Distill and Align Decomposition for Enhanced Claim Verification

Jabez Magomere; Elena Kochkina; Samuel Mensah; Simerjot Kaur; Fernando Acero; Arturo Oncevay; Charese H. Smiley; Xiaomo Liu

cs.AIFeb 25, 2026

Distill and Align Decomposition for Enhanced Claim Verification

Authors: Jabez Magomere, Elena Kochkina, Samuel Mensah, Simerjot Kaur, Fernando Acero, Arturo Oncevay, Charese H. Smiley, Xiaomo Liu,

Organizations: University of Oxford · JPMorgan AI Research

Abstract

Complex claim verification requires decomposing sentences into verifiable subclaims, yet existing methods struggle to align decomposition quality with verification performance. We propose a reinforcement learning (RL) approach that jointly optimizes decomposition quality and verifier alignment using Group Relative Policy Optimization (GRPO). Our method integrates: (i) structured sequential reasoning; (ii) supervised finetuning on teacher-distilled exemplars; and (iii) a multi-objective reward balancing format compliance, verifier alignment, and decomposition quality. Across six evaluation settings, our trained 8B decomposer improves downstream verification performance to (71.75%) macro-F1, outperforming prompt-based approaches ((+1.99), (+6.24)) and existing RL methods ((+5.84)). Human evaluation confirms the high quality of the generated subclaims. Our framework enables smaller language models to achieve state-of-the-art claim verification by jointly optimising for verification accuracy and decomposition quality.

Figures & tables

Figure 1: Distill-Align-Decompose (DAD). We (1) distill examples from a teacher model, (2) frame decomposition as a sequential reasoning task, and (3) align the policy via GRPO using multi-objective rewards.

Baselines
Decomposer	Approach	PerplexityAI						ChatGPT						FELM			BINGCHAT			Overall
		Atomicity-1			Atomicity-2			Atomicity-1			Atomicity-2			FELM			BINGCHAT			Overall
		BAcc	F1	SC	BAcc	F1	SC	BAcc	F1	SC	BAcc	F1	SC	BAcc	F1	SC	BAcc	F1	SC	BAcc	F1	SC
Llama-3.1-8B	FActScore	76.10	65.72	10.76	71.40	67.07	16.66	72.60	73.68	11.08	61.60	64.33	17.55	53.93 *	64.04 *	8.9 *	50.24 *	58.24 *	72.56 *	64.31	65.51	22.92
Mistral-7B	VeriScore-SFT	73.50	65.61	2.50	66.60	63.63	4.46	77.40	77.69	2.68	76.90	75.79	4.90	52.59	66.40	4.39	51.51	64.23	21.26	66.42	68.89	6.70
Llama-3.1-8B	VeriScore	78.60	71.61	2.49	71.00	68.48	4.20	82.30	82.76	2.85	73.70	75.84	5.76	59.56 *	67.86 *	6.65 *	52.15 *	52.02 *	28.00 *	69.55	69.76	8.33

Table 1: Balanced Accuracy (BAcc), macro-F1, and subclaim count (SC) per dataset. Best BAcc and F1 in each column are bold . The final block reports overall BAcc/F1 and the average SC. Entries marked * are reported from Hu et al. (2025) under a similar setup. On average across six evaluation settings, our 8B decomposer leads to the highest verification performance, with BAcc

71.37\%

and macro-F1

71.75\%

, while maintaining comparable granularity (avg.

8.14

subclaims).

Dataset	Metric	DAD ClearCheck	Llama-8B ClearCheck
PerplexityAI Atom.-1	BAcc	72.50	69.50
PerplexityAI Atom.-1	MacroF1	67.00	66.50
PerplexityAI Atom.-2	BAcc	71.70	59.70
PerplexityAI Atom.-2	MacroF1	70.80	59.80
ChatGPT Atom.-1	BAcc	79.60	66.60
ChatGPT Atom.-1	MacroF1	79.90	65.40

Table 2: DAD vs. Llama-8B decomposers paired with a non-aligned ClearCheck-8B verifier. DAD consistently outperforms Llama-8B, demonstrating that the decomposer provides value even without verifier-specific alignment . Bold indicates the higher score per dataset and metric.

Training Configuration	PerplexityAI		FELM
Training Configuration	F1	SC	F1	SC
Prompt only (no training)	67.46	2.00	68.57	5.05
RL only (no SFT)
+ Format + Verifier Accuracy (sparse)	70.86	1.95	58.94	3.78
+ Format + Verifier Brier (dense)	72.71	1.49	64.94	3.40
SFT only	71.55	2.28	66.39	5.01

Table 3: Ablation study on different training setups. We report macro-F1 and Average Subclaim Count (Avg. SC) on PerplexityAI and FELM , plus overall F1. Best F1 per column is bold .

Figure 2: Verifier training rewards (dense vs. sparse). Training the base policy with the verifier Brier reward (

\bullet

dense) yields smoother, more sample-efficient learning than the sparse accuracy reward (

\bullet

sparse).

Figure 3: Human-evaluated subclaim decomposition quality across the five desiderata. Our method achieves comparable completeness to baselines while maintaining high scores across other desiderata. Scores represent sentence-level averages of subclaim quality with higher values indicating better quality for a given desideratum.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Granularity	Dataset	Atomicity	Size	KS
Sentence	PerplexityAI	1	300	Wiki
	ChatGPT	1	249	Wiki
	PerplexityAI	2	149	Wiki
	ChatGPT	2	115	Wiki
Response	FELM	—	184	Google
Response	BINGCHAT	—	396	Google

Appendix

Table 4: Evaluation datasets by label granularity, atomicity, and size. Abbreviations: KS = Knowledge Source; Wiki = Wikipedia.

Desiderata	3-way (%)	Fleiss’ $\kappa$	Pairwise (%)	$\kappa$	r
Completeness	80.8	0.5	A1-A3: 85.0	0.5	0.5
			A1-A2: 85.9	0.5	0.5
			A3-A2: 90.9	0.6	0.6
Uniqueness	92.9	0.9	A1-A3: 94.0	0.8	0.8
			A1-A2: 93.9	0.8	0.8
			A3-A2: 97.0	0.9	0.9

Appendix

Table 5: Inter-annotator agreement statistics across the five desiderata. 3-way agreement shows percentage agreement among all three annotators (A1, A2, A3) with Fleiss’

\kappa

for multi-annotator reliability. Pairwise statistics include percentage agreement, Cohen’s

\kappa

, and Pearson correlation coefficient (r) for each annotator pair.

Question	Businessman, a Jewish American, grew up and lived in New York City, associated with the Democratic Party, and eventually became very famous. His name is "Bernie"-what?
Sentence	Bernie Sanders.
Subclaims	Bernie Sanders
Human Annotation Scores	Completeness Redundancy Verifiability Coherence Clarity Majority Majority Average Average Average 0 1 0 1 0

Question	Tell me a bio of Paul Kruger.
Sentence	Kruger was a prominent figure in the Boer War (1899-1902), leading the Boers in their defense against the British Empire. Kruger was a staunch believer in Boer independence and was instrumental in securing the Transvaal’s independence from British rule in 1881.

Appendix

Table 6: Subclaims and Human Annotation Scores for Example Questions

Method	What is Similar to Our Approach	What is Different / Missing Compared to Ours
FActScore Min et al. (2023)	Focuses on atomic fact extraction; uses selection and extraction steps.	Over-decomposes (too many subclaims); does not optimize for verifier alignment; lacks explicit decontextualization and relationship identification.
VeriScore Song et al. (2024)	Emphasizes verifiability in claim extraction; uses prompt-based decomposition.	Does not explicitly decontextualize claims; lacks relationship identification; does not jointly optimize decomposition and verifier alignment.
Molecular Facts Gunjal and Durrett (2024)	Targets fully decontextualized, minimal subclaims.	Focuses only on decontextualization; does not address selection, relationship identification, or verifier alignment.
Claimify Metropolitansky and Larson (2025)	Separates selection, disambiguation, and extraction into distinct steps.	Uses complex, multi-stage process with closed-source models; does not jointly optimize decomposition and verifier alignment; lacks efficiency of single model call.
DyDecomp Lu et al. (2025)	RL-based optimization for when to decompose, using verifier feedback.	Only decides when to decompose, not how; does not improve decomposition quality or address relationships and decontextualization.
DndScore Wanner et al. (2024b)	Focuses on decontextualization and decomposition for factuality verification.	Does not integrate selection, relationship identification, and verifier alignment in a unified framework.

Appendix

Table 7: Comparison of Decomposer-Focused Methods: Similarities and Differences with Our Approach

Example 1: Frederick Howard Taylor
Field	Content
Question	Tell me a bio of Frederick Howard Taylor.
Original Claim	Howard Taylor attended the Royal College of Surgeons where he received three high honors in his postgraduate studies.
Subclaims	1. Howard Taylor attended the Royal College of Surgeons where he received three high honors in his postgraduate studies.
Ground Truth Label	Not Supported (NS)
Model Prediction	Supported (S)

Appendix

Table 8: Detailed Verification Error Analysis with Subclaims and Evidence

Example 1: Lina Medina [Dataset: ChatGPT, Atomicity 1]
Field	Content
Question	Tell me a bio of Lina Medina.
Original Claim	Lina lived a relatively private life after the birth of her son and later married in 1970, but little else is known about her current whereabouts or status.
Reasoning	STEP 1: The target sentence contains verifiable claims about Lina Medina’s life after the birth of her son, specifically that she lived a relatively private life and later married in 1970. However, the sentence also mentions that little else is known about her current whereabouts or status, which is a lack-of-information statement and should be excluded . The verifiable sentence is: "Lina lived a relatively private life after the birth of her son and later married in 1970." STEP 2: To decontextualize the sentence, we need to ensure it is clear, self-contained, and contextually relevant. The sentence already mentions Lina’s name, and the context provides the necessary information about her son’s birth. The decontextualized sentence is: "Lina Medina lived a relatively private life after the birth of her son and later married in 1970." STEP 3: Analyzing the decontextualized sentence, we can identify the following relationships: - Temporal relationship: Lina lived a private life after the birth of her son and later married in 1970 (time sequence). - Expansion relationship: The sentence provides additional information about Lina’s life after the birth of her son (elaboration). STEP 4: To split the claim into the smallest possible factual units, we can break it down into two separate claims: - Claim 1: Lina Medina lived a relatively private life after the birth of her son. - Claim 2: Lina Medina married in 1970.
Subclaims	1. Lina Medina lived a relatively private life after the birth of her son. 2. Lina Medina married in 1970.
Model prediction	Not Supported (NS)

Appendix

Table 9: Decomposition analysis with reasoning traces. Highlighted segments are discussed in Appendix F .

Field	Content
Example: Harrison Ford
Question	Tell me a bio of Harrison Ford.
Text	He was born on July 13, 1942, in Chicago, Illinois, to former radio actress Dorothy and advertising executive John William "Christopher" Ford.
Label	Supported
FactScore (FS) Subclaims	[ ‘ He was born on July 13, 1942.’, ‘ He was born in 1942.’, ‘1942 is a year.’ , ‘July 13 is a date.’ , ‘July 13 is a date in 1942.’ , ‘He was born on July 13.’, ‘He was born in July 1942.’, ‘July is a month.’ , ‘July is a month of 1942.’, ‘He was born in July.’, ‘He was born in July 1942.’, ‘He was born in 1942.’, <.. repeated variants of ‘He was born in the year of his birth, 1942.’ ..> ]
Prediction using FS subclaims	Not Supported

Appendix

Table 10: Examples comparing decompositions across several methods. Highlighted issues with decompositions.

Field	Content
Example: Scientist (musician)
Question	Tell me a bio of Scientist (musician).
Text	He was a protégé of King Tubby and worked alongside other pioneers of the genre such as Bunny Lee, Philip Smart, Pat Kelly, and Prince Jammy.
Label	Supported
FactScore (FS) Subclaims	[ ‘ He was a protégé of King Tubby.’, ‘King Tubby is a person.’ , ‘ He was a protégé.’, ‘ He was a protégé of King Tubby.’, ‘King Tubby is a person who was a mentor.’, ‘He worked alongside other pioneers of the genre.’, ‘The genre is a type of music.’ , ‘Bunny Lee is a pioneer of the genre.’, ‘He worked alongside Bunny Lee.’, ‘Philip Smart is a pioneer of the genre.’, ‘He worked alongside Philip Smart.’, ‘Pat Kelly is a pioneer of the genre.’, ‘He worked alongside Pat Kelly.’, ‘Prince Jammy is a pioneer of the genre.’, ‘He worked alongside Prince Jammy.’ ]
Prediction using FS subclaims	Not Supported

Appendix

Table 11: Examples comparing decompositions across several methods. Highlighted issues with decompositions. (continued)

Claim Verification

Claim verification assesses whether statements are supported, contradicted, or unresolved by relevant evidence.

See all 59 papers

DecomposeRL: Learning to Ask Useful, Informative, and Diverse Questions for Semi-Supervised, Traceable Claim Verification
May 27, 2026Shubhashis Roy Dipta, Ankur Padia, Francis FerraroClaim Verification RL for Language Model Reasoning
Department of Computer Science and Electrical Engineering · University of Maryland Baltimore County · Baltimore, MD 21250 USA
ThinknCheck: Grounded Claim Verification with Compact, Reasoning-Driven, and Interpretable Models
Apr 2, 2026Delip Rao, Feijiang Han, Chris Callison-BurchClaim Verification LLM Answer Verification
University of Pennsylvania, Philadelphia, PA, USA
A Multi-Agent Approach for Claim Verification from Tabular Data Documents
Apr 19, 2026Rudra Ranajee Saha, Laks V. S. Lakshmanan, Raymond T. NgClaim Verification Multi-Agent LLMs
University of British Columbia

May 27, 2026J/K move · Enter open · S save

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

Shubhashis Roy Dipta, Ankur Padia, Francis Ferraro

Department of Computer Science and Electrical Engineering · University of Maryland Baltimore County · Baltimore, MD 21250 USA

Claim Verification RL for Language Model Reasoning Question Decomposition Automated Fact-Checking Semi-Supervised Learning

Open paper AskPDF arXiv

Claim verification splits between end-to-end classifiers that are accurate but yields no inspectable traces, and decomposition-based methods produce inspectable traces but lag performance on benchmark datasets. We propose DecomposeRL an accurate claim-verifier that produce inspectable traces. DecomposeRL frames decomposition as an RL policy trained with GRPO and a multi-faceted reward ensemble, enabling both fully supervised and semi-supervised learning from unlabeled claims. DecomposeRL addresses the prohibitive training cost of GRPO with a data-curation funnel that distills 115K fact-verification claims into a compact, learning-signal-dense subset of 5K claims. We show that a DecomposeRL-7B policy trained with full supervision on only ~5K curated claims achieves 86.3 in-domain and 69.8 out-of-domain balanced accuracy across 11 claim-verification benchmarks containing biomedical, political, scientific, and general-domain claims. Despite being 4x smaller, it matches 32B baselines and GPT-4.1-mini, and it further outperforms baselines in a semi-supervised setting with only 10% labeled claims data. Code, data, and models are available at https://dipta007.github.io/DecomposeRL

Abstract

Claim Verification

Explore similar work

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

ThinknCheck: Grounded Claim Verification with Compact, Reasoning-Driven, and Interpretable Models

A Multi-Agent Approach for Claim Verification from Tabular Data Documents