reliability

This paper challenges the 'Attention-Confidence Assumption' by demonstrating that attention map sharpness is a poor predictor of correctness in Vision-Language Models. Instead, it shows that reliability is better indicated by hidden-state geometry and self-consistency, with significant findings on architectural differences between late-fusion and early-fusion models.

This paper introduces NoisyCoconut, an inference-time method that improves LLM reliability by injecting noise into latent trajectories to generate diverse reasoning paths. The approach enables models to abstain when uncertain, significantly reducing error rates in mathematical reasoning tasks without requiring retraining.

This paper analyzes hallucination detection in LLMs, proposing a max-pooling approach that improves efficiency by eliminating costly semantic consistency computations while maintaining competitive performance.

This paper introduces AgentCollabBench, a diagnostic benchmark for multi-agent systems that evaluates behavioral risks like instruction decay and context leakage across four major LLMs. It argues that communication topology is a critical factor in multi-agent reliability, often overshadowing raw model capability.

This paper introduces Sem-ECE, a semantic-sampling framework for evaluating calibration in open-ended question answering by grouping model answers into semantic classes to estimate confidence.

The article highlights a lack of benchmarks for evaluating the reliability of agent harnesses, specifically focusing on how MCP implementations handle tool calls and errors compared to the models themselves.

The article argues that high autonomy in AI agents increases the cost of errors, advocating instead for constrained, reliable agents that prioritize safety and predictability over unrestricted capability.

The author argues that the hidden cost of unreliable AI agents lies in the cognitive overhead of constant human monitoring, emphasizing that predictability and environmental stability matter more than raw intelligence for real-world deployment. Practical workflows improve significantly when agents operate within controlled, validated environments rather than unpredictable ones.

This paper introduces a statistical framework for adaptively auditing AI systems using Safe Anytime-Valid Inference (SAVI) to draw rigorous conclusions with limited data. It proposes a 'testing by betting' approach to validate model robustness while controlling type-I errors during adaptive sampling.

The author describes improving AI agent reliability by replacing a single general-purpose agent with a four-agent workflow specializing in intake, research, action, and review. This shift prioritized system predictability and easier debugging over raw autonomy.

The author argues that in enterprise AI agent development, operational reliability and stability are more critical than high autonomy, advocating for controlled intelligence over fully autonomous systems.

The author argues that current AI agent evaluations often overlook execution efficiency, focusing only on final outputs while ignoring redundant actions and costly orchestration issues that arise in production.

A developer shares real-world experiences with AI orchestration frameworks (LangGraph, CrewAI, AutoGen), noting trade-offs between ease of prototyping and production reliability, and asks the community about handling failures, human-in-the-loop, and token costs.

This article discusses the growing gap between laboratory tests and production reality for semiconductor materials, highlighting how complex interactions in advanced packaging and heterogeneous integration lead to unexpected failures.

This paper introduces Partial-Evidence-Bench, a deterministic benchmark for measuring 'authorization-limited evidence' failures in agentic AI systems. It evaluates how models handle tasks where access control restricts visibility, assessing their ability to recognize and report incomplete information rather than silently producing seemingly complete but incomplete answers.

The article argues that increasing AI agent capability does not inherently improve reliability, emphasizing the need for robust control systems, audits, and human oversight similar to accounting standards to prevent convincing failures.

The article highlights a shift in the AI industry where the focus is moving from purely model benchmark performance to infrastructure challenges like latency, orchestration, and cost efficiency. It suggests that AI is maturing into a systems problem, with real-world experience becoming more important than raw model capability.

The article discusses the drop in reliability when AI agents move from sandboxed tests to production environments, highlighting that the orchestration layer often contains more bugs than the model itself.

The author observes that AI agents exhibit human-like failure patterns, such as overconfidence and skipping steps under context pressure, suggesting that system reliability depends more on robust validation and controlled environments than just model intelligence.

The article argues that reliable AI agents require deterministic control flow and programmatic verification in software, rather than relying solely on complex prompt chains.