Technical Summary: DECOMPOSERL

Problem Statement

Automated claim verification faces a fundamental trade-off between accuracy and interpretability. Current approaches generally fall into two categories:

1. End-to-end classifiers: These achieve high accuracy on benchmarks but operate as "black boxes," emitting a verdict without providing inspectable traces of how the conclusion was reached. This opacity is problematic in high-stakes domains like biomedical triage, political fact-checking, and scientific peer review, where users require visibility into which parts of a claim were verified and why.
2. Decomposition-based methods: These break claims into atomic sub-questions to generate inspectable traces. However, they often lag behind end-to-end classifiers in performance. Existing decomposition methods rely on brittle prompting or expensive imitation learning (fine-tuning on gold decompositions) without a training signal that measures the actual fitness of the decomposition for the verification task.

Methodology

The authors propose DECOMPOSERL, a reinforcement learning (RL) framework that trains a policy to decompose claims into useful, informative, and diverse questions, producing both high accuracy and auditable traces.

1. Data Curation Funnel

To address the prohibitive cost of training RL policies on large, noisy corpora, the authors designed a multi-stage data curation funnel. Starting from 155K claims across 14 heterogeneous corpora, they distilled a compact, learning-signal-dense subset of 5,464 claims (4% of the original data). The filtering stages include:

• Rule-based filtering: Removing claims with insufficient evidence length, trivial lexical overlap, or fewer than two named entities.
• Difficulty-based filtering: Using a MiniCheck-7B verifier to retain only claims where the model is uncertain (confidence $0.3 \le p \le 0.8$), ensuring the training data contains non-trivial reasoning challenges.
• Deduplication and Decontamination: Removing near-duplicates via MinHash and cosine similarity to prevent data leakage and ensure diversity.
• Silver Decomposition: Generating "silver" (pseudo-labeled) decompositions using GPT-5-mini for the remaining claims.
• Diversity Selection: Using a Facility-Location objective (submodular maximization) to select a final subset that maximizes embedding-space coverage, ensuring the training set represents the full distribution of claim types.
• Long-Evidence Augmentation: Adding claims with long evidence contexts to prevent distribution shift during evaluation.

2. Reinforcement Learning Framework

DECOMPOSERL frames claim verification as an iterative Question-Answer (QA) decomposition problem. The policy ($\pi_\theta$), based on Qwen2.5-7B-Instruct, generates a structured trace consisting of an initial analysis, $n$ question-answer cycles, and a final verdict ($SUPPORTED$ or $REFUTED$).

The policy is trained using Group Relative Policy Optimization (GRPO) with a novel multi-faceted reward ensemble. Unlike prior work that relies on a single outcome reward, DECOMPOSERL optimizes seven complementary signals:

1. Format ($R_{fmt}$): Ensures adherence to the XML-based structured output schema.
2. Verification ($R_{ver}$): Anchors the policy to the ground-truth verdict (only for labeled data).
3. Question Count ($R_{qc}$): Penalizes deviations from the optimal number of sub-questions.
4. Diversity ($R_{div}$): Uses Maximal Marginal Relevance (MMR) on embeddings to penalize redundant questions.
5. Coverage ($R_{cov}$): An LLM-as-a-judge evaluates if the set of answers alone can reconstruct the correct verdict.
6. Necessity via Leave-One-Out ($R_{nec}$): A critical component where the judge determines if removing a specific question changes the verdict. Questions that are necessary (removing them flips the verdict) receive high rewards; harmful questions (those that mislead the verdict) receive negative rewards.
7. Joint Multiplicative Quality ($R_{joint}$): A multiplicative reward combining Answerability, Atomicity (single-focus, no conjunctions), and Correctness. This ensures that a single failure in any dimension drives the question's score to zero.

3. Semi-Supervised Learning

To address the scarcity of gold-labeled data, DECOMPOSERL supports semi-supervised training. For unlabeled claims:

• Coverage is approximated via self-consistency: The majority vote of $G$ rollouts serves as a pseudo-label.
• Necessity is computed relative to the policy's own reconstructed verdict rather than a ground truth.
• The Verification reward is dropped for unlabeled data, while the other rewards remain label-agnostic.

Key Contributions

1. Multi-faceted Reward Ensemble: A novel reward system that captures decomposition quality across multiple dimensions (necessity, coverage, diversity, atomicity) using a combination of deterministic rules, embedding-based metrics, and LLM-as-a-judge evaluations.
2. Semi-Supervised Capability: A self-consistency mechanism that allows the model to learn from unlabeled claims by leveraging intra-prompt agreement, making the approach viable for domains with expensive annotations.
3. Efficient Data Curation: A distillation funnel that reduces training data requirements by ~96% (from 155K to ~5.5K claims) while improving signal density, enabling effective RL training with limited compute.
4. Performance and Scalability: Demonstrating that a 7B parameter model trained with this framework can match or exceed the performance of significantly larger models (32B) and proprietary frontier models (GPT-4.1-mini) while providing inspectable traces.

Experimental Results

The authors evaluated DECOMPOSERL on 11 claim-verification benchmarks (9 in-domain, 2 out-of-domain), including biomedical, political, and scientific datasets.

• Accuracy: The DECOMPOSERL-7B model achieved 86.3% balanced accuracy in-domain and 69.8% out-of-domain.
• Comparison to Baselines:
  • It outperformed all decomposition-based baselines (e.g., Decomposed Prompting, HiSS, ClaimDecomp) by significant margins (+4.4% in-domain, +2.8% out-of-domain over the strongest baseline).
  • It surpassed end-to-end classifiers like MiniCheck-7B.
  • It matched the performance of 32B parameter models and the GPT-4.1-mini frontier model on in-domain tasks, despite being 4x smaller than the 32B baseline and an order of magnitude smaller than the frontier model.
• Semi-Supervised Performance: When trained with only 10% labeled data ($s=0.1$), DECOMPOSERL-7B maintained an in-domain accuracy of 84.6% and out-of-domain accuracy of 69.7%, still outperforming all 7B baselines.
• Ablation Studies: Removing the Necessity reward caused the largest drop in out-of-domain performance (-4.8%), highlighting its critical role in generalization. Removing Coverage and Diversity also significantly impacted robustness, confirming the necessity of the full ensemble for out-of-domain performance.

Significance and Claims

The paper claims that DECOMPOSERL successfully bridges the gap between accuracy and interpretability in claim verification. By framing decomposition as an RL problem guided by a multi-faceted reward ensemble, the authors demonstrate that:

• Traceability need not come at the cost of accuracy: The model produces auditable, step-by-step verification traces without sacrificing performance compared to opaque classifiers.
• Parameter efficiency: High performance can be achieved with smaller models (7B) if the training signal (reward design and data curation) is optimized, rather than relying solely on scaling model size.
• Scalability to low-resource settings: The semi-supervised approach allows the system to adapt to new domains where gold-standard annotations are scarce or expensive, using self-consistency to bootstrap learning.

The authors position this work as a solution to the "traceability gap" in automated fact-checking, offering a method that is both accurate and transparent, suitable for high-stakes applications where users must inspect the reasoning process.