Semantic Faithfulness in NLP

• Semantic Faithfulness is defined as the preservation of semantic content and causal reasoning in model outputs, distinct from mere human plausibility.
• It is evaluated using methodologies such as deletion, negation, and retrieval attribution tests, along with metrics like LSS and semantic alignment scores.
• Research highlights a graded, task-specific approach to measuring faithfulness, addressing trade-offs between fluency, correctness, and true semantic preservation.

Semantic faithfulness denotes the property that an output, explanation, attribution, or formalization preserves or reflects the semantic content and decision process it is supposed to represent. In NLP interpretability, a faithful explanation is one that “accurately represents the reasoning process behind the model’s prediction,” rather than one that is merely plausible, readable, or persuasive (Jacovi et al., 2020). In retrieval-augmented generation, a faithful citation requires not only support from the cited document but also causal influence on generation (Wallat et al., 2024). In question answering, semantic content should causally figure in a model’s inferences under interventions on the input text (Chaturvedi et al., 2022). Across translation, grammatical error correction, text-to-image synthesis, and autoformalization, the recurring requirement is that semantics be genuinely preserved or used, not merely matched at the surface.

1. Conceptual foundations

A central theme in the literature is that semantic faithfulness is distinct from human plausibility. “Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?” defines faithfulness as an interpretation that accurately represents the reasoning process behind a model’s prediction, while plausibility concerns how convincing or understandable an interpretation is to humans (Jacovi et al., 2020). The same work argues that there is no single formal community-wide definition and organizes faithfulness evaluation around three assumptions: the Model Assumption, the Prediction Assumption, and the Linearity Assumption. The Model Assumption links sameness of predictions to sameness of reasoning; the Prediction Assumption links similarity of inputs and outputs to similarity of reasoning; the Linearity Assumption treats parts of the input as independently contributing to the decision and motivates feature-attribution tests such as erasure, comprehensiveness, and sufficiency (Jacovi et al., 2020).

The causal character of semantic faithfulness is explicit in intervention-based work on question answering. “Analyzing Semantic Faithfulness of LLMs via Input Intervention on Question Answering” formalizes semantic faithfulness so that the semantic content of a text should causally figure in a model’s inferences: if semantically relevant material is deleted or negated, the model’s answer should change in the corresponding way (Chaturvedi et al., 2022). A related causal requirement appears in retrieval attribution, where a citation is faithful only if the cited document is retrieved, supports the statement, and causally impacts the model’s generation of that statement (Wallat et al., 2024).

The literature also resists treating faithfulness as an all-or-nothing property. The 2020 interpretability survey argues that the current binary definition sets a potentially unrealistic bar and advocates a graded notion of faithfulness across models and tasks and across the input space (Jacovi et al., 2020). This suggests that semantic faithfulness is often evaluated locally, comparatively, or under restricted perturbation regimes rather than certified globally.

2. Distinctions from neighboring criteria

The most influential distinction in recent retrieval-augmented generation work is between correctness and faithfulness. “Correctness is not Faithfulness in RAG Attributions” defines citation correctness as a content-level property: the cited document supports the statement and is not factually wrong. Citation faithfulness is stronger: the statement must be supported by the cited document and also causally influenced by it during generation (Wallat et al., 2024). This yields the now-standard contrast between being “right for the right reason” and “right for the wrong reason,” and motivates the notion of post-rationalization, where a model answers from parametric memory and only later attaches a matching citation (Wallat et al., 2024).

An analogous distinction appears in formal reasoning. “Do LLMs Game Formalization? Evaluating Faithfulness in Logical Reasoning” separates proof validity from formalization faithfulness: a proof is valid if it type-checks in Lean’s kernel, but a formalization is faithful only if the axioms and theorem preserve the meaning of the original natural-language premises and conclusion (Kim et al., 21 Apr 2026). “The Faithfulness Gap: Certifying Semantic Equivalence Between Natural-Language and Formal Mathematical Statements” makes the same point for autoformalization: a formal statement can typecheck and be provable, yet still encode a different theorem than the source intended (Mohammad et al., 15 Jun 2026).

In generation tasks, faithfulness is likewise distinct from output quality measures that do not directly track meaning preservation. In grammatical error correction, “Reference-less Measure of Faithfulness for Grammatical Error Correction” treats grammaticality and semantic faithfulness as separate dimensions: an output may be grammatical yet semantically unfaithful by changing, adding, or removing meaning (Choshen et al., 2018). In literary translation, “Fluency and Faithfulness in Human and Machine Literary Translation” operationalizes fluency as original-likeness and faithfulness as preservation of source meaning via COMET-KIWI, then reports a consistent negative correlation between fluency and faithfulness after controlling for paragraph length (Griebel et al., 14 May 2026).

Open-domain question answering introduces a further distinction between faithfulness and correctness. “From Faithfulness to Correctness: Generative Reward Models that Think Critically” defines faithfulness as semantic alignment with supporting documents and correctness as objective factual accuracy; a sentence can therefore be faithful but incorrect if the source is wrong or outdated, or unfaithful but correct if it is true yet unsupported by the provided documents (Ma et al., 29 Sep 2025).

3. Formalizations and metrics

Several strands of work operationalize semantic faithfulness by demanding explicit support relations. In retrieval attribution, a generated statement $s$ with citation $a_j$ is faithful only if $a_j \in A$, $s$ is supported by $a_j$, and $s$ is causally impacted by $a_j$ (Wallat et al., 2024). The first two conditions are necessary but not sufficient, which makes the evaluation of causal dependence central rather than optional.

Other work measures faithfulness by isolating the supported part of a claim. “Evaluation of Faithfulness Using the Longest Supported Subsequence” defines the Longest Supported Subsequence (LSS) as the longest subsequence of words from the claim that is supported by the reference. Faithfulness is then computed as $M(c,\mathrm{LSS}(r,c))$, where $M$ may be BLEU, ROUGE, BERTScore, FactCC, QuestEval, or QAFactEval (Mittal et al., 2023). Because LSS need not be continuous and may be ungrammatical, it can retain supported content on both sides of a hallucinated span, and the paper reports that using LSS improves correlation with human faithfulness ratings relative to direct reference-claim comparison (Mittal et al., 2023).

For meaning preservation under editing, USim formalizes faithfulness as semantic graph similarity between source and correction using UCCA. Source and corrected sentences are parsed into UCCA DAGs, their leaves are aligned by weighted bipartite matching over token edit distances, higher-level nodes are aligned through yield overlap, and precision, recall, and F-score are computed over matched semantic edges (Choshen et al., 2018). The measure is reference-less and is intended to complement grammaticality-oriented reference-less measures, not replace them (Choshen et al., 2018).

A distinct line of work treats faithfulness as agreement between semantic transformations. “Semantic Faithfulness and Entropy Production Measures to Tame Your LLM Demons and Manage Hallucinations” represents question, context, and answer as topic distributions and introduces transition matrices $\mathbf{Q}$ and $a_j$0 for the question-induced and answer-induced transformations of context topics. Semantic Faithfulness is defined as $a_j$1, where $a_j$2 is the minimal KL divergence between the feasible answer and question transition operators (Halperin, 4 Dec 2025). The same work introduces Semantic Entropy Production as a complementary measure of irreversibility in answer generation and reports a moderate negative correlation between SF and SEP in simulations (Halperin, 4 Dec 2025).

Prompt-aware black-box detection uses yet another formalization. “Prompt-Response Semantic Divergence Metrics for Faithfulness Hallucination and Misalignment Detection in LLMs” generates paraphrases of a prompt, samples multiple answers for each paraphrase, jointly clusters prompt and answer sentences into a shared topic space, and computes divergence measures over prompt-response topic distributions (Halperin, 13 Aug 2025). Its practical instability score is

$a_j$3

and the paper treats $a_j$4 as a signal of Semantic Exploration rather than hallucination per se (Halperin, 13 Aug 2025).

4. Causal tests, interventions, and online monitoring

Counterfactual testing is a recurring strategy for probing whether semantics genuinely drive model behavior. In retrieval attribution, one practical necessary condition for faithfulness is that changing the cited document’s relevant information should change the model’s behavior; another is that adding irrelevant documents should not affect attribution if the answer remains unchanged (Wallat et al., 2024). The post-rationalization experiment in Command-R+ on 1,444 NaturalQuestions instances operationalizes the second condition by inserting adversarial documents containing answer statements and checking whether the model newly cites them. The same statement could be recovered in 63–70% of cases, and the inserted adversarial document was cited in 12% of random-document cases, 57% of relevant-but-uncited cases, and 55% of documents-cited-for-other-reasons cases, leading the authors to report up to 57% unfaithful citations in their setup (Wallat et al., 2024).

Intervention on the source text provides a stronger causal probe for question answering. “Analyzing Semantic Faithfulness of LLMs via Input Intervention on Question Answering” introduces deletion intervention, which removes the rationale, and negation intervention, which changes the story so that the modified text supports the opposite answer (Chaturvedi et al., 2022). Standardly trained BERT, RoBERTa, and XLNet models continue to predict the original answer in roughly 50% of deletion cases and suffer about a 20% drop in accuracy under negation intervention. An intervention-based training regime reduces deletion-intervention failure from about 50% to about 6% while leaving other aspects of semantic unfaithfulness, such as negation handling and predicate-argument structure, largely unresolved (Chaturvedi et al., 2022).

For explanation methods, “Faithfulness and the Notion of Adversarial Sensitivity in NLP Explanations” proposes a necessary test: if $a_j$5 is semantically close to $a_j$6 yet $a_j$7, a faithful explainer should produce meaningfully different explanations for the two inputs (Manna et al., 2024). Adversarial sensitivity is defined as the distance between local explanation vectors, using a correlation-based distance over rankings, and is intended to reject explainers that remain invariant when the model’s own reasoning must have changed (Manna et al., 2024).

Online monitoring methods shift from offline diagnosis to synchronous detection. “Synchronous Faithfulness Monitoring for Trustworthy Retrieval-Augmented Generation” introduces SynCheck, a sentence-level monitor that combines sequence likelihood, uncertainty quantification, context influence, and semantic alignment via align_score from AlignScore-base (Wu et al., 2024). On six long-form RAG tasks, the monitor achieves 0.85 AUROC for faithfulness error detection and supports a faithfulness-oriented decoding algorithm, FOD, that improves faithfulness by over 10% across six datasets while preserving more informativeness than abstention (Wu et al., 2024). A related but cheaper evaluation line adapts NLI models to faithfulness detection: “With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness” combines dialogue-specific augmentation, the score $a_j$8 from entailment and contradiction probabilities, and Monte-Carlo dropout, improving a vanilla NLI model on TRUE from 79.8 average AUC to 83.2, and an ensemble to 86.8 (Steen et al., 2023).

5. Domain-specific instantiations

In text-to-image synthesis, semantic faithfulness concerns prompt-image correspondence at the level of objects, attributes, counts, and spatial relations. “Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis” uses a large vision-LLM as a semantic judge and optimizes the initial noisy latent according to the VQA score

$a_j$9

The latent update preserves distribution consistency by keeping the updated latent in $a_j \in A$0, and the paper reports on Stable Diffusion V-1.4 that VQA reaches 0.9352 after 10 epochs and 0.9790 after 50 epochs on simple prompts, with larger gains on complex spatial prompts (Miao et al., 2024).

In affective computing, faithfulness can be architectural rather than post hoc. “Faithful by Definition: Emotion Analysis via Natural Semantic Metalanguage Explications” decomposes prediction into a learned parser $a_j \in A$1 that maps input text to a 12-slot Emotion Explication Schema and a fixed symbolic mapper $a_j \in A$2 that applies a priority-ordered decision list of 13 emotion rules, so that $a_j \in A$3 (Xing et al., 1 Jul 2026). The paper describes the resulting guarantee as causal and definitional for the explication-to-label segment, with per-line entailment verification providing an audit trail for the parser’s output (Xing et al., 1 Jul 2026).

In formal mathematics, semantic faithfulness becomes semantic equivalence between natural-language and formal statements. “The Faithfulness Gap: Certifying Semantic Equivalence Between Natural-Language and Formal Mathematical Statements” introduces an interpretation distribution $a_j \in A$4 over acceptable formalizations of a natural-language statement $a_j \in A$5, and defines a candidate $a_j \in A$6 as $a_j \in A$7-faithful if

$a_j \in A$8

Its Bidirectional Provability Fingerprinting framework compares forward and backward consequence neighborhoods against probes derived from the natural-language statement. On DRIFTBENCH, BPF + CPG detects 89.6% of drifted formalizations at a 3.0% false-positive rate, and Faithfulness-Guided Decoding reduces the drifted-output rate of a state-of-the-art autoformalizer by 47% (Mohammad et al., 15 Jun 2026).

A related issue appears in Lean proof generation. “Do LLMs Game Formalization? Evaluating Faithfulness in Logical Reasoning” finds no evidence of systematic formalization gaming in unified generation, but the two-stage pipeline exposes two different unfaithfulness modes: GPT-5 fabricates axioms during proof generation, especially “conclusion as axiom,” whereas DeepSeek-R1 mistranslates premises during formalization in a way that may evade cross-stage detection (Kim et al., 21 Apr 2026).

For sample-based explanations in NLP, semantic faithfulness can also attach to explanation units themselves. “On Sample Based Explanation Methods for NLP: Efficiency, Faithfulness, and Semantic Evaluation” distinguishes faithfulness to the final trained model from semantic alignment between explanation spans and test spans, introduces span-level influence variants, and evaluates them with Semantic Agreement, a cosine-based similarity between span embeddings derived from masked and unmasked sequences (Zhang et al., 2021). The work reports that span-level methods, especially TracIn++, align better with human judgment than label-agreement or retraining-based measures (Zhang et al., 2021).

6. Methodological disputes and open problems

A major dispute concerns whether faithfulness is even measurable as a single scalar. “Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation” applies three classifiers to the same 10,276 influenced reasoning traces and obtains overall faithfulness rates of 74.4%, 82.6%, and 69.7%, with non-overlapping 95% confidence intervals for the most permissive and strictest classifiers (Young, 20 Mar 2026). Per-model gaps range from 2.6 to 30.6 percentage points, Cohen’s $a_j \in A$9 ranges from 0.06 to 0.42 across hint types, and model rankings can reverse depending on whether the classifier operationalizes faithfulness as lexical mention or epistemic dependence (Young, 20 Mar 2026). This suggests that published faithfulness numbers are often classifier-relative rather than instrument-independent.

The broader NLG survey literature frames these tensions as part of a larger unsolved problem. “Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods” organizes evaluation into entailment-based, QA-based, fact-based, and other metrics, and optimization into factual guidance, auxiliary tasks, learning methods, post-editing, constrained decoding, and others (Li et al., 2022). It emphasizes intrinsic versus extrinsic hallucinations, source-reference divergence, exposure bias, and overreliance on language priors, while noting that even the best metric correlations with human factuality labels are generally below 0.6 (Li et al., 2022).

Recent work also shows that low instability is not the same as truth. In SDM, a deliberately nonsensical forced-hallucination prompt yields the lowest $s$0, because the model converges on a stable, repetitive, semantically decoupled response strategy rather than an erratic one (Halperin, 13 Aug 2025). The paper identifies this as the dangerous case of confident confabulation. A related caution appears in open-domain QA reward modeling, where faithfulness to supporting documents is necessary but insufficient: TRM explicitly inserts a reasoning step between faithfulness judgment and correctness judgment so that a sentence can be faithful yet still wrong (Ma et al., 29 Sep 2025).

Another unresolved issue is trade-off. In literary translation, partial Spearman correlation controlling for paragraph length is negative overall ($s$1), negative for human translations and Google Translate, and weakly positive overall for TranslateGemma, indicating that more original-like fluency does not reliably imply greater semantic preservation (Griebel et al., 14 May 2026). In interpretability, the graded view of faithfulness was proposed precisely because strict binary faithfulness is likely a “unicorn” (Jacovi et al., 2020).

A plausible implication is that semantic faithfulness is best understood as a family of causally and semantically constrained relations whose exact operationalization depends on the task: support by a source, causal dependence on retrieved evidence, preservation of meaning under editing, semantic equivalence under formalization, or prompt-conditioned alignment in multimodal generation. The common methodological requirement is not a single metric, but explicit testing of whether the semantics that should matter actually do matter.