Edit Fidelity Score (EFS)

• Edit Fidelity Score (EFS) is a context-dependent metric that measures whether a planned revision is accurately executed in a designated region.
• In domains like music generation and news captioning, EFS is operationalized through weighted precision metrics or entity F1 scores to balance localized edits with preservation of existing content.
• EFS techniques emphasize origin-aware comparisons and spillover suppression to isolate edit regions, addressing limitations of whole-output similarity evaluations.

Edit Fidelity Score (EFS) is a context-dependent label for evaluation constructs that measure whether an intended revision is realized faithfully while preserving what should remain unchanged. In recent arXiv usage, the term is not standardized: in editable music generation it denotes a metric for localized, plan-driven audio edits, whereas in news video captioning the same acronym denotes an entity-recovery F1 over named entities (Wang et al., 26 Sep 2025, Paredes et al., 29 Mar 2026). Several adjacent literatures do not define EFS as a named scalar, but they operationalize edit fidelity through origin-aware revision similarity, spillover suppression, evidence grounding, or downstream edited-output agreement, which together delineate a broader technical family of edit-fidelity evaluation (Gruzinov et al., 24 Oct 2025, Li et al., 19 Apr 2026).

1. Conceptual scope

The recurring motivation for EFS-like metrics is an evaluation gap left by whole-output similarity measures. In MusicWeaver, conventional metrics such as KL, IS, FD, FAD, PC, PQ, and alignment scores do not answer whether a requested edit actually happened in the specified region or whether the change remained localized (Wang et al., 26 Sep 2025). In revision similarity, pairwise measures such as BLEU are dominated by the large amount of content shared by the original document and its revisions, so they can assign high similarity to two revisions that differ precisely in the edited region (Gruzinov et al., 24 Oct 2025). In generative document parsing, CER, WER, IoU, and TEDS can penalize semantically correct yet structurally divergent outputs, thereby conflating interpretive diversity with error (Li et al., 16 Sep 2025).

These formulations converge on a common premise: edit fidelity is not identical to generic output fidelity. It is defined relative to an origin, a target region, a reference revision, or an explicit control interface. The corresponding score must therefore isolate changed content from inherited context, distinguish desired from undesired modifications, and, in some settings, tolerate alternative but semantically valid realizations. This suggests that EFS is best understood as a task-conditioned evaluation principle rather than a single universal metric.

2. Explicit definitions of EFS

Where EFS is defined explicitly, its formalization is domain-specific.

Domain	Meaning of EFS	Core formalism
Music generation	Edit Fidelity Score	Weighted combination of constraint satisfaction and edit localization
News video captioning	Entity Fidelity Score	Named-entity precision/recall F1 with fuzzy matching

In MusicWeaver, EFS is defined as

$$\mathrm{EFS} = 100\big[\alpha\,p_{\mathcal{E}} + (1-\alpha)\,r_{\Delta}\big],$$

where $p_{\mathcal{E}}$ is the success rate of edit constraints in the target region and $r_{\Delta}$ measures localization as the fraction of total change occurring inside the edited region (Wang et al., 26 Sep 2025). The components are

$$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$

with

$$\Delta_{\mathrm{in}} = \|\Phi(y')_{R} - \Phi(y)_{R}\|_1, \qquad \Delta_{\mathrm{out}} = \|\Phi(y')_{\overline{R}} - \Phi(y)_{\overline{R}}\|_1.$$

Here $y$ is original audio, $y'$ is edited audio, $R$ is the target edit region, $\overline{R}$ its complement, and $\Phi(\cdot)$ a feature map. By construction, EFS lies in $p_{\mathcal{E}}$0 and is complementary to Structure Coherence Score (SCS): SCS measures long-range structure and timing coherence, whereas EFS measures whether local, plan-driven edits are realized and remain localized.

In news video captioning, EFS instead means Entity Fidelity Score and is defined as an F1 score over named entities extracted from ground-truth and generated captions (Paredes et al., 29 Mar 2026). Let $p_{\mathcal{E}}$1 be the ground-truth entity set for clip $p_{\mathcal{E}}$2, $p_{\mathcal{E}}$3 the model entity set, $p_{\mathcal{E}}$4 the matched ground-truth entities, and $p_{\mathcal{E}}$5 the matched model entities. Then

$p_{\mathcal{E}}$6

and

$p_{\mathcal{E}}$7

This EFS uses spaCy en_core_web_sm NER over PERSON, GPE, ORG, LOC, NORP, FAC, and EVENT, with RapidFuzz token-ratio similarity threshold $p_{\mathcal{E}}$8. Clips with no ground-truth entities are excluded from EFS evaluation. The metric is therefore factual and reference-anchored rather than locality-based.

3. EFS-like constructions in adjacent literatures

Several papers do not define an EFS by name, but they instantiate the same evaluation logic. In Evidence-Diagnosed Intervention Training (EDIT), high-fidelity edits to grading reasoning are diagnosed through posterior belief over the final mark, step-wise error drift $p_{\mathcal{E}}$9, grounding scores $r_{\Delta}$0, $r_{\Delta}$1, and $r_{\Delta}$2, and rule-faithfulness metrics such as RSR and IB-RSR (Wu et al., 4 Jun 2026). The paper explicitly states that it does not define an “Edit Fidelity Score,” but its machinery is organized around correcting harmful local reasoning steps while preserving rubric-faithful, evidence-grounded steps.

In edit-aware RAW reconstruction, edit fidelity is likewise not a named score but is operationalized as agreement between edited sRGB renderings produced from reconstructed RAW and from ground-truth RAW under the same differentiable ISP or real editing pipeline (Punnappurath et al., 5 Dec 2025). The core loss is

$r_{\Delta}$3

which directly optimizes downstream edited-output fidelity rather than RAW-space pixel fidelity.

Speech editing adopts another EFS-like configuration. MAVE does not define an explicit EFS, but it evaluates edit fidelity through WER against target transcripts, MOS for naturalness and intelligibility, and pairwise indistinguishability judgments between original and edited speech (Mohammad et al., 6 Oct 2025). On a random 40-sample subset of RealEdit, 57.2% of judgments rated edited speech as perceptually equal to the original, 24.8% preferred the original, and 18.0% preferred MAVE, making “equal” and “edited better” together a direct proxy for edit invisibility.

Scene text editing reframes edit fidelity as the suppression of edit spillover into non-target text regions (Li et al., 19 Apr 2026). Spillover at region $r_{\Delta}$4 is defined by a binary decision over OCR text similarity and PSNR, and the proposed Edit Fidelity Field enforces a four-zone control field: Edit Core, Transition Zone, Protected Zone, and Background. The paper explicitly describes these ingredients as the basis for a potential EFS combining target correctness with non-target preservation.

In grammatical error correction, “Scoring Edit Impact in GEC” proposes an edit-level ranking framework based on an embedded association graph and marginal fluency gain (Xiao et al., 8 Apr 2026). The per-edit impact

$r_{\Delta}$5

is not named EFS, but it functions as an edit-importance score over grouped edits, which is precisely the kind of signal an edit-fidelity metric would require.

4. Recurrent mathematical components

Across these formulations, several components recur.

Target-side correctness: one term measures whether the requested or reference edit is actually realized. In MusicWeaver this is $r_{\Delta}$6 over plan-derived constraints; in scene text editing it corresponds to target OCR correctness or “Target Found”; in RAW reconstruction it is the agreement between edited renderings under the same ISP parameters (Wang et al., 26 Sep 2025, Li et al., 19 Apr 2026, Punnappurath et al., 5 Dec 2025).

Preservation outside the edit: a second term measures whether non-target content remains stable. MusicWeaver uses $r_{\Delta}$7 to quantify the fraction of change inside the target region; scene text editing measures spillover region by region and reports AvgPSNR, MinPSNR, and background PSNR; voice editing uses pairwise indistinguishability from the original as a perceptual version of the same principle (Wang et al., 26 Sep 2025, Li et al., 19 Apr 2026, Mohammad et al., 6 Oct 2025).

Origin-aware comparison: Excision Score formalizes revision similarity as a triplet problem $r_{\Delta}$8 and excises shared content before scoring only the divergent regions (Gruzinov et al., 24 Oct 2025). Its definition,

$r_{\Delta}$9

encodes the principle that edit fidelity should be invariant to shared context. This is the strongest explicit statement, in the set of papers considered here, that unchanged material must not dominate an edit score.

Factual grounding diagnostics: Entity Fidelity Score measures recovery of named entities rather than generic semantic similarity; SCORE for document parsing distinguishes TokensFound from TokensAdded to separate omissions from hallucinations; EDIT diagnoses low-fidelity reasoning steps through grounding to rubric, answer, and prefix (Paredes et al., 29 Mar 2026, Li et al., 16 Sep 2025, Wu et al., 4 Jun 2026). These mechanisms differ technically, but all isolate whether the edited output remains anchored to the intended source of truth.

5. Evaluation protocols and reported behavior

Reported values are task-specific and not directly comparable across domains, but they show how EFS-like metrics behave once the edit target is explicitly specified.

Setting	Reported outcome	Interpretation
MusicWeaver	MusicCaps T2M EFS 72.6; V2M-bench T2M 73.8; V2M 73.1; TV2M 70.1	Local plan edits are realized with substantial localization
BBC news captioning	Gemma 3 EFS 0.314; Qwen-VL 0.198	Entity-rich references make EFS discriminative
ChTV news captioning	EFS range 0.002 to 0.059	Sparse entity references make EFS weakly informative
Scene text editing with EFF	Spillover 94% to 25%; Avg Region PSNR 22.4 dB to 113.8 dB; Target Found 84% to 60%	Preservation can improve while target accuracy drops
Excision Score on HumanEvalFix	12% Pearson improvement over SARI; >21% over BLEU	Shared-context-invariant edit scoring correlates better with execution
MAVE pairwise study	57.2% equal, 24.8% original preferred, 18.0% edited preferred	Human indistinguishability functions as a perceptual edit-fidelity criterion

MusicWeaver further reports ablations showing that timing-related control signals materially affect EFS: removing phase channels lowers EFS from 72.6 to 68.4, removing boundary ramps lowers it to 69.5, removing BPM control yields 69.7, and removing control dropout lowers it to 67.9 (Wang et al., 26 Sep 2025). This ties edit fidelity to explicit temporal alignment mechanisms rather than to generic diffusion quality alone.

The news video captioning study demonstrates that EFS can expose behavior not visible in lexical metrics. On BBC, Gemma 3 leads on EFS with 0.314, while LLaVA-OneVision can lead on ROUGE-L and BERTScore yet remain low on EFS at 0.059 (Paredes et al., 29 Mar 2026). The result is a direct warning that token-level similarity and entity fidelity are separable dimensions.

Scene text editing provides the clearest preservation trade-off. EFF reduces spillover rate from 94% to 25% and improves Avg Region PSNR by +91.4 dB, but Target Found decreases from 84% to 60% (Li et al., 19 Apr 2026). This makes explicit that edit fidelity is often a balance between target realization and collateral preservation rather than a monotone single-objective optimization.

6. Limitations, ambiguities, and misconceptions

A central ambiguity is terminological. “EFS” is not a stable cross-domain standard: it can mean Edit Fidelity Score, Entity Fidelity Score, or merely a convenient label for a composite measure that a paper does not formally instantiate (Wang et al., 26 Sep 2025, Paredes et al., 29 Mar 2026, Wu et al., 4 Jun 2026). This suggests that the term should be interpreted through its operands and protocol rather than through the acronym alone.

The metric is also highly representation-dependent. MusicWeaver requires a set of edit constraints $$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$0, a region $$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$1, a feature map $$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$2, and a weighting factor $$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$3, and the paper does not state the exact $$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$4 used (Wang et al., 26 Sep 2025). Entity Fidelity Score depends on spaCy NER quality, a fuzzy-matching threshold $$p_{\mathcal{E}} = \frac{1}{|\mathcal{E}|}\sum_{c \in \mathcal{E}} \mathbf{1}\{f_c(y',R)\}, \qquad r_{\Delta} = \frac{\Delta_{\mathrm{in}}}{\Delta_{\mathrm{in}}+\Delta_{\mathrm{out}}},$$5, and entity-rich references; when ground truth is entity-sparse, as in ChTV, the metric becomes weakly informative (Paredes et al., 29 Mar 2026). Scene-text spillover metrics depend on OCR coverage, region padding, and a single-target setting, with very small text potentially escaping OCR protection (Li et al., 19 Apr 2026). Excision Score uses an approximate rather than exact 3-way LCS and remains fundamentally syntactic, only partially satisfying semantic-equivalence desiderata (Gruzinov et al., 24 Oct 2025).

A frequent misconception is that high whole-output similarity implies high edit fidelity. The cited papers repeatedly reject that equivalence. In revision similarity, BLEU can converge toward 1 as shared context grows, even when the actual edits are wrong (Gruzinov et al., 24 Oct 2025). In document parsing, traditional edit distance and TEDS can penalize semantically valid structural alternatives by 12–25% on ambiguous tables (Li et al., 16 Sep 2025). In news captioning, high ROUGE-L or BERTScore does not guarantee faithful entity recovery (Paredes et al., 29 Mar 2026). Edit fidelity therefore requires edit-aware isolation, not merely stronger generic similarity.

In this literature, EFS is best regarded as a family of origin-aware, region-aware, or control-aware scoring constructions whose common purpose is to measure fidelity of change rather than fidelity of the final artifact in the abstract. The exact score varies by task, but the underlying requirement is stable: the metric must identify what was supposed to change, what actually changed, and what should have remained intact.