All Articles
16670 articles total
Supporting Autonomous Process Execution within a Multi-Perspective Constraint Frame via Numeric Planning
Research introducing a tool for what-if analysis in AI-augmented Business Process Management Systems using multi-perspective constraint frames.
RELIC: Revealed Principles for Learning Interpretable Composable Skills in Multi-Agent Planning
Introduces RELIC, a framework for privacy-preserving multi-agent planning that uses 'revealed principles' instead of sharing executable policies.
FUSAR-R1: A Large-Scale Reasoning Model for Intelligent Interpretation of SAR Images
Presents FUSAR-R1, a large-scale reasoning model utilizing chain-of-thought and reinforcement learning for intelligent SAR image interpretation.
From Overload to Insights: How AI Agents Can Support Scientists in Analyzing Complex Data
Explores the design and evaluation of AI agents to assist scientists at European XFEL in analyzing complex, large-scale datasets.
AgentBrew: Lifelong Knowledge Brewing from Strong Teachers to Weak LLM Agents
Introduces AgentBrew, a training-free knowledge distillation method that converts teacher LLM experiences into external memory for weaker student agents.
Beyond Semantic Equivalence: Logical Graphs for LLM Uncertainty Quantification
Proposes Logical Graph Uncertainty (LGU), a framework for better LLM uncertainty quantification by modeling logical entailment and incompatibility.
Environment-free Synthetic Data Generation for API-Calling Agents
A method for generating synthetic training data for API-calling agents using LLMs as world models, removing the need for live executable environments.
Lomekwi: Resource-Bounded Tool Discovery in LLM Agents
Analyzes tool discovery in LLM agents, distinguishing between curiosity, recognition, and efficiency, and identifying an inverse scaling law for recognition.
Training Continuous Chain of Thought Models: A Tale of Two Regimes
Introduces C-MTP, a fast direct supervision approach for Continuous Chain-of-Thought models, while highlighting performance drops in long-trace reasoning.
A PHP server that can handle 10x as many concurrent requests as Nginx+PHP-fpm
A discussion about a new PHP server claiming to handle 10x more concurrent requests than the traditional Nginx and PHP-fpm stack.
Zoom is now Chrome only
Zoom has reportedly shifted its compatibility to be Chrome-only, potentially limiting user access on other browsers.
Just A Rather Very Intelligent Spoken Agent
Researchers introduce JarvisBench and a modular prototype for always-on, spoken AI mediators that facilitate better long-horizon agent-user interaction.
A Research Prototype for Closed-Loop Generative Design of Customized Foot Orthoses via Semantic-Physics Alignment
TANS-FO is a research prototype for automating the design of customized foot orthoses using a Text-Aligned Neural Surrogate and GNNs for real-time stress prediction.
TopoTuner: Topological Finetuning of Large Language Models
TopoTuner is a new framework for selective fine-tuning of LLMs using topological persistence diagrams to identify which weights to freeze, reducing parameters and training time.
Diversity-Oriented Fine-Tuning for Uncertainty-Based Hallucination Detection
This paper proposes diversity-oriented fine-tuning using SFT and DPO to make LLM hallucinations more detectable via semantic entropy.
DS@GT ARC at eRisk 2026: Hybrid Multi-Agent LLM System with Structured Algorithmic Guidance for Conversational Depression Screening
A hybrid multi-agent system combining Gemma 27B with algorithmic guidance for depression screening outperformed a paid GPT-5-nano baseline at lower cost.
Tractable Query Answering under Epistemic Confidentiality Policies in DL Ontologies (extended version)
The paper presents a new semantics based on minimal policy violation (MPV) for tractable query answering in DL ontologies under confidentiality policies.
RECON: Benchmarking Agent Memory for Compositional Reasoning over Long Contexts
RECON is a new benchmark evaluating compositional reasoning over long contexts (50k-100k tokens) to test agent memory and reliability.
Constraint-Anchored Reasoning Traces
CART is a neuro-symbolic framework that interleaves natural language reasoning with symbolic constraints to prevent 'error snowballing' in multimodal LLMs.
Reverse-engineering is cheap now
A discussion on the increasing accessibility and reduced cost of reverse-engineering tools and techniques.