💬 NLP

Compositional Multi-hop Factual Error Correction via Decomposition-and-Injection

The paper introduces CECoR, a reasoning-aware framework that employs a decomposition-and-injection paradigm to decompose multi-hop claims and synthesize training data, thereby significantly improving factual error correction performance on complex, multi-hop tasks compared to existing methods.

Original authors: Lei Zhu, Xiaobao Wang, Jianbiao Yang, Chenyang Wang, Dongxiao He, Longbiao Wang, Jianwu Dang

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lei Zhu, Xiaobao Wang, Jianbiao Yang, Chenyang Wang, Dongxiao He, Longbiao Wang, Jianwu Dang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a fact-checker for a newsroom. Your job is to fix stories that have gone wrong. But here's the catch: some stories are simple, like "The cat is on the mat," while others are complex puzzles, like "The cat that chased the dog that ran from the fire station is on the mat."

Most current AI tools are like a junior intern who only knows how to fix simple stories. If the story is complex, they get confused. They try to fix the whole sentence at once or just swap out one word, often making the story worse or leaving the deep logical errors untouched.

This paper introduces a new system called CECoR (Compositional Error Correction via Reasoning-aware Synthesis). Think of CECoR not as a junior intern, but as a master editor who uses a special three-step process to fix even the most tangled, multi-part stories.

1. The Problem: The "Atomic" Mistake

Current methods treat a story like a single, flat block of clay. If there's a mistake, they try to chip away at the surface.

The Flaw: In a complex story, the error isn't just one word; it's a chain of logic. If you change one part, the whole chain might collapse.
The Analogy: Imagine a Rube Goldberg machine (a complex chain of events). If the first ball misses the cup, the whole machine fails. Old methods try to fix the machine by just painting the last cup red. CECoR looks at the entire chain of balls and cups to find exactly where the logic broke.

2. The Solution: The "Decomposition and Injection" Workshop

CECoR works differently. Instead of looking at the messy story as one big block, it breaks it down into a step-by-step recipe.

Step A: The Blueprint (Decomposition)
First, CECoR takes a correct, complex story and breaks it down into a logical program, like a flowchart.

Example: Instead of just reading "The actor born in 1955 who acted in Ransom won an Oscar," it breaks it into:
1. Find the actor born in 1955.
2. Check if they were in Ransom.
3. Check if they won an Oscar.
4. Combine the answers.

Step B: The Sabotage (Injection)
Here is the clever part. Since there aren't enough real-world examples of "wrong stories" to teach the AI, CECoR creates its own practice problems.

It takes the logical blueprint and intentionally breaks one step at a time.
It swaps a name, flips a "yes" to a "no," or changes a date in just one step of the chain.
Then, it rebuilds the story from this broken blueprint.
The Result: It creates thousands of "fake" wrong stories that are perfectly paired with the "correct" answer. It's like a teacher creating a thousand math problems with one specific error in them to train a student.

Step C: The Filter
Not every broken story is useful. Some are gibberish. CECoR has a strict quality control filter. It throws away any fake story that doesn't make sense grammatically or doesn't actually contradict the facts. It keeps only the high-quality "wrong vs. right" pairs.

3. The Training: Two-Stage Learning

Once CECoR has its library of perfect practice problems, it trains the AI in two phases:

Phase 1: The Classroom (Supervised Fine-Tuning)
The AI studies the "fake" wrong stories and learns to fix them using the correct answers. It learns the patterns of how logic chains break.
Phase 2: The Real World (Reinforcement Learning)
After the classroom, the AI is thrown into the wild. It is given real, messy, naturally occurring wrong stories (not the ones it made up). It tries to fix them, and a "judge" (another AI) tells it if the fix is actually true and makes sense. If the AI does well, it gets a "reward." This teaches the AI to be robust and handle real-world chaos, not just textbook examples.

4. The Results: Why It Wins

The paper tested CECoR against other methods and found:

Better at Complex Puzzles: While other methods struggled with multi-step stories (like the Rube Goldberg machine), CECoR excelled because it understood the steps.
Better at Simple Stories Too: Even though it was trained on complex puzzles, it was so good at understanding logic that it also became excellent at fixing simple, one-step stories.
Resilient to Bad Clues: In the real world, the evidence you find might be slightly wrong or messy. CECoR remained stable even when the "clues" it was given were noisy, whereas other systems fell apart.

Summary

Think of CECoR as a detective who doesn't just look at the crime scene (the wrong sentence) but reconstructs the entire timeline of events (the reasoning chain). By breaking the story down, intentionally messing it up to learn, and then practicing on real-world messiness, it learns to fix facts with a level of understanding that previous tools simply didn't have. It turns the difficult task of "fixing a complex lie" into a manageable game of "fixing one step in a chain."

Technical Summary: CECoR (Compositional Error Correction via Reasoning-aware Synthesis)

1. Problem Definition

Factual Error Correction (FEC) aims to revise inaccurate text into statements consistent with external evidence. While recent methods perform adequately on single-hop corrections, they struggle with multi-hop factual errors. Existing paradigms, such as "Mask-then-Correct" and "Mask-then-Corrupt," rely on an implicit Atomic Fact Assumption: they treat claims as flat, undecomposed units where errors are localized to single entities or short spans.

This assumption fails for multi-hop claims, where truth conditions are distributed across reasoning chains involving multiple evidence pieces. Real-world errors in these contexts often exhibit non-local, structurally complex characteristics (e.g., coordinated entity substitutions or violations of temporal/relational constraints) that cannot be captured by independent surface-span masking. Furthermore, the scarcity of large-scale, gold-standard paired datasets (incorrect–correct claim pairs) hinders the training of robust correction models.

2. Methodology: The CECoR Framework

The authors propose CECoR (Compositional Error Correction via Reasoning-aware Synthesis), a framework that addresses multi-hop correction through a Decomposition-and-Injection paradigm. The framework operates in three main phases:

A. Multi-hop Claim Decomposition

Instead of treating a claim as a flat string, CECoR decomposes verified (correct) multi-hop claims into interpretable, step-wise reasoning programs. Inspired by program-guided reasoning, a planner generates a sequence of logical steps $P = [S_1, S_2, ..., S_k]$ , categorized into:

QUESTION: Queries to extract atomic facts from evidence.
VERIFY: Binary checks against retrieved facts.
PREDICT: Logical compositions of prior steps to form conclusions.

This transformation makes the latent reasoning structure explicit, providing a substrate for controlled error injection.

B. Reasoning-aware Error Injection

CECoR synthesizes high-quality training pairs by injecting controlled factual perturbations into the decomposed reasoning program:

Strategy: Errors are injected into only one step ( $S_j$ ) at a time to prevent uncontrolled cascading inconsistencies while ensuring semantic coherence upon recomposition.
Mechanisms:
- PREDICT steps: Substitute a shared entity with a semantically plausible but incorrect alternative.
- VERIFY steps: Replace a factual unit with an incorrect variant or rewrite the statement in its negated form.
- QUESTION steps: Substitute the correct answer with a plausible incorrect one and adjust the question accordingly.
Recomposition: The modified step is recomposed with the original steps to generate a fluent, structurally grounded, yet factually incorrect multi-hop claim.

C. Filtering and Two-Stage Learning

Filtering: To ensure training reliability, generated synthetic pairs undergo a multi-step filter. Retained pairs must satisfy criteria regarding length, fluency (perplexity), multi-hop integrity (requiring $\ge$ 2 reasoning steps), and confirmed factual inconsistency against the evidence.
Supervised Fine-Tuning (SFT): The model is first trained on the filtered synthetic pairs $(M_{incorrect}, M_{correct})$ to learn compositional error patterns.
Reinforcement Learning (RL): To improve robustness against real-world error distributions, the model undergoes a second stage using naturally occurring incorrect claims (from datasets like REFUTES). The RL objective maximizes a reward function balancing:
1. Factual Consistency ( $S_{correct}$ ): Alignment with evidence.
2. Semantic Similarity ( $S_{sim}$ ): Preservation of the original claim's intent.
3. Fluency ( $S_{flu}$ ): Grammatical correctness.

3. Key Contributions

Analysis of Limitations: The paper identifies the "Atomic Fact Assumption" in prior FEC methods as a fundamental bottleneck for multi-hop correction, where errors are coupled across reasoning steps rather than isolated.
CECoR Framework: The introduction of a decomposition-and-injection paradigm that leverages program-guided reasoning to synthesize high-fidelity, compositional training data without requiring manual annotations.
Training & Evaluation Strategy:
- A two-stage SFT + RL pipeline that combines synthetic supervision with real-world error alignment.
- An evaluation protocol that supplements surface-form metrics (SARI, ROUGE) with LLM-based judging to better capture semantic correctness and evidence support, addressing the limitations of reference-based metrics that may over-reward conservative outputs.

4. Experimental Results

Experiments were conducted on multi-hop benchmarks (HOVER, FEVEROUS) and a single-hop dataset (FECDATA).

Performance: CECoR achieves state-of-the-art performance on multi-hop benchmarks, outperforming both distantly supervised baselines (LIFE, VENCE) and few-shot Large Language Model (LLM) baselines (GPT-4o, LLaMA3).
Metric Nuance: While surface-form metrics sometimes favor "Do-Nothing" baselines or conservative edits, LLM-based evaluation confirms that CECoR's RL-enhanced variants (specifically CECoR-RL) produce semantically superior corrections that align with human preferences.
Generalization:
- Single-hop Transfer: Models trained exclusively on multi-hop data generalize effectively to single-hop correction tasks (FECDATA), outperforming baselines trained on single-hop data.
- Robustness: CECoR maintains strong performance when trained and tested with retrieved (noisy) evidence, whereas baselines like LIFE often fail to generate valid synthetic data under noisy conditions.
Ablation: The filtering pipeline is shown to be critical; removing it leads to significant performance drops, particularly in SARI-Add scores and LLM-based evaluations.

5. Significance and Claims

The paper claims that CECoR represents a shift from treating factual claims as flat text sequences to modeling them as structured reasoning programs. By explicitly exposing intermediate reasoning steps, the framework enables:

Scalable Supervision: Generating diverse, high-quality training pairs from unpaired correct claims, alleviating the data scarcity bottleneck.
Compositional Reasoning: The ability to correct complex, non-local errors that involve relational and temporal constraints across multiple evidence sources.
Practical Applicability: The model's robustness to noisy evidence and its ability to transfer across domains (multi-hop to single-hop) suggest strong potential for real-world deployment in misinformation correction and reliable AI systems.

The authors conclude that their approach yields correction models that are more interpretable, generalizable, and robust for complex factual revision tasks compared to prior methods relying on span-level masking.