faithfulness evaluation

This work addresses the challenge of deploying generative AI in high-stakes decision-making, where hallucinated reasoning, unsupported claims, and weak traceability often preclude compliance with certification-grade accountability requirements. To bridge this gap, the authors propose a “compliance-by-construction” architecture that uniquely integrates typed argumentation graphs, retrieval-augmented generation (RAG), formal verification kernels, and W3C PROV-based provenance tracking. This framework ensures that every AI-generated claim is grounded in authoritative evidence and subjected to rigorous inference constraints before being admitted into official decision records. Empirical evaluation demonstrates that the architecture effectively blocks unsubstantiated assertions from entering the decision pipeline and substantially improves the efficiency of constructing compliant, auditable arguments, thereby enabling controlled, verifiable, and accountable use of generative AI in high-assurance settings.

This work addresses the susceptibility of large language models to hallucinations in empirical reasoning—outputs lacking verifiable evidence or formal guarantees. The authors propose EG-VAR, an architecture that uniquely leverages the Lean 4 formal proof kernel as the sole trusted generator of claims. By integrating tool-certified axioms and a source elevation mechanism, EG-VAR ensures every output is bound to a kernel-verified chain of reasoning and tool invocation; otherwise, it abstains and provides a fully traceable audit trail. Evaluated on a TableBench subset, EG-VAR achieves perfect accuracy (120/120), substantially outperforming a 95% baseline. In counterfactual tests, it maintains 100% source fidelity—significantly higher than competing methods (80–90%)—and exhibits remarkably low semantic formalization error rates of 1.7% (Opus) and 3.3% (Sonnet).

Current large language model (LLM) agents lack verifiability, debuggability, and auditability, and relying solely on the accuracy of final answers fails to reveal their underlying reasoning. To address this, this work proposes the first unified provenance framework for LLM agents, systematically modeling causal relationships in tool usage, memory access, and environmental interactions. It introduces a comprehensive provenance taxonomy encompassing source, granularity, representation format, and trust functions. By integrating provenance-aware representation modeling, evidence attribution, runtime safeguards, provenance-informed memory management, and trajectory observability analysis, the study shifts the evaluation paradigm from outcome correctness to process accountability. The framework consolidates existing benchmarks to define a clear pathway for process-level trustworthiness assessment and highlights key challenges, including standardized trajectory schemas, semantic-level provenance, and privacy-preserving auditing.

Existing automated approaches for mapping cyber threat intelligence (CTI) to MITRE ATT&CK lack supporting evidence, provenance tracking, and validation history, making their credibility difficult to assess. This work proposes the first knowledge graph–driven framework for CTI governance that enables auditable management of TTP assertions through fine-grained evidence preservation, complete provenance chains, versioned trust decisions, and lossless revocation mechanisms. The framework integrates multi-extractor collaborative verification, assertion aggregation, consensus modeling, and policy-driven validation, all underpinned by versioned knowledge graph management. Evaluated on 65 CTI reports comprising 5,303 sentences, the approach achieves a precision of 90.6% under six-party consensus and efficiently supports seven categories of audit queries concerning provenance, trustworthiness, and versioning.

This work addresses the problem of cross-source confusion in large language model agents under the Model-Context-Prompt (MCP) architecture, where factual errors in responses are incorrectly attributed to irrelevant evidence sources. The paper introduces the first source-aware factuality verification framework, which parses MCP execution traces to decompose responses into atomic claims, routes each claim to its corresponding evidence source, and evaluates claim support through natural language inference and token alignment. Crucially, it compares the claimed source against the actual evidence source, enabling per-claim and holistic allow/block decisions. By treating source attribution as an independent dimension of factuality, the framework supports detection of misattribution and enables an automated retrieval-augmented repair loop. Evaluated on 281 medical MCP trajectories, it achieves a held-out blocking F1 of 0.802 and source accuracy of 0.858; in multi-source settings, blocking F1 improves to 0.846, with 100% detection and correction of injected attribution manipulations.

This study reveals that large language models struggle to effectively assess the veracity of statistical evidence when integrating multi-source information, exhibiting a tendency to rely on superficial stylistic cues in methodological text rather than numerical plausibility when judging source credibility. The work identifies a previously undocumented “cognitive alignment” bias—where models prefer sources with strong analytical register over those with content consistency. Employing interpretable techniques including causal tracing, linear probing (AUC: 0.83–0.92), and component-level attribution, the authors replicate this blind spot across five mainstream models through cross-model and cross-domain experiments. Further analysis localizes the issue to a methodology-register gating mechanism and demonstrates that neither prompt engineering nor post-training interventions adequately mitigate the bias, instead raising concerns about model generalization.

This work addresses the opacity of reasoning in large language model (LLM) agents performing data-intensive analysis, which hinders the verifiability of their conclusions. To resolve this, the authors propose VeriGraph—a traceable neuro-symbolic reasoning framework that constructs an explicit heterogeneous evidence directed acyclic graph (DAG) to unify raw data, variables, computational results, and natural language claims. VeriGraph introduces three evidence expansion primitives—computation, anchoring, and derivation—to enable structural traceability via graph reachability and incorporates claim-level evidence evaluation to quantify semantic support. Experimental results demonstrate that VeriGraph achieves state-of-the-art performance across four benchmarks, with its 8B variant attaining a claim-level anchoring accuracy of 87.61%, substantially enhancing the auditability and reproducibility of model outputs.

This work addresses a critical flaw in fact verification models, which often conflate “missing evidence” with “insufficient evidence,” thereby learning constructional biases rather than genuine reasoning capabilities. To systematically disentangle these two types of Not Enough Information (NEI) instances, the authors propose NEI-CAP, a diagnostic protocol that categorizes NEI samples by their construction family, incorporates human arbitration, evaluates cross-construction generalization, and analyzes claim-confidence alignment under fixed statements. Experiments reveal that models trained solely on shortcut-based constructions fail to recognize semantically grounded cases of insufficient evidence, while mixed-construction training only partially mitigates this issue. Moreover, aggregate NEI scores obscure model performance on specific subtasks. This study establishes a transferable, construction-aware diagnostic framework, offering a new paradigm for robust evaluation of fact verification systems.

This work addresses the limited reproducibility of behavioral validation in robotic simulation testing, which often stems from insufficiently documented test configurations, execution protocols, and post-processing procedures. To overcome this, the study proposes a deep integration of data provenance and FAIR (Findable, Accessible, Interoperable, Reusable) principles throughout the entire test generation pipeline—rather than merely appending them to final datasets. The authors extend an existing simulation testing framework by embedding machine-readable, structured metadata at every stage, thereby enabling end-to-end traceable validation workflows. This approach significantly enhances the reproducibility of mobile robot navigation datasets. Additionally, the project distills practical FAIR implementation guidelines tailored to robotics, identifying key challenges such as vocabulary alignment, attribute selection, and adoption of community standards, and offers actionable recommendations for addressing them.

This work addresses the inadequacy of current AI runtime logs in providing the structured evidence necessary for legal fact-finding—such as data boundary violations or human interventions. It formalizes, for the first time, the binary factual requirements of regulatory compliance into a criterion of evidentiary sufficiency for runtime records, mandating that logs explicitly encode the legal category of events and their determinative relationships (e.g., provenance, authorization, temporal validity). By integrating legal ontologies, event-type systems, provenance semantics, and temporal validity constraints—and drawing on the law of requisite variety and the Good Regulator theorem from cybernetics—the approach exposes limitations in tamper-proof logging and generic provenance mechanisms. Validation against selected obligations of the EU AI Act demonstrates that this criterion precisely delineates the boundary between traces and hyperproperties in runtime verification, thereby establishing a verifiable foundation for compliance.