Benchmarking AI Agents for Addressing Scientific Challenges Across Scales

AI agents are increasingly being developed to accelerate scientific discovery, yet their practical capabilities in real research settings remain poorly understood. Existing benchmarks for AI agents rarely capture the complexity, heterogeneity, and extended reasoning required by scientific work, whereas benchmarks for scientific tasks often reduce research to static, direct problems and provide limited support for interactive evaluation. Here, we introduce SciAgentArena, a systematic benchmark for evaluating AI agents in real-world scientific research scenarios drawn from emerging needs across multiple domains. SciAgentArena comprises approximately 200 tasks with stepwise verification and an interactive, agent-agnostic environment for assessing diverse AI agents. Using this benchmark, we find that current agents can contribute effectively to well-specified data-analysis workflows, particularly when the task structure and evaluation criteria are clear. However, their performance remains uneven across scientific contexts: agents struggle to generate genuinely novel insights, sustain self-directed exploration, and formulate robust solutions for open-ended research questions. We further characterize common failure modes across agents and identify opportunities for improving their reliability, autonomy, and scientific reasoning. Together, SciAgentArena provides a practical framework for measuring progress in AI agents for science and for guiding the design of future agents capable of addressing complex scientific challenges. Full codes, tasks, and datasets can be accessed via this link: https://sciagentarena.github.io/.

LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding

Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.

Rethinking VLM Representation for VLA Initialization

Vision-Language-Action (VLA) models widely adopt pretrained Vision-Language Models (VLMs) as policy backbones, yet it remains unclear what kind of pretrained VLM representation is useful as a VLA initialization. In this paper, we study VLA initialization as a controlled representation-design problem along three axes: capability-level embodied VQA supervision, parameter-update strategy, and robot-data pretraining. Our experiments show that the original pretrained VLM representation is a key source of action performance. However, embodied VQA adaptation does not yield uniform gains: its benefit depends on downstream bottlenecks, and gains from different capability domains are not simply additive. For update strategy, LoRA provides a more reliable initialization than Full Finetune, indicating that overly reshaping the pretrained representation can weaken VLA initialization. Robot-data pretraining further improves VLA initialization, with the strongest variant obtained by staged LoRA-based training. Together, these findings suggest that effective VLM-to-VLA adaptation should inject action-relevant embodied and robot-trajectory signals while preserving the pretrained VLM representation that remains useful for action learning.

UniRef-Image-Edit: Towards Scalable and Consistent Multi-Reference Image Editing

We present UniRef-Image-Edit, a high-performance multi-modal generation system that unifies single-image editing and multi-image composition within a single framework. Existing diffusion-based editing methods often struggle to maintain consistency across multiple conditions due to limited interaction between reference inputs. To address this, we introduce Sequence-Extended Latent Fusion (SELF), a unified input representation that dynamically serializes multiple reference images into a coherent latent sequence. During a dedicated training stage, all reference images are jointly constrained to fit within a fixed-length sequence under a global pixel-budget constraint. Building upon SELF, we propose a two-stage training framework comprising supervised fine-tuning (SFT) and reinforcement learning (RL). In the SFT stage, we jointly train on single-image editing and multi-image composition tasks to establish a robust generative prior. We adopt a progressive sequence length training strategy, in which all input images are initially resized to a total pixel budget of $1024^2$, and are then gradually increased to $1536^2$ and $2048^2$ to improve visual fidelity and cross-reference consistency. This gradual relaxation of compression enables the model to incrementally capture finer visual details while maintaining stable alignment across references. For the RL stage, we introduce Multi-Source GRPO (MSGRPO), to our knowledge the first reinforcement learning framework tailored for multi-reference image generation. MSGRPO optimizes the model to reconcile conflicting visual constraints, significantly enhancing compositional consistency. We will open-source the code, models, training data, and reward data for community research purposes.

GENIUS: Generative Fluid Intelligence Evaluation Suite

Unified Multimodal Models (UMMs) have shown remarkable progress in visual generation. Yet, existing benchmarks predominantly assess $\textit{Crystallized Intelligence}$, which relies on recalling accumulated knowledge and learned schemas. This focus overlooks $\textit{Generative Fluid Intelligence (GFI)}$: the capacity to induce patterns, reason through constraints, and adapt to novel scenarios on the fly. To rigorously assess this capability, we introduce $\textbf{GENIUS}$ ($\textbf{GEN}$ Fluid $\textbf{I}$ntelligence Eval$\textbf{U}$ation $\textbf{S}$uite). We formalize $\textit{GFI}$ as a synthesis of three primitives. These include $\textit{Inducing Implicit Patterns}$ (e.g., inferring personalized visual preferences), $\textit{Executing Ad-hoc Constraints}$ (e.g., visualizing abstract metaphors), and $\textit{Adapting to Contextual Knowledge}$ (e.g., simulating counter-intuitive physics). Collectively, these primitives challenge models to solve problems grounded entirely in the immediate context. Our systematic evaluation of 12 representative models reveals significant performance deficits in these tasks. Crucially, our diagnostic analysis disentangles these failure modes. It demonstrates that deficits stem from limited context comprehension rather than insufficient intrinsic generative capability. To bridge this gap, we propose a training-free attention intervention strategy. Ultimately, $\textbf{GENIUS}$ establishes a rigorous standard for $\textit{GFI}$, guiding the field beyond knowledge utilization toward dynamic, general-purpose reasoning. Our dataset and code will be released at: $\href{https://github.com/arctanxarc/GENIUS}{https://github.com/arctanxarc/GENIUS}$.

MICo-150K: A Comprehensive Dataset Advancing Multi-Image Composition

In controllable image generation, synthesizing coherent and consistent images from multiple reference inputs, i.e., Multi-Image Composition (MICo), remains a challenging problem, partly hindered by the lack of high-quality training data. To bridge this gap, we conduct a systematic study of MICo, categorizing it into 7 representative tasks and curate a large-scale collection of high-quality source images and construct diverse MICo prompts. Leveraging powerful proprietary models, we synthesize a rich amount of balanced composite images, followed by human-in-the-loop filtering and refinement, resulting in MICo-150K, a comprehensive dataset for MICo with identity consistency. We further build a Decomposition-and-Recomposition (De&Re) subset, where 11K real-world complex images are decomposed into components and recomposed, enabling both real and synthetic compositions. To enable comprehensive evaluation, we construct MICo-Bench with 100 cases per task and 300 challenging De&Re cases, and further introduce a new metric, Weighted-Ref-VIEScore, specifically tailored for MICo evaluation. Finally, we fine-tune multiple models on MICo-150K and evaluate them on MICo-Bench. The results show that MICo-150K effectively equips models without MICo capability and further enhances those with existing skills. Notably, our baseline model, Qwen-MICo, fine-tuned from Qwen-Image-Edit, matches Qwen-Image-2509 in 3-image composition while supporting arbitrary multi-image inputs beyond the latter's limitation. Our dataset, benchmark, and baseline collectively offer valuable resources for further research on Multi-Image Composition.

Retrieval Feedback Memory Enhancement Large Model Retrieval Generation Method

Large Language Models (LLMs) have shown remarkable capabilities across diverse tasks, yet they face inherent limitations such as constrained parametric knowledge and high retraining costs. Retrieval-Augmented Generation (RAG) augments the generation process by retrieving externally stored knowledge absent from the models internal parameters. However, RAG methods face challenges such as information loss and redundant retrievals during multi-round queries, accompanying the difficulties in precisely characterizing knowledge gaps for complex tasks. To address these problems, we propose Retrieval Feedback and Memory Retrieval Augmented Generation(RFM-RAG), which transforms the stateless retrieval of previous methods into stateful continuous knowledge management by constructing a dynamic evidence pool. Specifically, our method generates refined queries describing the models knowledge gaps using relational triples from questions and evidence from the dynamic evidence pool; Retrieves critical external knowledge to iteratively update this evidence pool; Employs a R-Feedback Model to evaluate evidence completeness until convergence. Compared to traditional RAG methods, our approach enables persistent storage of retrieved passages and effectively distills key information from passages to construct clearly new queries. Experiments on three public QA benchmarks demonstrate that RFM-RAG outperforms previous methods and improves overall system accuracy.

CEIDM: A Controlled Entity and Interaction Diffusion Model for Enhanced Text-to-Image Generation

In Text-to-Image (T2I) generation, the complexity of entities and their intricate interactions pose a significant challenge for T2I method based on diffusion model: how to effectively control entity and their interactions to produce high-quality images. To address this, we propose CEIDM, a image generation method based on diffusion model with dual controls for entity and interaction. First, we propose an entity interactive relationships mining approach based on Large Language Models (LLMs), extracting reasonable and rich implicit interactive relationships through chain of thought to guide diffusion models to generate high-quality images that are closer to realistic logic and have more reasonable interactive relationships. Furthermore, We propose an interactive action clustering and offset method to cluster and offset the interactive action features contained in each text prompts. By constructing global and local bidirectional offsets, we enhance semantic understanding and detail supplementation of original actions, making the model's understanding of the concept of interactive "actions" more accurate and generating images with more accurate interactive actions. Finally, we design an entity control network which generates masks with entity semantic guidance, then leveraging multi-scale convolutional network to enhance entity feature and dynamic network to fuse feature. It effectively controls entities and significantly improves image quality. Experiments show that the proposed CEIDM method is better than the most representative existing methods in both entity control and their interaction control.

Dynamic Embedding of Hierarchical Visual Features for Efficient Vision-Language Fine-Tuning

Large Vision-Language Models (LVLMs) commonly follow a paradigm that projects visual features and then concatenates them with text tokens to form a unified sequence input for Large Language Models (LLMs). However, this paradigm leads to a significant increase in the length of the input sequence, resulting in substantial computational overhead. Existing methods attempt to fuse visual information into the intermediate layers of LLMs, which alleviate the sequence length issue but often neglect the hierarchical semantic representations within the model and the fine-grained visual information available in the shallower visual encoding layers. To address this limitation, we propose DEHVF, an efficient vision-language fine-tuning method based on dynamic embedding and fusion of hierarchical visual features. Its core lies in leveraging the inherent hierarchical representation characteristics of visual encoders and language models. Through a lightweight hierarchical visual fuser, it dynamically selects and fuses hierarchical features corresponding to semantic granularity based on the internal representations of each layer in LLMs. The fused layer-related visual features are then projected and aligned before being directly embedded into the Feed-Forward Network (FFN) of the corresponding layer in LLMs. This approach not only avoids sequence expansion but also dynamically fuses multi-layer visual information. By fine-tuning only a small number of parameters, DEHVF achieves precise alignment and complementarity of cross-modal information at the same semantic granularity. We conducted experiments across various VL benchmarks, including visual question answering on ScienceQA and image captioning on COCO Captions. The results demonstrate that DEHVF achieves higher accuracy than existing parameter-efficient fine-tuning (PEFT) baselines while maintaining efficient training and inference.

Separation and Collaboration: Two-Level Routing Grouped Mixture-of-Experts for Multi-Domain Continual Learning

Multi-Domain Continual Learning (MDCL) acquires knowledge from sequential tasks with shifting class sets and distribution. Despite the Parameter-Efficient Fine-Tuning (PEFT) methods can adapt for this dual heterogeneity, they still suffer from catastrophic forgetting and forward forgetting. To address these challenges, we propose a Two-Level Routing Grouped Mixture-of-Experts (TRGE) method. Firstly, TRGE dynamically expands the pre-trained CLIP model, assigning specific expert group for each task to mitigate catastrophic forgetting. With the number of experts continually grows in this process, TRGE maintains the static experts count within the group and introduces the intra-group router to alleviate routing overfitting caused by the increasing routing complexity. Meanwhile, we design an inter-group routing policy based on task identifiers and task prototype distance, which dynamically selects relevant expert groups and combines their outputs to enhance inter-task collaboration. Secondly, to get the correct task identifiers, we leverage Multimodal Large Language Models (MLLMs) which own powerful multimodal comprehension capabilities to generate semantic task descriptions and recognize the correct task identifier. Finally, to mitigate forward forgetting, we dynamically fuse outputs for unseen samples from the frozen CLIP model and TRGE adapter based on training progress, leveraging both pre-trained and learned knowledge. Through extensive experiments across various settings, our method outperforms other advanced methods with fewer trainable parameters.