Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering

The advancement of artificial intelligence toward agentic science is currently bottlenecked by the challenge of ultra-long-horizon autonomy, the ability to sustain strategic coherence and iterative correction over experimental cycles spanning days or weeks. While Large Language Models (LLMs) have demonstrated prowess in short-horizon reasoning, they are easily overwhelmed by execution details in the high-dimensional, delayed-feedback environments of real-world research, failing to consolidate sparse feedback into coherent long-term guidance. Here, we present ML-Master 2.0, an autonomous agent that masters ultra-long-horizon machine learning engineering (MLE) which is a representative microcosm of scientific discovery. By reframing context management as a process of cognitive accumulation, our approach introduces Hierarchical Cognitive Caching (HCC), a multi-tiered architecture inspired by computer systems that enables the structural differentiation of experience over time. By dynamically distilling transient execution traces into stable knowledge and cross-task wisdom, HCC allows agents to decouple immediate execution from long-term experimental strategy, effectively overcoming the scaling limits of static context windows. In evaluations on OpenAI's MLE-Bench under 24-hour budgets, ML-Master 2.0 achieves a state-of-the-art medal rate of 56.44%. Our findings demonstrate that ultra-long-horizon autonomy provides a scalable blueprint for AI capable of autonomous exploration beyond human-precedent complexities.

Source-Free Domain Adaptation for Geospatial Point Cloud Semantic Segmentation

Semantic segmentation of 3D geospatial point clouds is pivotal for remote sensing applications. However, variations in geographic patterns across regions and data acquisition strategies induce significant domain shifts, severely degrading the performance of deployed models. Existing domain adaptation methods typically rely on access to source-domain data. However, this requirement is rarely met due to data privacy concerns, regulatory policies, and data transmission limitations. This motivates the largely underexplored setting of source-free unsupervised domain adaptation (SFUDA), where only a pretrained model and unlabeled target-domain data are available. In this paper, we propose LoGo (Local-Global Dual-Consensus), a novel SFUDA framework specifically designed for geospatial point clouds. At the local level, we introduce a class-balanced prototype estimation module that abandons conventional global threshold filtering in favor of an intra-class independent anchor mining strategy. This ensures that robust feature prototypes can be generated even for sample-scarce tail classes, effectively mitigating the feature collapse caused by long-tailed distributions. At the global level, we introduce an optimal transport-based global distribution alignment module that formulates pseudo-label assignment as a global optimization problem. By enforcing global distribution constraints, this module effectively corrects the over-dominance of head classes inherent in local greedy assignments, preventing model predictions from being severely biased towards majority classes. Finally, we propose a dual-consistency pseudo-label filtering mechanism. This strategy retains only high-confidence pseudo-labels where local multi-augmented ensemble predictions align with global optimal transport assignments for self-training.

The RoboSense Challenge: Sense Anything, Navigate Anywhere, Adapt Across Platforms

Autonomous systems are increasingly deployed in open and dynamic environments -- from city streets to aerial and indoor spaces -- where perception models must remain reliable under sensor noise, environmental variation, and platform shifts. However, even state-of-the-art methods often degrade under unseen conditions, highlighting the need for robust and generalizable robot sensing. The RoboSense 2025 Challenge is designed to advance robustness and adaptability in robot perception across diverse sensing scenarios. It unifies five complementary research tracks spanning language-grounded decision making, socially compliant navigation, sensor configuration generalization, cross-view and cross-modal correspondence, and cross-platform 3D perception. Together, these tasks form a comprehensive benchmark for evaluating real-world sensing reliability under domain shifts, sensor failures, and platform discrepancies. RoboSense 2025 provides standardized datasets, baseline models, and unified evaluation protocols, enabling large-scale and reproducible comparison of robust perception methods. The challenge attracted 143 teams from 85 institutions across 16 countries, reflecting broad community engagement. By consolidating insights from 23 winning solutions, this report highlights emerging methodological trends, shared design principles, and open challenges across all tracks, marking a step toward building robots that can sense reliably, act robustly, and adapt across platforms in real-world environments.

Unlocking the Pre-Trained Model as a Dual-Alignment Calibrator for Post-Trained LLMs

Post-training improves large language models (LLMs) but often worsens confidence calibration, leading to systematic overconfidence. Recent unsupervised post-hoc methods for post-trained LMs (PoLMs) mitigate this by aligning PoLM confidence to that of well-calibrated pre-trained counterparts. However, framing calibration as static output-distribution matching overlooks the inference-time dynamics introduced by post-training. In particular, we show that calibration errors arise from two regimes: (i) confidence drift, where final confidence inflates despite largely consistent intermediate decision processes, and (ii) process drift, where intermediate inference pathways diverge. Guided by this diagnosis, we propose Dual-Align, an unsupervised post-hoc framework for dual alignment in confidence calibration. Dual-Align performs confidence alignment to correct confidence drift via final-distribution matching, and introduces process alignment to address process drift by locating the layer where trajectories diverge and realigning the stability of subsequent inference. This dual strategy learns a single temperature parameter that corrects both drift types without sacrificing post-training performance gains. Experiments show consistent improvements over baselines, reducing calibration errors and approaching a supervised oracle.

MoniRefer: A Real-world Large-scale Multi-modal Dataset based on Roadside Infrastructure for 3D Visual Grounding

3D visual grounding aims to localize the object in 3D point cloud scenes that semantically corresponds to given natural language sentences. It is very critical for roadside infrastructure system to interpret natural languages and localize relevant target objects in complex traffic environments. However, most existing datasets and approaches for 3D visual grounding focus on the indoor and outdoor driving scenes, outdoor monitoring scenarios remain unexplored due to scarcity of paired point cloud-text data captured by roadside infrastructure sensors. In this paper, we introduce a novel task of 3D Visual Grounding for Outdoor Monitoring Scenarios, which enables infrastructure-level understanding of traffic scenes beyond the ego-vehicle perspective. To support this task, we construct MoniRefer, the first real-world large-scale multi-modal dataset for roadside-level 3D visual grounding. The dataset consists of about 136,018 objects with 411,128 natural language expressions collected from multiple complex traffic intersections in the real-world environments. To ensure the quality and accuracy of the dataset, we manually verified all linguistic descriptions and 3D labels for objects. Additionally, we also propose a new end-to-end method, named Moni3DVG, which utilizes the rich appearance information provided by images and geometry and optical information from point cloud for multi-modal feature learning and 3D object localization. Extensive experiments and ablation studies on the proposed benchmarks demonstrate the superiority and effectiveness of our method. Our dataset and code will be released.

Improve Contrastive Clustering Performance by Multiple Fusing-Augmenting ViT Blocks

In the field of image clustering, the widely used contrastive learning networks improve clustering performance by maximizing the similarity between positive pairs and the dissimilarity of negative pairs of the inputs. Extant contrastive learning networks, whose two encoders often implicitly interact with each other by parameter sharing or momentum updating, may not fully exploit the complementarity and similarity of the positive pairs to extract clustering features from input data. To explicitly fuse the learned features of positive pairs, we design a novel multiple fusing-augmenting ViT blocks (MFAVBs) based on the excellent feature learning ability of Vision Transformers (ViT). Firstly, two preprocessed augmentions as positive pairs are separately fed into two shared-weight ViTs, then their output features are fused to input into a larger ViT. Secondly, the learned features are split into a pair of new augmented positive samples and passed to the next FAVBs, enabling multiple fusion and augmention through MFAVBs operations. Finally, the learned features are projected into both instance-level and clustering-level spaces to calculate the cross-entropy loss, followed by parameter updates by backpropagation to finalize the training process. To further enhance ability of the model to distinguish between similar images, our input data for the network we propose is preprocessed augmentions with features extracted from the CLIP pretrained model. Our experiments on seven public datasets demonstrate that MFAVBs serving as the backbone for contrastive clustering outperforms the state-of-the-art techniques in terms of clustering performance.

NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding

Underwater exploration offers critical insights into our planet and attracts increasing attention for its broader applications in resource exploration, national security, etc. We study the underwater scene understanding methods, which aim to achieve automated underwater exploration. The underwater scene understanding task demands multi-task perceptions from multiple granularities. However, the absence of large-scale underwater multi-task instruction-tuning datasets hinders the progress of this research. To bridge this gap, we construct NautData, a dataset containing 1.45 M image-text pairs supporting eight underwater scene understanding tasks. It enables the development and thorough evaluation of the underwater scene understanding models. Underwater image degradation is a widely recognized challenge that interferes with underwater tasks. To improve the robustness of underwater scene understanding, we introduce physical priors derived from underwater imaging models and propose a plug-and-play vision feature enhancement (VFE) module, which explicitly restores clear underwater information. We integrate this module into renowned baselines LLaVA-1.5 and Qwen2.5-VL and build our underwater LMM, NAUTILUS. Experiments conducted on the NautData and public underwater datasets demonstrate the effectiveness of the VFE module, consistently improving the performance of both baselines on the majority of supported tasks, thus ensuring the superiority of NAUTILUS in the underwater scene understanding area. Data and models are available at https://github.com/H-EmbodVis/NAUTILUS.

InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents

Information seeking is a fundamental requirement for humans. However, existing LLM agents rely heavily on open-web search, which exposes two fundamental weaknesses: online content is noisy and unreliable, and many real-world tasks require precise, domain-specific knowledge unavailable from the web. The emergence of the Model Context Protocol (MCP) now allows agents to interface with thousands of specialized tools, seemingly resolving this limitation. Yet it remains unclear whether agents can effectively leverage such tools -- and more importantly, whether they can integrate them with general-purpose search to solve complex tasks. Therefore, we introduce InfoMosaic-Bench, the first benchmark dedicated to multi-source information seeking in tool-augmented agents. Covering six representative domains (medicine, finance, maps, video, web, and multi-domain integration), InfoMosaic-Bench requires agents to combine general-purpose search with domain-specific tools. Tasks are synthesized with InfoMosaic-Flow, a scalable pipeline that grounds task conditions in verified tool outputs, enforces cross-source dependencies, and filters out shortcut cases solvable by trivial lookup. This design guarantees both reliability and non-triviality. Experiments with 14 state-of-the-art LLM agents reveal three findings: (i) web information alone is insufficient, with GPT-5 achieving only 38.2% accuracy and 67.5% pass rate; (ii) domain tools provide selective but inconsistent benefits, improving some domains while degrading others; and (iii) 22.4% of failures arise from incorrect tool usage or selection, highlighting that current LLMs still struggle with even basic tool handling.

Ultra-Efficient Decoding for End-to-End Neural Compression and Reconstruction

Image compression and reconstruction are crucial for various digital applications. While contemporary neural compression methods achieve impressive compression rates, the adoption of such technology has been largely hindered by the complexity and large computational costs of the convolution-based decoders during data reconstruction. To address the decoder bottleneck in neural compression, we develop a new compression-reconstruction framework based on incorporating low-rank representation in an autoencoder with vector quantization. We demonstrated that performing a series of computationally efficient low-rank operations on the learned latent representation of images can efficiently reconstruct the data with high quality. Our approach dramatically reduces the computational overhead in the decoding phase of neural compression/reconstruction, essentially eliminating the decoder compute bottleneck while maintaining high fidelity of image outputs.

Enhancing Physical Consistency in Lightweight World Models

A major challenge in deploying world models is the trade-off between size and performance. Large world models can capture rich physical dynamics but require massive computing resources, making them impractical for edge devices. Small world models are easier to deploy but often struggle to learn accurate physics, leading to poor predictions. We propose the Physics-Informed BEV World Model (PIWM), a compact model designed to efficiently capture physical interactions in bird's-eye-view (BEV) representations. PIWM uses Soft Mask during training to improve dynamic object modeling and future prediction. We also introduce a simple yet effective technique, Warm Start, for inference to enhance prediction quality with a zero-shot model. Experiments show that at the same parameter scale (400M), PIWM surpasses the baseline by 60.6% in weighted overall score. Moreover, even when compared with the largest baseline model (400M), the smallest PIWM (130M Soft Mask) achieves a 7.4% higher weighted overall score with a 28% faster inference speed.