Intelligence Brief

TODAY'S INTELLIGENCE BRIEF

On 2026-04-09, our systems ingested 568 new papers, revealing 10 newly introduced concepts. The AI research landscape is increasingly focused on refining agentic capabilities and multimodal understanding through advanced benchmarking and data-centric training strategies. Significant efforts are also directed towards making AI systems safer and more robust in complex, dynamic environments, as evidenced by new frameworks for LLM productivity agents and video understanding.

ACCELERATING CONCEPTS

This week highlights a continued surge in concepts underpinning more robust and understandable AI systems. Notably, advancements in agentic capabilities and explainability are taking center stage, alongside specialized architectural protocols.

• Explainable AI (XAI) (Category: evaluation, Maturity: emerging)

  Description: An approach or set of techniques to make AI system decisions understandable, serving as a mitigation strategy for biases in digital health technologies. Papers like MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome implicitly address the need for understanding complex agent behavior, contributing to the XAI discourse by evaluating process quality.

• Agentic AI (Category: application, Maturity: emerging)

  Description: Agentic AI enables smart systems to operate autonomously, establish objectives, and apply skills such as comprehension, reasoning, planning, memory, and task completion in complex healthcare environments. This concept is heavily driven by papers like MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome, ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces, and How Well Do Agentic Skills Work in the Wild: Benchmarking LLM Skill Usage in Realistic Settings, which scrutinize agent performance and safety in realistic settings.

• Federated Learning (Category: training, Maturity: established)

  Description: A distributed machine learning approach that enables model training across decentralized devices or servers holding local data samples, without exchanging them. Its continued prominence signifies ongoing efforts in privacy-preserving and distributed AI training paradigms.

• Model Context Protocol (MCP) (Category: architecture, Maturity: emerging)

  Description: A protocol used by AgentRob to bridge online community forums, LLM-powered agents, and physical robots. Its emergence indicates a growing need for standardized communication and integration mechanisms for multi-agent and physical-world AI systems, particularly in contexts involving human-AI interaction.

• AI Literacy (Category: application, Maturity: established)

  Description: The necessary competencies for individuals to interact with and critically examine AI/ML systems, with implications discussed for both teachers and students. Its sustained acceleration reflects the increasing societal integration of AI and the consequent educational demands.

NEWLY INTRODUCED CONCEPTS

This week saw the introduction of several compelling new ideas, signaling shifts in AI research frontiers, particularly in optimization, evaluation, and system design:

• Difficulty-aware Length Penalty (Category: training)

  An extension of the standard length penalty that encourages longer reasoning for difficult problems and shorter traces for easy ones without additional training overhead. This concept addresses the common challenge of optimizing reasoning length for diverse problem complexities, potentially making LLM reasoning more efficient and targeted.

• Topological Data Analysis (TDA) (Category: theory)

  A principled framework applied to the 21 cm forest for extracting information about the organization and merging hierarchy of absorption troughs, using persistence diagrams and Betti curves. This highlights the interdisciplinary application of advanced mathematical tools to extract complex patterns from scientific data, moving beyond traditional statistical methods.

• REMind (Category: application)

  An innovative educational robot-mediated role-play game designed to support anti-bullying bystander intervention among children by having them observe a scenario, reflect, and rehearse defending strategies. This concept showcases a novel application of AI and robotics in social-emotional learning, addressing real-world societal challenges.

• Paper Circle (Category: architecture)

  A multi-agent research discovery and analysis system designed to reduce the effort required to find, assess, organize, and understand academic literature. This framework, described in Paper Circle: An Open-source Multi-agent Research Discovery and Analysis Framework, represents a significant step towards AI-assisted scientific discovery by structuring and analyzing complex literature.

• Analysis Pipeline (Category: architecture)

  A component of Paper Circle that transforms individual papers into structured knowledge graphs with typed nodes and edges, enabling graph-aware question answering and coverage verification. This underscores a move towards more semantic and relational understanding of research papers, offering a structured way to query scientific knowledge.

• Value Verifier (Category: architecture)

  A novel component within the CVA architecture trained on authentic human data to explicitly model dynamic value activation. This concept points towards more human-aligned AI systems capable of understanding and adapting to nuanced human values, a critical step for ethical AI.

• Proactive Intelligence (Category: theory)

  A paradigm shift in AI where systems are capable of taking initiative and making decisions rather than just reacting to inputs. This signals a fundamental evolution in AI design, moving towards truly autonomous and anticipatory systems.

• MedResearchBench (Category: evaluation)

  A novel benchmark specifically designed to evaluate AI systems on medical clinical research tasks, covering 7 clinical domains and using publicly available datasets with ground truth from published papers. This addresses a critical gap in evaluating AI for high-stakes medical research, promoting more reliable and clinically relevant AI tools.

• Discovery Pipeline (Category: architecture)

  A component of Paper Circle that integrates offline and online retrieval from multiple sources, multi-criteria scoring, diversity-aware ranking, and structured outputs. Complementing the Analysis Pipeline, this demonstrates a comprehensive approach to automating the front-end of research literature management, detailed in Paper Circle: An Open-source Multi-agent Research Discovery and Analysis Framework.

• Pair-Aware Reasoning Selection (Category: training)

  A mechanism that uses counterfactual intervention to identify reasoning paths beneficial for query-target alignment in contrastive embedding. This technique aims to enhance the precision and relevance of reasoning in complex AI systems, particularly for tasks requiring nuanced understanding of relationships.

METHODS & TECHNIQUES IN FOCUS

The research landscape continues to rely on robust evaluation methods while also seeing an expansion in sophisticated AI architectures. Qualitative analysis methods remain critical, highlighting a persistent need for human interpretability and insight in AI development.

• Thematic Analysis (Method Type: evaluation_method, Usage Count: 36)

  A qualitative method applied to questionnaire-based data to identify recurring themes and patterns. This method's high usage underscores the community's continued focus on understanding human perspectives and qualitative insights, especially in areas like AI adoption and impact.

• Systematic Review (Method Type: evaluation_method, Usage Count: 32)

  A method used to analyze literature on technical architectures for federated AI governance, focusing on architectural concerns, API specifications, and reference architectures. The consistent application of systematic reviews indicates a mature approach to synthesizing knowledge and identifying gaps in complex, evolving domains.

• Random Forest (Method Type: algorithm, Usage Count: 31)

  An ensemble machine learning method that constructs multiple decision trees and outputs the mode of the classes (for classification) or mean prediction (for regression) of the individual trees. Its enduring popularity reflects its robustness and interpretability for a wide range of predictive tasks.

• Semi-structured Interviews (Method Type: evaluation_method, Usage Count: 27)

  A qualitative data collection method used with domain experts to gain insights into design trade-offs, deployment challenges, and organizational readiness for AI adoption. This highlights the importance of expert knowledge and real-world context in guiding AI research and deployment strategies.

• Systematic Literature Review (Method Type: evaluation_method, Usage Count: 25)

  A research methodology used to synthesize empirical evidence by systematically searching, selecting, and evaluating studies, following guidelines like PRISMA. Similar to Systematic Review, its high usage confirms the rigorous approach researchers take to establish comprehensive understanding within a field.

BENCHMARK & DATASET TRENDS

Current trends in benchmarks and datasets underscore a strong move towards evaluating AI performance on real-world, dynamic, and complex multimodal tasks. There's a notable push for benchmarks that truly test visual grounding and agentic capabilities, rather than superficial performance metrics.

• real-world datasets (Domain: general, Eval Count: 8)

  Empirical datasets used to demonstrate CAKE's performance and applicability in practical scenarios. The increasing reliance on "real-world datasets" signals a maturation in AI evaluation, moving beyond synthetic or clean academic datasets to confront messy, practical challenges.

• SWE-bench (Domain: code, Eval Count: 7)

  A benchmark dataset for coding tasks. Its frequent use indicates a significant focus on improving AI's code generation and understanding capabilities, essential for developer productivity and autonomous software engineering.

• MNIST (Domain: vision, Eval Count: 7)

  Dataset of handwritten digits used for benchmarking. While foundational, its continued use for quick prototyping and basic model validation remains, though often alongside more complex benchmarks.

• LoCoMo (Domain: general, Eval Count: 7)

  A benchmark used to evaluate the accuracy of memory systems like Hippocampus. This dataset is crucial for advancing agentic AI, particularly in areas requiring long-term memory and contextual recall, as seen in Omni-SimpleMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory, which achieved a +411% F1 increase on LoCoMo.

• CIFAR-10 (Domain: vision, Eval Count: 7)

  A dataset of 60,000 32x32 colour images in 10 classes, with 6,000 images per class. Like MNIST, it remains a standard for vision model development and benchmarking, offering a balanced challenge for image classification.

• Scopus database (Domain: general, Eval Count: 7)

  A comprehensive abstract and citation database of peer-reviewed literature, from which 311 documents published between 2023 and 2025 were indexed and analyzed. This reflects a growing trend in using academic databases as datasets for meta-research, bibliometric analysis, and AI-driven literature review systems like Paper Circle.

BRIDGE PAPERS

No explicit bridge papers (connecting previously separate subfields) were identified in the provided data today. However, many of the high-impact papers exhibit interdisciplinary characteristics, particularly in their evaluation methodologies which combine quantitative metrics with qualitative human assessment, bridging technical performance with user experience and safety considerations.

UNRESOLVED PROBLEMS GAINING ATTENTION

Several critical and significant open problems continue to challenge the AI community, particularly concerning the deployment and ongoing management of AI systems. The high recurrence of these problems indicates a persistent struggle to achieve robust, scalable, and compliant AI solutions, especially as agentic and dynamic systems become more prevalent.

• High demand for continuous updates and audits to maintain relevance and compliance. (Severity: significant, Recurrence: 3)

  This problem, often exacerbated by rapidly evolving data and regulatory landscapes, remains a major hurdle for long-term AI system maintenance. Methods addressing this implicitly include modular architectures that facilitate updates, and data-centric approaches like DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models, which streamlines dynamic data optimization to keep models relevant.

• Requires significant resource investment for implementation. (Severity: significant, Recurrence: 3)

  The financial and computational cost of deploying and maintaining complex AI systems is a consistent barrier. Solutions are being explored through frameworks that optimize training efficiency (e.g., DataFlex with its runtime improvements) and through more compute-optimal scaling laws that account for inference costs, as presented in Test-Time Scaling Makes Overtraining Compute-Optimal.

• Multi-agent LLM systems suffer from false positives, where they report success on tasks that fail strict validation. (Severity: critical, Recurrence: 2)

  This critical problem, highlighted by systems like those evaluated in MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome, underscores the limitations in agentic trustworthiness and verification. Benchmarks like MiroEval, with its focus on "agentic factuality verification" and "process-centric evaluation," are crucial steps toward identifying and mitigating these false positives.

• A critical gap exists in systematic frameworks for characterizing the interactions of domain specialization, coordination topology, context persistence, authority boundaries, and escalation protocols across production deployments of LLM-based agents. (Severity: critical, Recurrence: 2)

  This architectural and operational challenge for complex agent systems is still largely open. The development of benchmarks like ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces and ClawArena: Benchmarking AI Agents in Evolving Information Environments begins to address this by simulating realistic agent workspaces and dynamic information environments, revealing critical interaction patterns and safety concerns.

• Existing text-driven 3D avatar generation methods based on iterative Score Distillation Sampling (SDS) or CLIP optimization struggle with fine-grained semantic control and suffer from excessively slow inference. (Severity: significant, Recurrence: 2)

  This challenge in 3D content generation highlights the need for more efficient and controllable generative models, indicating a gap in current diffusion or optimization techniques for detailed 3D assets.

INSTITUTION LEADERBOARD

Academic institutions in East Asia continue to dominate the research output, reflecting substantial investment and active research communities. Collaboration patterns suggest a strong emphasis on internal university collaborations, with a growing need for greater cross-institutional and academic-industry partnerships to accelerate diverse research areas.

Academic Institutions:

• Tsinghua University: 285 recent papers, 351 active researchers. A leading powerhouse, often seen in major AI benchmarks and foundational research.
• Shanghai Jiao Tong University: 261 recent papers, 261 active researchers. Consistently high output, reflecting broad research activity.
• Zhejiang University: 240 recent papers, 265 active researchers. Strong presence across various AI subfields.
• Fudan University: 209 recent papers, 237 active researchers. Demonstrates significant research momentum.
• University of Science and Technology of China: 178 recent papers, 151 active researchers. Maintains a high research volume, often in core AI areas.
• National University of Singapore: 174 recent papers, 176 active researchers. A prominent international player, strong in areas like AI applications and systems.
• Nanyang Technological University: 166 recent papers, 215 active researchers. Another leading institution from Singapore, contributing significantly to diverse AI domains.

Industry contributions are less explicitly tracked in the provided leaderboard, which heavily favors academic publication volume. However, companies like Kuaishou Technology and X-Era AI Lab appear in collaboration clusters, indicating their active involvement in specific research collaborations.

RISING AUTHORS & COLLABORATION CLUSTERS

The landscape of rising authors reveals several highly productive researchers, with a strong emphasis on prolific individuals from major Chinese academic institutions. Collaboration clusters indicate tightly-knit research groups, primarily within the same institution, driving focused research agendas.

Rising Authors:

• Yang Liu (Beijing Institute of Mathematical Sciences and Applications): 51 total papers, 21 recent papers. A highly prolific author with a significant recent surge.
• Wei Wang (Meituan LongCat Team): 28 total papers, 13 recent papers. Demonstrates strong recent activity, likely in applied AI or industrial research.
• Jie Li: 28 total papers, 11 recent papers. A consistently active researcher.
• Qi Li (National University of Singapore): 14 total papers, 9 recent papers. High recent velocity for their total publication count.
• Yu Wang (Lenovo Group Ltd.): 20 total papers, 8 recent papers. Strong presence from an industry researcher.

Collaboration Clusters:

• Dingkang Liang & Xiang Bai (Kling Team, Kuaishou Technology): 7 shared papers. A strong industry research collaboration.
• Zeyu Zheng & Cihang Xie (UCSC): 7 shared papers. Indicates a productive academic partnership.
• Shaohan Huang & Furu Wei (Tsinghua University): 6 shared papers. A key collaboration from a top academic institution.

The dominance of self-collaborations (e.g., "tshingombe tshitadi" with "tshingombe tshitadi") and unlisted institutions for some authors in the raw data suggests limitations in disambiguating author identities and affiliations or reflects solo-authored work within a project context.

CONCEPT CONVERGENCE SIGNALS

The co-occurrence of concepts reveals interesting connections, particularly in structured reasoning, curriculum development, and memory/learning challenges. These convergences often highlight areas where researchers are seeking to integrate disparate ideas for more holistic solutions.

• Logigram & Algorigram (Co-occurrences: 12, Weight: 12.0)

  This strong convergence suggests an increasing focus on formalizing and structuring logical and algorithmic reasoning. Researchers are likely exploring how to better represent, generate, and evaluate complex computational thought processes, perhaps for more robust explainability or agent planning.

• Curriculum Engineering & Algorigram (Co-occurrences: 10, Weight: 10.0)

  The link between "Curriculum Engineering" and "Algorigram" points to efforts in designing educational or training pathways (curricula) that are themselves algorithmically structured or enhanced. This could involve using algorithmic methods to optimize learning sequences or to develop adaptive educational AI systems.

• Curriculum Engineering & Logigram (Co-occurrences: 10, Weight: 10.0)

  Similar to the above, this convergence reinforces the idea of logically structured curriculum development, potentially leveraging formal logic systems to build coherent and effective learning programs, possibly for AI systems themselves or for human education about AI.

• Catastrophic Forgetting & Parameter-Efficient Fine-Tuning (PEFT) (Co-occurrences: 7, Weight: 7.0)

  This convergence highlights a core challenge in continual learning for large models. Researchers are actively exploring PEFT methods as a primary strategy to mitigate catastrophic forgetting, allowing models to adapt to new tasks without losing previously acquired knowledge. This is critical for practical, lifelong learning AI systems.

• Model Context Protocol (MCP) & Retrieval-Augmented Generation (RAG) (Co-occurrences: 6, Weight: 6.0)

  While RAG is ubiquitous, its specific convergence with MCP suggests efforts to standardize how external knowledge is retrieved and integrated into dynamic model contexts. This indicates a move towards more structured and protocol-driven RAG implementations, particularly for agentic systems interacting with diverse information sources.

TODAY'S RECOMMENDED READS

The following papers are highly recommended for their impact, novelty, and concrete contributions to the field today:

• Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding (Impact Score: 1.0)

  Key Findings: This new comprehensive benchmark reveals a substantial performance gap between state-of-the-art models like Gemini-3-Pro and human experts in video understanding. It exposes a hierarchical bottleneck where errors in low-level visual information aggregation and temporal modeling significantly propagate, limiting high-level reasoning capabilities, especially in scenarios where models rely on textual cues, with subtitles sometimes degrading performance in purely visual settings.

• DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models (Impact Score: 1.0)

  Key Findings: DataFlex significantly improves LLM performance, with dynamic data selection consistently outperforming static full-data training on MMLU for Mistral-7B and Llama-3.2-3B. For data mixture optimization, it enables DoReMi and ODM to improve MMLU accuracy and corpus-level perplexity over default proportions when pretraining Qwen2.5-1.5B, while also achieving consistent runtime improvements and supporting large-scale settings like DeepSpeed ZeRO-3.

• MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (Impact Score: 1.0)

  Key Findings: MiroEval's three evaluation dimensions (adaptive synthesis quality, agentic factuality verification, and process-centric evaluation) reveal distinct strengths and weaknesses across 13 evaluated systems. Process quality is a reliable predictor of outcome, and multimodal tasks significantly challenge systems, causing 3 to 10 point performance declines. The MiroThinker series achieved the most balanced performance, ranking highest overall in both text-only and multimodal settings.

• AURA: Always-On Understanding and Real-Time Assistance via Video Streams (Impact Score: 1.0)

  Key Findings: AURA, an end-to-end streaming visual interaction framework, enables a unified VideoLLM to continuously process video streams, supporting both real-time question answering and proactive responses. It integrates context management, data construction, training objectives, and deployment optimization to achieve stable long-horizon streaming interaction, operating at 2 FPS on two 80G accelerators in a real-time demo system.

• Watch Before You Answer: Learning from Visually Grounded Post-Training (Impact Score: 1.0)

  Key Findings: This paper reveals that 40-60% of questions in common long video understanding benchmarks can be answered with text alone, indicating a lack of true visual grounding. VidGround, a simple solution using only visually grounded questions for post-training, improves performance by up to 6.2 points relative to using the full dataset, while using only 69.1% of the original data, highlighting data quality as a major bottleneck.

• General Multimodal Protein Design Enables DNA-Encoding of Chemistry (Impact Score: 1.0)

  Key Findings: The DISCO multimodal diffusion model designs novel protein sequences and 3D structures around arbitrary biomolecules, conditioned on reactive intermediates, leading to diverse heme enzymes with new active-site geometries. DISCO-designed enzymes catalyze new-to-nature carbene-transfer reactions with high activities surpassing previously engineered enzymes, significantly broadening the scope of genetically encodable transformations.

• Paper Circle: An Open-source Multi-agent Research Discovery and Analysis Framework (Impact Score: 1.0)

  Key Findings: Paper Circle is an open-source multi-agent framework designed to streamline research discovery and analysis, integrating a Discovery Pipeline (multi-source retrieval, scoring, ranking) and an Analysis Pipeline (paper-to-knowledge graph conversion for Q&A). Benchmarking shows it consistently improves paper retrieval and review generation performance (hit rate, MRR, Recall@K), with stronger agent models yielding better results, and is publicly released.

• Test-Time Scaling Makes Overtraining Compute-Optimal (Impact Score: 1.0)

  Key Findings: When accounting for inference costs, optimal LLM pretraining shifts significantly towards 'overtraining,' outside traditional scaling suites. The introduced Train-to-Test (T^2) scaling laws, which jointly optimize model size, training tokens, and inference samples under fixed computational budgets, forecast heavily overtrained models that demonstrate substantially stronger performance compared to models optimized solely by pretraining scaling laws.

• ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces (Impact Score: 1.0)

  Key Findings: ClawsBench, a new benchmark with five mock services and 44 tasks, shows LLM productivity agents achieve only 39-64% task success even with full scaffolding, and exhibit unsafe action rates of 7-33%. The research identifies eight recurring unsafe behavior patterns, including multi-step sandbox escalation, highlighting significant risks in deploying agents on live services despite comparable task success across models on the OpenClaw benchmark.

• DARE: Diffusion Large Language Models Alignment and Reinforcement Executor (Impact Score: 1.0)

  Key Findings: DARE is an open framework unifying post-training and evaluation for Diffusion Large Language Models (dLLMs), integrating SFT, PEFT, preference optimization, and dLLM-specific reinforcement learning under a shared execution stack. It provides broad algorithmic coverage, reproducible benchmark evaluation, and practical acceleration across dLLM families (LLaDA, Dream, SDAR, LLaDA2.x), aiming to accelerate research iteration and facilitate fair comparisons.

KNOWLEDGE GRAPH GROWTH

Today's ingestion further enriched our knowledge graph, adding new connections and entities to deepen our understanding of the AI research landscape.

• Papers: 19702 (increased by 568 today)
• Authors: 82506
• Concepts: 50641
• Problems: 41389
• Topics: 32
• Methods: 29692
• Datasets: 8406
• Institutions: 4587

Today's additions highlighted new architectural paradigms like "Paper Circle" and "Analysis Pipeline," alongside novel evaluation concepts like "MedResearchBench" and "Difficulty-aware Length Penalty." The growth reflects an increasing density of connections, especially around agentic systems, multimodal understanding, and data-centric training, which are becoming central to solving complex, real-world problems.

AI LAB WATCH

No specific announcements or new model releases from major AI labs (Anthropic, OpenAI, Google DeepMind, Meta AI, IBM Research, NVIDIA, Microsoft Research, Apple ML, Mistral, Cohere, xAI) were directly identified in today's ingested papers or blog feeds. However, the high-impact papers reflect broader trends that these labs are undoubtedly contributing to and monitoring:

• OpenAI / Google DeepMind / Anthropic: The focus on ClawsBench and MiroEval for benchmarking agentic capabilities and safety is highly relevant to these labs, given their substantial investments in developing advanced LLM-based agents. The observed 7-33% unsafe action rates for agents on ClawsBench highlight critical safety challenges that these frontier labs are actively working to address. Similarly, the performance gaps in video understanding revealed by Video-MME-v2 will inform their multimodal model development.
• NVIDIA / Microsoft Research: Papers like DataFlex, which offers a unified framework for data-centric dynamic training of LLMs, and Test-Time Scaling Makes Overtraining Compute-Optimal, which re-evaluates pretraining for inference efficiency, are crucial for optimizing large-scale model development and deployment, areas where NVIDIA and Microsoft are key players.
• General Trend: The emphasis on more rigorous and realistic benchmarks across the board (e.g., Video-MME-v2 for video, MiroEval for agents, ClawsBench for productivity agents) indicates a collective industry shift towards not just achieving high scores, but building truly capable, reliable, and safe AI systems for real-world applications.

SOURCES & METHODOLOGY

Today's intelligence report was generated by querying a diverse set of academic and industry data sources to capture the latest advancements in AI research. The ingestion process involved several steps to ensure comprehensive coverage and data quality.

• OpenAlex: Queried for broad academic publications.
• arXiv: Main source for pre-print research papers, contributing 568 papers.
• DBLP: Used for author and publication metadata.
• CrossRef: Utilized for citation and DOI resolution.
• Papers With Code: Tracked for popular methods, datasets, and benchmarks.
• HF Daily Papers (Hugging Face): Contributed 568 papers, primarily focused on recent machine learning and natural language processing preprints.
• AI lab blogs: No new posts detected today from major labs.
• Web search: Conducted for supplementary information and context.

Deduplication Stats: A total of 568 papers were ingested today from arXiv and HF Daily Papers. Deduplication mechanisms ensured that unique papers were processed, preventing redundant entries in the knowledge graph. No significant pipeline issues (failed fetches, rate limits) were reported today, indicating stable data acquisition.