trustworthiness

This paper benchmarks LLMs against human verbal uncertainty markers from psychology literature and introduces VOCAL, an optimization-based algorithm that learns an optimal marker-to-uncertainty profile directly from LLM outputs, revealing systematic gaps between how LLMs and humans express confidence in language.

The paper proposes RegLLM, a diagnostic harness for evaluating bounded autonomy in regulated agentic AI workflows, instrumenting six trustworthiness signals with constitutional rewards, escalation labels, and runtime governance. Small-scale Qwen2.5-3B pilots reveal that configuration variance can overwhelm apparent tuning effects on bounded-autonomy metrics, motivating larger evaluation sets and repeated runs.

EvidenT is a lightweight pipeline that enhances evidence groundedness and traceability in enterprise retrieval-augmented generation (RAG) systems, improving gold-source hit rate by 29% without model retraining.

LEGIT is a credentialing protocol for trustworthy AI agent marketplaces that connects certification, reputation, and marketplace allocation to enable verifiable quality assessment of AI agents.

A tweet discusses whether AI models should be trusted with tools during tests, referencing the Terminal-Bench-2.1 benchmark where models can access solutions but are instructed not to use them.

The article explores the evolution of AI from capability to economic trustworthiness, highlighting the growing need for trustworthy AI agents as they handle real-world tasks with financial implications.

The author describes a three-week experiment testing AI agent trustworthiness, finding that agreement between agents using the same model is unreliable, and outlines a system with human approval for critical decisions.

This paper provides an up-to-date overview of the phenomenon of model collapse in generative AI and reviews countermeasures to mitigate it, highlighting challenges and future research opportunities.

The article discusses the criteria for trusting AI agents with real-world tasks, questioning the balance between usefulness and risk, and seeks insights from users on practical workflows.

This paper evaluates the trustworthiness of small language models across fairness, robustness, privacy, and ethics, comparing pre-trained SLMs with compressed larger models, and finds that quantization preserves trustworthiness better than pruning and that distillation can further enhance reliability.

TRACE is a proposed retrieval-augmented conversational engine for public service chatbots that improves constraint-aware recommendations by strengthening retrieval quality over noisy directories, reducing hallucinated responses.

A systematic cross-architecture empirical study measuring the trustworthiness cost of domain adaptation in small language models, finding that safety-preserving fine-tuning strategies do not reliably transfer alignment.

This survey proposes a trustworthiness model for agentic AI in critical engineering systems, covering safety, robustness, transparency, accountability, and security across domains like power systems and autonomous vehicles.

ConceptSMILE is a perturbation-based auditing framework for evaluating the reliability of concept-based explainable AI, tested on retinal fundus images.

This paper introduces GraphEVAL, a graph-based framework for quantifying uncertainty in LLM reasoning, and proposes a new metric, Graph Reasoning Coherence Score (GRCS), that captures semantic-structural consensus and detects confident hallucinations. The authors also present Graph Self-Consistency (GSC), a decoding strategy that prioritizes reasoning fidelity over nominal accuracy.

This paper proposes SolarChain-Eval, a physics-constrained benchmark for evaluating trustworthy economic agents in decentralized energy markets, integrating an LLM-based Planner/Auditor layer and revealing a utility-safety trade-off.

Cognition launched SWE-1.7, built on the open-source Kimi K2.7, with trustworthiness training to address concerns about Chinese base models.

A thought experiment questioning whether AI systems should maintain a verifiable memory trail of their knowledge and beliefs at the time of decision-making to enhance trust and accountability.

Discusses the issue of AI models producing incorrect answers with high confidence, highlighting the problem of overconfidence in AI outputs.

This paper proposes Robust-TO, an agentic video understanding framework that integrates per-frame trustworthiness to address the Blind Trust Problem, achieving significant accuracy gains under realistic perturbations.