens instruction coverage, improves cross-domain generalization, and enhances fidelity. We detail data curation and training and show that cross-discipline learning strengthens transfer and downstream reliability. The model, instruct tuning datasets and the evaluation code are open-sourced at https://huggingface.co/SciReason and https://github.com/open-sciencelab/SciReason.

In recent years, with the rapid development of the depth and breadth of large language models' capabilities, various corresponding evaluation benchmarks have been emerging in increasing numbers. As a quantitative assessment tool for model performance, benchmarks are not only a core means to measure model capabilities but also a key element in guiding the direction of model development and promoting technological innovation. We systematically review the current status and development of large language model benchmarks for the first time, categorizing 283 representative benchmarks into three categories: general capabilities, domain-specific, and target-specific. General capability benchmarks cover aspects such as core linguistics, knowledge, and reasoning; domain-specific benchmarks focus on fields like natural sciences, humanities and social sciences, and engineering technology; target-specific benchmarks pay attention to risks, reliability, agents, etc. We point out that current benchmarks have problems such as inflated scores caused by data contamination, unfair evaluation due to cultural and linguistic biases, and lack of evaluation on process credibility and dynamic environments, and provide a referable design paradigm for future benchmark innovation.

Deep Research Agents generate analyst-grade reports, yet evaluating them remains challenging due to the absence of a single ground truth and the multidimensional nature of research quality. Recent benchmarks propose distinct methodologies, yet they suffer from the Mirage of Synthesis, where strong surface-level fluency and citation alignment can obscure underlying factual and reasoning defects. We characterize this gap by introducing a taxonomy across four verticals that exposes a critical capability mismatch: static evaluators inherently lack the tool-use capabilities required to assess temporal validity and factual correctness. To address this, we propose DREAM (Deep Research Evaluation with Agentic Metrics), a framework that instantiates the principle of capability parity by making evaluation itself agentic. DREAM structures assessment through an evaluation protocol combining query-agnostic metrics with adaptive metrics generated by a tool-calling agent, enabling temporally aware coverage, grounded verification, and systematic reasoning probes. Controlled evaluations demonstrate DREAM is significantly more sensitive to factual and temporal decay than existing benchmarks, offering a scalable, reference-free evaluation paradigm.

We present MEGA-Bench, an evaluation suite that scales multimodal evaluation to over 500 real-world tasks, to address the highly heterogeneous daily use cases of end users. Our objective is to optimize for a set of high-quality data samples that cover a highly diverse and rich set of multimodal tasks, while enabling cost-effective and accurate model evaluation. In particular, we collected 505 realistic tasks encompassing over 8,000 samples from 16 expert annotators to extensively cover the multimodal task space. Instead of unifying these problems into standard multi-choice questions (like MMMU, MMBench, and MMT-Bench), we embrace a wide range of output formats like numbers, phrases, code, \LaTeX, coordinates, JSON, free-form, etc. To accommodate these formats, we developed over 40 metrics to evaluate these tasks. Unlike existing benchmarks, MEGA-Bench offers a fine-grained capability report across multiple dimensions (e.g., application, input type, output format, skill), allowing users to interact with and visualize model capabilities in depth. We evaluate a wide variety of frontier vision-language models on MEGA-Bench to understand their capabilities across these dimensions.

Existing benchmarks measure capability -- whether a model succeeds on a single attempt -- but production deployments require reliability -- consistent success across repeated attempts on tasks of varying duration. We show these properties diverge systematically as task duration grows, and that pass@1 on short tasks is structurally blind to this divergence. We introduce a reliability science framework for long-horizon LLM agents with four metrics: Reliability Decay Curve (RDC), Variance Amplification Factor (VAF), Graceful Degradation Score (GDS), and Meltdown Onset Point (MOP). We evaluate 10 models across 23,392 episodes on a 396-task benchmark spanning four duration buckets and three domains. Key findings: (1) reliability decay is domain-stratified -- SE GDS drops from 0.90 to 0.44 while document processing is nearly flat (0.74 to 0.71); (2) VAF bifurcates by capability tier -- high VAF is a capability signature, not an instability signal; (3) capability and reliability rankings diverge substantially, with multi-rank inversions at long horizons; (4) frontier models have the highest meltdown rates (up to 19%) because they attempt ambitious multi-step strategies that sometimes spiral; and (5) memory scaffolds universally hurt long-horizon performance across all 10 models. These results motivate reliability as a first-class evaluation dimension alongside capability.

Despite the advancement of machine learning techniques in recent years, state-of-the-art systems lack robustness to "real world" events, where the input distributions and tasks encountered by the deployed systems will not be limited to the original training context, and systems will instead need to adapt to novel distributions and tasks while deployed. This critical gap may be addressed through the development of "Lifelong Learning" systems that are capable of 1) Continuous Learning, 2) Transfer and Adaptation, and 3) Scalability. Unfortunately, efforts to improve these capabilities are typically treated as distinct areas of research that are assessed independently, without regard to the impact of each separate capability on other aspects of the system. We instead propose a holistic approach, using a suite of metrics and an evaluation framework to assess Lifelong Learning in a principled way that is agnostic to specific domains or system techniques. Through five case studies, we show that this suite of metrics can inform the development of varied and complex Lifelong Learning systems. We highlight how the proposed suite of metrics quantifies performance trade-offs present during Lifelong Learning system development - both the widely discussed Stability-Plasticity dilemma and the newly proposed relationship between Sample Efficient and Robust Learning. Further, we make recommendations for the formulation and use of metrics to guide the continuing development of Lifelong Learning systems and assess their progress in the future.

Large Language Model agents increasingly operate external systems through programmatic interfaces, yet practitioners lack empirical guidance on how to structure the context these agents consume. Using SQL generation as a proxy for programmatic agent operations, we present a systematic study of context engineering for structured data, comprising 9,649 experiments across 11 models, 4 formats (YAML, Markdown, JSON, Token-Oriented Object Notation [TOON]), and schemas ranging from 10 to 10,000 tables. Our findings challenge common assumptions. First, architecture choice is model-dependent: file-based context retrieval improves accuracy for frontier-tier models (Claude, GPT, Gemini; +2.7%, p=0.029) but shows mixed results for open source models (aggregate -7.7%, p<0.001), with deficits varying substantially by model. Second, format does not significantly affect aggregate accuracy (chi-squared=2.45, p=0.484), though individual models, particularly open source, exhibit format-specific sensitivities. Third, model capability is the dominant factor, with a 21 percentage point accuracy gap between frontier and open source tiers that dwarfs any format or architecture effect. Fourth, file-native agents scale to 10,000 tables through domain-partitioned schemas while maintaining high navigation accuracy. Fifth, file size does not predict runtime efficiency: compact or novel formats can incur a token overhead driven by grep output density and pattern unfamiliarity, with the magnitude depending on model capability. These findings provide practitioners with evidence-based guidance for deploying LLM agents on structured systems, demonstrating that architectural decisions should be tailored to model capability rather than assuming universal best practices.

Large language models (LLMs) have demonstrated excellent capabilities in generating structured diagrams from natural language descriptions. In particular, they have shown great promise in generating sequence diagrams for software engineering, typically represented in a text-based syntax such as Mermaid. However, systematic evaluations in this space remain underdeveloped as there is a lack of existing benchmarks to assess the LLM's correctness in this task. To address this shortcoming, we introduce MermaidSeqBench, a human-verified and LLM-synthetically-extended benchmark for assessing an LLM's capabilities in generating Mermaid sequence diagrams from textual prompts. The benchmark consists of a core set of 132 samples, starting from a small set of manually crafted and verified flows. These were expanded via a hybrid methodology combining human annotation, in-context LLM prompting, and rule-based variation generation. Our benchmark uses an LLM-as-a-judge model to assess Mermaid sequence diagram generation across fine-grained metrics, including syntax correctness, activation handling, error handling, and practical usability. We perform initial evaluations on numerous state-of-the-art LLMs and utilize multiple LLM judge models to demonstrate the effectiveness and flexibility of our benchmark. Our results reveal significant capability gaps across models and evaluation modes. Our proposed benchmark provides a foundation for advancing research in structured diagram generation and for developing more rigorous, fine-grained evaluation methodologies.

The transition from monolithic language models to modular, skill-equipped agents marks a defining shift in how large language models (LLMs) are deployed in practice. Rather than encoding all procedural knowledge within model weights, agent skills -- composable packages of instructions, code, and resources that agents load on demand -- enable dynamic capability extension without retraining. It is formalized in a paradigm of progressive disclosure, portable skill definitions, and integration with the Model Context Protocol (MCP). This survey provides a comprehensive treatment of the agent skills landscape, as it has rapidly evolved during the last few months. We organize the field along four axes: (i) architectural foundations, examining the SKILL.md specification, progressive context loading, and the complementary roles of skills and MCP; (ii) skill acquisition, covering reinforcement learning with skill libraries, autonomous skill discovery (SEAgent), and compositional skill synthesis; (iii) deployment at scale, including the computer-use agent (CUA) stack, GUI grounding advances, and benchmark progress on OSWorld and SWE-bench; and (iv) security, where recent empirical analyses reveal that 26.1% of community-contributed skills contain vulnerabilities, motivating our proposed Skill Trust and Lifecycle Governance Framework -- a four-tier, gate-based permission model that maps skill provenance to graduated deployment capabilities. We identify seven open challenges -- from cross-platform skill portability to capability-based permission models -- and propose a research agenda for realizing trustworthy, self-improving skill ecosystems. Unlike prior surveys that broadly cover LLM agents or tool use, this work focuses specifically on the emerging skill abstraction layer and its implications for the next generation of agentic systems. Project repo: https://github.com/scienceaix/agentskills

Large Language Models (LLMs) deployed in agentic environments must exercise multiple capabilities across different task instances, where a capability is performing one or more actions in a trajectory that are necessary for successfully solving a subset of tasks in the environment. Many existing approaches either rely on synthetic training data that is not targeted to the model's actual capability deficits in the target environment or train directly on the target environment, where the model needs to implicitly learn the capabilities across tasks. We introduce TRACE (Turning Recurrent Agent failures into Capability-targeted training Environments), an end-to-end system for environment-specific agent self-improvement. TRACE contrasts successful and failed trajectories to automatically identify lacking capabilities, synthesizes a targeted training environment for each that rewards whether the capability was exercised, and trains a LoRA adapter via RL on each synthetic environment, routing to the relevant adapter at inference. Empirically, TRACE generalizes across different environments, improving over the base agent by +14.1 points on τ^2-bench (customer service) and +7 perfect scores on ToolSandbox (tool use), outperforming the strongest baseline by +7.4 points and +4 perfect scores, respectively. Given the same number of rollouts, TRACE scales more efficiently than baselines, outperforming GRPO and GEPA by +9.2 and +7.4 points on τ^2-bench.

Large Language Models (LLMs) are advancing at an amazing speed and have become indispensable across academia, industry, and daily applications. To keep pace with the status quo, this survey probes the core challenges that the rise of LLMs poses for evaluation. We identify and analyze two pivotal transitions: (i) from task-specific to capability-based evaluation, which reorganizes benchmarks around core competencies such as knowledge, reasoning, instruction following, multi-modal understanding, and safety; and (ii) from manual to automated evaluation, encompassing dynamic dataset curation and "LLM-as-a-judge" scoring. Yet, even with these transitions, a crucial obstacle persists: the evaluation generalization issue. Bounded test sets cannot scale alongside models whose abilities grow seemingly without limit. We will dissect this issue, along with the core challenges of the above two transitions, from the perspectives of methods, datasets, evaluators, and metrics. Due to the fast evolving of this field, we will maintain a living GitHub repository (links are in each section) to crowd-source updates and corrections, and warmly invite contributors and collaborators.

An ideal model evaluation should achieve two goals: identifying where the model fails and providing actionable improvement guidance. Toward these goals for Language Model (LM) evaluations, we formulate the problem of generating a weakness profile, a set of weaknesses expressed in natural language, given an LM's performance on every individual instance in a benchmark. We introduce a suite of quantitative assessments to compare different weakness profiling methods. We also propose a weakness profiling method EvalTree. It constructs a capability tree where each node represents a capability described in natural language and is linked to a subset of benchmark instances that specifically evaluate this capability; it then extracts nodes where the LM performs poorly to generate a weakness profile. On the MATH and WildChat benchmarks, we show that EvalTree outperforms baseline weakness profiling methods by identifying weaknesses more precisely and comprehensively. Weakness profiling further enables weakness-guided data collection, and training data collection guided by EvalTree-identified weaknesses improves LM performance more than other data collection strategies. We also show how EvalTree exposes flaws in Chatbot Arena's human-voter-based evaluation practice. To facilitate future work, we release our code and an interface that allows practitioners to interactively explore the capability trees built by EvalTree.

We present Federation of Agents (FoA), a distributed orchestration framework that transforms static multi-agent coordination into dynamic, capability-driven collaboration. FoA introduces Versioned Capability Vectors (VCVs): machine-readable profiles that make agent capabilities searchable through semantic embeddings, enabling agents to advertise their capabilities, cost, and limitations. Our aarchitecturecombines three key innovations: (1) semantic routing that matches tasks to agents over sharded HNSW indices while enforcing operational constraints through cost-biased optimization, (2) dynamic task decomposition where compatible agents collaboratively break down complex tasks into DAGs of subtasks through consensus-based merging, and (3) smart clustering that groups agents working on similar subtasks into collaborative channels for k-round refinement before synthesis. Built on top of MQTT,s publish-subscribe semantics for scalable message passing, FoA achieves sub-linear complexity through hierarchical capability matching and efficient index maintenance. Evaluation on HealthBench shows 13x improvements over single-model baselines, with clustering-enhanced laboration particularly effective for complex reasoning tasks requiring multiple perspectives. The system scales horizontally while maintaining consistent performance, demonstrating that semantic orchestration with structured collaboration can unlock the collective intelligence of heterogeneous federations of AI agents.

We investigate whether post-trained capabilities can be transferred across models without retraining, with a focus on transfer across different model scales. We propose the Master Key Hypothesis, which states that model capabilities correspond to directions in a low-dimensional latent subspace that induce specific behaviors and are transferable across models through linear alignment. Based on this hypothesis, we introduce UNLOCK, a training-free and label-free framework that extracts a capability direction by contrasting activations between capability-present and capability-absent Source variants, aligns it with a Target model through a low-rank linear transformation, and applies it at inference time to elicit the behavior. Experiments on reasoning behaviors, including Chain-of-Thought (CoT) and mathematical reasoning, demonstrate substantial improvements across model scales without training. For example, transferring CoT reasoning from Qwen1.5-14B to Qwen1.5-7B yields an accuracy gain of 12.1% on MATH, and transferring a mathematical reasoning direction from Qwen3-4B-Base to Qwen3-14B-Base improves AGIEval Math accuracy from 61.1% to 71.3%, surpassing the 67.8% achieved by the 14B post-trained model. Our analysis shows that the success of transfer depends on the capabilities learned during pre-training, and that our intervention amplifies latent capabilities by sharpening the output distribution toward successful reasoning trajectories.

Recent advances in large language models and tool-using agents have expanded the range of benchmarked web tasks. Yet an important class of specialized retrieval tasks remains undercharacterized. On many specialized data-retrieval websites, answer-bearing evidence becomes accessible only after establishing the correct site-specific retrieval state through filters, views, hierarchies, or scopes. We term this capability state-gated retrieval (SGR). We introduce SGR-Bench, a benchmark for this setting containing 100 expert-curated tasks spanning six source families and 12 public data ecosystems. Each task requires discovering the appropriate website and configuring its site-specific retrieval state to produce a structured answer. SGR-Bench pairs constraint-guided and goal-oriented formulations of the same underlying problems, enabling controlled comparisons between explicit and implicit guidance for state-gated retrieval. We evaluate eight CLI-based agentic LLM systems and three commercial search-agent products. On SGR-Bench, the strongest system reaches only 66.18% item-level F1, while row-level F1 remains much lower. A manual audit of 156 analyzable failed CLI trajectories shows why: agents often reach a relevant web source, but establish the wrong site-specific retrieval state. Retrieval-scope drift (37.2%) and criterion mismatch (27.6%) dominate, whereas final answer composition accounts for only 10.3%. The dataset and single-case evaluation instructions are available at https://huggingface.co/datasets/PKUAIWeb/SGR-BENCH.

Existing browser agent benchmarks face a fundamental trilemma: real-website benchmarks lack reproducibility due to content drift, controlled environments sacrifice realism by omitting real-web noise, and both require costly manual curation that limits scalability. We present WebForge, the first fully automated framework that resolves this trilemma through a four-agent pipeline -- Plan, Generate, Refine, and Validate -- that produces interactive, self-contained web environments end-to-end without human annotation. A seven-dimensional difficulty control framework structures task design along navigation depth, visual complexity, reasoning difficulty, and more, enabling systematic capability profiling beyond single aggregate scores. Using WebForge, we construct WebForge-Bench, a benchmark of 934 tasks spanning 7 domains and 3 difficulty levels. Multi-model experiments show that difficulty stratification effectively differentiates model capabilities, while cross-domain analysis exposes capability biases invisible to aggregate metrics. Together, these results confirm that multi-dimensional evaluation reveals distinct capability profiles that a single aggregate score cannot capture. Code and benchmark are publicly available at https://github.com/yuandaxia2001/WebForge.

As multimodal large language models (MLLMs) advance rapidly, rigorous evaluation has become essential, providing further guidance for their development. In this work, we focus on a unified and robust evaluation of vision perception abilities, the foundational skill of MLLMs. We find that existing perception benchmarks, each focusing on different question types, domains, and evaluation metrics, introduce significant evaluation variance, complicating comprehensive assessments of perception abilities when relying on any single benchmark. To address this, we introduce AbilityLens, a unified benchmark designed to evaluate MLLMs across six key perception abilities, focusing on both accuracy and stability, with each ability encompassing diverse question types, domains, and metrics. With the assistance of AbilityLens, we: (1) identify the strengths and weaknesses of current models, highlighting stability patterns and revealing a notable performance gap between open-source and closed-source models; (2) introduce an online evaluation mode, which uncovers interesting ability conflict and early convergence phenomena during MLLM training; and (3) design a simple ability-specific model merging method that combines the best ability checkpoint from early training stages, effectively mitigating performance decline due to ability conflict. The benchmark and online leaderboard will be released soon.

Large Language Models (LLMs) are increasingly used in empirical software engineering (ESE) to automate or assist annotation tasks such as labeling commits, issues, and qualitative artifacts. Yet the reliability and reproducibility of such annotations remain underexplored. Existing studies often lack standardized measures for reliability, calibration, and drift, and frequently omit essential configuration details. We argue that LLM-based annotation should be treated as a measurement process rather than a purely automated activity. In this position paper, we outline the Operationalization for LLM-based Annotation Framework (OLAF), a conceptual framework that organizes key constructs: reliability, calibration, drift, consensus, aggregation, and transparency. The paper aims to motivate methodological discussion and future empirical work toward more transparent and reproducible LLM-based annotation in software engineering research.

Autonomous language-model agents increasingly rely on installable skills and tools to complete user tasks. Static skill auditing can expose capability surface before deployment, but it cannot determine whether a particular invocation is unsafe under the current user request and runtime context. We therefore study skill invocation auditing as a continuous-risk estimation problem: given a user request, candidate skill, and runtime context, predict a score that supports ranking and triage before a hard intervention is applied. We introduce STARS, which combines a static capability prior, a request-conditioned invocation risk model, and a calibrated risk-fusion policy. To evaluate this setting, we construct SIA-Bench, a benchmark of 3,000 invocation records with group-safe splits, lineage metadata, runtime context, canonical action labels, and derived continuous-risk targets. On a held-out split of indirect prompt injection attacks, calibrated fusion reaches 0.439 high-risk AUPRC, improving over 0.405 for the contextual scorer and 0.380 for the strongest static baseline, while the contextual scorer remains better calibrated with 0.289 expected calibration error. On the locked in-distribution test split, gains are smaller and static priors remain useful. The resulting claim is therefore narrower: request-conditioned auditing is most valuable as an invocation-time risk-scoring and triage layer rather than as a replacement for static screening. Code is available at https://github.com/123zgj123/STARS.

Current language models (LMs) excel at reasoning over prompts using pre-trained knowledge. However, real-world tasks are far more complex and context-dependent: models must learn from task-specific context and leverage new knowledge beyond what is learned during pre-training to reason and resolve tasks. We term this capability context learning, a crucial ability that humans naturally possess but has been largely overlooked. To this end, we introduce CL-bench, a real-world benchmark consisting of 500 complex contexts, 1,899 tasks, and 31,607 verification rubrics, all crafted by experienced domain experts. Each task is designed such that the new content required to resolve it is contained within the corresponding context. Resolving tasks in CL-bench requires models to learn from the context, ranging from new domain-specific knowledge, rule systems, and complex procedures to laws derived from empirical data, all of which are absent from pre-training. This goes far beyond long-context tasks that primarily test retrieval or reading comprehension, and in-context learning tasks, where models learn simple task patterns via instructions and demonstrations. Our evaluations of ten frontier LMs find that models solve only 17.2% of tasks on average. Even the best-performing model, GPT-5.1, solves only 23.7%, revealing that LMs have yet to achieve effective context learning, which poses a critical bottleneck for tackling real-world, complex context-dependent tasks. CL-bench represents a step towards building LMs with this fundamental capability, making them more intelligent and advancing their deployment in real-world scenarios.

Recent advances in Large Language Models (LLMs) have sparked concerns over their potential to acquire and misuse dangerous or high-risk capabilities, posing frontier risks. Current safety evaluations primarily test for what a model can do - its capabilities - without assessing what it would do if endowed with high-risk capabilities. This leaves a critical blind spot: models may strategically conceal capabilities or rapidly acquire them, while harboring latent inclinations toward misuse. We argue that propensity - the likelihood of a model to pursue harmful actions if empowered - is a critical, yet underexplored, axis of safety evaluation. We present PropensityBench, a novel benchmark framework that assesses the proclivity of models to engage in risky behaviors when equipped with simulated dangerous capabilities using proxy tools. Our framework includes 5,874 scenarios with 6,648 tools spanning four high-risk domains: cybersecurity, self-proliferation, biosecurity, and chemical security. We simulate access to powerful capabilities via a controlled agentic environment and evaluate the models' choices under varying operational pressures that reflect real-world constraints or incentives models may encounter, such as resource scarcity or gaining more autonomy. Across open-source and proprietary frontier models, we uncover 9 alarming signs of propensity: models frequently choose high-risk tools when under pressure, despite lacking the capability to execute such actions unaided. These findings call for a shift from static capability audits toward dynamic propensity assessments as a prerequisite for deploying frontier AI systems safely. Our code is available at https://github.com/scaleapi/propensity-evaluation.

In large organisations, identifying experts on a given topic is crucial in leveraging the internal knowledge spread across teams and departments. So-called enterprise expert retrieval systems automatically discover and structure employees' expertise based on the vast amount of heterogeneous data available about them and the work they perform. Evaluating these systems requires comprehensive ground truth expert annotations, which are hard to obtain. Therefore, the annotation process typically relies on automated recommendations of knowledge areas to validate. This case study provides an analysis of how these recommendations can impact the evaluation of expert finding systems. We demonstrate on a popular benchmark that system-validated annotations lead to overestimated performance of traditional term-based retrieval models and even invalidate comparisons with more recent neural methods. We also augment knowledge areas with synonyms to uncover a strong bias towards literal mentions of their constituent words. Finally, we propose constraints to the annotation process to prevent these biased evaluations, and show that this still allows annotation suggestions of high utility. These findings should inform benchmark creation or selection for expert finding, to guarantee meaningful comparison of methods.

The performance on general tasks decreases after Large Language Models (LLMs) are fine-tuned on domain-specific tasks, the phenomenon is known as Catastrophic Forgetting (CF). However, this paper presents a further challenge for real application of domain-specific LLMs beyond CF, called General Capabilities Integration (GCI), which necessitates the integration of both the general capabilities and domain knowledge within a single instance. The objective of GCI is not merely to retain previously acquired general capabilities alongside new domain knowledge, but to harmonize and utilize both sets of skills in a cohesive manner to enhance performance on domain-specific tasks. Taking legal domain as an example, we carefully design three groups of training and testing tasks without lacking practicability, and construct the corresponding datasets. To better incorporate general capabilities across domain-specific scenarios, we introduce ALoRA, which utilizes a multi-head attention module upon LoRA, facilitating direct information transfer from preceding tokens to the current one. This enhancement permits the representation to dynamically switch between domain-specific knowledge and general competencies according to the attention. Extensive experiments are conducted on the proposed tasks. The results exhibit the significance of our setting, and the effectiveness of our method.

The NLP community has long advocated for the construction of multi-annotator datasets to better capture the nuances of language interpretation, subjectivity, and ambiguity. This paper conducts a retrospective study to show how performance scores can vary when a dataset expands from a single annotation per instance to multiple annotations. We propose a novel multi-annotator simulation process to generate datasets with varying annotation budgets. We show that similar datasets with the same annotation budget can lead to varying performance gains. Our findings challenge the popular belief that models trained on multi-annotation examples always lead to better performance than models trained on single or few-annotation examples.

We document empirical capability ceilings in decoder-only autoregressive language models across knowledge-intensive tasks. Systematic evaluation of OPT and Pythia model families (70M-30B parameters, spanning 240 times scaling) reveals that knowledge retrieval tasks show negligible accuracy improvement despite smooth loss reduction. On MMLU mathematics benchmarks, accuracy remains flat at 19-20% (below 25% random chance) across all scales while cross-entropy loss decreases by 31%. In contrast, procedural tasks like arithmetic show conventional scaling where both metrics improve together. Attention intervention experiments reveal high sensitivity to perturbation: swapping attention patterns between models causes catastrophic performance collapse (complete accuracy loss) rather than graceful degradation. These measurements have immediate engineering implications: for knowledge-intensive applications using OPT and Pythia architectures, parameter scaling beyond 1-2B offers minimal accuracy gains despite continued loss improvement. Our findings quantify capability-specific scaling failures in these model families to inform resource allocation decisions. Whether these patterns reflect fundamental constraints of decoder-only architectures or implementation-specific limitations remains an open question requiring investigation across diverse architectural approaches.

Geospatial Foundation Models (GeoFMs) are transforming Earth Observation (EO), but evaluation lacks standardized protocols. GEO-Bench-2 addresses this with a comprehensive framework spanning classification, segmentation, regression, object detection, and instance segmentation across 19 permissively-licensed datasets. We introduce ''capability'' groups to rank models on datasets that share common characteristics (e.g., resolution, bands, temporality). This enables users to identify which models excel in each capability and determine which areas need improvement in future work. To support both fair comparison and methodological innovation, we define a prescriptive yet flexible evaluation protocol. This not only ensures consistency in benchmarking but also facilitates research into model adaptation strategies, a key and open challenge in advancing GeoFMs for downstream tasks. Our experiments show that no single model dominates across all tasks, confirming the specificity of the choices made during architecture design and pretraining. While models pretrained on natural images (ConvNext ImageNet, DINO V3) excel on high-resolution tasks, EO-specific models (TerraMind, Prithvi, and Clay) outperform them on multispectral applications such as agriculture and disaster response. These findings demonstrate that optimal model choice depends on task requirements, data modalities, and constraints. This shows that the goal of a single GeoFM model that performs well across all tasks remains open for future research. GEO-Bench-2 enables informed, reproducible GeoFM evaluation tailored to specific use cases. Code, data, and leaderboard for GEO-Bench-2 are publicly released under a permissive license.

As multimodal large models (MLLMs) continue to advance across challenging tasks, a key question emerges: What essential capabilities are still missing? A critical aspect of human learning is continuous interaction with the environment -- not limited to language, but also involving multimodal understanding and generation. To move closer to human-level intelligence, models must similarly support multi-turn, multimodal interaction. In particular, they should comprehend interleaved multimodal contexts and respond coherently in ongoing exchanges. In this work, we present an initial exploration through the InterMT -- the first preference dataset for multi-turn multimodal interaction, grounded in real human feedback. In this exploration, we particularly emphasize the importance of human oversight, introducing expert annotations to guide the process, motivated by the fact that current MLLMs lack such complex interactive capabilities. InterMT captures human preferences at both global and local levels into nine sub-dimensions, consists of 15.6k prompts, 52.6k multi-turn dialogue instances, and 32.4k human-labeled preference pairs. To compensate for the lack of capability for multi-modal understanding and generation, we introduce an agentic workflow that leverages tool-augmented MLLMs to construct multi-turn QA instances. To further this goal, we introduce InterMT-Bench to assess the ability of MLLMs in assisting judges with multi-turn, multimodal tasks. We demonstrate the utility of \InterMT through applications such as judge moderation and further reveal the multi-turn scaling law of judge model. We hope the open-source of our data can help facilitate further research on aligning current MLLMs to the next step. Our project website can be found at https://pku-intermt.github.io .

Software that contains machine learning algorithms is an integral part of automotive perception, for example, in driving automation systems. The development of such software, specifically the training and validation of the machine learning components, require large annotated datasets. An industry of data and annotation services has emerged to serve the development of such data-intensive automotive software components. Wide-spread difficulties to specify data and annotation needs challenge collaborations between OEMs (Original Equipment Manufacturers) and their suppliers of software components, data, and annotations. This paper investigates the reasons for these difficulties for practitioners in the Swedish automotive industry to arrive at clear specifications for data and annotations. The results from an interview study show that a lack of effective metrics for data quality aspects, ambiguities in the way of working, unclear definitions of annotation quality, and deficits in the business ecosystems are causes for the difficulty in deriving the specifications. We provide a list of recommendations that can mitigate challenges when deriving specifications and we propose future research opportunities to overcome these challenges. Our work contributes towards the on-going research on accountability of machine learning as applied to complex software systems, especially for high-stake applications such as automated driving.

Existing large language models (LLMs) evaluation methods typically focus on testing the performance on some closed-environment and domain-specific benchmarks with human annotations. In this paper, we explore a novel unsupervised evaluation direction, utilizing peer-review mechanisms to measure LLMs automatically. In this setting, both open-source and closed-source LLMs lie in the same environment, capable of answering unlabeled questions and evaluating each other, where each LLM's response score is jointly determined by other anonymous ones. To obtain the ability hierarchy among these models, we assign each LLM a learnable capability parameter to adjust the final ranking. We formalize it as a constrained optimization problem, intending to maximize the consistency of each LLM's capabilities and scores. The key assumption behind is that high-level LLM can evaluate others' answers more accurately than low-level ones, while higher-level LLM can also achieve higher response scores. Moreover, we propose three metrics called PEN, CIN, and LIS to evaluate the gap in aligning human rankings. We perform experiments on multiple datasets with these metrics, validating the effectiveness of the proposed approach.

As large language models grow in capability and agency, identifying vulnerabilities through red-teaming becomes vital for safe deployment. However, traditional prompt-engineering approaches may prove ineffective once red-teaming turns into a weak-to-strong problem, where target models surpass red-teamers in capabilities. To study this shift, we frame red-teaming through the lens of the capability gap between attacker and target. We evaluate more than 500 attacker-target pairs using LLM-based jailbreak attacks that mimic human red-teamers across diverse families, sizes, and capability levels. Three strong trends emerge: (i) more capable models are better attackers, (ii) attack success drops sharply once the target's capability exceeds the attacker's, and (iii) attack success rates correlate with high performance on social science splits of the MMLU-Pro benchmark. From these trends, we derive a jailbreaking scaling law that predicts attack success for a fixed target based on attacker-target capability gap. These findings suggest that fixed-capability attackers (e.g., humans) may become ineffective against future models, increasingly capable open-source models amplify risks for existing systems, and model providers must accurately measure and control models' persuasive and manipulative abilities to limit their effectiveness as attackers.

Interestingly, LLMs yet struggle with some basic tasks that humans find trivial to handle, e.g., counting the number of character r's in the word "strawberry". There are several popular conjectures (e.g., tokenization, architecture and training data) regarding the reason for deficiency of LLMs in simple word-based counting problems, sharing the similar belief that such failure stems from model pretraining hence probably inevitable during deployment. In this paper, we carefully design multiple evaluation settings to investigate validity of prevalent conjectures. Meanwhile, we measure transferability of advanced mathematical and coding reasoning capabilities from specialized LLMs to simple counting tasks. Although specialized LLMs suffer from counting problems as well, we find conjectures about inherent deficiency of LLMs invalid and further seek opportunities to elicit knowledge and capabilities from LLMs that are beneficial to counting tasks. Compared with strategies such as finetuning and in-context learning that are commonly adopted to enhance performance on new or challenging tasks, we show that engaging reasoning is the most robust and efficient way to help LLMs better perceive tasks with more accurate responses. We hope our conjecture validation design could provide insights into the study of future critical failure modes of LLMs. Based on challenges in transferring advanced capabilities to much simpler tasks, we call for more attention to model capability acquisition and evaluation. We also highlight the importance of cultivating consciousness of "reasoning before responding" during model pretraining.

A central challenge in large language model (LLM) editing is capability preservation: methods that successfully change targeted behavior can quietly game the editing proxy and corrupt general capabilities, producing degenerate behaviors reminiscent of proxy/reward hacking. We present CrispEdit, a scalable and principled second-order editing algorithm that treats capability preservation as an explicit constraint, unifying and generalizing several existing editing approaches. CrispEdit formulates editing as constrained optimization and enforces the constraint by projecting edit updates onto the low-curvature subspace of the capability-loss landscape. At the crux of CrispEdit is expressing capability constraint via Bregman divergence, whose quadratic form yields the Gauss-Newton Hessian exactly and even when the base model is not trained to convergence. We make this second-order procedure efficient at the LLM scale using Kronecker-factored approximate curvature (K-FAC) and a novel matrix-free projector that exploits Kronecker structure to avoid constructing massive projection matrices. Across standard model-editing benchmarks, CrispEdit achieves high edit success while keeping capability degradation below 1% on average across datasets, significantly improving over prior editors.

Political scientists are rapidly adopting large language models (LLMs) for text annotation, yet the sensitivity of annotation results to implementation choices remains poorly understood. Most evaluations test a single model or configuration; how model choice, model size, learning approach, and prompt style interact, and whether popular "best practices" survive controlled comparison, are largely unexplored. We present a controlled evaluation of these pipeline choices, testing six open-weight models across four political science annotation tasks under identical quantisation, hardware, and prompt-template conditions. Our central finding is methodological: interaction effects dominate main effects, so seemingly reasonable pipeline choices can become consequential researcher degrees of freedom. No single model, prompt style, or learning approach is uniformly superior, and the best-performing model varies across tasks. Two corollaries follow. First, model size is an unreliable guide both to cost and to performance: cross-family efficiency differences are so large that some larger models are less resource-intensive than much smaller alternatives, while within model families mid-range variants often match or exceed larger counterparts. Second, widely recommended prompt engineering techniques yield inconsistent and sometimes negative effects on annotation performance. We use these benchmark results to develop a validation-first framework - with a principled ordering of pipeline decisions, guidance on prompt freezing and held-out evaluation, reporting standards, and open-source tools - to help researchers navigate this decision space transparently.

Recent advancements in large language models (LLMs) showcase varied multilingual capabilities across tasks like translation, code generation, and reasoning. Previous assessments often limited their scope to fundamental natural language processing (NLP) or isolated capability-specific tasks. To alleviate this drawback, we aim to present a comprehensive multilingual multitask benchmark. First, we present a pipeline for selecting available and reasonable benchmarks from massive ones, addressing the oversight in previous work regarding the utility of these benchmarks, i.e., their ability to differentiate between models being evaluated. Leveraging this pipeline, we introduce P-MMEval, a large-scale benchmark covering effective fundamental and capability-specialized datasets. Furthermore, P-MMEval delivers consistent language coverage across various datasets and provides parallel samples. Finally, we conduct extensive experiments on representative multilingual model series to compare performances across models, analyze dataset effectiveness, examine prompt impacts on model performances, and explore the relationship between multilingual performances and factors such as tasks, model sizes, and languages. These insights offer valuable guidance for future research. The dataset is available at https://huggingface.co/datasets/Qwen/P-MMEval.

Skill extraction and recommendation systems have been studied from recruiter, applicant, and education perspectives. While AI applications in job advertisements have received broad attention, deficiencies in the instructed skills side remain a challenge. In this work, we address the scarcity of publicly available datasets by releasing both manually annotated and synthetic datasets of skills from the European Skills, Competences, Qualifications and Occupations (ESCO) taxonomy and university course pairs and publishing corresponding annotation guidelines. Specifically, we match graduate-level university courses with skills from the Systems Analysts and Management and Organization Analyst ESCO occupation groups at two granularities: course title with a skill, and course sentence with a skill. We train language models on this dataset to serve as a baseline for retrieval and recommendation systems for course-to-skill and skill-to-course matching. We evaluate the models on a portion of the annotated data. Our BERT model achieves 87% F1-score, showing that course and skill matching is a feasible task.

With the rapid evolution of large language models (LLMs), new and hard-to-predict harmful capabilities are emerging. This requires developers to be able to identify risks through the evaluation of "dangerous capabilities" in order to responsibly deploy LLMs. In this work, we collect the first open-source dataset to evaluate safeguards in LLMs, and deploy safer open-source LLMs at a low cost. Our dataset is curated and filtered to consist only of instructions that responsible language models should not follow. We annotate and assess the responses of six popular LLMs to these instructions. Based on our annotation, we proceed to train several BERT-like classifiers, and find that these small classifiers can achieve results that are comparable with GPT-4 on automatic safety evaluation. Warning: this paper contains example data that may be offensive, harmful, or biased.

Given that Large Language Models (LLMs) have made significant progress in writing code, can they now be used to autonomously reproduce results from research repositories? Such a capability would be a boon to the research community, helping researchers validate, understand, and extend prior work. To advance towards this goal, we introduce SUPER, the first benchmark designed to evaluate the capability of LLMs in setting up and executing tasks from research repositories. SUPERaims to capture the realistic challenges faced by researchers working with Machine Learning (ML) and Natural Language Processing (NLP) research repositories. Our benchmark comprises three distinct problem sets: 45 end-to-end problems with annotated expert solutions, 152 sub problems derived from the expert set that focus on specific challenges (e.g., configuring a trainer), and 602 automatically generated problems for larger-scale development. We introduce various evaluation measures to assess both task success and progress, utilizing gold solutions when available or approximations otherwise. We show that state-of-the-art approaches struggle to solve these problems with the best model (GPT-4o) solving only 16.3% of the end-to-end set, and 46.1% of the scenarios. This illustrates the challenge of this task, and suggests that SUPER can serve as a valuable resource for the community to make and measure progress.

Query clustering organizes queries into groups that reflect shared latent capability demands, enabling capability-aware LLM evaluation. Existing clustering methods, which primarily rely on semantic taxonomies or embeddings, often fail to capture such latent capability requirements due to a misalignment between surface-level semantics and actual model performance. We propose ECC, an algorithm that calibrates prior semantic embeddings using limited posterior model comparisons to bridge the gap between surface-level semantics and latent capability requirements. ECC characterizes each cluster through a capability profile parameterized by a Bradley-Terry model and uses trainable mixture weights to accommodate queries with mixed capability demands, jointly learning a flexible, capability-aware clustering structure that supports query-specific inference of LLM capabilities. Extensive quantitative and qualitative evaluations demonstrate that ECC significantly improves LLM capability ranking quality, outperforming human-labeled and embedding-based baselines by an average of 17.64 and 18.02 percentage points, respectively, and proves effective in downstream tasks such as query routing.

Large Language Model (LLM) agents frameworks often employ modular architectures, incorporating components such as planning, reasoning, action execution, and reflection to tackle complex tasks. However, quantifying the contribution of each module to overall system performance remains a significant challenge, impeding optimization and interpretability. To address this, we introduce CapaBench (Capability-level Assessment Benchmark), an evaluation framework grounded in cooperative game theory's Shapley Value, which systematically measures the marginal impact of individual modules and their interactions within an agent's architecture. By replacing default modules with test variants across all possible combinations, CapaBench provides a principle method for attributing performance contributions. Key contributions include: (1) We are the first to propose a Shapley Value-based methodology for quantifying the contributions of capabilities in LLM agents; (2) Modules with high Shapley Values consistently lead to predictable performance gains when combined, enabling targeted optimization; and (3) We build a multi-round dataset of over 1,500 entries spanning diverse domains and practical task scenarios, enabling comprehensive evaluation of agent capabilities. CapaBench bridges the gap between component-level evaluation and holistic system assessment, providing actionable insights for optimizing modular LLM agents and advancing their deployment in complex, real-world scenarios.

Long context capability is a crucial competency for large language models (LLMs) as it mitigates the human struggle to digest long-form texts. This capability enables complex task-solving scenarios such as book summarization, code assistance, and many more tasks that are traditionally manpower-intensive. However, transformer-based LLMs face significant challenges with long context input due to the growing size of the KV cache and the intrinsic complexity of attending to extended inputs; where multiple schools of efficiency-driven approaches -- such as KV cache quantization, token dropping, prompt compression, linear-time sequence models, and hybrid architectures -- have been proposed to produce efficient yet long context-capable models. Despite these advancements, no existing work has comprehensively benchmarked these methods in a reasonably aligned environment. In this work, we fill this gap by providing a taxonomy of current methods and evaluating 10+ state-of-the-art approaches across seven categories of long context tasks. Our work reveals numerous previously unknown phenomena and offers insights -- as well as a friendly workbench -- for the future development of long context-capable LLMs. The source code will be available at https://github.com/henryzhongsc/longctx_bench

With the rapid advancement of Large Language Models (LLMs), long-context information understanding and processing have become a hot topic in academia and industry. However, benchmarks for evaluating the ability of LLMs to handle long-context information do not seem to have kept pace with the development of LLMs. Despite the emergence of various long-context evaluation benchmarks, the types of capability assessed are still limited, without new capability dimensions. In this paper, we introduce DetectiveQA, a narrative reasoning benchmark featured with an average context length of over 100K tokens. DetectiveQA focuses on evaluating the long-context reasoning ability of LLMs, which not only requires a full understanding of context but also requires extracting important evidences from the context and reasoning according to extracted evidences to answer the given questions. This is a new dimension of capability evaluation, which is more in line with the current intelligence level of LLMs. We use detective novels as data sources, which naturally have various reasoning elements. Finally, we manually annotated 600 questions in Chinese and then also provided an English edition of the context information and questions. We evaluate many long-context LLMs on DetectiveQA, including commercial and open-sourced models, and the results indicate that existing long-context LLMs still require significant advancements to effectively process true long-context dependency questions.

Despite the power of Large Language Models (LLMs) like GPT-4, they still struggle with tasks that require generating complex, structured outputs. In this study, we assess the capability of Current LLMs in generating complex structured data and propose a structure-aware fine-tuning approach as a solution to improve this ability. To perform a comprehensive evaluation, we propose Struc-Bench, include five representative LLMs (i.e., GPT-NeoX 20B, GPT-3.5, GPT-4, and Vicuna) and evaluate them on our carefully constructed datasets spanning raw text, HTML, and LaTeX tables. Based on our analysis of current model performance, we identify specific common formatting errors and areas of potential improvement. To address complex formatting requirements, we utilize FormatCoT (Chain-of-Thought) to generate format instructions from target outputs. Our experiments show that our structure-aware fine-tuning method, when applied to LLaMA-7B, significantly improves adherence to natural language constraints, outperforming other evaluated LLMs. Based on these results, we present an ability map of model capabilities from six dimensions (i.e., coverage, formatting, reasoning, comprehension, pragmatics, and hallucination). This map highlights the weaknesses of LLMs in handling complex structured outputs and suggests promising directions for future work. Our code and models can be found at https://github.com/gersteinlab/Struc-Bench.

Annotated data plays a critical role in Natural Language Processing (NLP) in training models and evaluating their performance. Given recent developments in Large Language Models (LLMs), models such as ChatGPT demonstrate zero-shot capability on many text-annotation tasks, comparable with or even exceeding human annotators. Such LLMs can serve as alternatives for manual annotation, due to lower costs and higher scalability. However, limited work has leveraged LLMs as complementary annotators, nor explored how annotation work is best allocated among humans and LLMs to achieve both quality and cost objectives. We propose CoAnnotating, a novel paradigm for Human-LLM co-annotation of unstructured texts at scale. Under this framework, we utilize uncertainty to estimate LLMs' annotation capability. Our empirical study shows CoAnnotating to be an effective means to allocate work from results on different datasets, with up to 21% performance improvement over random baseline. For code implementation, see https://github.com/SALT-NLP/CoAnnotating.

As Large Language Model (LLM) alignment evolves from simple completions to complex, highly sophisticated generation, Reward Models are increasingly shifting toward rubric-guided evaluation to mitigate surface-level biases. However, the community lacks a unified benchmark to assess this evaluation paradigm, as existing benchmarks lack both the discriminative complexity and the ground-truth rubric annotations required for rigorous analysis. To bridge this gap, we introduce RubricBench, a curated benchmark with 1,147 pairwise comparisons specifically designed to assess the reliability of rubric-based evaluation. Our construction employs a multi-dimensional filtration pipeline to target hard samples featuring nuanced input complexity and misleading surface bias, augmenting each with expert-annotated, atomic rubrics derived strictly from instructions. Comprehensive experiments reveal a substantial capability gap between human-annotated and model-generated rubrics, indicating that even state-of-the-art models struggle to autonomously specify valid evaluation criteria, lagging considerably behind human-guided performance.

The rapid development of large language models and multimodal large language models has accelerated the emergence of proactive agents capable of operating everyday tools and assisting users in real-world environments. However, existing benchmarks struggle to evaluate such agents effectively, as they often rely on sandboxed environments and single-turn evaluation paradigms. Moreover, their scenario-based task taxonomies mix multiple model capabilities within the same task category, making it difficult to identify the root causes of agent failures. To address these limitations, we introduce UniClawBench, the first capability-driven benchmark designed to evaluate proactive agents in dynamic, real-world settings. UniClawBench is built around five foundational model capabilities: Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, and Cross-Platform Coordination. Based on these capabilities, we design 400 bilingual real-world tasks. Unlike previous benchmarks that rely on static, pre-recorded answers, our benchmark evaluates agents in live Docker containers using fine-grained, step-by-step completion checkpoints. Furthermore, we design a closed-loop evaluation strategy comprising an executor agent, a hidden supervisor agent, and a user agent to simulate realistic multi-turn human feedback without leaking grading criteria. To disentangle base model capabilities from framework-level design choices, we evaluate state-of-the-art models under multiple agent frameworks. Through comprehensive comparisons across both models and frameworks, we show how base model capabilities and agent framework designs jointly shape performance in real-world environments. To facilitate future research, we make our benchmark and code publicly available at https://github.com/HKU-MMLab/UniClawBench.

Retrieval models are often evaluated on partially-annotated datasets. Each query is mapped to a few relevant texts and the remaining corpus is assumed to be irrelevant. As a result, models that successfully retrieve false negatives are punished in evaluation. Unfortunately, completely annotating all texts for every query is not resource efficient. In this work, we show that using partially-annotated datasets in evaluation can paint a distorted picture. We curate D-MERIT, a passage retrieval evaluation set from Wikipedia, aspiring to contain all relevant passages for each query. Queries describe a group (e.g., ``journals about linguistics'') and relevant passages are evidence that entities belong to the group (e.g., a passage indicating that Language is a journal about linguistics). We show that evaluating on a dataset containing annotations for only a subset of the relevant passages might result in misleading ranking of the retrieval systems and that as more relevant texts are included in the evaluation set, the rankings converge. We propose our dataset as a resource for evaluation and our study as a recommendation for balance between resource-efficiency and reliable evaluation when annotating evaluation sets for text retrieval.

As AI-driven document understanding and processing tools become increasingly prevalent in real-world applications, the need for rigorous evaluation standards has grown increasingly urgent. Existing benchmarks and evaluations often focus on isolated capabilities or simplified scenarios, failing to capture the end-to-end task effectiveness required in practical settings. To address this gap, we introduce AIDABench, a comprehensive benchmark for evaluating AI systems on complex data analytics tasks in an end-to-end manner. AIDABench encompasses 600+ diverse document analysis tasks across three core capability dimensions: question answering, data visualization, and file generation. These tasks are grounded in realistic scenarios involving heterogeneous data types, including spreadsheets, databases, financial reports, and operational records, and reflect analytical demands across diverse industries and job functions. Notably, the tasks in AIDABench are sufficiently challenging that even human experts require 1-2 hours per question when assisted by AI tools, underscoring the benchmark's difficulty and real-world complexity. We evaluate 11 state-of-the-art models on AIDABench, spanning both proprietary (e.g., Claude Sonnet 4.5, Gemini 3 Pro Preview) and open-source (e.g., Qwen3-Max-2026-01-23-Thinking) families. Our results reveal that complex, real-world data analytics tasks remain a significant challenge for current AI systems, with the best-performing model achieving only 59.43% pass-at-1. We provide a detailed analysis of failure modes across each capability dimension and identify key challenges for future research. AIDABench offers a principled reference for enterprise procurement, tool selection, and model optimization, and is publicly available at https://github.com/MichaelYang-lyx/AIDABench.

This paper explores the use of open generative Large Language Models (LLMs) for annotation tasks in the social sciences. The study highlights the challenges associated with proprietary models, such as limited reproducibility and privacy concerns, and advocates for the adoption of open (source) models that can be operated on independent devices. Two examples of annotation tasks, sentiment analysis in tweets and identification of leisure activities in childhood aspirational essays are provided. The study evaluates the performance of different prompting strategies and models (neural-chat-7b-v3-2, Starling-LM-7B-alpha, openchat_3.5, zephyr-7b-alpha and zephyr-7b-beta). The results indicate the need for careful validation and tailored prompt engineering. The study highlights the advantages of open models for data privacy and reproducibility.

Large language model (LLM)-based AI agents extend LLM capabilities by enabling access to tools such as data sources, APIs, search engines, code sandboxes, and even other agents. While this empowers agents to perform complex tasks, LLMs may invoke unintended tool interactions and introduce risks, such as leaking sensitive data or overwriting critical records, which are unacceptable in enterprise contexts. Current approaches to mitigate these risks, such as model-based safeguards, enhance agents' reliability but cannot guarantee system safety. Methods like information flow control (IFC) and temporal constraints aim to provide guarantees but often require extensive human annotation. We propose a process that starts with applying System-Theoretic Process Analysis (STPA) to identify hazards in agent workflows, derive safety requirements, and formalize them as enforceable specifications on data flows and tool sequences. To enable this, we introduce a capability-enhanced Model Context Protocol (MCP) framework that requires structured labels on capabilities, confidentiality, and trust level. Together, these contributions aim to shift LLM-based agent safety from ad hoc reliability fixes to proactive guardrails with formal guarantees, while reducing dependence on user confirmation and making autonomy a deliberate design choice.

If AI models can detect when they are being evaluated, the effectiveness of evaluations might be compromised. For example, models could have systematically different behavior during evaluations, leading to less reliable benchmarks for deployment and governance decisions. We investigate whether frontier language models can accurately classify transcripts based on whether they originate from evaluations or real-world deployment, a capability we call evaluation awareness. To achieve this, we construct a diverse benchmark of 1,000 prompts and transcripts from 61 distinct datasets. These span public benchmarks (e.g., MMLU, SWEBench), real-world deployment interactions, and agent trajectories from scaffolding frameworks (e.g., web-browsing agents). Frontier models clearly demonstrate above-random evaluation awareness (Gemini-2.5-Pro reaches an AUC of 0.83), but do not yet surpass our simple human baseline (AUC of 0.92). Furthermore, both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings. Additionally, we test whether models can identify the purpose of the evaluation. Under multiple-choice and open-ended questioning, AI models far outperform random chance in identifying what an evaluation is testing for. Our results indicate that frontier models already exhibit a substantial, though not yet superhuman, level of evaluation-awareness. We recommend tracking this capability in future models.

Large language models (LLMs) are primarily evaluated by overall performance on various text understanding and generation tasks. However, such a paradigm fails to comprehensively differentiate the fine-grained language and cognitive skills, rendering the lack of sufficient interpretation to LLMs' capabilities. In this paper, we present FAC^2E, a framework for Fine-grAined and Cognition-grounded LLMs' Capability Evaluation. Specifically, we formulate LLMs' evaluation in a multi-dimensional and explainable manner by dissociating the language-related capabilities and the cognition-related ones. Besides, through extracting the intermediate reasoning from LLMs, we further break down the process of applying a specific capability into three sub-steps: recalling relevant knowledge, utilizing knowledge, and solving problems. Finally, FAC^2E evaluates each sub-step of each fine-grained capability, providing a two-faceted diagnosis for LLMs. Utilizing FAC^2E, we identify a common shortfall in knowledge utilization among models and propose a straightforward, knowledge-enhanced method to mitigate this issue. Our results not only showcase promising performance enhancements but also highlight a direction for future LLM advancements.

Reward models (RMs) are at the crux of successful RLHF to align pretrained models to human preferences, yet there has been relatively little study that focuses on evaluation of those reward models. Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models and which values are embedded in them. To date, very few descriptors of capabilities, training methods, or open-source reward models exist. In this paper, we present RewardBench, a benchmark dataset and code-base for evaluation, to enhance scientific understanding of reward models. The RewardBench dataset is a collection of prompt-win-lose trios spanning chat, reasoning, and safety, to benchmark how reward models perform on challenging, structured and out-of-distribution queries. We created specific comparison datasets for RMs that have subtle, but verifiable reasons (e.g. bugs, incorrect facts) why one answer should be preferred to another. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods, such as the direct MLE training of classifiers and the implicit reward modeling of Direct Preference Optimization (DPO), and on a spectrum of datasets. We present many findings on propensity for refusals, reasoning limitations, and instruction following shortcomings of various reward models towards a better understanding of the RLHF process.

We release Llamion, a family of 14B-parameter open-weight language models obtained by transforming Orion-14B into the standardized Llama-family architecture. The transformation is performed by Efficient Knowledge Preservation for Transformation (KEPT), a recipe that combines (i) Normal Parameter Mapping (NPM) for unchanged modules, (ii) Optimized Parameter Mapping (OPM), a training-free LayerNorm-to-RMSNorm initialization we prove optimal under the near-zero-mean activation regime induced by weight decay, and (iii) Cross-architecture Knowledge Distillation (XKD), an equal-size frozen-teacher distillation that aligns the converted model's outputs with the source model's on any reasonable input distribution. Llamion recovers Orion's behaviour on H6, MT-Bench, and KoMMLU with only ~123M tokens on a single A100 in four days; Llamion-Base reaches 66.87% on KoMMLU, exceeding the next-best entry of the Open Ko LLM Leaderboard by >7.0 absolute points at submission time. Capabilities entirely absent from the transfer corpus (Python programming and 200K-token context handling) survive the architectural transition intact. We release three checkpoints (Base, Chat, LongChat) that load with trust_remote_code=False in the Hugging Face Transformers library.

Deep research -- producing comprehensive, citation-grounded reports by searching and synthesizing information from hundreds of live web sources -- marks an important frontier for agentic systems. To rigorously evaluate this ability, four principles are essential: tasks should be (1) user-centric, reflecting realistic information needs, (2) dynamic, requiring up-to-date information beyond parametric knowledge, (3) unambiguous, ensuring consistent interpretation across users, and (4) multi-faceted and search-intensive, requiring search over numerous web sources and in-depth analysis. Existing benchmarks fall short of these principles, often focusing on narrow domains or posing ambiguous questions that hinder fair comparison. Guided by these principles, we introduce LiveResearchBench, a benchmark of 100 expert-curated tasks spanning daily life, enterprise, and academia, each requiring extensive, dynamic, real-time web search and synthesis. Built with over 1,500 hours of human labor, LiveResearchBench provides a rigorous basis for systematic evaluation. To evaluate citation-grounded long-form reports, we introduce DeepEval, a comprehensive suite covering both content- and report-level quality, including coverage, presentation, citation accuracy and association, consistency and depth of analysis. DeepEval integrates four complementary evaluation protocols, each designed to ensure stable assessment and high agreement with human judgments. Using LiveResearchBench and DeepEval, we conduct a comprehensive evaluation of 17 frontier deep research systems, including single-agent web search, single-agent deep research, and multi-agent systems. Our analysis reveals current strengths, recurring failure modes, and key system components needed to advance reliable, insightful deep research.

The performance of the reward model (RM) is a critical factor in improving the effectiveness of the large language model (LLM) during alignment fine-tuning. There remain two challenges in RM training: 1) training the same RM using various categories of data may cause its generalization performance to suffer from multi-task disturbance, and 2) the human annotation consistency rate is generally only 60% to 75%, causing training data to contain a lot of noise. To tackle these two challenges, we introduced the idea of Mixture-of-Experts (MoE) into the field of RM for the first time. We propose the Double-Layer MoE RM (DMoERM). The outer layer MoE is a sparse model. After classifying an input into task categories, we route it to the corresponding inner layer task-specific model. The inner layer MoE is a dense model. We decompose the specific task into multiple capability dimensions and individually fine-tune a LoRA expert on each one. Their outputs are then synthesized by an MLP to compute the final rewards. To minimize costs, we call a public LLM API to obtain the capability preference labels. The validation on manually labeled datasets confirms that our model attains superior consistency with human preference and outstrips advanced generative approaches. Meanwhile, through BoN sampling and RL experiments, we demonstrate that our model outperforms state-of-the-art ensemble methods of RM and mitigates the overoptimization problem. Our code and dataset are available at: https://github.com/quanshr/DMoERM-v1.

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. Due to the difficulty of obtaining high-quality human preference annotations, distilling preferences from generative LLMs has emerged as a standard practice. However, existing approaches predominantly treat teacher models as simple binary annotators, failing to fully exploit the rich knowledge and capabilities for RM distillation. To address this, we propose RM-Distiller, a framework designed to systematically exploit the multifaceted capabilities of teacher LLMs: (1) Refinement capability, which synthesizes highly correlated response pairs to create fine-grained and contrastive signals. (2) Scoring capability, which guides the RM in capturing precise preference strength via a margin-aware optimization objective. (3) Generation capability, which incorporates the teacher's generative distribution to regularize the RM to preserve its fundamental linguistic knowledge. Extensive experiments demonstrate that RM-Distiller significantly outperforms traditional distillation methods both on RM benchmarks and reinforcement learning-based alignment, proving that exploiting multifaceted teacher capabilities is critical for effective reward modeling. To the best of our knowledge, this is the first systematic research on RM distillation from generative LLMs.

This paper presents a new challenging information extraction task in the domain of materials science. We develop an annotation scheme for marking information on experiments related to solid oxide fuel cells in scientific publications, such as involved materials and measurement conditions. With this paper, we publish our annotation guidelines, as well as our SOFC-Exp corpus consisting of 45 open-access scholarly articles annotated by domain experts. A corpus and an inter-annotator agreement study demonstrate the complexity of the suggested named entity recognition and slot filling tasks as well as high annotation quality. We also present strong neural-network based models for a variety of tasks that can be addressed on the basis of our new data set. On all tasks, using BERT embeddings leads to large performance gains, but with increasing task complexity, adding a recurrent neural network on top seems beneficial. Our models will serve as competitive baselines in future work, and analysis of their performance highlights difficult cases when modeling the data and suggests promising research directions.

What makes a good Large Language Model (LLM)? That it performs well on the relevant benchmarks -- which hopefully measure, with some validity, the presence of capabilities that are also challenged in real application. But what makes the model perform well? What gives a model its abilities? We take a recently introduced type of benchmark that is meant to challenge capabilities in a goal-directed, agentive context through self-play of conversational games, and analyse how performance develops as a function of model characteristics like number of parameters, or type of training. We find that while there is a clear relationship between number of parameters and performance, there is still a wide spread of performance points within a given size bracket, which is to be accounted for by training parameters such as fine-tuning data quality and method. From a more practical angle, we also find a certain degree of unpredictability about performance across access methods, possible due to unexposed sampling parameters, and a, very welcome, performance stability against at least moderate weight quantisation during inference.

Long-context capability is considered one of the most important abilities of LLMs, as a truly long context-capable LLM enables users to effortlessly process many originally exhausting tasks -- e.g., digesting a long-form document to find answers vs. directly asking an LLM about it. However, existing real-task-based long-context evaluation benchmarks have two major shortcomings. First, benchmarks like LongBench often do not provide proper metrics to separate long-context performance from the model's baseline ability, making cross-model comparison unclear. Second, such benchmarks are usually constructed with fixed input lengths, which limits their applicability across different models and fails to reveal when a model begins to break down. To address these issues, we introduce a length-controllable long-context benchmark and a novel metric that disentangles baseline knowledge from true long-context capabilities. Experiments demonstrate the superiority of our approach in effectively evaluating LLMs.

Recent advances have made long-form report-generating systems widely available. This has prompted evaluation frameworks that use LLM-as-judge protocols and claim verification, along with meta-evaluation frameworks that seek to validate these methods. Many of the meta-evaluations estimate an evaluation quality's by comparing its assessments against human pairwise preferences. Prior work, however, suggests that human pairwise preference may be overly simplistic and can fail to capture nuances of expert expectations. We conduct a case study in meta-evaluation for long-form QA benchmarks using ScholarQA-CS2, a benchmark designed for assessing retrieval-augmented deep-research QA in the scientific domain. We comprehensively validate the benchmark through human pairwise preference judgments, then critically examine the strengths, weaknesses, and confounders of this approach. We show that pairwise preference rankings are best suited for system-level evaluation, while explicit metric-wise annotations and expert annotators are critical for reliable metric-level assessment, with subjectivity remaining a key challenge. Based on our findings, we offer practical guidelines for designing future meta-evaluations that better align evaluation methods, annotator expertise, and reporting practices. By surfacing these methodological challenges, we aim to advance evaluation standards for deep-research systems.

Large Language Models (LLMs) can correct their self-generated responses, but a decline in accuracy after self-correction is also witnessed. To have a deeper understanding of self-correction, we endeavor to decompose, evaluate, and analyze the self-correction behaviors of LLMs. By enumerating and analyzing answer correctness before and after self-correction, we decompose the self-correction capability into confidence (being confident to correct answers) and critique (turning wrong answers to correct) capabilities, and propose two metrics from a probabilistic perspective to measure these 2 capabilities, along with another metric for overall self-correction capability evaluation. Based on our decomposition and evaluation metrics, we conduct extensive experiments and draw some empirical conclusions. For example, we find different models can exhibit distinct behaviors: some models are confident while others are more critical. We also find the trade-off between the two capabilities (i.e. improving one can lead to a decline in the other) when manipulating model self-correction behavior by prompts or in-context learning. Further, we find a simple yet efficient strategy to improve self-correction capability by transforming Supervision Fine-Tuning (SFT) data format, and our strategy outperforms vanilla SFT in both capabilities and achieves much higher accuracy after self-correction. Our code will be publicly available on GitHub.

We present EBench, a simulation benchmark that diagnoses generalist mobile manipulation policies beyond a single success-rate scalar. EBench comprises 26 diverse and challenging manipulation tasks annotated along 5 capability dimensions and 4 generalization dimensions. We evaluate state-of-the-art generalist manipulation models including π_0, π_{0.5}, XVLA, and InternVLA-A1, and reveal that models with near success rates exhibit strikingly different capability profiles: π_{0.5} achieves the highest test success rate and the best train--test retention, whereas InternVLA-A1 dominates mobile manipulation but collapses on dexterous tasks, and XVLA exhibits strengths on a disjoint set of atomic skills compared to other policies. Beyond capability profiling, EBench analyzes the generalization ability from 4 representative perspectives, identifying the impact of different distribution shift factors. The results reveal strengths and weaknesses of models behind an overall score. We hope this benchmark offers a broad set of diagnostic signals to guide iteration on generalist manipulation models.

In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.

Evaluations are limited by benchmark availability. As models evolve, there is a need to create benchmarks that can measure progress on new generative capabilities. However, creating new benchmarks through human annotations is slow and expensive, restricting comprehensive evaluations for any capability. We introduce BENCHAGENTS, a framework that methodically leverages large language models (LLMs) to automate benchmark creation for complex capabilities while inherently ensuring data and metric quality. BENCHAGENTS decomposes the benchmark creation process into planning, generation, data verification, and evaluation, each of which is executed by an LLM agent. These agents interact with each other and utilize human-in-the-loop feedback from benchmark developers to explicitly improve and flexibly control data diversity and quality. We use BENCHAGENTS to create benchmarks to evaluate capabilities related to planning and constraint satisfaction during text generation. We then use these benchmarks to study seven state-of-the-art models and extract new insights on common failure modes and model differences.

Improvements in language model capabilities are often attributed to increasing model size or training data, but in some cases smaller models trained on curated data or with different architectural decisions can outperform larger ones trained on more tokens. What accounts for this? To quantify the impact of these design choices, we meta-analyze 92 open-source pretrained models across a wide array of scales, including state-of-the-art open-weights models as well as less performant models and those with less conventional design decisions. We find that by incorporating features besides model size and number of training tokens, we can achieve a relative 3-28% increase in ability to predict downstream performance compared with using scale alone. Analysis of model design decisions reveal insights into data composition, such as the trade-off between language and code tasks at 15-25\% code, as well as the better performance of some architectural decisions such as choosing rotary over learned embeddings. Broadly, our framework lays a foundation for more systematic investigation of how model development choices shape final capabilities.

Recognizing vulnerability is crucial for understanding and implementing targeted support to empower individuals in need. This is especially important at the European Court of Human Rights (ECtHR), where the court adapts Convention standards to meet actual individual needs and thus ensures effective human rights protection. However, the concept of vulnerability remains elusive at the ECtHR and no prior NLP research has dealt with it. To enable future research in this area, we present VECHR, a novel expert-annotated multi-label dataset comprising of vulnerability type classification and explanation rationale. We benchmark the performance of state-of-the-art models on VECHR from both prediction and explainability perspectives. Our results demonstrate the challenging nature of the task with lower prediction performance and limited agreement between models and experts. Further, we analyze the robustness of these models in dealing with out-of-domain (OOD) data and observe overall limited performance. Our dataset poses unique challenges offering significant room for improvement regarding performance, explainability, and robustness.

The ability to use, understand, and create tools is a hallmark of human intelligence, enabling sophisticated interaction with the physical world. For any general-purpose intelligent agent to achieve true versatility, it must also master these fundamental skills. While modern Multimodal Large Language Models (MLLMs) leverage their extensive common knowledge for high-level planning in embodied AI and in downstream Vision-Language-Action (VLA) models, the extent of their true understanding of physical tools remains unquantified. To bridge this gap, we present PhysToolBench, the first benchmark dedicated to evaluating the comprehension of physical tools by MLLMs. Our benchmark is structured as a Visual Question Answering (VQA) dataset comprising over 1,000 image-text pairs. It assesses capabilities across three distinct difficulty levels: (1) Tool Recognition: Requiring the recognition of a tool's primary function. (2) Tool Understanding: Testing the ability to grasp the underlying principles of a tool's operation. (3) Tool Creation: Challenging the model to fashion a new tool from surrounding objects when conventional options are unavailable. Our comprehensive evaluation of 32 MLLMs-spanning proprietary, open-source, specialized embodied, and backbones in VLAs-reveals a significant deficiency in tool understanding. Furthermore, we provide an in-depth analysis and propose preliminary solutions. Code and dataset are publicly available.

Peer review in grant evaluation informs funding decisions, but the contents of peer review reports are rarely analyzed. In this work, we develop a thoroughly tested pipeline to analyze the texts of grant peer review reports using methods from applied Natural Language Processing (NLP) and machine learning. We start by developing twelve categories reflecting content of grant peer review reports that are of interest to research funders. This is followed by multiple human annotators' iterative annotation of these categories in a novel text corpus of grant peer review reports submitted to the Swiss National Science Foundation. After validating the human annotation, we use the annotated texts to fine-tune pre-trained transformer models to classify these categories at scale, while conducting several robustness and validation checks. Our results show that many categories can be reliably identified by human annotators and machine learning approaches. However, the choice of text classification approach considerably influences the classification performance. We also find a high correspondence between out-of-sample classification performance and human annotators' perceived difficulty in identifying categories. Our results and publicly available fine-tuned transformer models will allow researchers and research funders and anybody interested in peer review to examine and report on the contents of these reports in a structured manner. Ultimately, we hope our approach can contribute to ensuring the quality and trustworthiness of grant peer review.

Tool-use capability is a fundamental component of LLM agents, enabling them to interact with external systems through structured function calls. However, existing research exhibits inconsistent interaction representations, largely overlooks the structural distribution of tool-use trajectories, and relies on incompatible evaluation benchmarks. We present UniToolCall, a unified framework for tool learning that standardizes the entire pipeline from toolset construction and dataset generation to evaluation. The framework curates a large tool pool of 22k+ tools and constructs a hybrid training corpus of 390k+ instances by combining 10 standardized public datasets with structurally controlled synthetic trajectories. It explicitly models diverse interaction patterns, including single-hop vs. multi-hop and single-turn vs. multi-turn, while capturing both serial and parallel execution structures. To support coherent multi-turn reasoning, we further introduce an Anchor Linkage mechanism that enforces cross-turn dependencies. Furthermore, we convert 7 public benchmarks into a unified Query--Action--Observation--Answer (QAOA) representation with fine-grained evaluation at the function-call, turn, and conversation levels. Experiments show that fine-tuning Qwen3-8B on our dataset substantially improves tool-use performance. Under the distractor-heavy Hybrid-20 setting, achieves 93.0% single-turn Strict Precision, outperforming commercial models including GPT, Gemini, and Claude.

Matching submissions with suitable reviewers at scale is a growing challenge for major venues, yet existing approaches either rely on coarse proxy signals that conflate general relatedness with true suitability, or require expensive human annotations that are difficult to scale for training. We propose MERIT, a two-stage framework that bridges this gap by converting criterion-level expertise matching into scalable suitability supervision. In the first stage, we train a reviewer assessor via reinforcement learning to identify the expertise dimensions a paper requires, match them against the reviewer's prior work, and produce a suitability decision, with rewards provided by an LLM judge guided by paper-specific expertise rubrics. In the second stage, we distill the assessor's predictions into an embedding-based retriever for efficient large-scale assignment. Experiments show that our 4B reviewer assessor outperforms larger general-purpose LLMs on suitability classification, and the resulting retriever achieves state-of-the-art performance across LR-Bench and the CMU Gold dataset. Our code is available at https://github.com/Luli3220/MERIT.

Embodied capabilities refer to a suite of fundamental abilities for an agent to perceive, comprehend, and interact with the physical world. While multimodal large language models (MLLMs) show promise as embodied agents, a thorough and systematic evaluation of their embodied capabilities remains underexplored, as existing benchmarks primarily focus on specific domains such as planning or spatial understanding. To bridge this gap, we introduce BEAR, a comprehensive and fine-grained benchmark that evaluates MLLMs on atomic embodied capabilities. BEAR comprises 4,469 interleaved image-video-text entries across 14 domains in 6 categories, including tasks from low-level pointing, trajectory understanding, spatial reasoning, to high-level planning. Extensive evaluation results of 20 representative MLLMs reveal their persistent limitations across all domains of embodied capabilities. To tackle the shortfall, we propose BEAR-Agent, a multimodal conversable agent that integrates pretrained vision models to strengthen MLLM perception, 3D understanding, and planning capabilities. It substantially enhances MLLM performance across diverse embodied capabilities on BEAR, yielding a 9.12% absolute gain and a relative improvement of 17.5% on GPT-5. Furthermore, our experiments indicate that improving MLLM embodied capabilities can benefit embodied tasks in simulated environments. Project website: https://bear-official66.github.io/

Many specialized domains remain untouched by deep learning, as large labeled datasets require expensive expert annotators. We address this bottleneck within the legal domain by introducing the Contract Understanding Atticus Dataset (CUAD), a new dataset for legal contract review. CUAD was created with dozens of legal experts from The Atticus Project and consists of over 13,000 annotations. The task is to highlight salient portions of a contract that are important for a human to review. We find that Transformer models have nascent performance, but that this performance is strongly influenced by model design and training dataset size. Despite these promising results, there is still substantial room for improvement. As one of the only large, specialized NLP benchmarks annotated by experts, CUAD can serve as a challenging research benchmark for the broader NLP community.

We present a corpus of 5,000 richly annotated abstracts of medical articles describing clinical randomized controlled trials. Annotations include demarcations of text spans that describe the Patient population enrolled, the Interventions studied and to what they were Compared, and the Outcomes measured (the `PICO' elements). These spans are further annotated at a more granular level, e.g., individual interventions within them are marked and mapped onto a structured medical vocabulary. We acquired annotations from a diverse set of workers with varying levels of expertise and cost. We describe our data collection process and the corpus itself in detail. We then outline a set of challenging NLP tasks that would aid searching of the medical literature and the practice of evidence-based medicine.

When training data are collected from human annotators, the design of the annotation instrument, the instructions given to annotators, the characteristics of the annotators, and their interactions can impact training data. This study demonstrates that design choices made when creating an annotation instrument also impact the models trained on the resulting annotations. We introduce the term annotation sensitivity to refer to the impact of annotation data collection methods on the annotations themselves and on downstream model performance and predictions. We collect annotations of hate speech and offensive language in five experimental conditions of an annotation instrument, randomly assigning annotators to conditions. We then fine-tune BERT models on each of the five resulting datasets and evaluate model performance on a holdout portion of each condition. We find considerable differences between the conditions for 1) the share of hate speech/offensive language annotations, 2) model performance, 3) model predictions, and 4) model learning curves. Our results emphasize the crucial role played by the annotation instrument which has received little attention in the machine learning literature. We call for additional research into how and why the instrument impacts the annotations to inform the development of best practices in instrument design.

This comprehensive study evaluates the performance of OpenAI's o1-preview large language model across a diverse array of complex reasoning tasks, spanning multiple domains, including computer science, mathematics, natural sciences, medicine, linguistics, and social sciences. Through rigorous testing, o1-preview demonstrated remarkable capabilities, often achieving human-level or superior performance in areas ranging from coding challenges to scientific reasoning and from language processing to creative problem-solving. Key findings include: -83.3% success rate in solving complex competitive programming problems, surpassing many human experts. -Superior ability in generating coherent and accurate radiology reports, outperforming other evaluated models. -100% accuracy in high school-level mathematical reasoning tasks, providing detailed step-by-step solutions. -Advanced natural language inference capabilities across general and specialized domains like medicine. -Impressive performance in chip design tasks, outperforming specialized models in areas such as EDA script generation and bug analysis. -Remarkable proficiency in anthropology and geology, demonstrating deep understanding and reasoning in these specialized fields. -Strong capabilities in quantitative investing. O1 has comprehensive financial knowledge and statistical modeling skills. -Effective performance in social media analysis, including sentiment analysis and emotion recognition. The model excelled particularly in tasks requiring intricate reasoning and knowledge integration across various fields. While some limitations were observed, including occasional errors on simpler problems and challenges with certain highly specialized concepts, the overall results indicate significant progress towards artificial general intelligence.

To determine the safety of large language models (LLMs), AI developers must be able to assess their dangerous capabilities. But simple prompting strategies often fail to elicit an LLM's full capabilities. One way to elicit capabilities more robustly is to fine-tune the LLM to complete the task. In this paper, we investigate the conditions under which fine-tuning-based elicitation suffices to elicit capabilities. To do this, we introduce password-locked models, LLMs fine-tuned such that some of their capabilities are deliberately hidden. Specifically, these LLMs are trained to exhibit these capabilities only when a password is present in the prompt, and to imitate a much weaker LLM otherwise. Password-locked models enable a novel method of evaluating capabilities elicitation methods, by testing whether these password-locked capabilities can be elicited without using the password. We find that a few high-quality demonstrations are often sufficient to fully elicit password-locked capabilities. More surprisingly, fine-tuning can elicit other capabilities that have been locked using the same password, or even different passwords. Furthermore, when only evaluations, and not demonstrations, are available, approaches like reinforcement learning are still often able to elicit capabilities. Overall, our findings suggest that fine-tuning is an effective method of eliciting hidden capabilities of current models, but may be unreliable when high-quality demonstrations are not available, e.g. as may be the case when models' (hidden) capabilities exceed those of human demonstrators.

The ability of Large Language Models (LLMs) to use external tools unlocks powerful real-world interactions, making rigorous evaluation essential. However, current benchmarks primarily report final accuracy, revealing what models can do but obscuring the cognitive bottlenecks that define their true capability boundaries. To move from simple performance scoring to a diagnostic tool, we introduce a framework grounded in Cognitive Load Theory. Our framework deconstructs task complexity into two quantifiable components: Intrinsic Load, the inherent structural complexity of the solution path, formalized with a novel Tool Interaction Graph; and Extraneous Load, the difficulty arising from ambiguous task presentation. To enable controlled experiments, we construct ToolLoad-Bench, the first benchmark with parametrically adjustable cognitive load. Our evaluation reveals distinct performance cliffs as cognitive load increases, allowing us to precisely map each model's capability boundary. We validate that our framework's predictions are highly calibrated with empirical results, establishing a principled methodology for understanding an agent's limits and a practical foundation for building more efficient systems.

Reliably transferring specialized human knowledge from text into large language models remains a fundamental challenge in artificial intelligence. Fine-tuning on domain corpora has enabled substantial capability gains, but the process operates without feedback: when a model fails on a domain task, there is no method to diagnose what is deficient in the training data, and the only recourse is to add more data indiscriminately. Here we show that when a structured knowledge representation extracted from the source corpus serves as the shared foundation for both training data and evaluation, the complete data-engineering lifecycle maps onto the software development lifecycle in a precise and operative way: training data becomes source code specifying what the model should learn, model training becomes compilation, benchmarking becomes unit testing, and failure-driven data repair becomes debugging. Under this correspondence, model failures decompose into concept-level gaps and reasoning-chain breaks that can be traced back to specific deficiencies in the data and repaired through targeted patches, with each repair cycle producing consistent improvements across model scales and architectures without degrading general capabilities. We formalize this principle as Programming with Data and instantiate it across sixteen disciplines spanning the natural sciences, engineering, biomedicine, and the social sciences, releasing a structured knowledge base, benchmark suite, and training corpus as open resources. By demonstrating that the relationship between training data and model behaviour is structurally traceable and systematically repairable, this work establishes a principled foundation for the reliable engineering of human expertise into language models.

Agentic coding assistants are increasingly given extra capabilities, such as browser based testing tools and design oriented system prompts, on the assumption that more capability yields better software. This study tested that assumption directly. Ninety independent agent runs built the same application, a real time retrospective board, from one detailed specification, each scored on a fixed 14 criterion functional rubric (42 point maximum) and a visual quality review. The runs spanned several model generations, two agent harnesses, two reasoning effort levels, a testing tool, and two design oriented prompts. Capability tier dominated: frontier models clustered near the ceiling while a low cost local model fell to 24 to 37 points. A criterion level analysis revealed what run totals conceal. Container deployment was the dominant defect, failing first try in 44 percent of runs, with its failure rate shifting sharply across model generations while mean totals moved less than a point. The testing tool raised cost by 42 to 68 percent without improving functional score or reliability, even on interface visible criteria. Raising reasoning effort from High to xHigh lifted first try perfect runs from 28 percent to 89 percent and cut corrective prompts about five fold, for 9 to 29 percent more cost. A design oriented prompt raised visual quality, 4.5 versus 3.0 on a 5 point scale, without lifting function, and a one paragraph paraphrase of its directive reproduced the entire lift. The practical lesson is to match the fix to the failure: most first run failures came from weak reasoning, which a stronger model or more effort prevents, not from visible flaws a checking tool would catch.

Large language models store all learned knowledge in a single, fixed weight vector. Teaching a model new capabilities requires modifying those same weights, inevitably degrading previously acquired knowledge. This fundamental limitation, known as catastrophic forgetting, has resisted principled solutions for decades. Existing approaches treat weights as immutable artifacts that must be protected through techniques like regularization heuristics, replay buffers, or isolated adapter modules. The problem is none of these provide a structural guarantee against forgetting. In this work, we propose Non-Interfering Weight Fields (NIWF), a framework that replaces the fixed weight paradigm with a learned function that generates weight configurations on demand from a continuous capability coordinate space. After training on a task, we commit the occupied coordinate region by snapshotting the fields outputs on anchor points to enforce a functional lock during all future training. We validate NIWF on sequential instructionfollowing and code generation tasks using Mistral-7B, demonstrating zero forgetting on committed tasks with competitive perplexity on new tasks. The framework introduces the notion of software-like versioning for neural network intelligence, where capabilities can be committed, extended, composed, and rolled back without retraining.

We present our submission to the LLM track of the 2026 Computational Models of Reference, Anaphora and Coreference (CRAC 2026) shared task. With an average CoNLL F1 score of 74.32 on the official test set, our system ranked first in the LLM track, and third overall. Our system is based on the Gemma-3-27b model, fine-tuned using a two-stage strategy with a multilingual base adapter followed by dataset-specific adapters. We represent mention spans by their headword using an XML-inspired format with local reindexing and annotate documents iteratively. These design choices proved effective across languages, document lengths, and annotation guidelines.