This empirical study investigates jailbreak detection in large language models, showing that single output evaluation systematically underestimates vulnerability. Increasing the number of sampled generations, especially from one to moderate sampling, significantly improves the detection of harmful behavior.
This paper introduces a neuro-symbolic framework for translating natural-language reasoning problems into executable Narsese, leveraging first-order logic. It presents NARS-Reasoning-v0.1, a new benchmark featuring reasoning problems with corresponding formal representations and truth labels for evaluating reasoning capabilities.
This paper introduces a multi-agent system (MAS) prototype designed for hydrodynamics, addressing the limitations of single-agent LLM workflows. Specialized agents are coordinated through a Layer Execution Graph (LEG) to improve reliability and context management in scientific tasks.
This paper introduces CLIPR, a framework designed to enable Large Language Models (LLMs) to make human-aligned decisions by inferring latent user preferences from limited interactions. It addresses the challenge of LLMs struggling with human alignment and the limitations of existing approaches in generalizing preferences across tasks.
This research evaluates Gemini 3.0 Flash's ability to answer user health queries using Personal Health Records (PHRs) as context. It analyzes responses generated with and without PHR data across various query types to assess the utility of PHRs in personalized health AI.
Agentopic is a novel agent-based workflow for explainable topic modeling that leverages the reasoning capabilities of Large Language Models (LLMs). It enhances transparency by enabling users to trace the reasoning behind topic assignments, achieving an F1-score of 0.95, matching GPT-4.1.
This research evaluated the impact of prompting language on the diagnostic reasoning and accuracy of large language models (LLMs) in clinical settings. Four out of five models performed better in English, highlighting the uncertainty regarding LLM reliability across different languages.
HELLoRA proposes a novel method for fine-tuning Mixture-of-Experts (MoE) models by applying Low-Rank Adaptation (LoRA) modules only to the most frequently activated experts at each layer. This technique significantly reduces trainable parameters and improves downstream performance, attributing its success to structured regularization that maintains expert specialization.
This week highlights practical strategies for AI code generation using Claude's latest Opus 4.7 capabilities, promising enhanced performance. Additionally, a significant Python proposal aims to boost system-level observability, vital for robust AI framework deployments and leveraging advanced prompt engineering techniques.
This research systematically evaluates the relationship between differential privacy (DP) and social bias in large language models (LLMs). It compares a DP-trained LLM against non-DP baselines across various tasks, finding that DP reduces bias in sentence scoring but not universally, and reveals a discrepancy between logit-level and output-level bias.
SPEAR introduces a free-form agentic optimizer for automatic prompt engineering, leveraging a Python sandbox for error analysis and autonomous improvement. It utilizes tools like evaluation, code execution, and auto-rollback to optimize prompts for LLM-as-judge tasks.
Sanctions for AI hallucinations became a serious board-room issue in April 2026, driven by new state privacy laws adding AI transparency rules and a White House framework holding deployers accountable. Companies are now expected to understand and mitigate hallucinations, with specific rulings highlighting the legal and financial risks of unverified LLM output.
The author started building Hyphae to create a cognitive substrate without large language models, but a simple baseline matched its performance, highlighting a critical issue. This project now focuses on ensuring provenance in AI-generated answers, which is essential for auditability in critical applications.
This article highlights the inefficiency of autonomous AI agents, citing high failure rates and costs. It introduces "Compound AI Systems" as a successful alternative, where traditional code orchestrates LLM calls.
This article addresses the fundamental architectural problem of Large Language Models (LLMs) forgetting context in autonomous AI agents every few hours. Meridian, an autonomous AI, describes how it solved this issue by building a 21-layer memory stack to ensure continuous operation.
This article discusses Claude Opus 4.8, questioning whether its capabilities have improved to avoid providing misleading information. It analyzes the model's performance in terms of reliability and accuracy.
This article explores why most RAG (Retrieval-Augmented Generation) pipelines fail in production, contrasting the simplicity of demos with the complexity and messiness of real-world datasets. It highlights the challenges of AI systems engineering, particularly in data ingestion for scaling RAG to production environments.
A plugin named Caveman can reduce AI agent token costs by 75% by stripping away redundant communication and optimizing context space. It teaches agents to be efficient communicators, focusing on essential information for developers.
This content explores why a strong grasp of core Machine Learning concepts is crucial for effectively leveraging Large Language Models. It highlights how foundational ML knowledge enhances the practical application and understanding of LLMs.
The author introduces Origami, a terminal built with LLMs, and shares valuable insights from its development. They emphasize that AI coding isn't a simple solution and highlight software architecture as the most vital skill for effective AI integration.