Rationalizing Errors for Text Refinement

29 concepts19 papers1 proposal

When a Large Language Model (LLM) generates text containing a factual hallucination, a grammatical slip, or a conversational contradiction, our instinct is to ask the model to simply "fix it." However, direct text editing often fails.

Direct correction forces a model to perform two complex cognitive tasks simultaneously within a single forward pass: it must diagnose exactly what is wrong with the text, and it must rewrite the sequence to resolve the issue. By forcing the model to generate a natural language explanation of the error before attempting a correction, we decouple diagnosis from execution. The generated explanation acts as a "reasoning bridge"—an explicit scratchpad that alters the conditional probabilities of the subsequent tokens, steering the model toward a precise and reliable edit.


The Explanation as a Diagnostic Bridge

Why does writing out an explanation make a model a better editor? The answer lies in how autoregressive models generate text. Without an intermediate explanation, a model must compute the tokens for a corrected sentence without explicitly grounding its generation in the specific nature of the failure.

[Flawed Text] ───(Direct Edit)───> [Often Fails / Misses the Error]

[Flawed Text] ───(Articulate Error)───> [Diagnostic Report] ───(Guided Edit)───> [Precise Correction]

By forcing the model to verbalize the error first, the explanation serves as a diagnostic report. In the context of abstractive summarization, Yixin Liu et al. demonstrated that requiring a model to articulate a factual error before producing a corrected summary significantly improves editing accuracy. This sequential approach mirrors chain-of-thought reasoning; the verbalized explanation ensures the model understands the underlying contradiction before it modifies a single word.

This mechanism acts as a form of internal supervision across diverse domains:

  • Factual Consistency: Manya Wadhwa et al. formalized this into a "Detect, Critique, Refine" pipeline, showing that span-level natural language critiques allow even smaller open-source models to successfully resolve hallucinations. Similarly, Juyeon Kim et al. showed that explaining the discrepancy between an initial output and retrieved external evidence serves as a critical bridge for accurate post-editing.
  • Dialogue Coherence: When a chatbot contradicts its own persona, Xiaofei Wen et al. and Lifeng Jin et al. found that generating a natural language explanation of the logical conflict provides the necessary strategy for the model to formulate a coherent recovery response.
  • Machine Translation: In automated post-editing, Marcos Vinicius Treviso et al. showed that verbalizing translation errors acts as a catalyst for higher-quality multilingual refinement.

Turning Evaluations into Actionable Prompts

Beyond improving a single generation, explicit rationalization transforms text evaluation from a passive scoring mechanism into an active optimization tool.

Traditionally, evaluation models output binary labels or scalar scores (e.g., "Inconsistent" or "7/10"). These scores are useless as feedback for rewriting text. However, when an evaluation model generates a structured, fine-grained natural language explanation of an inconsistency, that explanation doubles as an actionable prompt for a refinement model.

Wendi Cui et al. leveraged this property in their Divide-Conquer-Reasoning framework, demonstrating that providing explicit justifications for detected inconsistencies allows models to correct roughly 90% of factual errors in a single iteration. To scale this capability, datasets like the one introduced by Ge Luo 0002 et al. have moved away from binary consistency labels, instead pairing text with source-verified explanations to explicitly train models to bridge the gap between error detection and error repair.

This loop is powerful enough to enable self-improvement. For instance, Haoxiang Sun et al. utilized an explain-then-correct workflow to generate high-quality self-supervised preference data, allowing a model to learn from its own corrected mistakes without relying on expensive human annotations.


The Order of Operations Dilemma

While generating an explanation before a correction is highly effective for semantic and factual errors, researchers have uncovered a critical architectural nuance: the optimal order of operations depends heavily on the task.

If a model's intermediate reasoning is flawed, an "explain-then-correct" workflow can poison the generation context, derailing the subsequent correction. In fields like Grammatical Error Correction (GEC), the cognitive load of diagnosing a complex linguistic rule can sometimes cause the model to hallucinate false rules, which then forces an incorrect edit.

Explain-Then-Correct (Best for Factual/Semantic Errors):
[Flawed Text] ──> [Reasoning / Explanation] ──> [Final Correction]
                                  │
                                  └── (Flawed reasoning can derail correction)

Correct-Then-Explain (Best for Syntactic/Grammatical Errors):
[Flawed Text] ──> [Final Correction] ──> [Post-Hoc Explanation]

To circumvent this, Masahiro Kaneko et al. proposed a "Prompt Insertion" method for GEC, where the model first performs the full text correction and subsequently generates the natural language explanations post-hoc. By grounding the explanation in a finalized target text, the model avoids the risk of cascading reasoning errors.

Jingheng Ye et al. confirmed this tension in Chinese GEC benchmarks, noting that while joint learning of correction and explanation remains a challenge, generating explanations after the correction ensures that the reasoning remains anchored to a valid target sequence.

Ultimately, whether the explanation serves as a precursor to guide the edit or a post-hoc anchor to justify it, explicitly translating abstract errors into structured natural language text remains one of the most reliable levers for improving LLM generation quality.

Key Foundational Work

On Improving Summarization Factual Consistency from Natural Language Feedback (2023) — contributed 2 concept elements to this direction.
Findings of the WMT 2023 Shared Task on Automatic Post-Editing (2023) — contributed 1 concept elements to this direction.
An Investigation of Language Model Interpretability via Sentence Editing (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_explanation_guided_correction_d854ecd1,
  title = {explanation-guided correction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explanation-guided-correction/rdr_d854ecd10e7c45d496703f2cf10aa6bc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_explanation_guided_correction_d854ecd1_v1,
  title = {explanation-guided correction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explanation-guided-correction/rdr_d854ecd10e7c45d496703f2cf10aa6bc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake