reliability

An analysis of the termination problem in recursive AI agent systems, highlighting how agents can lose verified position while still selecting valid actions, and questioning where termination logic should reside in the agent stack.

This paper introduces MisKnow-Agent, a framework for generating misleading knowledge to test DeepResearch agents, showing that limited exposure to credible-looking false information can lead to false conclusions in final reports.

This paper identifies and characterizes two failure modes in LLM-based autonomous agents—Safety Drift and Operational Hallucination—and proposes a lightweight architectural layer to intercept violations without false positives.

This paper discovers that larger language models have a hidden auto-regressive risk regime where they commit to low-probability tokens and then snowball errors, causing reliability to degrade faster with scale. It shows that this failure mode is causal, dominant, and invisible to the model's own self-monitoring.

A reminder that an AI agent can appear to be running ("up") while actually being broken or malfunctioning, highlighting the need for better monitoring and validation.

This paper evaluates whether multilingual sentence embeddings can replace translation for linguistic-integrated reliability auditing across multiple languages in educational assessments, finding that native-language embeddings reproduce translation-based reliability estimates closely.

This paper evaluates the predictive accuracy, cross-task generalizability, and test-retest reliability of multimodal features for measuring conversational states like cognitive load and power in dyadic remote collaborative tasks. Findings show that linguistic features predict well but generalize poorly, acoustic reliability degrades when controlling for speaker identity, and interaction features provide the most reliable signal.

An article arguing that multi-agent setups are often overused, providing guidelines on when to use single vs multi-agent architectures based on task parallelism and coordination needs.

This article argues that failures in AI agent systems can be understood through the lens of distributed systems, drawing parallels between agent behavior and classic distributed system problems.

Proposes a Center of Gravity guided weight correction method for fault-tolerant deep neural networks, achieving significant fault tolerance improvements on LSTM and CNN models without retraining.

EduPanel is a rubric-grounded, learner-conditioned LLM judge that uses three specialized agents to evaluate teaching video quality, achieving reliability comparable to human experts and improving scoring accuracy while maintaining detectability of unreliable outputs.

This paper introduces ProofAgent-Harness, an open-source evaluation tool that measures context engineering quality for AI agents across seven criteria, validating it as a leading indicator of agent reliability.

Discusses a common failure mode in AI agents where the model claims to have executed a tool call without actually firing it, and advocates for trusting execution receipts over agent narration to ensure reliability in production.

This blog post by Ai2 describes the architecture and lessons learned from building Shippy, a maritime AI agent for real-time domain awareness, covering aspects like agent anatomy (soul, skills, config), deterministic tools, sandboxed hosting, and evaluation.

Oodle launches Agent Observability on Hacker News, offering agent traces at $10 per million spans to help AI-native teams detect silent failures and improve reliability.

A discussion asking how developers handle changes in tools, APIs, or model versions that their AI agents depend on in production, including detection, fixes, and costs.

This paper shows that LLM judges tend to over-credit incorrect answers when no reference answer is provided, and adding a reference can flip verdicts by up to 85%, aligning more with human judgments. The authors propose calibration steps for using LLM judges in reference-free settings.

This systematic review evaluates methodological reliability in machine learning models for early Chronic Kidney Disease prediction, revealing that data leakage inflates reported accuracy by over 15% and that more than 80% of predictors lack stability across studies.

This paper develops a minimal theory for partially correlated verifier cascades in LLM harnesses, showing concave log-odds, polynomial reliability decay, blind-spot ceilings, and that decorrelation is more effective than adding more gates.

The article shares production learnings for reliably generating structured JSON output from LLMs, covering methods like JSON mode, schema validation, and retry loops, achieving 99.5% validity.