A GitHub repository for agentic conversational memory implementation.
The repository contains a hands-on ladybugdb implementation of agentic conversational memory, including design explanations and schema files. It is based on the author's book chapter on temporal aware memory. The code is written in Makefile, Rust, and Cypher.
Based on: GitHub - Volland/agentic-memory · github.com
Cloudflare introduces Code Mode, a technique to reduce context window usage during agent tool use.
Cloudflare's Code Mode collapses the entire Cloudflare API into two tools and 1,000 tokens of context. This reduces the number of input tokens used by 99.9% compared to an equivalent MCP server without Code Mode. The new technique is open-sourced as a Code Mode SDK in the Cloudflare Agents SDK.
Based on: Code Mode: give agents an entire API in 1,000 tokens · blog.cloudflare.com
A system of agents that performs real work and learns an organization's operations.
Hyperagent is a platform that enables organizations to deploy a fleet of agents that perform various tasks, such as research, building, and delivering content. The agents learn from the organization's data and tools, allowing them to adapt to changing needs. Hyperagent supports multiple formats, including websites, slides, documents, and dashboards.
Based on: Hyperagent · hyperagent.com
Paper proposing a taxonomy for distributional shifts in reinforcement learning.
The paper develops a unified causal-origin taxonomy to characterize sources of distributional shift in RL. It relates ID/OOD generalization to non-stationary settings and introduces an evaluation framework for measuring shift impact and adaptation. The proposed taxonomy grounds distributional shift in the causal-origin structure of RL, supporting systematic analysis of robustness under distributional shift.
Based on: A Unified Causal-Origin Taxonomy of Distributional Shifts in Reinforcement Learning · arXiv
A framework for improving the reliability of reasoning language models through error checking and correction.
CheckRLM proposes a framework to improve the reliability of reasoning language models by timely checking and correcting factual errors. It extracts claims from the reasoning chain, identifies inconsistencies, and performs minimal-cost corrections using external knowledge. The framework demonstrates strong capability in mitigating error accumulation in long-horizon reasoning with lower costs.
Based on: CheckRLM: Effective Knowledge-Thought Coherence Checking in Retrieval-Augmented Reasoning · arXiv
Paper exploring the integration of knowledge graphs and explainable AI in urban mining pre-demolition assessment.
The paper proposes four modes of integrating knowledge graphs and explainable AI to enhance defensibility in urban mining. It provides a complementarity-theoretic interpretation grounded in the IS resource-based tradition. The authors illustrate their approach using a fire-door example from the urban-mining process.
Based on: Knowledge Graphs and Explainable AI as Complementary Resources for Urban Mining · arXiv
A framework for retrieval-augmented generation that conditions language models on individual retrieved documents.
The paper proposes a new approach to question answering, called BERAG, which uses Bayesian ensemble methods to condition language models on individual documents. This allows for probabilistic re-ranking and clear attribution of document contribution. The authors evaluate BERAG on knowledge-based visual question answering tasks and demonstrate substantial improvements over standard RAG.
Based on: BERAG: Bayesian Ensemble Retrieval-Augmented Generation for Knowledge-based Visual Question Answering · arXiv
A taxonomy linking AD/ADRD caregiver needs with technology-based interventions.
This study introduces a taxonomy that systematically links caregiver needs with corresponding classes of technology-based interventions. The framework identifies mismatches between caregiver priorities and existing technological support, highlights under-served domains, and proposes design directions for adaptive systems. It offers a shared vocabulary to guide clinicians, researchers, and technology designers in developing more person-centered innovation.
Based on: A Taxonomy of Mental Health and Technology Needs for Alzheimer's and Dementia Caregivers · arXiv
A research paper investigating the interpretability and knowledge representation in large language models.
The authors analyze whether large language models (LMs) can be viewed as task-specific knowledge bases. They find that LMs encode knowledge in a task-specific manner, undermining the 'knowledge base' analogy. The study also explores the implications for the reliability and controllability of factual knowledge in LMs.
Based on: LMs as Task-Specific Knowledge Bases: An Interpretability Analysis · arXiv
Paper exploring citation faithfulness in Agentic GraphRag systems.
The authors investigate how traversal context and provenance affect citation faithfulness in Agentic GraphRag. They conduct controlled ablation experiments to analyze the impact of cited and uncited graph entities on answer accuracy. The results suggest that citation evaluation should consider the broader retrieval trajectory, not just source support.
Based on: Why Neighborhoods Matter: Traversal Context and Provenance in Agentic GraphRAG · arXiv
A paper proposing a runtime representation for neuro-symbolic learning.
The authors present Native Differentiable Virtual Machine (NDVM), a runtime representation that differentiates executable programs without compiling each candidate into a separate graph. NDVM separates symbolic structure from differentiable numeric state, allowing for efficient evaluation of large populations of parameter vectors. The paper demonstrates the effectiveness of NDVM in co-search over LLM-proposed programs.
Based on: Differentiate the Evaluator, Not the Program: An Efficient Runtime Representation for Neuro-Symbolic Learning · arXiv
Paper on fraud detection in payment networks using an observation-mechanism taxonomy.
The paper introduces a taxonomy that partitions fraud into five classes, each with distinct censorship and labeling pipelines. It proves that estimating fraud rates separately by class is more efficient than pooled estimation. The authors derive theoretical constraints on detection for each class, including label corruption and structural non-observability.
Based on: Fraud Type Decomposition and the Observation-Mechanism Taxonomy:Class-Specific Detection Limits in Payment Networks · arXiv
Highlights, analysis, and commentary on external articles — with links back to the original source.