Andrii Bidochko

• Updated: June 20, 2026
• 6 min read

DecomposeRL: Learning to Ask Useful, Informative, and Diverse Questions for Semi-Supervised, Traceable Claim Verification

Direct Answer

DecomposeRL introduces a reinforcement‑learning (RL) policy that learns to ask useful, informative, and diverse sub‑questions when verifying claims, while still delivering state‑of‑the‑art accuracy. By combining a multi‑faceted reward system with a data‑curation funnel, the approach produces human‑traceable verification paths without sacrificing performance, even in semi‑supervised settings.

Background: Why This Problem Is Hard

Claim verification sits at the intersection of natural language understanding, factual retrieval, and logical reasoning. Two dominant paradigms have emerged:

• End‑to‑end classifiers—large language models (LLMs) that directly label a claim as true or false. They achieve high accuracy but provide no insight into *how* the decision was reached, making them unsuitable for high‑stakes domains such as medicine or law.
• Decomposition‑based pipelines—systems that break a claim into sub‑questions, retrieve evidence, and aggregate answers. These pipelines are transparent but historically lag behind end‑to‑end models on benchmark scores, largely because the sub‑question generation step is brittle and often redundant.

The core difficulty stems from three intertwined challenges:

1. Question relevance: Generated sub‑questions must target the factual gaps in a claim, not wander into tangential topics.
2. Diversity vs. redundancy: A good verification trace explores multiple evidence strands without repeating the same query.
3. Training signal scarcity: High‑quality, fully annotated claim‑verification datasets are expensive; most real‑world data is unlabeled.

Addressing these pain points is critical as enterprises increasingly demand AI systems that are both accurate *and* auditable.

What the Researchers Propose

DecomposeRL reframes claim decomposition as a sequential decision‑making problem. An RL agent observes a claim and the evidence collected so far, then decides which sub‑question to ask next. The policy is trained with Generalized Reward‑Based Policy Optimization (GRPO), a variant of policy gradient that can handle sparse, multi‑objective rewards.

The reward ensemble captures four desiderata:

• Usefulness: Does the answer to the sub‑question reduce uncertainty about the claim?
• Informativeness: Does the retrieved evidence contain novel facts?
• Diversity: Are the generated questions sufficiently different from previous ones?
• Correctness: Does the final aggregation of answers match the ground‑truth label (when available)?

To make training tractable, the authors introduce a “data‑curation funnel” that distills a massive 115 K‑claim pool into a dense 5 K‑claim subset rich in learning signal. This curated set fuels both fully supervised and semi‑supervised training, allowing the policy to benefit from unlabeled claims while still grounding its behavior in high‑quality examples.

How It Works in Practice

The DecomposeRL pipeline consists of three tightly coupled components:

1. Claim Encoder: A transformer‑based model converts the raw claim into a contextual embedding that captures its semantic nuances.
2. RL Question Generator: Conditioned on the claim embedding and the current evidence state, the policy outputs a natural‑language sub‑question. The policy’s action space is constrained by a vocabulary of plausible interrogatives, ensuring grammaticality.
3. Evidence Retriever & Aggregator: Each generated question is sent to a retrieval engine (e.g., BM25 or dense vector search). Retrieved passages are fed to a reading‑comprehension module that extracts concise answers, which are then stored in the evidence state.

The loop repeats until a termination condition is met—either a confidence threshold is reached or a maximum number of questions is issued. At that point, a final classifier aggregates the collected answers to produce the claim label.

What sets DecomposeRL apart is the *closed‑loop* feedback between the RL policy and the evidence state. Each new answer reshapes the policy’s belief about what information is still missing, guiding the next question toward the most informative gap.

In practice, the system can be deployed as a microservice that receives a claim via an API, returns a traceable verification report, and optionally streams intermediate sub‑questions for human review.

Evaluation & Results

Researchers benchmarked DecomposeRL on eleven claim‑verification datasets spanning biomedical, political, scientific, and general‑domain domains. The evaluation protocol included:

• In‑domain balanced accuracy: Performance on the same distribution as the training data.
• Out‑of‑domain balanced accuracy: Generalization to unseen claim types.
• Traceability metrics: Coverage of evidence, redundancy rate, and question diversity.
• Semi‑supervised scenario: Training with only 10 % labeled claims while leveraging the remaining unlabeled pool.

Key findings:

• A 7 B‑parameter DecomposeRL policy trained on just ~5 K curated claims achieved **86.3 % in‑domain** and **69.8 % out‑of‑domain** balanced accuracy, matching or surpassing 32 B‑parameter baselines and even GPT‑4.1‑mini.
• In the semi‑supervised setting, DecomposeRL outperformed all baselines by a margin of 4–6 % despite using only a tenth of the labeled data.
• Traceability analysis showed a 35 % reduction in redundant questions and a 22 % increase in novel evidence compared to prior decomposition methods.

These results demonstrate that a well‑designed RL reward structure can close the performance gap between transparent pipelines and black‑box classifiers, while also delivering robust generalization.

Why This Matters for AI Systems and Agents

For practitioners building AI agents that must justify their decisions—think compliance bots, medical triage assistants, or financial audit tools—DecomposeRL offers a blueprint for marrying accuracy with interpretability. The traceable question‑answer chain can be logged, visualized, and audited, satisfying regulatory requirements that are increasingly demanding evidence of *why* an AI made a particular claim.

Moreover, the semi‑supervised capability reduces the reliance on costly annotation pipelines, enabling faster iteration on domain‑specific verification tasks. Companies can bootstrap a verification service with a modest labeled dataset and let the RL policy harvest signal from the vast sea of unlabeled claims that naturally accrue in production.

Integrating DecomposeRL‑style agents into existing orchestration platforms is straightforward. For example, the Enterprise AI platform by UBOS can host the RL policy as a containerized service, while the platform’s workflow automation studio coordinates retrieval, reading, and aggregation steps. This modularity accelerates deployment across use cases ranging from content moderation to fact‑checking news feeds.

What Comes Next

While DecomposeRL marks a significant step forward, several open challenges remain:

• Scalability of the reward computation: GRPO’s multi‑objective rewards are computationally intensive; future work could explore surrogate models or curriculum learning to reduce overhead.
• Domain adaptation: Extending the policy to highly specialized vocabularies (e.g., legal statutes) may require additional pre‑training or prompt‑engineering.
• Human‑in‑the‑loop refinement: Allowing experts to intervene on sub‑question selection could further improve both accuracy and trustworthiness.

Potential applications are broad. In the marketing arena, AI marketing agents could use DecomposeRL to verify product claims before launching campaigns, reducing the risk of misinformation. In education, the framework could power automated fact‑checking assistants that guide students through evidence‑based reasoning.

Finally, the authors have open‑sourced the code, data, and models (DecomposeRL paper on arXiv), inviting the community to build on this foundation, explore alternative reward designs, and integrate the approach with emerging retrieval‑augmented generation (RAG) architectures.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.