<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"><channel><title>Crew Briefing</title><link>https://koolo.app/research/</link><description>Public AI engineering and Crew project research</description><language>zh-TW</language><item><title>Curriculum-aligned analysis of assessment feedback: a retrieval-augmented large language model approach for revealing fine-grained student error patterns in higher educat</title><link>https://koolo.app/research/2026-11-17/2042b36438c27e2d</link><guid isPermaLink="false">2042b36438c27e2d</guid><pubDate>Sun, 15 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Assessment feedback in higher education provides fine-grained evidence about students’ specific errors but remains underutilized as a data source for teaching improvement. This study proposes a method based on a large language model (LLM) that transforms unstructured feedback into curriculum-aligned error anal</description></item><item><title>VANTAGE-V-AI: Bounded-Memory KV-Cache Compression and Streaming Inference Runtime for Large Language Models</title><link>https://koolo.app/research/2026-11-17/2eb144820d0cb8cf</link><guid isPermaLink="false">2eb144820d0cb8cf</guid><pubDate>Sun, 15 Nov 2026 00:00:00 GMT</pubDate><description>Autoregressive generation in Large Language Models (LLMs) deployed within extended multi-turn dialogs and autonomous agentic workflows suffers from an acute memory wall. Because conventional Key-Value (KV) caching retains representations for all preceding tokens, memory consumption scales linearly O(N) with sequence le</description></item><item><title>Securing Autonomous Multi-Agent AI Systems: A Zero-Trust Framework for Threat Detection, Behavioral Monitoring, and Resilience in Enterprise Environments</title><link>https://koolo.app/research/2026-11-17/7376db199d43cc71</link><guid isPermaLink="false">7376db199d43cc71</guid><pubDate>Sun, 15 Nov 2026 00:00:00 GMT</pubDate><description>This is an important issue in developing multi-agent systems in the context of distributed, autonomous, large-language models. Large-language model (LLM) agents delegate tasks, call tools and act on real-world agents. Generally, the most dangerous inputs arrived as natural-language messages from trusted peers, and one</description></item><item><title>LLM-agent-based cross-modal retrieval method for detecting clues of ecological environment damage</title><link>https://koolo.app/research/2026-11-17/04b76b545b9466d5</link><guid isPermaLink="false">04b76b545b9466d5</guid><pubDate>Sun, 15 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Within Ecological Conservation Redline zones, ecological environmental damage caused by non-compliant production and construction activities has become increasingly severe. Existing methods typically leverage remote sensing imagery and deep learning techniques to automatically identify unauthorized constructio</description></item><item><title>AgentMesh: Toward Redundancy- and Barrier-Free Dataflow in Multi-Agent Systems</title><link>https://koolo.app/research/2026-11-17/7695a3f136f7c356</link><guid isPermaLink="false">7695a3f136f7c356</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Multi-agent systems (MAS) now power long-running agentic applications like deep research and coding agents, yet their end-to-end latency reaches minutes to hours. The runtime executes each dataflow as an opaque, indivisible block at coarse granularity, wasting end-to-end time on redundancy (acquiring content already he</description></item><item><title>Build specialized multimodal models with LFM3</title><link>https://koolo.app/research/2026-11-17/1c5b51464767c2ad</link><guid isPermaLink="false">1c5b51464767c2ad</guid><pubDate>Tue, 27 Oct 2026 17:16:52 GMT</pubDate><description>近期公開研究線索：Build specialized multimodal models with LFM3。需先閱讀原文核實方法及結果。</description></item><item><title>Don’t Trust, Verify: Adaptive Multi-Agent Feature Engineering for Financial Fraud Detection</title><link>https://koolo.app/research/2026-11-16/e0d05eabdf80ef01</link><guid isPermaLink="false">e0d05eabdf80ef01</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Feature engineering in tabular fraud detection is critical yet challenging to automate due to extreme class imbalance, which heavily penalizes spurious features. While Large Language Models (LLMs) can generate candidate features from dataset schemas, schema-level plausibility correlates poorly with actual predictive va</description></item><item><title>Enhancing AI Feedback Systems: Lessons Learned from a RAG Implementation Study</title><link>https://koolo.app/research/2026-11-16/c947e289663b68df</link><guid isPermaLink="false">c947e289663b68df</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>This study investigates the impact of Retrieval-Augmented Generation (RAG) on graduate students&apos; perceptions of AI-generated feedback in a cyber-social learning environment. Using a non-equivalent cohort comparison quasi-experimental design, the research compared survey responses from Fall 2023 (n=96) and Spring 2024 (</description></item><item><title>FlexMARL: Graph-Driven Rollout-Training Co-Design for LLM-Based Multi-Agent Reinforcement Learning</title><link>https://koolo.app/research/2026-11-16/0968157e7f51185d</link><guid isPermaLink="false">0968157e7f51185d</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Despite algorithm-level innovations for multi-agent reinforcement learning (MARL), the underlying infrastructure for large-scale MARL remains underexplored. Existing training frameworks primarily optimize for single-agent scenarios and fail to capture the graph-structured nature of MARL. This fundamental mismatch leads</description></item><item><title>From Code Completion to Agentic Development: A Survey of AI Coding Agents, Ecosystems, and Evolving Software Engineering Practice</title><link>https://koolo.app/research/2026-11-16/8c16080c3a1a9ef0</link><guid isPermaLink="false">8c16080c3a1a9ef0</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Large language models have extended software-development assistance from token-level completion to systems that can navigate repositories, edit files, execute tools, and iterate on test feedback. This survey studies the design, use, and evaluation of these AI coding agents, together with their commercial and s</description></item><item><title>AI and Data Science in Education: Reimagining Pedagogy, Personalized Learning, and Open Educational Resources through Intelligent Systems</title><link>https://koolo.app/research/2026-11-16/ad18b8ef569bd06c</link><guid isPermaLink="false">ad18b8ef569bd06c</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Artificial Intelligence (AI) and Data Science are transforming education by reshaping teaching methodologies, personalizing learning experiences, and expanding access to educational resources. The integration of intelligent systems into educational environments enables institutions to move beyond traditional one-size-f</description></item><item><title>From Tool to Agent: How Worker Needs Reorganize Across the Agentic Roles of Workplace AI</title><link>https://koolo.app/research/2026-11-15/0c0bdbb0b06fb2fc</link><guid isPermaLink="false">0c0bdbb0b06fb2fc</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>As AI systems gain agency in workplaces, workers increasingly interact with systems that do more than support tasks: they make decisions, allocate work, and shape how workers communicate with others. We interviewed 16 workers in healthcare, finance, and management about their daily interactions with workplace AI system</description></item><item><title>AssemPlanner: a multi-agent based task planning framework for flexible assembly system</title><link>https://koolo.app/research/2026-11-15/21071ee454e66b4b</link><guid isPermaLink="false">21071ee454e66b4b</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>研究論文索引：AssemPlanner: a multi-agent based task planning framework for flexible assembly system。詳細方法及結果請閱讀原文。</description></item><item><title>Claude Code AI agent update: - Fixed sessions on many third-party gateways and Foundry stopping with a context length error on every turn instead of compacting</title><link>https://koolo.app/research/2026-11-15/e58a3ce1f17ad846</link><guid isPermaLink="false">e58a3ce1f17ad846</guid><pubDate>Fri, 13 Nov 2026 20:40:01 GMT</pubDate><description>- Fixed sessions on many third-party gateways and Foundry stopping with a context length error on every turn instead of compacting - Fixed the welcome header showing control characters or splitting across lines when an agent&apos;s name or the working folder contains them or a line break - Fixed `/voice` saying voice mode changed when managed settings, `--settings`, or the project&apos;s `.claude/settings.j</description></item><item><title>LangGraph AI agent update: * fix(langgraph): stream a child graph on its own thread to a caller streaming subgraphs (#9698)</title><link>https://koolo.app/research/2026-11-15/1b011fa040f80893</link><guid isPermaLink="false">1b011fa040f80893</guid><pubDate>Sat, 14 Nov 2026 01:43:55 GMT</pubDate><description>* fix(langgraph): stream a child graph on its own thread to a caller streaming subgraphs (#9698)</description></item><item><title>MIRAGE: Misleading Retrieval-Augmented Generation via Black-box and Query-agnostic Poisoning Attacks</title><link>https://koolo.app/research/2026-11-14/aeb91968d9e30da7</link><guid isPermaLink="false">aeb91968d9e30da7</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>Retrieval-Augmented Generation (RAG) systems enhance LLMs with external knowledge but introduce a critical attack surface: corpus poisoning. While recent studies have demonstrated the potential of such attacks, they typically rely on impractical assumptions, such as white-box access or known user queries, thereby under</description></item><item><title>MemPoison: Bypassing Selective Memory Mechanisms to Plant Backdoors in LLM Agents</title><link>https://koolo.app/research/2026-11-14/4c6393a334f55c81</link><guid isPermaLink="false">4c6393a334f55c81</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>Large language model (LLM) agents increasingly leverage long-term memory to support persistent and autonomous task execution. However, this capability also introduces a new attack surface: memory poisoning, where adversaries can inject malicious information to influence future behavior. Existing memory poisoning attack</description></item><item><title>AI Agents and Agentic AI in Medicine: An Autonomy-Centred Scoping Review and Research Agenda</title><link>https://koolo.app/research/2026-11-14/5aaaf5b5fff8a04a</link><guid isPermaLink="false">5aaaf5b5fff8a04a</guid><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Artificial intelligence agents—software systems that perceive, reason, plan and act over multiple steps with limited human intervention—are moving from research prototypes towards clinical deployment. Recent reviews map what has been built; this one asks what would have to hold for such systems to be deployed</description></item><item><title>LiveKit Agents AI agent update: You can recover caller answers that GPT-Live transcribes but never delegates. After a quiet grace period, a detector checks missed</title><link>https://koolo.app/research/2026-11-14/45402fb9d67b00c1</link><guid isPermaLink="false">45402fb9d67b00c1</guid><pubDate>Fri, 13 Nov 2026 17:37:40 GMT</pubDate><description>You can recover caller answers that GPT-Live transcribes but never delegates. After a quiet grace period, a detector checks missed turns and starts backend delegation when needed, with a note for answers spoken over the agent. GPTLiveModel(delegation_recovery=True) - **xAI**: `xai/stt-2` through Inference, with word alignment for adaptive interruption. - **Google**: Gemini 4 Carbon, Gemini 4 Hydro</description></item><item><title>Multimodal Resource-Exhaustion Attacks on Vision-Language Models via Joint Pixel-Prompt Optimization</title><link>https://koolo.app/research/2026-11-14/971c3d9593dac9ae</link><guid isPermaLink="false">971c3d9593dac9ae</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>Resource-exhaustion attacks against autoregressive vision-language models (VLMs) typically assume unimodal threat models, treating the image branch as the primary optimization surface while holding user-visible prompts fixed. Even recent loop-centric variants remain confined to this single-channel paradigm, leaving the</description></item><item><title>Do Hackers Dream of Electric Teachers?: A Large-Scale, In-Situ Measurement of Cybersecurity Student Behaviors and Educational Performance with AI Tutors</title><link>https://koolo.app/research/2026-11-14/b2016a69e01bd34b</link><guid isPermaLink="false">b2016a69e01bd34b</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>To meet the ever-increasing demands of the cybersecurity workforce, AI tutors have been proposed for personalized, scalable education. But, while AI tutors have shown promise in introductory programming courses, no work has evaluated their use in hands-on exploration and exploitation exercises (e.g., &quot;Capture the Flag&quot;</description></item><item><title>Durable Memory on a Fixed Context Budget: Architecture and Measurements of a Long-Lived Coding Agent</title><link>https://koolo.app/research/2026-11-13/d7ece0fe3ddc565b</link><guid isPermaLink="false">d7ece0fe3ddc565b</guid><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><description>Memory for language-model agents is usually treated as a retrieval problem. We also treat it as a budget: a context window keeps nothing between sessions, and all of it is read again on every call. We describe the memory architecture one operator and one coding-agent harness built and used for 140 days, the argument th</description></item><item><title>Jacobian-Causal Memory Control for Persistent Memory in Language Model Systems</title><link>https://koolo.app/research/2026-11-13/6d6c9c26405ceaea</link><guid isPermaLink="false">6d6c9c26405ceaea</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Persistent-memory systems choose not only which records are retrieved but which prior information can alter current inference. We propose Jacobian-causal memory control (JCMC), which represents each authorized memory intervention by a predicted influence signature in a shared behavioral measurement space, selects inter</description></item><item><title>Digital twin–guided graph-transformer with meta-reinforcement learning for real-time adaptive cyber-attack mitigation in smart grids</title><link>https://koolo.app/research/2026-11-13/339b0a64442366c5</link><guid isPermaLink="false">339b0a64442366c5</guid><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><description>Smart grids are rapidly changing the current face of power systems by the use of state-of-the-art sensing and communication technology and renewable energy resources to improve the efficiency, reliability and sustainability of the transmission system. However, with the growing cyber–physical connectivity, new and more</description></item><item><title>Design, Implementation, and Empirical Evaluation of a Human–AI Collaborative Adaptive Assessment Model in Higher Education</title><link>https://koolo.app/research/2026-11-13/c08c94dd2e52fb6e</link><guid isPermaLink="false">c08c94dd2e52fb6e</guid><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><description>AI-supported assessment increasingly enables adaptive tasks, automated analysis, and personalized feedback, yet their integration into coherent, human-supervised formative assessment remains insufficiently established. This study developed and evaluated the Human–AI Collaborative Adaptive Assessment Model (HA-CAAM) usi</description></item><item><title>Claude Code AI agent update: - Fixed running `/compact` again, or after a command such as `/context`, summarizing away a message that Claude had not answered y</title><link>https://koolo.app/research/2026-11-13/8427d4fee9a7d1b9</link><guid isPermaLink="false">8427d4fee9a7d1b9</guid><pubDate>Thu, 12 Nov 2026 22:12:13 GMT</pubDate><description>- Fixed running `/compact` again, or after a command such as `/context`, summarizing away a message that Claude had not answered yet - Fixed a memory spike at the next prompt after Claude changed a very large file - Fixed Edit and Bash running out of memory on certain `/proc` files, and Edit hanging or running out of memory on devices and named pipes - Fixed Claude Code failing to start, or to cle</description></item><item><title>On Evaluating Performance of LLM Inference Systems</title><link>https://koolo.app/research/2026-11-12/34522ac5f8576c4d</link><guid isPermaLink="false">34522ac5f8576c4d</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>The rapid evolution of Large Language Model (LLM) inference systems has led to significant improvements in serving efficiency. However, our systematic analysis reveals that current evaluation methodologies suffer from fundamental flaws that mask critical performance issues and impede scientific progress. Through a comp</description></item><item><title>Retrieval-augmented Large Language Models for automated discharge summary generation in post-surgical ICU cancer patients</title><link>https://koolo.app/research/2026-11-12/c58687452c6095b8</link><guid isPermaLink="false">c58687452c6095b8</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Introduction Discharge summary generation is an essential yet time-consuming component of clinical documentation, particularly for post-surgical cancer patients admitted to the Intensive Care Unit (ICU), where clinically relevant information is distributed across structured electronic health records (EHRs) and unstruct</description></item><item><title>When Sources Disagree: Accuracy-Preserving Batch Verification and the Limits of Source-Age Screening in Retrieval-Augmented Generation</title><link>https://koolo.app/research/2026-11-12/aa216fcc294034bb</link><guid isPermaLink="false">aa216fcc294034bb</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Retrieval-augmented generation (RAG) can attach citations that contradict a generated claim or no longer reflect current facts. We revisit an experimentally evaluated citation-verification design using SciFact for semantic contradiction and a small, independently reviewed FreshQA case study for temporal change</description></item><item><title>When does online vision-language inference help? A survey of socially aware robot navigation</title><link>https://koolo.app/research/2026-11-12/f1dede3a6bf3a065</link><guid isPermaLink="false">f1dede3a6bf3a065</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Vision-language models (VLMs) and large language models are increasingly built into the navigation stacks of robots that share space with people, where the models read cues that occupancy maps omit, such as a raised hand or a conversing group. This survey organizes 41 cases reported from 2024 to 2026, 34 on so</description></item><item><title>Self-Healing JSON Generation in Small Language Models: A Systematic Evaluation of Syntax Repair, Schema Coercion, and Validator Feedback</title><link>https://koolo.app/research/2026-11-12/8871e53c935d51d9</link><guid isPermaLink="false">8871e53c935d51d9</guid><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Small language models offer an affordable way to extract structured information, including on devices with limited computing resources. However, their responses often contain JSON that cannot be parsed or does not satisfy the required schema. We investigate a self-healing approach that operates on generated te</description></item><item><title>Typicality biases in interpreting unmarked sentences: an artificial language learning experiment on differential argument marking</title><link>https://koolo.app/research/2026-11-12/1a211b11d0ee7957</link><guid isPermaLink="false">1a211b11d0ee7957</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Many languages exhibit differential argument marking (DAM), in which nouns are marked for their grammatical role only in certain contexts. DAM manifests as differential object marking (DOM) and differential subject marking (DSM), both of which are conditioned by factors such as animacy and hypothesized to arise via amb</description></item><item><title>Trojan Hippo Bench: A Dynamic Benchmark for Persistent Memory Attacks and Defenses in LLM Agents</title><link>https://koolo.app/research/2026-11-11/439bc15c1bd58db9</link><guid isPermaLink="false">439bc15c1bd58db9</guid><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><description>Memory systems enable otherwise-stateless LLM agents to persist user information across sessions, but also introduce a new attack surface. The Trojan Hippo attack is a class of persistent memory attacks that operates under a more realistic threat model than prior memory poisoning work. The attacker plants a dormant pay</description></item><item><title>Evaluating the use of LLM-based multi-agent systems for internal information sharing and access within manufacturing</title><link>https://koolo.app/research/2026-11-11/278e8c0220d7331f</link><guid isPermaLink="false">278e8c0220d7331f</guid><pubDate>Fri, 06 Nov 2026 00:00:00 GMT</pubDate><description>Abstract Internal information sharing in manufacturing remains constrained by data silos, heterogeneous systems, poor data quality, and limited interoperability across ISA-95 levels. While Large Language Models (LLMs) are increasingly proposed for manufacturing information access and for integrating information with op</description></item><item><title>ProAct: Harnessing Streaming Motion Generation and Agentic Reasoning for Realtime Embodied Social Interaction</title><link>https://koolo.app/research/2026-11-11/9d83a3a97a94d7a7</link><guid isPermaLink="false">9d83a3a97a94d7a7</guid><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><description>Real-time embodied social interaction places two equally demanding requirements on an agent: continuously generating fluent multimodal interaction behavior, and proactively reasoning over accumulated dialogue and visual context to decide when to take initiative. These requirements must both be satisfied under a strict</description></item></channel></rss>
