Repeated Claim-Level Evidence (SRE)

• Repeated Claim-Level Evidence (SRE) is a setup where a complex claim is decomposed into sub-claims, each verified using the same claim-level evidence.
• It exposes an alignment bottleneck by using coarse, redundant evidence that does not match the granularity of the decomposed sub-claims.
• Empirical results indicate that SRE offers only marginal improvements over vanilla methods and may even degrade performance compared to sub-claim aligned setups like SAE.

Repeated Claim-level Evidence (SRE) is a decomposition-based claim-verification setup in which a complex claim is split into sub-claims, but each sub-claim is verified using the same full evidence set originally associated with the whole claim rather than evidence aligned to that specific sub-claim. The term is introduced in “The Alignment Bottleneck in Decomposition-Based Claim Verification” (Akhter et al., 11 Feb 2026), where SRE is treated as a common but under-examined verification condition that helps explain why structured claim decomposition has produced inconsistent empirical results. In that account, SRE embodies an alignment bottleneck: the reasoning structure is refined into sub-claims, but the evidence remains coarse, claim-level, and repeated.

1. Definition and formal formulation

In the formulation of (Akhter et al., 11 Feb 2026), a complex claim $C$ has claim-level evidence $E_C=\{E_{Ci}\}_{i=1}^{n}$, and is decomposed into sub-claims $\{s_j\}_{j=1}^{m}$, each ideally paired with evidence $E_j$ and label $y_j$. Claim-level inference is written as

$Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$

with temporally bounded evidence

$$E_{Ci} \subseteq \{ e \in \mathcal{D} : t(e) \leq t(C) \}.$$

The paper defines three evidence configurations for claim verification:

$$(i)\ \text{Vanilla:} \quad \mathcal{E} = E_{C}=\{E_{Ci}\}_{i=1}^{n}$$

$$(ii)\ \text{SRE:} \quad \mathcal{E} = (s_j,\{E_{Ci}\}_{i=1}^{n},y_j)_{j=1}^{m}$$

$$(iii)\ \text{SAE:} \quad \mathcal{E} = \{(s_j, E_j, y_j)\}_{j=1}^{m}.$$

The operational distinction is explicit: “In practice, SRE repeats the same claim-level evidence block after each sub-claim in the prompt, whereas SAE inserts the sub-claim-specific evidence block $E_C=\{E_{Ci}\}_{i=1}^{n}$0 next to its sub-claim $E_C=\{E_{Ci}\}_{i=1}^{n}$1” (Akhter et al., 11 Feb 2026). SRE therefore differs from Sub-claim Aligned Evidence (SAE) not in whether decomposition is used, but in whether evidence granularity matches decomposition granularity.

A concise comparison of the three evidence regimes is useful because the term SRE is defined relative to two contrasting baselines.

Configuration	Evidence supplied	Structural property
Vanilla	Claim-level evidence only	No decomposition
SRE	Same claim-level evidence repeated for each sub-claim	Decomposition without evidence alignment
SAE	Sub-claim-specific evidence paired with each sub-claim	Decomposition with evidence alignment

Within this framework, evidence alignment refers to whether the evidence given to a sub-claim is actually the evidence relevant to that sub-claim. SRE is therefore not simply a prompting variant; it is a particular mismatch between inference structure and evidence packaging. The paper’s central hypothesis is that decomposition helps only when “(i) how evidence is aligned and aggregated at the sub-claim level, and (ii) whether sub-claim veracity signals are reliable or noisy” are both handled properly (Akhter et al., 11 Feb 2026).

2. Experimental instantiation and operational characteristics

The main experimental testbed for SRE in (Akhter et al., 11 Feb 2026) is a new decomposition dataset derived from PHEME/PHEMEPlus. The construction process starts from PHEME rumours, filters for complex claims, decomposes them using a fixed LLM-based procedure from FactScore with gpt-3.5-turbo-instruct, and then manually validates the resulting sub-claims. The final dataset contains 399 complex claims and 1169 sub-claims. Crucially, the dataset also includes manually annotated sub-claim evidence spans defined as “the minimal set of sentences justifying the label,” which makes SAE evaluation possible.

The same paper also evaluates generalization on MMM-Fact and COVID-Fact, but those datasets do not contain human-annotated sub-claim evidence spans or sub-claim labels. As a result, SAE cannot be run there; only SRE-like ablations are possible. This is important because SRE is presented not as an artificial control condition, but as a practical approximation of many real deployments where sub-claim-aligned evidence is unavailable (Akhter et al., 11 Feb 2026).

For claim-level verification, the model is always Qwen3-14B, with prediction

$E_C=\{E_{Ci}\}_{i=1}^{n}$2

The claim-level label space is restricted to $E_C=\{E_{Ci}\}_{i=1}^{n}$3, dropping Unverified. Evaluation uses macro-F1 and balanced accuracy, with paired bootstrap and McNemar testing.

The paper also defines ablations that remove sub-claim labels:

$E_C=\{E_{Ci}\}_{i=1}^{n}$4

$E_C=\{E_{Ci}\}_{i=1}^{n}$5

The paper states that Ablation SRE “mirrors real-world use cases where we may have sub-claims available, with the parent claim and corresponding evidence but no sub-claim labels or sub-claim level evidence” (Akhter et al., 11 Feb 2026). At the sub-claim level, SRE is instantiated even more directly:

$E_C=\{E_{Ci}\}_{i=1}^{n}$6

Predicted sub-claim labels can then be aggregated for noisy claim verification as

$E_C=\{E_{Ci}\}_{i=1}^{n}$7

The implementation details expose a practical cost of SRE. Because the same evidence block is repeated after every sub-claim, input length grows sharply. Qwen3-14B is run in zero-shot mode with fixed prompts, temperature $E_C=\{E_{Ci}\}_{i=1}^{n}$8, top_p $E_C=\{E_{Ci}\}_{i=1}^{n}$9, top_k $\{s_j\}_{j=1}^{m}$0, and max new tokens 8172. The input context is 16384 for SAE and 40960 for SRE, with truncation above 40960 in SRE (Akhter et al., 11 Feb 2026). The paper does not isolate truncation as a causal variable, but it explicitly presents SRE as context-heavier and operationally noisier.

3. Empirical performance profile

The central empirical result of (Akhter et al., 11 Feb 2026) is that decomposition does not reliably help under SRE. On the main PHEMEPlus benchmark, claim-level results are:

• Vanilla: F1 $\{s_j\}_{j=1}^{m}$1, balanced accuracy $\{s_j\}_{j=1}^{m}$2
• Oracle_SRE: F1 $\{s_j\}_{j=1}^{m}$3, balanced accuracy $\{s_j\}_{j=1}^{m}$4
• Oracle_SAE: F1 $\{s_j\}_{j=1}^{m}$5, balanced accuracy $\{s_j\}_{j=1}^{m}$6

Thus SRE yields only a small F1 gain over Vanilla ($\{s_j\}_{j=1}^{m}$7) and almost no balanced-accuracy gain ($\{s_j\}_{j=1}^{m}$8); the associated bootstrap significance is weak or absent, with $\{s_j\}_{j=1}^{m}$9 for F1 and $E_j$0 for balanced accuracy. By contrast, SAE yields $E_j$1 F1 and $E_j$2 balanced accuracy, both with $E_j$3 (Akhter et al., 11 Feb 2026). The paper therefore characterizes SRE as negligible improvement at best and substantially inferior to aligned-evidence decomposition.

The pattern becomes more negative on datasets where only SRE-style setups are feasible. On COVID-Fact, Vanilla achieves F1 $E_j$4, while Ablation_SRE achieves F1 $E_j$5. On MMM-Fact, Vanilla reaches F1 $E_j$6, whereas Ablation_SRE falls to F1 $E_j$7, a degradation of $E_j$8 with $E_j$9 (Akhter et al., 11 Feb 2026). The paper treats this as evidence that repeated claim-level evidence can actively harm claim verification when decomposition is imposed without aligned evidence.

An internal ablation on PHEMEPlus further clarifies the weakness of SRE. Oracle_SRE obtains F1 $y_j$0, whereas Ablation_SRE obtains F1 $y_j$1. The drop of only $y_j$2 indicates that even oracle sub-claim labels add little when the evidence remains repeated and coarse. In the paper’s interpretation, SRE “underuses the decomposition structure because the evidence never becomes discriminative at sub-claim level” (Akhter et al., 11 Feb 2026).

These findings are consistent with several adjacent lines of work that emphasize the importance of evidence packaging and evidence granularity. “Read it Twice: Towards Faithfully Interpretable Fact Verification by Revisiting Evidence” (Hu et al., 2023) improves claim verification by replacing a full noisy document with a retriever-refined evidence subset in a second pass, while “Retrieve-Refine-Calibrate: A Framework for Complex Claim Fact-Checking” (Sun et al., 23 Jan 2026) similarly treats claim-centric evidence refinement as a remedy for noise introduced by decomposition. These papers do not define SRE, but they reinforce the distinction between merely having multiple evidence items and having evidence structured in a way that supports the target inference.

4. The alignment bottleneck

The paper’s explanation for SRE’s weak performance is that it fails to solve the actual problem decomposition is supposed to solve: granular evidence attribution. Its summary statement is concise: “When evidence is repeated at the claim level (SRE), decomposition adds structure without resolving ambiguity” (Akhter et al., 11 Feb 2026).

Under SRE, each sub-claim receives the full evidence package relevant to the whole claim, including information intended for other sub-claims or information irrelevant to the current one. The verifier must therefore perform an implicit alignment step before it can judge veracity. Because the experiments use manually controlled evidence rather than open-web retrieval noise, the paper argues that poor SRE performance cannot be reduced to retrieval failure. The problem is instead an evidence granularity mismatch: decomposition produces atomic reasoning units, but the evidence stays broad, redundant, and only loosely linked to those units.

The paper repeatedly describes SRE evidence as redundant claim-level evidence and coarse evidence that “introduce[s] noise that harms the final inference” (Akhter et al., 11 Feb 2026). A practical manifestation of this appears in prompting and context construction: SRE repeats the same evidence block after every sub-claim, which makes prompts much longer than in SAE and may amplify distraction and redundancy. The paper does not claim that context length alone causes failure, but it treats the inflation from 16384 to 40960 context tokens as consistent with SRE’s operational inefficiency.

A useful illustrative example in the paper is a Lufthansa/Germanwings crash claim decomposed into four sub-claims, one of which—“The jet was carrying 148 people”—is false while others are true. This shows why decomposition can in principle help: a mostly correct complex claim can still fail because one atomic detail is wrong. But the broader result of the paper is that this potential is only realized when each sub-claim is paired with the right evidence span. Under SRE, the model repeatedly sees the entire bundle of crash reports for every sub-claim, which makes it harder to isolate which sentence answers which factual component (Akhter et al., 11 Feb 2026).

The same diagnosis is visible, in different form, in other work on claim–evidence alignment. “The Warrant Gap: Claim-Conditioned Re-scoring for Fact-Checking” (Dey et al., 23 Jun 2026) shows that correct Supports verdicts can still rest on warrants that support only local facets or patchworks of partial evidence, and “Complex Claim Verification with Evidence Retrieved in the Wild” (Chen et al., 2023) shows that claim decomposition improves retrieval coverage only when the retrieved evidence package is subsequently compressed into a usable claim-level representation. These are not SRE papers in the strict sense, but they support the same broader point: evidence organization and claim–evidence alignment are separable from verdict prediction and can determine whether decomposition is actually beneficial.

5. Interaction with noisy sub-claim labels

A major contribution of (Akhter et al., 11 Feb 2026) is the claim that not all sub-claim errors are equally harmful. Once oracle labels are replaced by predicted sub-claim labels, SRE becomes especially brittle. In the noisy-label setting, the results are:

• Qwen Noisy_SAE: F1 $y_j$3
• Qwen Noisy_SRE: F1 $y_j$4
• GNN Noisy_SAE: F1 $y_j$5
• GNN Noisy_SRE: F1 $y_j$6

The paper states that under SRE “the models collapse,” with roughly 0.19 F1 loss relative to Oracle_SAE (Akhter et al., 11 Feb 2026). The interpretation is that coarse repeated evidence amplifies the effect of noisy intermediate labels because there is too little aligned evidence to correct wrong local decisions during aggregation.

The two sub-claim predictors analyzed in the paper have notably different error profiles:

• Qwen: $y_j$7, $y_j$8, $y_j$9, $Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$0, Cov$Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$1, Acc$Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$2, Acc$Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$3
• GNN: $Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$4, $Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$5, $Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$6, $Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$7, Cov$Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$8, Acc$Y_C = f(C,\{(s_j, E_j, y_j)\}_{j=1}^{m})$9, Acc$$E_{Ci} \subseteq \{ e \in \mathcal{D} : t(e) \leq t(C) \}.$$0

The paper characterizes Qwen as more aggressive, especially in predicting false/refuted sub-claims, and the GNN as more conservative, often predicting Unverified. Its substantive conclusion is that “the type of labeling error matters more than overall accuracy,” and that conservative “abstention” significantly reduces error propagation compared to aggressive but incorrect predictions (Akhter et al., 11 Feb 2026).

This point is especially consequential under SRE. Because the evidence supplied to each sub-claim is already coarse and ambiguous, a wrong polar intermediate label can dominate downstream claim aggregation more strongly than an abstention. Under aligned evidence, the same upstream accuracy can lead to better downstream robustness because sub-claim-specific evidence can help constrain aggregation. Under SRE, by contrast, the evidence never becomes discriminative enough to absorb or counteract mislabels.

This interaction between evidence multiplicity and aggregation strategy also appears in adjacent work. “SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation” (Qiu et al., 5 May 2026) argues that evidence sufficiency is a set-level property and that conflict, disagreement, and coverage deficits cannot be detected by independent passage scoring alone. That work differs from SRE because it usually treats one short answer claim rather than a decomposed complex claim, but its result is compatible with the observation that simply repeating or pooling evidence without calibrated aggregation can increase error rather than reduce it.

6. Relation to adjacent frameworks and broader significance

SRE names a specific decomposition setup, but it also functions as a diagnostic category for a broader family of systems in which evidence remains claim-level even after claims have been structurally decomposed. The reason it matters is that many practical and prior decomposition pipelines operate in SRE-like conditions: sub-claims are available, but sub-claim-aligned evidence spans are not (Akhter et al., 11 Feb 2026).

This makes SRE a useful conceptual contrast to several adjacent approaches. SAE is the most immediate contrast because it asks whether decomposition helps when evidence is actually aligned at sub-claim level. ClaimVer performs per-claim evidence attribution over a knowledge graph rather than repeating the same evidence package for each claim, producing outputs of the form $$E_{Ci} \subseteq \{ e \in \mathcal{D} : t(e) \leq t(C) \}.$$1 (Dammu et al., 2024). eTracer treats generated text as a set of atomic claims and builds a signed claim–sentence grounding matrix $$E_{Ci} \subseteq \{ e \in \mathcal{D} : t(e) \leq t(C) \}.$$2, thereby making claim-level evidence relations explicit rather than repeated wholesale (Chu et al., 7 Jan 2026). EvidenceLens likewise reframes auditing as a claim–evidence alignment problem and represents support, contradiction, and modality imbalance in a claim–evidence matrix (Gu et al., 19 Jun 2026). These systems are not implementations of SRE, but they illustrate alternative ways of preserving claim-level granularity at the evidence layer.

Historically, SRE also sits against older approaches that either concatenate all evidence or aggregate sentence-level decisions without explicit evidence-set structure. “Hierarchical Evidence Set Modeling for Automated Fact Extraction and Verification” (Subramanian et al., 2020) argues that concatenating all evidence introduces redundant and noisy information, while independent claim-sentence scoring misses early composition of related evidence. In that sense, SRE can be read as a decomposition-era analogue of the same structural concern: simply repeating coarse claim-level evidence after each sub-claim does not create the evidential conditions that decomposition presupposes.

The practical implication drawn in (Akhter et al., 11 Feb 2026) is direct. Systems should not assume that decomposition helps whenever sub-claims can be generated. What must be prioritized is “precise evidence synthesis and alignment,” along with label calibration at the sub-claim level. If aligned evidence is unavailable, SRE should be treated as a weak fallback rather than as evidence that decomposition-based claim verification is working. The paper’s own bottom line is that the real bottleneck is “not decomposition itself, but the capability for precise evidence synthesis and alignment” (Akhter et al., 11 Feb 2026).

In that sense, Repeated Claim-level Evidence is significant not because it is a desirable final design, but because it identifies a pervasive failure mode. It explains why decomposition-based verification can appear theoretically appealing yet empirically inconsistent: the claim structure was decomposed, but the evidence was not.