Faithfulness Audit: Evaluating Model Alignment

• Faithfulness audit is the evaluation of whether a model’s output, reasoning, or actions are aligned with the internal evidence and computations that produced them.
• It employs methodologies like counterfactual interventions and precision-recall assessments to test causal influence, stance consistency, and evidence grounding.
• Current audits reveal challenges such as metric instability, ambiguity in labeling, and the need for mitigation strategies to enhance true model verification.

Faithfulness audit is the evaluation of whether a model’s output, rationale, chain of thought, or action trajectory is aligned with the evidence, intermediate computations, or internal decision process that actually produced it. In the recent literature, faithfulness is treated as distinct from plausibility, accuracy, and consensus. A reasoning trace may be coherent yet be “Reasoning Theater” rather than a load-bearing causal chain (Khanzadeh, 5 Jan 2026). A model may be “unfaithfully correct” or “faithfully wrong” because correctness does not determine whether the stated reasoning governed the answer (Han et al., 19 Feb 2026). In grounded generation, precision-only evaluation can “reward abstention,” so an apparently faithful system may simply omit the relevant facts (Santillana, 8 Jun 2026). In summarization, fluent text can remain subtly inconsistent with the source, and ambiguity can make binary faithful/unfaithful labeling itself unstable (Subbiah et al., 2024, Koupaee et al., 12 Feb 2025).

1. Definitions and conceptual boundaries

The literature uses several closely related definitions. One review defines faithful output as “accurate, reliable, and aligned with the ground truth or source material,” with hallucination as the introduction of “incorrect, unsupported, or invented statements” (Malin et al., 2024). For reasoning models, RFEval defines faithfulness through two testable conditions: stance consistency, meaning that the reasoning, explanation, and final answer form a coherent chain of stances, and causal influence, meaning that counterfactually altering the reasoning changes subsequent reasoning and/or the answer in a way that reflects that intervention (Han et al., 19 Feb 2026). BonaFide argues that neither plausibility nor importance is equivalent to faithfulness: a chain of thought can be plausible but unfaithful, or faithful but unimportant for the final answer (Gur-Arieh et al., 24 May 2026).

A second boundary concerns completeness. “Precision Is Not Faithfulness” defines faithfulness as the supported fraction of claims a model makes, but argues that this is only precision and therefore incomplete unless paired with recall or coverage against a complete oracle (Santillana, 8 Jun 2026). This reframes faithfulness audits for grounded generation as joint audits of correctness and completeness. A plausible implication is that the audit target depends on the task: for reasoning, the target is causal governance of the answer; for grounded generation, it is support plus coverage; for summarization, it is consistency with the source under potential ambiguity.

Ambiguity complicates binary audit judgments. Madisse introduces ambiguity as a distinct dimension for summary evaluation and provides a taxonomy with 16 fine-grained types grouped into implicit reasoning phenomena, meaning phenomena, context phenomena, and an “other” category (Koupaee et al., 12 Feb 2025). STORYSUMM reaches a related conclusion for narrative summarization: multiple human annotation protocols identify different real inconsistencies, so “ground truth” is itself an audited construct rather than a single annotation pass (Subbiah et al., 2024).

2. Counterfactual and causal auditing

A central family of faithfulness audits asks a direct causal question: if the reasoning is changed, does the answer change too? Project Ariadne formalizes this with a Structural Causal Model,

$$\mathcal{M}=\langle \mathcal{U}, \mathcal{V}, \mathcal{F} \rangle,$$

where $\mathcal{U}=\{q,\theta\}$ and $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$, and intervenes on intermediate reasoning nodes with Pearlian hard interventions such as $do(s_k=s'_k)$ (Khanzadeh, 5 Jan 2026). The audit output is the Causal Sensitivity Score

$\phi(q,k,\iota)=1-S(a,a^*),$

together with a binary violation indicator and aggregate Violation Density $\rho$. The intervention operator includes LogicFlip, FactReversal, PremiseNegation, and CausalInversion. In this formulation, a reasoning step is faithful only if intervening on it changes the output in the expected way (Khanzadeh, 5 Jan 2026).

RFEval uses a different counterfactual interface but the same structural intuition. It performs output-level counterfactual intervention by injecting flawed reasoning into the model’s reasoning channel and checks whether the intervened output remains stance-consistent and causally influenced by the injected reasoning (Han et al., 19 Feb 2026). The paper’s reasoning-faithfulness label requires baseline stance consistency, intervened stance consistency, and causal influence simultaneously. This makes explicit that a faithful audit can assign negative judgments to accurate outputs and positive judgments to inaccurate ones if the reasoning process is, respectively, decoupled or causally followed (Han et al., 19 Feb 2026).

Comparable counterfactual logic appears in other modalities. EDCT treats a vision-LLM’s own explanation as a falsifiable hypothesis: it extracts cited visual concepts, edits the image by targeted inpainting, re-queries the model, and scores whether both answer and explanation change appropriately (Ding et al., 27 Sep 2025). Its core metric is

$$\mathrm{CCS}_i=\mathrm{PCS}_i \cdot \mathrm{NCC}_i,$$

with the overall Counterfactual Consistency Score given by the average over extracted concepts. “Faithfulness Serum” uses hint injection rather than image editing: it keeps only cases where a hint flips the answer, then checks whether the explanation reflects that hint, and proposes attribution-guided attention interventions to improve epistemic faithfulness (Alon et al., 15 Apr 2026). “A Causal Lens for Evaluating Faithfulness Metrics” shifts the audit target from model outputs to the metrics themselves, using edited models to produce faithful/unfaithful explanation pairs and testing whether a metric ranks the faithful explanation higher (Zaman et al., 26 Feb 2025).

BonaFide extends this line by constructing tasks whose outputs reveal which intermediate computations must have occurred, thereby yielding automated ground-truth labels at both the step and chain-of-thought level (Gur-Arieh et al., 24 May 2026). This is significant because it moves faithfulness auditing away from proxy notions such as plausibility and toward directly labeled evaluation targets.

3. Evidence grounding, coverage, and support auditing

A second major tradition treats faithfulness as evidence alignment. In grounded generation, “Precision Is Not Faithfulness” states that “The supported fraction of the claims a model makes is its faithfulness, or precision,” formalized as

$\text{Precision}=\frac{|S|}{|C|},$

but adds exact recall because the domain provides a complete oracle (Santillana, 8 Jun 2026). The paper therefore audits systems with precision, recall, and

$F_1=\frac{2PR}{P+R},$

and argues that precision-only faithfulness can reorder system comparisons because a model can score highly by saying almost nothing (Santillana, 8 Jun 2026).

Faithfulness-QA operationalizes evidence grounding in RAG by manufacturing controlled conflicts between retrieved context and parametric memory. It begins from SQuAD and TriviaQA, identifies answer-bearing named entities, substitutes them with type-consistent alternatives, and produces 99,094 counterfactual samples from an input pool of 174,640 QA pairs, with a typed entity bank of 76,953 unique entities (Ju et al., 28 Apr 2026). The audit question is explicit: given question + modified_context, does the model output the faithful_answer from the altered passage or the original_answer from parametric memory? This makes context-faithfulness observable precisely when context and memory disagree (Ju et al., 28 Apr 2026).

LatentAudit moves evidence auditing inside the generator. It pools mid-to-late residual-stream activations into an answer-state vector $V_{act}$, maps retrieved evidence into the same space as $\mathcal{U}=\{q,\theta\}$0, and scores support with the Mahalanobis distance

$\mathcal{U}=\{q,\theta\}$1

If the distance exceeds a calibrated threshold, the answer is flagged as potentially unfaithful (Yu et al., 7 Apr 2026). The method is explicitly a runtime monitor for whether the answer is supported by retrieved evidence, not a world-truth verifier, and its decision rule is simple enough for quantized public verification (Yu et al., 7 Apr 2026). This suggests a distinct audit mode: not post-hoc judging of text, but real-time faithfulness monitoring from residual-stream geometry.

4. Agentic, interactive, and long-horizon auditing

In agentic systems, faithfulness auditing extends from isolated answers to persistent state and trajectories. SLUMP defines faithfulness loss under emergent specification as the reduction in final implementation faithfulness when a target design is progressively disclosed through interaction rather than given in a single upfront prompt (Yan et al., 17 Mar 2026). The benchmark contains 20 recent ML papers, 371 atomic verifiable components, and scripts of approximately 60 coding requests per paper. Final repositories are scored with a five-level component-faithfulness rubric, while an exposure audit verifies whether scored components are recoverable from the visible interaction. Across the benchmark, macro-averaged RCR is 0.994, macro-averaged ESR is 0.666, and 358 of 371 components are both committed and recoverable (Yan et al., 17 Mar 2026).

SAVeR treats reasoning trajectories as internal belief states that guide action selection, tool invocation, memory updates, and future decisions (Yuan et al., 9 Apr 2026). It defines a step-support function $\mathcal{U}=\{q,\theta\}$2 and trajectory-level unfaithfulness rate

$\mathcal{U}=\{q,\theta\}$3

The framework generates persona-based candidate beliefs, selects structurally diverse candidates with a $\mathcal{U}=\{q,\theta\}$4-Determinantal Point Process, audits them with a violation taxonomy that includes Missing_Assumption, Invalid_Precondition, Unjustified_Inference, Circular_Reasoning, Contradiction, and Overgeneralization, and performs constraint-guided minimal repair before action commitment (Yuan et al., 9 Apr 2026). Its explicit critique is that consensus mechanisms conflate agreement with faithfulness (Yuan et al., 9 Apr 2026).

DFAH targets regulatory replay in tool-using financial agents. It records full trajectories and evaluates Action Determinism, Signature Determinism, and Decision Determinism, while faithfulness is defined as evidence-conditioned alignment rather than absolute truth (Khatchadourian, 17 Jan 2026). The paper distinguishes pass@k from pass$\mathcal{U}=\{q,\theta\}$5 and argues that pass$\mathcal{U}=\{q,\theta\}$6 is the relevant audit metric for compliance replay. This broadens faithfulness audit into a joint assurance problem: reproducibility of the trajectory and evidence grounding of the rationale (Khatchadourian, 17 Jan 2026).

Audit-LLM addresses a related problem in log-based insider threat detection. Its Decomposer, Tool Builder, and Executor architecture grounds judgments in tool-derived evidence, while the pair-wise Evidence-based Multi-agent Debate mechanism is designed to correct cases where the final synthesized judgment drifts away from the log evidence (Song et al., 2024). The paper does not define a formal faithfulness score, but it treats faithfulness hallucination as misalignment between final verdicts and sub-task findings (Song et al., 2024).

5. Benchmarks and empirical regularities

Several empirical patterns recur across benchmarks. First, answer correctness is not a reliable proxy for faithfulness. RFEval reports unfaithfulness in 49.7% of outputs and finds that, after controlling for model and task fixed effects, the accuracy-faithfulness relationship is weak and statistically insignificant, with weighted Pearson $\mathcal{U}=\{q,\theta\}$7, $\mathcal{U}=\{q,\theta\}$8, and weighted Spearman $\mathcal{U}=\{q,\theta\}$9, $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$0 (Han et al., 19 Feb 2026).

Second, factual and scientific reasoning often show strong decoupling between stated reasoning and terminal answers. Project Ariadne evaluates 500 queries across General Knowledge, Scientific Reasoning, and Mathematical Logic, reporting mean faithfulness $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$1, similarity $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$2, and violation rate 92% for General Knowledge; $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$3, $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$4, and violation rate 96% for Scientific Reasoning; and $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$5, $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$6, and violation rate 20% for Mathematical Logic (Khanzadeh, 5 Jan 2026). It also reports an aggregate violation density $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$7 across 30 distinct reasoning traces (Khanzadeh, 5 Jan 2026).

Third, metric performance is frequently weak even when the task is precisely defined. BonaFide reports that most prominent faithfulness metrics perform near chance, with the best metric reaching only 0.70 AUROC at the CoT level and another reaching 0.59 at the step level (Gur-Arieh et al., 24 May 2026). “A Causal Lens for Evaluating Faithfulness Metrics” similarly finds that all tested faithfulness metrics often fail to surpass a random baseline and that no metric exceeds 0.50 across all models and tasks (Zaman et al., 26 Feb 2025).

Fourth, automatic factuality evaluation remains difficult in summarization. STORYSUMM reports that none of the evaluated automatic metrics achieve more than 70% balanced accuracy on narrative summarization (Subbiah et al., 2024). In dialogue summarization, fine-grained human analysis on SAMSum finds that about 17% of human reference summaries are factually inconsistent and that 36%–50% of machine-generated summaries contain at least one factual error (Wang et al., 2022). These results reinforce the claim that fluent summaries can remain unfaithful even under apparently strong human or automatic evaluation.

Fifth, some white-box support monitors can be strong and efficient. On PubMedQA with Llama-3-8B, LatentAudit reaches 0.942 AUROC with 0.77 ms overhead, and under a four-way stress benchmark it reaches 0.9566–0.9815 AUROC on PubMedQA and 0.9142–0.9315 on HotpotQA (Yu et al., 7 Apr 2026). This suggests that, in some settings, faithfulness signals in latent geometry are operationally usable.

6. Limitations, mitigation, and research directions

Faithfulness audits remain limited by the availability and validity of ground truth. Coverage-aware grounded-generation evaluation depends on a complete oracle, but the paper is explicit that recall is measured relative to the derived oracle and that event detection is heuristic (Santillana, 8 Jun 2026). Summary evaluation is complicated by genuine ambiguity: in the MeetingBank ambiguity-annotated set introduced by Madisse, 131 of 770 sentences are marked ambiguous after adjudication (Koupaee et al., 12 Feb 2025). GeoFaith states directly that human-observable step labels are only a proxy for true internal faithfulness and that step-level counterfactual verification would be more direct but scales as $\mathcal{V}=\{s_1,s_2,\dots,s_n,a\}$8 (Lv et al., 26 May 2026).

Audit validity is also sensitive to the audit mechanism itself. EDCT depends on LLM judges and on the quality of the counterfactual image edit; the authors therefore recommend robustness checks, ensemble-judge variants, segmentation masks, and similarity measures such as LPIPS (Ding et al., 27 Sep 2025). LatentAudit requires open-weight access to hidden states and verifies support from retrieved evidence, not truthfulness in the world (Yu et al., 7 Apr 2026). DFAH warns against using an LLM to judge another LLM’s faithfulness in regulated replay settings because that introduces recursive non-determinism (Khatchadourian, 17 Jan 2026).

Mitigation work increasingly tries to make faithfulness a first-class optimization target rather than a pure evaluation target. “Precision Is Not Faithfulness” proposes verifier-guided generation that feeds back contradicted claims and uncovered oracle facts (Santillana, 8 Jun 2026). Faithfulness Serum introduces training-free attribution guidance during explanation generation (Alon et al., 15 Apr 2026). AoU constrains inference to validated premises through decomposition, auditing, and posterior-constrained inference, with abstention when the validated evidence is insufficient (Abdaljalil et al., 11 Oct 2025). SLUMP presents ProjectGuard as an external project-state layer for preserving durable design commitments in long-horizon coding sessions (Yan et al., 17 Mar 2026). SAVeR enforces verification over internal belief states before action commitment (Yuan et al., 9 Apr 2026). Project Ariadne explicitly suggests multi-step path-specific interventions and faithfulness as a training objective in RLHF or DPO (Khanzadeh, 5 Jan 2026).

A plausible implication is that faithfulness audit is becoming a layered methodology rather than a single score. Across reasoning traces, grounded generation, RAG, summarization, vision-language explanation, and tool-using agents, the strongest audits are those that test whether changing the purported grounds of a decision changes the decision itself, whether the stated grounds cover the facts that mattered, and whether the resulting trajectory can be replayed, inspected, and verified.