REVIEW 3 major objections 8 minor 1 cited by
The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content
T0 review · 3 major / 8 minor · reviewed 2026-07-05 · glm-5.2
Pith's one-line read KG triples steal 2-3x more attention than equivalent prose, regardless of relevance
desk verdict The paper identifies a real phenomenon — KG triples capturing 2–3× more attention per token than equivalent natural-language text — but the 'structural' attribution doesn't hold cleanly for both models, and the performance-effect claims are statistically underpowered. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The structural attention tax decomposition (Eq. 2): attention allocated to a knowledge region splits into a semantic relevance component and a structural capture component λ·σ(K). Proposition 1 provides a compression bound showing the structural term enters exponentially, and Definition 1 formalises σ(K) as a function of structured-pattern token density and repetitiveness. The framework generates four testable predictions: source dominance (P1), confidence modulation (P2), format-invariant capture (P3), and compression–performance decoupling (P4).
What would settle it
If attention masking or activation patching interventions (acknowledged as future work) fail to show that reducing KG-region attention restores demonstration attention and changes performance accordingly, the causal claim that format-driven attention capture compresses demonstrations would be weakened to a correlational observation.
Extended reading notes
Core claim
The central object is the structural attention tax: a format-driven, content-independent attention bias formalised as λ·σ(K), where σ(K) quantifies the structural intensity of a prompt region (triple density, delimiter frequency, slot repetitiveness) and λ is a model-intrinsic bias coefficient. The decomposition A_K = A_sem + λ·σ(K) separates attention into a semantic term (governing whether captured attention carries useful signal or noise) and a structural term (governing how much attention is captured). The structural term enters the demonstration compression bound exponentially (Proposition 1), meaning even modest format bias produces substantial compression of demonstration attention. A
Load-bearing premise
The framework assumes that attention scores can be cleanly separated into an additive semantic component and an additive structural component, where the structural bias is independent of the semantic content. If structural patterns interact with specific semantic categories — for example, if certain types of facts are more distracting when presented as triples than others — the clean decoupling between 'how much attention is taxed' and 'whether it helps or hurts' would break.
Editorial extensions
If this is right
- Format normalisation — converting structured data (KG triples, SQL, JSON, code blocks) to natural prose before injection — may be a standard preprocessing step for RAG pipelines, not an optional optimisation.
- The structural attention tax likely extends beyond KG triples to any prompt region with distinctive formatting patterns, including SQL queries, JSON objects, and code blocks, though this is untested in the paper.
- Retrieval quality (what you retrieve) and retrieval format (how you present it) are independent levers; improving one does not compensate for neglecting the other, and the semantic axis currently dominates by an order of magnitude.
- Training-time regularisation that penalises attention to noise-structured regions (Strategy S5) could durably reduce model-intrinsic format bias λ, but this remains untested.
- Gating strategies that decide whether to inject knowledge based on confidence (≤2 pp gains) are far less impactful than matching the knowledge source to the task (>30 pp gains), suggesting the field's focus on gating sophistication may be misplaced.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper introduces the 'structural attention tax': the claim that knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2–3× more attention per token than semantically equivalent natural-language text, compressing demonstration attention by up to 42% regardless of whether the triples are relevant or noise. The authors formalize this via a decomposition of attention scores into semantic and structural components (Eq. 2), derive a compression bound (Proposition 1), and test four predictions across two models (Mistral-7B, LLaMA-3-8B) and three QA benchmarks using a seven-condition study. They also compare KG retrieval against BM25 Wikipedia passage retrieval, finding a >30 pp source-alignment gap, and propose five mitigation strategies (two empirically tested). The paper is clearly written, the experimental design is thoughtful (particularly the C2/C3/C5 contrast and the C5b verbalization control), and the authors are commendably transparent about limitations. However, two issues bear on the central claim: the C5b control partially contradicts the format-driven (vs. content-driven) attribution for Mistral, and the statistical evidence for performance effects is weak (1/6 Bonferroni-corrected comparisons significant).
Significance. The paper addresses a real and underexplored question: whether the format of retrieved content, independent of its semantic relevance, systematically distorts attention in retrieval-augmented ICL. The C2/C3/C5 contrast is a clean experimental design for isolating format from content effects, and the multi-layer attention analysis (Appendix C.2) strengthens the claim beyond last-layer artifacts. The source-alignment dominance result (>30 pp gap) is a useful empirical anchor. The five-strategy mitigation taxonomy, grounded in the decomposition, provides a structured design space even though only two strategies are tested. The authors provide a full experiment-to-code mapping (Appendix H), which is a positive reproducibility signal. The theoretical framework is acknowledged as heuristic, which is appropriate given the correlational nature of the attention evidence.
major comments (3)
- §5.1, Table 21 (C5b): The verbalized-triple control is the cleanest test of the paper's central claim that attention capture is format-driven rather than content-driven. For LLaMA-3-8B, C5b reduces KG-region attention by 17–29% (ratios 0.71–0.83), supporting the format-driven attribution. However, for Mistral-7B, C5b attention barely changes from C2 (ratios 0.90, 0.90, 1.14 — TriviaQA actually increases), and C5b attention (7.4–9.4%) remains far above C5 neutral-text attention (~3%). This means that for Mistral, the elevated attention persists even when KG structural format is removed and the same content is presented as natural language. The paper acknowledges Mistral shows 'weaker and less consistent effects' but still claims the structural tax as a general phenomenon across both model families. The authors should either (a) scope the central claim to LLaMA-3-8B (or architectures with
- §5.2, Table 2: The statistical evidence for the confidence-dependent performance effect (Prediction 2) is weak. Only 1 of 6 Bonferroni-corrected McNemar comparisons is significant (LLaMA HotpotQA, p_Bonf = 0.001); the sign test across all six outcomes yields p = 0.69. The authors acknowledge this in §8 ('the confidence-dependent pattern is a directional trend requiring replication'), but the abstract and conclusion still present the confidence modulation as an established finding ('the semantic term determines whether this helps or hurts'). The framing in the abstract and conclusion should be tempered to match the statistical evidence — e.g., 'directional trend consistent with' rather than 'determines' — so that readers do not over-interpret the performance claims.
- §3.2, Definition 1 and Eq. 2: The structural capture potential σ(K) (Definition 1) is defined in terms of structural pattern tokens and repetitiveness, and the paper then validates the structural tax by showing that high-σ(K) regions capture more attention. This is somewhat definitional: regions with more structural-pattern tokens are defined as having high σ(K), and the validation shows they capture more attention. The paper would benefit from an independent operationalization of σ(K) — e.g., computing σ(K) from surface features alone and then predicting attention capture in held-out data — to demonstrate that the measure has predictive value beyond its definition. Currently, the C2/C3 similarity (both high σ, different relevance) and C5 contrast (low σ) provide circumstantial evidence, but a direct test of σ(K) as a predictor would strengthen the claim.
minor comments (8)
- §3.3, Eq. 3: The competition coefficient η is introduced with range [0.5, 1.0] but no justification is given for this range. A brief note on why this interval is appropriate would help readers.
- §4.2: The 4-bit NF4 quantization is noted to produce a +10 pp gap on HotpotQA C1 (Appendix A.2). This is a substantial perturbation relative to the 1–3 pp effect sizes in the main results. The paper should more prominently flag that fine-grained accuracy comparisons (e.g., C2 vs. C1 on CSQA: −2.0 pp) are within the quantization noise floor and should be interpreted as directional only.
- §5.3, Table 3: The C7 (BM25) vs. C2 (ConceptNet) comparison conflates retrieval source, retrieval unit (passages vs. triples), token budget, and text coherence. The authors acknowledge this in §8 ('confounded by retrieval-unit differences'), but the >30 pp gap is still presented as the paper's 'strongest finding' in the discussion (§7). The caveat should appear alongside the result where it is first cited.
- §6, Table 4: Three of five mitigation strategies (S2, S4, S5) are untested. The table marks these as 'Untested,' which is transparent, but the abstract's claim of 'five structure-aware mitigation strategies' could be read as implying empirical validation. Consider clarifying in the abstract that only two are empirically evaluated.
- Figure 1: The caption states 'Illustrative values; see Section 5 for measured data.' This is good, but the figure is visually prominent and the values (25%→48%, 35%→20%) are more dramatic than the measured data (e.g., Mistral C2 KG attention ~10%, not 48%). Consider using measured values or adding a note that the figure is schematic and exaggerates the effect for visibility.
- §3.2: The notation λ(l,h) suggests λ varies by layer and head, but the empirical estimates (λ ≈ 0.07–0.10 for Mistral, 0.03–0.06 for LLaMA) appear to be single scalar values. Clarify whether these are averages over layers/heads or from a specific layer.
- Appendix F.1.3: The calibration of the compression bound uses T₀ ≈ 350 and m ≈ 30 without explaining how these values were obtained. A brief note on the derivation would improve transparency.
- Table 1: The C6 column for Mistral TriviaQA (72.1) and LLaMA TriviaQA (73.1) appear to differ from the C6 values in Table 10 (72.1 and 73.1 respectively — consistent on re-reading, but the formatting makes cross-referencing difficult). A footnote or consistent formatting in the table itself would help.
Simulated Author's Rebuttal
We thank the referee for a careful and constructive review. The referee identifies three substantive issues; we agree with all three and will revise accordingly. Two require scoping/tempering of claims, and one suggests an additional experiment that we will conduct. We address each below.
read point-by-point responses
-
Referee: §5.1, Table 21 (C5b): For Mistral-7B, C5b attention barely changes from C2 (ratios 0.90, 0.90, 1.14), and C5b attention (7.4–9.4%) remains far above C5 neutral-text attention (~3%). The paper claims the structural tax as a general phenomenon across both model families but the C5b control partially contradicts the format-driven attribution for Mistral.
Authors: The referee is correct, and we will revise the manuscript to address this. We agree that the C5b results for Mistral-7B do not support the format-driven attribution as cleanly as the LLaMA-3-8B results do. For LLaMA-3-8B, verbalization reduces KG-region attention by 17–29% (ratios 0.71–0.83), which is consistent with the structural tax being format-driven. For Mistral-7B, the ratios (0.90, 0.90, 1.14) show that verbalization does not reliably reduce attention capture, and the residual attention (7.4–9.4%) remains well above the C5 neutral-text baseline (~3%). This means that for Mistral, some component of the elevated attention persists even when the KG structural format is removed and the same content is presented as natural language — which could reflect a content-driven component, a model-specific sensitivity to the verbalized triple content, or an interaction between verbalization style and Mistral's attention patterns that our control does not fully isolate. We do not have a definitive explanation for why Mistral behaves differently, and we will not speculate beyond the data. We will make the following changes: (1) In the abstract, we will scope the format-driven claim to note that the C5b verbalization control confirms the format-driven attribution for LLaMA-3-8B, while Mistral-7B shows the format-invariance pattern (C2≈C3 attention) but weaker evidence from the verbalization control. (2) In §5.1 and §8, we will explicitly state that the structural attention tax is most cleanly demonstrated in LLaMA-3-8B, and that the Mistral results support the format-invariance prediction (P3: noise and relevant triples capture similar attention) but not the format-flattening prediction (S3) as strongly. (3) We will adjust the conclusion to reflect this model-dependent evidence. revision: yes
-
Referee: §5.2, Table 2: Only 1 of 6 Bonferroni-corrected McNemar comparisons is significant; sign test p = 0.69. The abstract and conclusion still present the confidence modulation as an established finding ('the semantic term determines whether this helps or hurts'). The framing should be tempered.
Authors: The referee is correct. The statistical evidence does not support presenting the confidence-dependent performance effect as an established finding. Only LLaMA-3-8B on HotpotQA survives Bonferroni correction (p_Bonf = 0.001); the sign test across all six outcomes is non-significant (p = 0.69). We already acknowledge this in §8 ('the confidence-dependent pattern is a directional trend requiring replication'), but the abstract and conclusion overstate the result. We will revise the abstract to replace 'the semantic term governs whether this helps or hurts' with language such as 'the semantic term is hypothesized to govern whether this helps or hurts, though the performance evidence remains a directional trend (1/6 comparisons significant after Bonferroni correction).' We will make a corresponding change in the conclusion, replacing 'the semantic term determines whether augmentation helps' with 'the semantic term is expected to determine whether augmentation helps, consistent with a directional trend that requires replication.' We will also add a brief caveat in §5.2 noting that the theoretical framework predicts the confidence-dependent pattern but the empirical evidence is underpowered to confirm it. revision: yes
-
Referee: §3.2, Definition 1 and Eq. 2: σ(K) is defined in terms of structural pattern tokens and repetitiveness, and the paper validates the structural tax by showing that high-σ(K) regions capture more attention. This is somewhat definitional. The paper would benefit from an independent operationalization of σ(K) — e.g., computing σ(K) from surface features alone and then predicting attention capture in held-out data.
Authors: We agree that the current validation is circumstantial rather than a direct predictive test of σ(K), and that this is a legitimate methodological gap. The referee's suggestion — computing σ(K) from surface features alone and predicting attention capture on held-out data — is a clean way to demonstrate that the measure has predictive value beyond its definition. We will conduct this analysis in revision. Specifically, we will: (1) Compute σ(K) for each knowledge region in our existing conditions using only surface features (structural-pattern token density and repetitiveness as defined in Definition 1), without reference to attention values. (2) Additionally generate a set of held-out prompt variants with intermediate σ(K) values — e.g., partially structured formats mixing triples with prose at varying ratios — to provide a continuous range of σ(K) rather than just the binary high/low contrast (C2/C3 vs. C5). (3) Test whether σ(K) predicts KG-region attention capture in these held-out variants using regression or rank correlation. If σ(K) predicts attention capture on held-out data with formats not used to define the measure, this would address the circularity concern. If the predictive relationship is weak or absent, we will report that honestly and discuss what it implies for the framework. We note that even if σ(K) proves to be a strong predictor, this remains correlational evidence; causal validation via attention masking is important future work that we already flag in §8. revision: yes
Circularity Check
σ(K) is defined by structural pattern tokens; validating that high-σ(K) regions capture more attention is partially definitional. λ is estimated from the same attention data it explains.
-
self definitional
[Definition 1 (Section 3.2) and Section 5.1 validation]
"σ(R) = γ · (1/m) Σ_{j∈R} I[token_j ∈ P_struct] + β_rep · rep(R), where P_struct is the structured-pattern token set (relation keywords, delimiters, slot markers) and rep(R) quantifies repetitiveness."
σ(K) is defined as the density of structural-pattern tokens (delimiters, relation keywords, slot markers) in a region. The paper then validates the 'structural attention tax' by showing that high-σ(K) regions (KG triples, which are by construction composed of these delimiters and slot markers) capture more attention than low-σ(K) regions (neutral text, which lacks these tokens by construction). The prediction that 'regions with more structural-pattern tokens capture more attention' is partially entailed by the definition of σ(K) as a count of those very tokens. The empirical observation that KG triples (high σ) get more attention than neutral text (low σ) is consistent with the definition but does not independently verify that the mechanism is structural rather than semantic, because the σ
-
fitted input called prediction
[Section 5.1 ('Quantifying the tax') and Appendix F.1.3]
"KG tokens receive ≈ 0.34%/token (Mistral) vs. 0.13%/token for demonstrations (2.6×), with estimated structural bias λ ≈ 0.07–0.10 (Mistral) and 0.03–0.06 (LLaMA). The exponential amplification in Eq. 5 explains why even modest λ·σ(K) ≈ 0.05–0.07 produces substantial compression."
The parameter λ is described as a 'model-intrinsic bias coefficient' in Eq. 2, but it is estimated from the same attention data (the ratio of KG-token attention to demonstration-token attention) that the framework purports to explain. In Appendix F.1.3, λ·σ ≈ 0.05–0.07 is derived from the observed compression ratio A(K)_D / A(0)_D = 0.58. The framework then 'explains' the observed compression by plugging this fitted λ back into Eq. 5's exponential. The prediction that 'λ·σ(K) produces substantial compression' is forced by construction: λ was calibrated from the compression ratio it is said to explain. The framework organizes the observation but does not independently predict the compression magnitude.
full rationale
The paper contains two partially circular elements. (1) σ(K) (Definition 1) is defined as the density of structural-pattern tokens (delimiters, relation keywords, slot markers), and the central empirical validation shows that regions high in these tokens capture more attention — the prediction is partially entailed by the definition. (2) The structural bias parameter λ is estimated from the same attention compression data that Eq. 5 then 'explains,' making the compression bound's explanatory power partly a restatement of its calibration input. However, the paper is not deeply circular: the C2/C3 contrast (relevant vs. noise triples showing similar attention) is a genuine empirical finding not forced by the definition, and the C5b verbalization control provides an independent (if model-dependent) test. The framework is best understood as a heuristic organizing tool — which the authors acknowledge ('Eq. 2 assumes additive separation; Eq. 10 is a heuristic') — rather than a derivation whose outputs are fully determined by its inputs. Score 4 reflects partial circularity in the σ(K) definition and λ estimation, while the central empirical claims retain independent content.
Assumptions & free parameters
free parameters (6)
- λ (structural bias coefficient) =
0.07-0.10 (Mistral), 0.03-0.06 (LLaMA)
- σ(K) (structural intensity) =
≈0.70 (KG), ≈0.25 (neutral)
- η (competition coefficient) =
[0.5, 1.0]
- τ (gating threshold for C4) =
-0.3
- c (logit suppression constant for S2) =
[0.5, 1.5]
- α_reg (adversarial regularization weight for S5) =
unspecified
assumptions (5)
- ad hoc to paper Attention scores can be additively decomposed into semantic and structural components: s_ij = s_sem + b_j (Eq. 1-2).
- ad hoc to paper The structural term λ·σ(K) and semantic term s̄_sem_K are orthogonal improvement axes.
- domain assumption Last-layer attention distribution is a valid proxy for the model's internal processing of prompt regions.
- domain assumption 4-bit NF4 quantisation preserves the relative attention patterns needed to observe the structural tax.
- standard math Jensen's inequality applied to attention logits yields a meaningful compression bound (Proposition 1).
invented entities (2)
-
Structural attention tax (λ·σ(K))
independent evidence
-
Structural capture potential σ(R)
Cite this review
Pith. "Pith review of The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content." pith.science (2026). https://pith.science/paper/JR6QTSCB
@misc{pith2026260611198,
author = {Pith},
title = {Pith review of: The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content},
year = {2026},
howpublished = {\url{https://pith.science/paper/JR6QTSCB}},
note = {Machine review of arXiv:2606.11198}
}
abstract
Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution. We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3x more attention per token than semantically equivalent natural-language text ($\hat{o}$(KG) $\approx$ 0.70 vs. $\hat{o}$(neutral) $\approx$ 0.25), compressing demonstration attention by up to 42% -- regardless of whether the triples are relevant or noise. We develop a formal framework decomposing attention scores into semantic and structural components (Eq. 2), derive a compression bound (Proposition 1) connecting token-level format bias to demonstration attention loss, and show that the structural term governs how much attention is diverted while the semantic term governs whether this helps or hurts. This decoupling reveals two orthogonal axes for improving retrieval-augmented ICL: optimising retrieval quality (semantic axis) and reducing format-driven attention capture (structural axis). Empirically, across two model families (Mistral-7B, LLaMA-3-8B) and three QA benchmarks, we observe that source-task alignment dominates: task-matched BM25 retrieval achieves 58-62% on HotpotQA vs. ConceptNet's 25-27%, a >30 pp gap that dwarfs all gating strategies ($\leq$2 pp). We derive five structure-aware mitigation strategies from the framework, ranging from zero-cost prompt modifications to training-time regularisation; format flattening (S3) is validated by both accuracy and attention-level evidence from a verbalized-triple control, while structural dispersal (S1) yields mixed results that illuminate the challenges of format-level intervention.
Figures
Figures from the paper (2 more)
Forward citations
Cited by 1 Pith paper
-
The Announcement Carries the Cue: Markup, Boundaries, and the Notation of Pre-Training Corpora
Deleting structural announcements makes following prose harder for LLMs to predict, swapping notation does nothing, and the paper proposes a pure-frame format that strips announcements into sidecars.
Reference graph
Works this paper leans on
-
[1]
Joshi, Mandar and Choi, Eunsol and Weld, Daniel and Zettlemoyer, Luke , booktitle=. Trivia
-
[2]
Dubey, Abhimanyu and others , journal=. The
-
[3]
Note on the sampling error of the difference between correlated proportions or percentages , author=. Psychometrika , volume=
-
[4]
Advances in Neural Information Processing Systems , volume=
Language models are few-shot learners , author=. Advances in Neural Information Processing Systems , volume=
-
[5]
Talmor, Alon and Herzig, Jonathan and Lourie, Nicholas and Berant, Jonathan , booktitle=. Commonsense
-
[6]
Yang, Zhilin and others , booktitle=. Hotpot
-
[7]
Speer, Robyn and Chin, Joshua and Havasi, Catherine , booktitle=. Concept
- [8]
Show all 31 references
-
[9]
Proceedings of ACL , pages=
Few-shot in-context learning on knowledge base question answering , author=. Proceedings of ACL , pages=
-
[10]
Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive
Khattab, Omar and others , journal=. Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive
-
[11]
Huang, Qian and others , booktitle=
-
[12]
Proceedings of EMNLP , pages=
Rethinking the role of demonstrations: What makes in-context learning work? , author=. Proceedings of EMNLP , pages=
-
[13]
Proceedings of EMNLP , pages=
Skill-based few-shot selection for in-context learning , author=. Proceedings of EMNLP , pages=
-
[14]
How much can
Liu, Jingyu and others , journal=. How much can
-
[15]
Advances in Neural Information Processing Systems , volume=
Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in Neural Information Processing Systems , volume=
-
[16]
Proceedings of EMNLP , pages=
Active retrieval augmented generation , author=. Proceedings of EMNLP , pages=
-
[17]
Kazemi, Seyed Mehran and others , booktitle=
-
[18]
and others , journal=
Sun, Jiaxin A. and others , journal=. Search-on-Graph: Iterative informed navigation for
-
[19]
Sui, Yuan and He, Yufei and Liu, Nian and He, Xiaoxin and Wang, Kun and Hooi, Bryan , booktitle=. Fi
-
[20]
Proceedings of ACL , pages=
Disentangling memory and reasoning ability in large language models , author=. Proceedings of ACL , pages=
-
[21]
arXiv preprint arXiv:2310.04680 , year=
The cost of down-scaling language models: Fact recall deteriorates before in-context learning , author=. arXiv preprint arXiv:2310.04680 , year=
-
[22]
Proceedings of EMNLP , pages=
Can we edit factual knowledge by in-context learning? , author=. Proceedings of EMNLP , pages=
-
[23]
Proceedings of ICML , pages=
Large language models can be easily distracted by irrelevant context , author=. Proceedings of ICML , pages=
-
[24]
arXiv preprint arXiv:2510.13905 , year=
Schema for in-context learning , author=. arXiv preprint arXiv:2510.13905 , year=
-
[25]
Findings of ACL , pages=
Concept-aware data construction improves in-context learning of language models , author=. Findings of ACL , pages=
-
[26]
Proceedings of ICLR , year=
Selective annotation makes language models better few-shot learners , author=. Proceedings of ICLR , year=
-
[27]
Towards a unified view of answer calibration for multi-step reasoning , author=. Proc. 2nd Workshop on Natural Language Reasoning and Structured Explanations (NLRSE @ ACL 2024) , pages=
2024
-
[28]
Proceedings of ECAI , year=
Investigating the learning behaviour of in-context learning: A comparison with supervised learning , author=. Proceedings of ECAI , year=
-
[29]
``In-context learning'' or: How
Parry, Alistair and Ganguly, Debasis and Chandra, Mandar , booktitle=. ``In-context learning'' or: How
-
[30]
Proceedings of NAACL-HLT , pages=
Learning to retrieve prompts for in-context learning , author=. Proceedings of NAACL-HLT , pages=
-
[31]
How faithful are
Wu, Kevin and Wu, Eric and Zou, James , journal=. How faithful are
Reviewed July 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.