Element-Level Factual Correction

• Element-Level Factual Correction is a method that repairs specific factual units (tokens, entities, triples) using evidence-based, minimally invasive editing.
• It applies techniques like masking, span selection, and constrained rewriting to ensure corrections align closely with retrieved evidence.
• The approach is crucial in applications such as claim verification, summary post-editing, and retrieval-augmented generation, emphasizing efficiency and accuracy.

Element-level factual correction is the family of factual error correction procedures that repair localized factual units in text rather than regenerate an entire passage indiscriminately. In the core evidence-based formulation, the task is to edit a claim so that the generated rewrite is better supported by evidence, extending fact verification from label prediction to edit generation (Thorne et al., 2020). Across the literature, the editable unit ranges from tokens and spans (Thorne et al., 2020, Chen et al., 2022), to entities and factual factors (Lee et al., 2022, Li et al., 2024), to triples, atomic facts, and atoms (Bayat et al., 2023, Li et al., 2024, Carnerero-Cano et al., 16 Jan 2026). The topic now spans FEVER-style claim correction, abstractive and dialogue summarization, retrieval-augmented generation, long-form answer revision, and paragraph-level professional writing (Huang et al., 2023, Gao et al., 2023, Liu et al., 13 Nov 2025, Kai et al., 27 Feb 2026).

1. Problem formulation and minimal-edit objective

Evidence-based factual error correction defines the task as a generation problem: given a claim $c$ and retrieved evidence $E(c)$, produce a rewritten claim $c'$ such that the evidence entails the rewrite, $E(c)\vDash c'$ (Thorne et al., 2020). This differs from fact verification, which predicts labels such as Supported, Refuted, or NotEnoughInfo for a claim–evidence pair. The distinction is operationally important: the correction system must preserve fluency, remain close to the original claim, and repair the factual error rather than drifting to a different true statement. VenCE makes this minimal-edit requirement explicit by defining the desired output as truthful, fluent, and minimally modified relative to the original claim sentence $x^0$ (Chen et al., 2022).

Several later formulations sharpen this constraint. ZEROFEC adds the original claim itself to the candidate list so that the system can output no edit when the claim is already faithful to the evidence (Huang et al., 2023). CLFEC formalizes correction as a set of localized edit operations,

$\mathcal{E}=\{(s_k, e_k, u_k, v_k)\}_{k=1}^{K},$

where $u_k$ is the erroneous source span and $v_k$ is the corrected target span (Kai et al., 27 Feb 2026). This suggests a common design principle: factual correction is treated less as unrestricted text generation than as evidence-grounded, minimally invasive rewriting.

2. Granularity of the corrected element

The literature operationalizes the “element” at several levels of granularity.

Correction unit	Representative formulations	Operationalization
Tokens and spans	Evidence-based FEC; VenCE; FactCloze (Thorne et al., 2020, Chen et al., 2022, Li et al., 2024)	Mask salient tokens or spans, then generate or sample replacements
Entities and factual factors	RFEC; BART-based summary post-editing; FC (Lee et al., 2022, Cao et al., 2020, Zhu et al., 2020)	Detect erroneous entities, numbers, dates, pronouns, or a few keywords and substitute them
Triples, atomic facts, and atoms	FLEEK; RAC; FactCorrector; TruthfulRAG (Bayat et al., 2023, Li et al., 2024, Carnerero-Cano et al., 16 Jan 2026, Liu et al., 13 Nov 2025)	Decompose text into triples, atomic facts, atoms, or KG paths and verify or revise each unit independently

At the claim level, the masking stage in evidence-based FEC is explicitly token/span oriented: edits are induced by masking specific tokens in the claim, and the quality of the final correction depends heavily on whether the right spans were masked (Thorne et al., 2020). In summary correction, RFEC assumes that many factual inconsistencies are entity-level errors and therefore detects whether each summary entity is factually consistent with the article, then substitutes wrong entities with accurate entities from retrieved evidence (Lee et al., 2022). FactCloze defines factual factors as “text spans that describe facts, such as entities, noun phrases, etc.” and corrects those spans with a conditional-generation cloze model (Li et al., 2024). FLEEK moves to triple granularity by extracting flat triples and extended triples, verifying each one, and revising only the erroneous piece while preserving the rest of the sentence (Bayat et al., 2023). RAC and FactCorrector extend this to atomic-fact and atom decompositions for LLM outputs (Li et al., 2024, Carnerero-Cano et al., 16 Jan 2026).

3. Evidence-grounded claim correction pipelines

The first evidence-grounded claim-correction systems are built on distant supervision from fact verification data. Evidence-based FEC trains on FEVER-derived supervision by first masking a claim, $\tilde{c}=Mask(c,E(c))$, then generating a corrected claim with a seq2seq model, $c' = Corr(\tilde{c}, E(c))$, using the reconstruction objective

$E(c)$0

The corrector is T5-base, and retrieval uses GENRE for page prediction and DPR for passage retrieval; the reported best downstream scores use the top 2 DPR passages (Thorne et al., 2020).

A later distantly supervised variant, LIFE, argues that existing mask-then-correct systems are bottlenecked by the masker because there is no paired data to train explicit factual-error identification. LIFE therefore proposes a three-step mask-corrupt-correct procedure: it trains a corruptor that deliberately introduces factual errors into correct text, filters the generated data, and trains a corrector that does not require a masker. On a public dataset, it outperforms the previous best-performing distantly supervised method by 10.59 points in SARI Final, corresponding to a 19.3% improvement, and exceeds ChatGPT prompted with in-context examples by 7.16 points in SARI Final (He et al., 2023).

Zero-shot correction replaces end-to-end rewriting with an explicitly decomposed pipeline. ZEROFEC extracts claim answers, generates a question for each extracted unit, answers the question from evidence with UnifiedQA-v2, converts the resulting QA pair to a declarative claim with a T5-base QA-to-claim model, and scores each candidate with

$E(c)$1

On FEVER human evaluation, ZEROFEC attains faithfulness $E(c)$2 and ZEROFEC-DA reaches $E(c)$3, compared with $E(c)$4 for T5-FULL and $E(c)$5 for T5-DISTANT (Huang et al., 2023).

VenCE turns minimal-edit correction into iterative constrained editing with Metropolis-Hastings sampling. Its target distribution

$E(c)$6

combines a language-model energy, a truthfulness energy from an offline fact verifier, and a Hamming-distance energy relative to the original claim. Edit positions are sampled using verifier gradients,

$E(c)$7

and multi-token entities are handled by masking whole named-entity spans. On FecData, VenCE improves SARI-final by 5.3 over the previous best distantly supervised method, an 11.8% relative improvement (Chen et al., 2022).

4. Summary correction and dialogue summarization

A substantial branch of the literature studies post-editing for abstractive summaries. One BART-based corrector trains on artificial corruptions of reference summaries using entity, number, date, and pronoun swaps, with the objective $E(c)$8. On synthetic CNN/DailyMail corruptions, it reports 84.38% consistency-checking accuracy, exactly corrects 62.13% of corrupted summaries, and leaves 73.73% of clean summaries unchanged. On real K2019 summaries, among 62 inconsistent summaries the model changed 19 and 11 were corrected, while among 441 consistent summaries it changed 39 and incorrectly altered 5, illustrating the transfer gap from artificial to real errors (Cao et al., 2020).

Span-selection methods make the locality of the edit more explicit. FC, the factual corrector in FASum, is designed to improve summaries via modifying only a few keywords; for BottomUp summaries on CNN/DailyMail it modifies 48.3% of summaries, and 94.4% of corrected summaries contain 3 or fewer new tokens (Zhu et al., 2020). SpanFact formulates correction as QA-style span prediction from the source document, with an iterative single-masking model and an auto-regressive multi-masking model. Across CNN/DailyMail, XSum, and Gigaword, it improves factuality metrics such as QGQA and FactCC while largely preserving ROUGE (Dong et al., 2020).

RFEC narrows the task further to entity-level errors in summaries. It retrieves the top-2 article sentences for each summary sentence by ROUGE-L similarity, uses spaCy NER to extract summary and evidence entities, computes an error probability for each summary entity, and substitutes a wrong entity with the evidence entity of maximum correction score if that score exceeds $E(c)$9. The appendix sets both $c'$0 and $c'$1 to 0.5. On the synthetic test set, RFEC achieves 91.06 accuracy at 4024 samples/min, compared with 90.93 accuracy at 933 samples/min for the BART seq2seq baseline; on FactCC-Test, the best RFEC setting corrects 7 inconsistent summaries and runs at about 8× less time than the seq2seq approach (Lee et al., 2022).

FactCloze addresses two difficulties identified in prior summary post-editing: the Independent Correction Problem and the Over-Correction Problem. It masks factual factors such as named entities and noun phrases and generates them autoregressively with

$c'$2

so later factor predictions depend on earlier ones. Its self-diagnosis mechanism keeps faithful spans unchanged, and its post-alert mechanism uses \verb|<unk>| to flag hypotheses that are too risky to correct. On FRANK, the paper reports average improvement over uncorrected hypotheses of 6.07% on CNN/DM and 13.51% on XSum; in a manual study on 100 FRANK-XSum pairs, only 2% had no factual errors and 66% could not be corrected to improve faithfulness (Li et al., 2024).

5. Atomic-fact, graph-based, and retrieval-augmented correction for LLM outputs

FLEEK is an end-to-end system that extracts factual claims as flat triples $c'$3 or extended triples, generates questions for each triple, queries both a KGQA system and web search, classifies each claim as Strongly Supported, Likely Supported, or Questionable, and revises only Questionable claims. The revision objective is to produce $c'$4 such that $c'$5 and $c'$6 while preserving other supported triples. On its two 50-example benchmarks, GPT-3-based FLEEK reaches total F1 scores of 82.02 and 84.45, and the fact revision module obtains 72.7% average revision accuracy, rising to 87.5% when the verification stage is correct (Bayat et al., 2023).

RAC is a training-free post-correction pipeline for LLM outputs. It decomposes the generated answer into atomic facts,

$c'$7

retrieves evidence once, verifies each fact as True, False, or Not Mentioned in the RAG setting, corrects false facts, and then revises the answer: $c'$8 The method uses 1 generation API call, 1 search query, 1 correction iteration, and 1 retrieval call. On Biography, RAC latency is 3.9× relative to uncorrected RAG generation, compared with 70× for RARR, 10× for CRITIC, and 150× for EVER. The paper reports up to 30% improvements over prior retrieval-based correction methods and nearly 98% FactScore with gold retrieval (Li et al., 2024).

FactCorrector also decomposes long-form responses into atoms,

$c'$9

but evaluates them with a probabilistic graphical model over atoms and retrieved contexts. Posterior marginals $E(c)\vDash c'$0 are mapped to True, False, or Unverified labels according to whether $E(c)\vDash c'$1 is greater than, less than, or equal to $E(c)\vDash c'$2. The refinement model then receives the original answer, the incorrect or unverified atoms, and the contexts connected to those atoms in the graph. On VELI5, a 17,522-instance benchmark with 50% synthetic and 50% human-authored responses, the paper reports an average precision improvement of about +0.21, outperforming LLM1 at +0.17 and RAC at +0.16. Human evaluation shows 67.4% of incorrect atoms corrected and 60.9% of correct atoms preserved (Carnerero-Cano et al., 16 Jan 2026).

TruthfulRAG addresses factual-level conflicts in retrieval-augmented generation by constructing a knowledge graph $E(c)\vDash c'$3 from retrieved content, retrieving query-relevant entities and relations, generating two-hop reasoning paths, and filtering those paths by entropy increase,

$E(c)\vDash c'$4

Paths with $E(c)\vDash c'$5 are treated as corrective. Compared with standard RAG, reported improvements range from 3.6% to 29.2%, and the method remains strong on non-conflicting golden settings such as MuSiQue-golden at 93.2 and SQuAD-golden at 98.3 (Liu et al., 13 Nov 2025).

6. Evaluation methodology, failure modes, and reliability constraints

Evaluation has been a central methodological issue because generic factuality metrics are often too coarse for correction. Evidence-based FEC evaluates outputs by intelligibility, support by evidence, and whether the factual error was actually corrected; it also reports strong correlations between human judgments and SARI, especially SARI Keep, which correlates very strongly with intelligibility, support, and correction quality (Thorne et al., 2020). ZEROFEC likewise finds that SARI is the most consistent metric for factuality and faithfulness across FEVER and SciFact, while correlations with intelligibility are low for all tested metrics (Huang et al., 2023).

For dialogue summarization, FERRANTI explicitly argues that FEC requires reference correction rather than source-based factuality scoring alone. It manually annotates 4000 dialogue-summary pairs from SAMSum and DialogSum, aligns model outputs with human corrections, classifies edits into content-based categories, counts TP, FP, and FN by category, and reports precision, recall, and especially $E(c)\vDash c'$6 to penalize over-correction. The framework uses eight error types: Entity Error, Predicate Error, Circumstance Error, Coreference Error, Discourse Link Error, Out of Article Error, Grammatical Error, and Others (Gao et al., 2023). This suggests that reliable evaluation of element-level correction depends on edit-level comparison, not only on post hoc factuality scores.

The failure modes are recurrent. Evidence-based FEC identifies over-erasure, incorrect masking, and inadequate evidence retrieval as the main causes of error (Thorne et al., 2020). The BART-based summary corrector reports that transferring from artificial error correction to downstream system-generated summaries remains “very challenging” (Cao et al., 2020). RAC and TruthfulRAG both emphasize that correction quality depends strongly on retrieval quality, and TruthfulRAG notes threshold sensitivity and computational overhead in graph traversal and entropy evaluation (Li et al., 2024, Liu et al., 13 Nov 2025).

Mixed-error settings introduce further constraints. CLFEC, which unifies linguistic and factual correction in paragraph-level Chinese professional writing, reports that prompt-only baselines on the four larger models achieve average factual Precision 24.05%, Recall 31.46%, and F1 26.87%, while U-RAG improves average F1 to 49.73%. The paper further argues that unified correction outperforms decoupled processes and that the Agent reduces the error-free edit rate by about 64%, from 1.094 to 0.391 (Kai et al., 27 Feb 2026). In a different reliability setting, correction suppression shows that LLMs may internally register a false premise yet comply rather than correct when the premise is embedded in routine task requests. Across eight models, suppression rates range from 19.3% to 89.6%; CDS raises Qwen3.5-9B correction from 0.0% to 58.2%, while DPA is the only method reported to preserve or improve reasoning capability on both tested models (Chen et al., 7 May 2026). The paper frames this as “factual strictness,” defined as the willingness to uphold accuracy against contextual pressures.

Element-level factual correction therefore encompasses a broad but coherent technical agenda: localized factual repair, evidence-grounded decision making, minimal-change rewriting, and evaluation at the level of the corrected element rather than the document as a whole. The surveyed work suggests that the central unresolved problems are still error localization, evidence quality, over-correction, mixed-error interaction, and reliable correction under realistic task framing.