Relevance–Robustness Gap

• Relevance–Robustness Gap is the mismatch between what is considered truly relevant and what a system actually uses, leading to performance instability under perturbations.
• It spans various domains including language models, zero-shot summarization, computer vision, and search relevance, each revealing domain-specific failure modes.
• Overcoming the gap requires coupling relevance with measures like invariance, explicit abstention, and counterfactual risk minimization, alongside robust evaluation metrics.

The relevance–robustness gap denotes a family of failures in which a system’s criterion for selecting, weighting, or explaining “relevant” information does not align with the conditions required for stable performance under perturbation, distribution shift, noisy evidence, or genuinely irrelevant additions. In retrieval-augmented LLMs, it is explicitly defined as the mismatch between “what is truly relevant and reliable” and “what the system actually uses” (Raj et al., 28 Jul 2025). In zero-shot summarization, the same phrase is used for the consistent degradation that appears when only the most relevant source sentences are minimally paraphrased (Askari et al., 2024). In computer vision, closely related work studies a substantial gap between clean and corrupted performance under naturally induced alterations (Drenkow et al., 2021), while in structured argumentation relevance desiderata are framed as robustness under the addition of irrelevant information (Borg et al., 2018). Across these settings, relevance alone is insufficient unless it is coupled to invariance, calibration, or explicit abstention.

1. Cross-domain formulations

The term is not used identically across all fields, but the documented formulations are tightly related.

Setting	Relevance notion	Robustness failure
RALMs and RAG	Semantic fit, source reliability, evidential utility	Hallucination, sycophantic retrieval, failure to abstain
Zero-shot summarization	Sentences most relevant to the gold summary	Output instability under meaning-preserving paraphrase
Computer vision	Features, foreground regions, or clean-distribution cues	Clean-to-corrupt or ID-to-OOD performance drop
Structured argumentation	Information sharing propositional variables or derived consequences	Instability under irrelevant additions
Goodness-of-fit testing	Practical adequacy of a model	Rejection of mild perturbations that may be task-irrelevant

In structured relevance assessment for RALMs, the gap is introduced as the consequence of three specific failures: retrieved documents are often treated as equally useful, external retrieval is over-trusted even when intrinsic knowledge suffices, and the system lacks a mechanism to “admit ignorance” (Raj et al., 28 Jul 2025). In summarization, relevance paraphrasing exposes a different failure mode: changing only the sentences that directly explain the gold summary, while preserving meaning, still causes large score drops (Askari et al., 2024). In non-adversarial vision robustness, the central issue is the difference between expected performance on clean data and on naturally corrupted data, often with degradation of the same order of magnitude as adversarial conditions (Drenkow et al., 2021). In structured argumentation, relevance postulates are explicitly formulated as robustness of consequence relations under the addition of information sharing no propositional variables with the original database (Borg et al., 2018).

This suggests that the phrase functions as a cross-disciplinary descriptor for a common mismatch: systems often optimize for an operational notion of relevance that is too narrow, too surface-level, or too brittle for deployment conditions.

2. Retrieval-augmented LLMs and robust retrieval

A direct formulation appears in “Structured Relevance Assessment for Robust Retrieval-Augmented LLMs,” which defines the gap as the mismatch between “what is truly relevant and reliable” and “what the system actually uses” (Raj et al., 28 Jul 2025). The proposed framework uses a two-component score

$$S_i = \alpha\,\mathrm{Sim}(E_Q,E_{d_i})+\beta\,R(d_i),\qquad \alpha+\beta=1$$

and a normalized variant

$$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$

Here, $\mathrm{Sim}(E_Q,E_{d_i})$ is cosine similarity between query and document embeddings, while $R(d_i)\in[0,1]$ is a source-reliability score obtained by mapping discrete source classes from A to E, plus F for indeterminate, into a continuous interval. Embeddings are extracted from DeepSeek-R1-1.5B, Llama3.2-1B, and Qwen2.5-1.5B, with vectors stored in a GPU-accelerated vector database such as FAISS. The same framework adds synthetic mixed-quality training data, a four-way knowledge-integration switch between intrinsic knowledge and retrieval, and an explicit “Unknown” protocol for unanswerable queries. The paper reports a 40% reduction in hallucination rate, 100% accuracy in identifying when a document was retrieval vs. parametric, 57.1% success in detecting hallucinations prior to generation, and correct abstention on >90% of truly unanswerable questions, compared with <5% in baselines (Raj et al., 28 Jul 2025).

A more explicitly causal formulation appears in “Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation,” which argues that semantic similarity is an inadequate proxy for utility when queries contain false premises, confirmation bias, or topical distraction (Liu et al., 2 May 2026). The paper defines a robustness utility

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$

contrasting standard retrieval based on $P(d\mid X)$ with a counterfactual objective approximating $P(Y=Y^*\mid do(X=x),D=d)$. Training uses a Cognitive Perturbation Protocol with three perturbation types—False-Premise, Confirmation-Bias, and Topical Distraction—and distills the resulting supervision into an Evidence Critic. In the reported experiments, CoRM-RAG achieves 52.6% on Biased-NQ and a clean-to-biased gap of 1.3%, compared with 6.3% for Contriever and Cross-Encoder (PA-CE); it is also reported as approximately 93× faster than GPT-4o reranking while slightly more accurate on the adversarial benchmark (Liu et al., 2 May 2026).

A complementary evaluation perspective is given by SCORE, which treats robustness and answer relevance as separate axes rather than a single scalar quality measure (Shomee et al., 10 Feb 2026). Robustness is decomposed into paraphrase invariance,

$R_{\mathit{para}}(q)=\mathrm{sim}(a,a_p),$

and semantic perturbation sensitivity,

$R_{\mathit{pert}}(q)=\mathrm{sim}(a,a_h),$

while answer relevance is estimated by inverse-question generation and reranking:

$$R(q,a)=\frac{1}{n}\sum_{i=1}^n \mathrm{sim}(q,\hat q_i).$$

The paper’s results show that no single metric sufficiently captures answer quality in isolation, reinforcing the idea that robust retrieval and generation require structured, multi-metric assessment rather than semantic relevance alone (Shomee et al., 10 Feb 2026).

3. Minimal perturbations in zero-shot summarization

In zero-shot abstractive summarization, the relevance–robustness gap is operationalized through relevance paraphrasing (Askari et al., 2024). Let $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$0 denote the article corpus, $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$1 the gold summaries, and $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$2 the zero-shot summarizer. For each gold summary sentence $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$3, a mapping function $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$4 selects the most relevant source sentence using tf–idf similarity:

$$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$5

The index set of the most relevant article sentences is

$$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$6

A minimally perturbed article $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$7 is then formed by paraphrasing exactly those sentences:

$$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$8

Robustness is measured as the relative change in standard summarization metrics when moving from $$S_i^{\mathrm{norm}} = Z_1\,\alpha\,\mathrm{Sim}(E_Q,E_{d_i})+Z_2\,\beta\,R(d_i).$$9 to $\mathrm{Sim}(E_Q,E_{d_i})$0:

$\mathrm{Sim}(E_Q,E_{d_i})$1

The reported metrics are ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore. Across CNN/DM, XSum, News, and Reddit, and across GPT-3.5-Turbo, Llama-2-13B, Mistral-7B, and Dolly-v2-7B, the paper reports substantial drops, often 20–50% on ROUGE-2. On CNN/DM, for example, GPT-3.5-Turbo shows –23.73% on ROUGE-2 and Dolly-v2-7B shows –31.66%; on News, Dolly-v2-7B shows –50.30% on ROUGE-2 (Askari et al., 2024).

The authors attribute the gap to two mechanisms. First, sentence selection brittleness: after paraphrasing the relevant sentences, the model shifts to different source sentences and often loses the salient content. Second, reliance on surface cues: because paraphrasing preserves meaning, a fully robust summarizer should remain invariant, but the observed drops indicate dependence on exact wording or phraseology rather than abstract semantic content (Askari et al., 2024). A plausible implication is that benchmark scores on unperturbed inputs can conceal a substantial instability precisely at the locations most responsible for the summary.

4. Vision: foreground relevance, corruption robustness, and shortcut suppression

The vision literature documents a closely related gap between feature relevance and out-of-distribution stability. A systematic review defines non-adversarial robustness as preserving model performance under naturally induced image corruptions or alterations, and formalizes it via expected loss on clean and corrupted distributions:

$\mathrm{Sim}(E_Q,E_{d_i})$2

A model is robust if $\mathrm{Sim}(E_Q,E_{d_i})$3. The review reports that even with state-of-the-art tactics, the absolute clean-to-corrupt accuracy drop remains on the order of 20–40%; for example, Adversarial Noise Training improves corrupt accuracy on ImageNet-C from 39.2% to 50.4%, but the gap to clean accuracy 76.1% is still 25.7% (Drenkow et al., 2021). The associated structural causal model represents image formation through task, environment, distractors, concept, sensor, renderer, image, and label nodes, with corruptions modeled as soft interventions on nuisance factors.

Several methods then attempt to close the gap by regularizing or redistributing relevance. “Optimizing Relevance Maps of Vision Transformers Improves Robustness” computes Generic Attention Explainability relevance maps and finetunes ViTs with three losses: background suppression,

$\mathrm{Sim}(E_Q,E_{d_i})$4

foreground coverage,

$\mathrm{Sim}(E_Q,E_{d_i})$5

and confidence maximization,

$\mathrm{Sim}(E_Q,E_{d_i})$6

With $\mathrm{Sim}(E_Q,E_{d_i})$7, $\mathrm{Sim}(E_Q,E_{d_i})$8, and $\mathrm{Sim}(E_Q,E_{d_i})$9, the paper reports average top-1 changes across seven architectures of +5.8 on ImageNet-A, +2.7 on ImageNet-R, +0.9 on ImageNet-Sketch, +5.0 on ObjectNet, with only –0.8 on ImageNet validation (Chefer et al., 2022).

“Relevance-driven Input Dropout” instead uses Layer-wise Relevance Propagation (LRP) to identify the most relevant pixels or points and selectively occlude them during training (Gururaj et al., 27 May 2025). The augmented input is

$R(d_i)\in[0,1]$0

and training minimizes expected classification loss on $R(d_i)\in[0,1]$1. On ResNet-50 finetuned on CIFAR-100, the reported accuracies are 79.01% for the baseline, 79.72% for Random Erasing, and 80.35% for RelDrop; on ImageNet-R, zero-shot accuracy increases from 37.85% to 38.65%, while Mean Relevance Rank Accuracy rises from 61.74% to 62.47%. The paper also reports a compute overhead of approximately 2–2.5× slower per epoch because of the extra backward pass required for LRP (Gururaj et al., 27 May 2025).

“Improving deep neural network generalization and robustness to background bias via layer-wise relevance propagation optimization” treats the gap as a mismatch between apparently plausible relevance maps and poor OOD generalization caused by shortcut learning (Bassi et al., 2022). The total training objective is

$R(d_i)\in[0,1]$2

The method, ISNet, penalizes relevance over background and stabilizes foreground relevance. Reported OOD results include macro-F1 0.773 \pm 0.009 for COVID-19 detection, compared with 0.645 \pm 0.009 for a U-Net+DenseNet pipeline and 0.546 \pm 0.010 for DenseNet alone; for tuberculosis detection, ISNet reports macro-F1 0.738 \pm 0.044, compared with 0.576 \pm 0.052 for the pipeline and 0.566 \pm 0.050 for DenseNet (Bassi et al., 2022).

Taken together, these results indicate that robustness improvements in vision often arise not from generic regularization alone, but from explicit constraints on where a model is allowed to place relevance.

5. Search relevance modeling under distribution shift

In e-commerce search, the relevance–robustness gap is stated as the inability of lightweight representation-based models to match the accuracy and generalization of interaction-based models, especially on domain-specific terms, adversarial or noisy inputs, and long-tail queries and items not seen in training (Chen et al., 2024). “Robust Interaction-Based Relevance Modeling for Online e-Commerce Search” addresses this with three components. The dynamic-length representation scheme truncates each batch to the maximum non-pad length actually present, reducing attention complexity from $R(d_i)\in[0,1]$3 to $R(d_i)\in[0,1]$4; in production, the paper reports approximately 36% latency drop, and a hot-query cache gives a further ~48% reduction. The professional-terms recognition method extends BERT’s vocabulary with the top 20 K raw words from 12 months’ of Alibaba search logs and adds NER-style tag embeddings, reducing average tokens per query-item pair from ~41.6 to 36.6 and improving AUC by +1.5%. The contrastive adversarial training protocol combines binary cross-entropy, adversarial CE, and bidirectional KL regularization:

$R(d_i)\in[0,1]$5

On 100 K manually annotated query-item pairs, the reported best AUC is 0.9033 for the full 3-layer system, compared with 0.9011 for StructBERT_base and 0.8964 for RoBERTa_base. In online A/B testing, incremental gains include CTR +1.54%, CVR +1.02%, and PAY +1.87% for the full CAT stage (Chen et al., 2024).

“Boosting LLM-based Relevance Modeling with Distribution-Aware Robust Learning” addresses a related gap in Alipay Search, where naive LLM fine-tuning is described as overconfident, weak at distinguishing fine-grained relevance levels, and vulnerable to data distribution shift (Liu et al., 2024). DaRL uses three labels—strongly relevant, weakly relevant, and irrelevant—and augments binary cross-entropy with a KL term,

$R(d_i)\in[0,1]$6

where $R(d_i)\in[0,1]$7 imposes a prior spread across relevance levels so that weakly relevant examples are not collapsed toward $R(d_i)\in[0,1]$8. The Distribution-Aware Sample Augmentation module selects OOD samples with both Mahalanobis distance and $R(d_i)\in[0,1]$9-NN distance in pooled hidden-state space, and the LPFT schedule applies linear probing, full fine-tuning, and final interpolation

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$0

with the paper stating that $$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$1 worked best. Reported results on AntGLM-0.3B show F1 rising from 0.9252 to 0.9349 on Test_Tid and from 0.6431 to 0.6678 on Test_Tood; live A/B testing reports 1.2% lift in click-through rate on valid PVs, 3.0% lift in conversion rate, 4% drop in irrelevant results, and 3% rise in highly relevant results (Liu et al., 2024).

These search papers converge on a specific conclusion: robustness is not only a matter of adversarial training, but also of representing graded relevance, domain vocabulary, and distributional coverage more faithfully.

6. Formal extensions beyond neural retrieval

Outside neural retrieval and generation, the same conceptual tension appears in statistics, probabilistic modeling, and logic. In robust goodness-of-fit testing, the issue is that classical null hypotheses are often statistically exact but practically irrelevant. “On the Robustness of Kernel Goodness-of-Fit Tests” replaces the point null $$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$2 with a composite null $$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$3 and distinguishes qualitative robustness

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$4

from quantitative robustness

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$5

The paper shows that standard kernel Stein discrepancy tests are not robust under common robustness notions, even with tilted kernels, and proposes a KSD-ball uncertainty set

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$6

with robust statistic

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$7

The resulting robust-KSD test is proved to control Type I error over the entire KSD ball while remaining consistent outside it (Liu et al., 2024). Here the relevance–robustness tension concerns whether a deviation from the model is practically meaningful rather than merely detectable.

In Gaussian process regression, “Robust Gaussian Processes via Relevance Pursuit” frames the problem as an accuracy–robustness trade-off under sparse outliers (Ament et al., 2024). Standard GPs with homoskedastic Gaussian noise use every datum fully when the model is correct, but can be dominated by a few corruptions. Relevance Pursuit introduces point-specific noise variances $$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$8 through

$$U(d,X)=\mathbb{E}_{\delta\sim P(\delta)}\bigl[I(M(X\oplus\delta,d)=Y)\bigr],$$9

with $P(d\mid X)$0. The paper shows that a reparameterized log marginal likelihood can be made strongly concave, implying weak submodularity of the associated subset selection problem and enabling a greedy algorithm with guarantee

$P(d\mid X)$1

Empirically, the paper reports that RRP nearly matches standard GP performance with no outliers, while heavy-tailed GPs only dominate when outliers exceed approximately 40% (Ament et al., 2024).

In structured argumentation, relevance is formalized directly as robustness under irrelevant additions (Borg et al., 2018). Non-Interference requires that if $P(d\mid X)$2, then

$P(d\mid X)$3

The underlying deducibility relation is Pre-Relevant if whenever $P(d\mid X)$4 and $P(d\mid X)$5, there exists $P(d\mid X)$6 with $P(d\mid X)$7. The paper proves that Pre-Relevance yields Non-Interference for induced consequence relations under grounded, complete, and preferred semantics, while Cut and pointedness yield cumulativity for grounded semantics (Borg et al., 2018). In this setting, the relevance–robustness gap is not empirical but logical: a consequence relation fails if conclusions can be changed by information that is formally irrelevant.

7. Persistent trade-offs and open directions

Across the surveyed work, closing the relevance–robustness gap introduces explicit trade-offs rather than eliminating them. In structured relevance assessment for RALMs, source reliability is described as a hard classification task, especially for newly emerging domains or quietly published but factually correct content; careful tuning of $P(d\mid X)$8, $P(d\mid X)$9, and thresholds $P(Y=Y^*\mid do(X=x),D=d)$0 is required to avoid excessive abstentions or blind trust, and latency must balance the retrieval cost $P(Y=Y^*\mid do(X=x),D=d)$1 against integration cost $P(Y=Y^*\mid do(X=x),D=d)$2 with

$P(Y=Y^*\mid do(X=x),D=d)$3

(Raj et al., 28 Jul 2025). In CoRM-RAG, robustness is improved by replacing purely observational retrieval with counterfactual risk minimization, but inference still depends on a safety threshold $P(Y=Y^*\mid do(X=x),D=d)$4 that determines whether the system abstains (Liu et al., 2 May 2026).

Evaluation itself remains multi-dimensional. SCORE reports that no single metric sufficiently captures answer quality in high-stakes, domain-specific settings, and its human study also documents uneven agreement across annotation dimensions: hazard and location show high agreement, whereas timeline is substantially harder (Shomee et al., 10 Feb 2026). This supports the view that relevance and robustness should be evaluated jointly with specificity and context utilization rather than collapsed into a single score.

Vision robustness remains far from solved. The systematic review argues for more precise mathematical definitions of allowed interventions, better evaluation protocols that avoid training on the same tail distributions later used for testing, and new metrics that account for corruption likelihoods and downstream safety consequences (Drenkow et al., 2021). Explanation-guided methods also impose operational constraints: RelDrop adds approximately 2–2.5× training cost per epoch (Gururaj et al., 27 May 2025), whereas ISNet requires foreground masks during training and scales training cost with the number of classes (Bassi et al., 2022).

In statistical testing, the trade-off is explicit: larger KSD-ball radii $P(Y=Y^*\mid do(X=x),D=d)$5 improve immunity to mild deviations but reduce power against subtle alternatives (Liu et al., 2024). In Gaussian processes, pointwise relevance pursuit gives strong resilience to sparse corruptions, but the paper notes that heavy-tailed alternatives may be preferable once outlier prevalence becomes large (Ament et al., 2024). In structured argumentation, strengthening syntactic relevance through Pre-Relevance does not automatically secure semantic stability under all semantics; preferred and stable semantics require additional conditions such as Cut, contraposition, and consistency restrictions (Borg et al., 2018).

A consistent implication across these literatures is that robustness cannot be recovered by post hoc calibration of a narrow relevance signal. The successful approaches instead expand the relevance notion itself—by adding trustworthiness, foreground constraints, counterfactual utility, OOD-aware sampling, or composite nulls—and then couple that richer notion to abstention, causal intervention, or formal stability guarantees.