Pith. sign in

REVIEW 3 major objections 4 minor 27 references

Source Attribution in Retrieval-Augmented Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Shapley-based attribution can identify influential RAG documents, but every tested method under-credits the synthesis document in multi-hop answers.

desk verdict A solid empirical benchmark for Shapley-style attribution in RAG; the synergy result is suggestive but needs verification of its core premise. read the letter →

arxiv 2507.04480 v1 pith:HTFXWN6S submitted 2025-07-06 cs.LG cs.AI

classification cs.LGcs.AI
keywords Retrieval-AugmentedGenerationShapleyvaluesdocumentattributionLLMinterpretabilityKernelSHAPContextCitemulti-hopreasoningexplainability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether Shapley values—the standard game-theoretic way to divide credit among inputs—can be adapted to say which retrieved documents actually drove an LLM's answer in a Retrieval-Augmented Generation system. The authors define a document subset's utility as the teacher-forced log-likelihood of the original response given only that subset, then compare exact Shapley against cheaper approximations (Kernel SHAP, ContextCite, TMC-Shapley, Beta-Shapley, and leave-one-out) across three instruction-tuned LLMs and two question-answering datasets. They find that surrogate-model approximations, especially Kernel SHAP with ContextCite close behind, reproduce exact Shapley rankings accurately while using far fewer LLM calls. But on synthetic multi-hop questions, every tested method systematically over-credits the document containing the direct answer and under-credits the document needed to synthesize it, showing that current attribution is reliable for simple lookups and unreliable exactly where RAG reasoning is hardest.

What carries the argument

The load-bearing object is the Shapley value of a retrieved document under a coalitional utility function $v(S) = \sum_t \log P(\text{token}_t \mid \text{token}_{<t}, Q, S)$ (Equation 2), the teacher-forced log-likelihood of the originally generated response given the query and a subset $S$ of documents. Shapley's formula averages each document's marginal contribution $v(S \cup \{j\}) - v(S)$ over all coalitions, which makes exact Shapley the reference attribution. Because every utility evaluation is an LLM call, the practical machinery is surrogate modelling: Kernel SHAP and ContextCite sample coalitions, evaluate $v$, and fit a weighted linear model whose coefficients become the attribution scores, achieving near-exact Shapley replication with 32 to 100 evaluations instead of $2^{10}$ for a ten-document set.

What would settle it

Compute exact Shapley values on the same queries with a different utility, for example whether the subset's response entails the ground-truth answer according to an LLM judge, and check whether Kernel SHAP's rank correlation with exact Shapley survives; if the correlation collapses or the synergy document suddenly receives credit, the central claim is tied to the likelihood proxy rather than to document influence as such. A second check is to present the synergy scenario with the synthesis document placed first instead of the direct-answer document; if any method credits the synthesis document at least as highly as the direct-answer document in either ordering, the claimed uniform under-crediting would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that utility-based Shapley attribution is both feasible and informative for RAG document attribution, and that two surrogate-model approximations—Kernel SHAP and ContextCite—are the most accurate and sample-efficient ways to compute it. The discovery that matters most is a systematic failure mode: in synergy scenarios, where an answer requires synthesizing information from two documents, all evaluated methods disproportionately favor the document containing the direct answer and underestimate the contribution of the document required for synthesis. The authors confirm this effect by swapping document order and by averaging normalized scores across 20 synthetic query-document pairs and three LLMs, and they conclude that attribution scores alone are insufficient; interpreting interactions among highly attributed documents is essential for a complete understanding.

Load-bearing premise

The entire evaluation rests on treating $v(S)$, the teacher-forced log-likelihood of the original response given a document subset, as the faithful measure of how much that subset supports the answer; if this likelihood does not track true causal support, then every Shapley score and comparison in the paper measures likelihood fit rather than document influence.

Editorial extensions

If this is right

  • Kernel SHAP and ContextCite reproduce exact Shapley rankings with high correlation (Pearson above 0.95 for Kernel SHAP at 100 samples) and identify the single most important document with precision above 0.8, so cheap surrogate approximations can replace exhaustive Shapley in practical RAG explainability.
  • Sampling-based approximations improve monotonically with sample size for most methods, meaning users can trade LLM-call budget against attribution fidelity.
  • In redundancy scenarios, the first-position duplicate receives higher attribution, revealing a position bias that persists when the order of the duplicates is swapped.
  • In synergy scenarios, all methods systematically assign higher scores to the direct-answer document and under-credit the synthesis document, so attribution scores alone can mislead about which documents are necessary for the answer.
  • Exact Shapley achieves the highest precision in exhaustive top-k removal tests, confirming that it remains the reference standard while approximations approach it closely.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A practical extension the paper leaves implicit: averaging attribution scores over random shuffles of document order could test whether the observed redundancy position bias is purely positional, and if so, order-robust attribution could be achieved by ensembling over permutations.
  • Because the synergy failure is uniform across methods and models, retrieval pipelines that use these attribution scores as training signal risk learning to demote the very documents needed for multi-hop answers; a testable fix would add pairwise interaction terms to the surrogate model and check whether multi-hop recall improves.
  • The teacher-forced log-likelihood proxy could be swapped for an answer-correctness utility, such as entailment of the ground truth according to an LLM judge; if Shapley rankings change materially, the paper's conclusions are specific to the likelihood utility rather than to document influence in general.
  • The paper evaluates only ten-document sets; scaling to larger retrieved sets would require learned utility predictors, and it remains open whether the surrogate-model advantage persists when exhaustive Shapley is no longer computable as a reference.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper adapts Shapley-based attribution to document-level retrieval-augmented generation (RAG). It defines the utility of a document subset as the teacher-forced log-likelihood of the original full-context response given the query and that subset, v(S) = log P(Rtarget | Q, S) (Eq. 2). Using this utility, the authors compare exact Shapley values with approximations—TMC-Shapley, Beta Shapley, Kernel SHAP, and ContextCite—and with leave-one-out, across three instruction-tuned LLMs and two standard datasets (BIOASQ, NQ). Three experiments address: (RQ1) how well approximations replicate exact Shapley; (RQ2) how well methods identify the k documents whose removal causes the largest utility drop; and (RQ3) how methods behave under redundancy, complementarity, and synergy among retrieved documents. The central claimed findings are that Kernel SHAP and ContextCite are the most accurate approximations of exact Shapley, and that in synergy scenarios all utility-based methods systematically undervalue the document needed for synthesis while favoring the document containing the direct answer.

Significance. If the claims hold, the paper provides useful practical guidance for practitioners who need attribution in RAG under limited LLM call budgets, and it identifies a limitation of current utility-based attribution under multi-hop reasoning. Strengths of the manuscript include: the internal benchmark of approximation methods against exact Shapley computed with the same utility function; the inclusion of three models, two standard datasets, and multiple approximation budgets; the construction of synthetic scenarios for redundancy, complementarity, and synergy; and a public codebase. The replication-quality comparison (Experiment 1) is internally coherent, and the paper correctly notes that Shapley axioms are violated by the non-additive likelihood utility. However, the significance hinges on whether the chosen likelihood utility is a valid proxy for document influence. That assumption is not validated, and the synergy conclusion in Experiment 3 is not supported by the evidence presented. The paper would be strengthened by validating Eq. (2) against human judgments or counterfactual answer quality and by reporting uncertainty estimates for the headline comparisons.

major comments (3)
  1. [§3.4] The evaluation is entirely internal to the unvalidated utility v(S) = log P(Rtarget | Q, S). Experiment 1 uses exact Shapley under this v as the reference, and Experiment 2 uses removal-based utility drops under the same v as ground truth (Eq. 3). The paper never shows that this likelihood is a faithful proxy for the causal support of a document set for the generated answer; it is an assumption inherited from ContextCite and stated without independent evidence. A concrete test is needed: for a sample of queries, compare the utility-based ranking against human judgments or against the factual correctness of answers produced from document subsets. Without such validation, the replication-quality results and the removal-effectiveness results measure fidelity to a proxy whose real-world meaning is undemonstrated.
  2. [§4.3] The synergy claim that all methods 'disproportionately favor the document containing the direct answer' rests on the premise, stated in Section 3.4, that the two positive documents 'individually offer no utility towards Rtarget.' This premise is never verified against the actual utility in Eq. (2). In the displayed synergy example (Figure 4), Positive B already contains the entity name 'Lyra Vael' and the requested salary, so v({B}) is very likely to be high for the response tokens that mention 'Lyra Vael commands a salary of approximately 50 million Credits'; the response's initial phrase 'most popular actor' also appears in the query. If v({B}) is indeed high, then assigning B a higher score is the correct behavior of the utility, not evidence of systematic underestimation. To substantiate the synergy conclusion, the authors should report v({A}), v({B}), and v({A,B}) for the shown example and, ideally, across the 20 synthetic pairs. Figure 5 only shows average normalized attribution scores, which cannot establish that the utility's ground truth favors B.
  3. [§4.1] All Experiment 1 and Experiment 2 results are reported as point averages without standard deviations, confidence intervals, or statistical significance tests. Statements such as 'KernelSHAP consistently emerges as the top-performing attribution method' (Section 4.1) and the cross-method comparisons in Table 1 are therefore not substantiated beyond qualitative observation. The observed differences might be within noise, especially for Precision@k where several methods differ by only 0.01–0.02. The authors should report variance across the 100 queries (and, where relevant, across models) or run paired significance tests. This is load-bearing because the main RQ1/RQ2 claims are comparative rankings.
minor comments (4)
  1. [§4.1] The sentence 'TMC-Shapley and Beta Shapley require more ground truth utilities (n × T where T = |D|!)' appears to be a technical error: TMC-Shapley samples permutations, and the number of iterations T is a user-chosen budget, not the factorial of the number of documents. Please correct this description.
  2. [§3.4] The construction of the synthetic datasets is described only briefly: 20 pairs per scenario were built using ChatGPT, Gemini, and DeepSeek and manually inspected. Please provide more detail on the generation procedure, the manual inspection criteria, and whether the final 20 pairs are the ones used in Figures 2–5; this matters for reproducibility.
  3. [Fig. 1] The subplot captions do not clearly distinguish the datasets or metrics for each panel. In particular, the rightmost column is labeled 'Precision@k' with x-axis 'Top k Values' but the caption does not state which dataset each row belongs to (BIOASQ vs NQ), or which panel corresponds to which metric. Please label the panels explicitly or expand the caption.
  4. [Appendix A] The LLM-as-a-judge evaluation relies on Gemini-2.0-Flash with a custom prompt but reports no judge reliability checks (e.g., agreement with human labels on a sample). Since the cosine similarity metric is shown to be misleading, the judge's consistency should be assessed, especially for the synthetic scenarios where judge scores are near 1.0.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper benchmarks Shapley approximations against exact Shapley and removal-based utility drops; no fitted parameter is renamed as a prediction and no self-citation chain supports the central claims.

full rationale

The paper's load-bearing comparisons are replication benchmarks rather than circular derivations. Experiment 1 compares KernelSHAP, ContextCite, TMC-Shapley, and Beta-Shapley against exact Shapley values computed from the same utility v(S)=log P(Rtarget|Q,S) (Eq. 2); that is the definitionally correct reference target for Shapley approximations. Experiment 2 compares attribution rankings against the exhaustive top-k utility-loss set S*_k defined in Eq. 3, which is an independent criterion derived from v and not from any method's output. The utility definition is explicitly inherited from ContextCite, an external prior work by different authors, so no load-bearing self-citation is present. No uniqueness theorem from the authors' own prior work is invoked, and no result is renamed from known work. The synergy discussion in Section 4.3 relies on the qualitative premise that both synthetic documents are individually useless for the answer; that premise is not derived from Eq. 2 and may even conflict with the displayed Aethelon example, where document B already contains the direct answer. This is a validity or ground-truth-construction concern, not a case of the paper deriving its conclusion from the definition of that conclusion. Therefore the circularity score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The ledger is modest because this is an empirical evaluation rather than a derivation. The key assumptions are the utility definition (Eq. 2) and the use of exact Shapley as a reference. No invented entities are introduced, and the only hand-chosen numbers are the sampling budgets (32, 64, 100), which are swept rather than fitted.

free parameters (1)
  • Sampling budget for approximation methods = 32, 64, 100
    The number of utility evaluations used by TMC-Shapley, Beta-Shapley, Kernel SHAP, and ContextCite is swept across three values. It is a hand-chosen experimental hyperparameter, not fitted to data, but the monotonic improvement conclusion depends on comparing these discrete budgets.
assumptions (3)
  • domain assumption Teacher-forced log-likelihood log P(Rtarget | Q, S) measures the contribution of document subset S to the generated response.
    Defined in Eq. (2) and used for all utility values in all three experiments. If this is not a faithful proxy for causal document influence, the entire attribution comparison is off-target.
  • domain assumption Exact Shapley values computed under this utility are the appropriate gold standard for document attribution.
    Experiment 1 measures approximation quality by correlation with exact Shapley. The Shapley axioms are acknowledged to be violated in this nonlinear setting, so Shapley is used as a pragmatic reference rather than a proven ground truth.
  • domain assumption The generated response Rtarget is a fixed target, and teacher-forcing evaluates the model's likelihood of that same response under different document subsets.
    All utilities condition on the same Rtarget. This assumes the likelihood of the original response is a stable, comparable quantity across subsets, which may not hold if the model would have produced a different answer under a different subset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Source Attribution in Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/HTFXWN6S

@misc{pith2026250704480,
  author       = {Pith},
  title        = {Pith review of: Source Attribution in Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTFXWN6S}},
  note         = {Machine review of arXiv:2507.04480}
}
read the original abstract

While attribution methods, such as Shapley values, are widely used to explain the importance of features or training data in traditional machine learning, their application to Large Language Models (LLMs), particularly within Retrieval-Augmented Generation (RAG) systems, is nascent and challenging. The primary obstacle is the substantial computational cost, where each utility function evaluation involves an expensive LLM call, resulting in direct monetary and time expenses. This paper investigates the feasibility and effectiveness of adapting Shapley-based attribution to identify influential retrieved documents in RAG. We compare Shapley with more computationally tractable approximations and some existing attribution methods for LLM. Our work aims to: (1) systematically apply established attribution principles to the RAG document-level setting; (2) quantify how well SHAP approximations can mirror exact attributions while minimizing costly LLM interactions; and (3) evaluate their practical explainability in identifying critical documents, especially under complex inter-document relationships such as redundancy, complementarity, and synergy. This study seeks to bridge the gap between powerful attribution techniques and the practical constraints of LLM-based RAG systems, offering insights into achieving reliable and affordable RAG explainability.

Figures

Figures reproduced from arXiv: 2507.04480 by the authors.

Figure 1
Figure 1. Experiment 1 with Mistral: Performance on BIOASQ and NQ. Procedure For each synthetic dataset scenario: 1. Generate Rtarget = LLM(Q, D). 2. Apply all attribution methods. 3. Analyze qualitatively the behavior of methods in inter-relation consideration for attribution computing. 4 Results and Analysis 4.1 Experiment 1: Shapley Replication Quality The performance of the attribution methods was evaluated on the BIOASQ … view at source ↗
Figure 2
Figure 2. (Top) Sample question belonging to the redundancy scenario (Bottom) [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. (Top) Sample question belonging to the complementary scenario (Bottom) [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (Top) Sample question belonging to the synergy scenario (Bottom) At [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Average normalized attribution scores for synergetic documents A and B, [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Correlation to Shapley values with Llama-8B using BIOASQ and NQ [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Correlation to Shapley values with Qwen-3B using BIOASQ and NQ [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 19 canonical work pages

  1. [1]

    In: In- ternational Conference on Learning Representations (2021),https://openreview

    Basu, S., Pope, P., Feizi, S.: Influence functions in deep learning are fragile. In: In- ternational Conference on Learning Representations (2021),https://openreview. net/forum?id=xHKVVHGDOEk

  2. [2]

    In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C

    Cohen-Wang, B., Shah, H., Georgiev, K., Mądry, A.: Contextcite: Attribut- ing model generation to context. In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (eds.) Advances in Neural In- formation Processing Systems. vol. 37, pp. 95764–95807. Curran Associates, Source Attribution in Retrieval-Augmented Generation 15 ...

  3. [3]

    In: Chaudhuri, K., Salakhutdinov, R

    Ghorbani, A., Zou, J.: Data shapley: Equitable valuation of data for machine learn- ing. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th Inter- national Conference on Machine Learning. Proceedings of Machine Learning Re- search, vol. 97, pp. 2242–2251. PMLR (09–15 Jun 2019),https://proceedings. mlr.press/v97/ghorbani19c.html

  4. [4]

    Hu, L., Liu, Y., Liu, N., Huai, M., Sun, L., Wang, D.: Seat: stable and explainable attention. In: Proceedings of the Thirty-Seventh AAAI Confer- ence on Artificial Intelligence and Thirty-Fifth Conference on Innovative Ap- plications of Artificial Intelligence and Thirteenth Symposium on Educa- tional Advances in Artificial Intelligence. AAAI’23/IAAI’23/...

  5. [5]

    In: Proceedings of the 5th ACM International Conference on Multimedia in Asia

    Huang, Y., Jia, A., Zhang, X., Zhang, J.: Generic attention-model explainability by weighted relevance accumulation. In: Proceedings of the 5th ACM International Conference on Multimedia in Asia. MMAsia ’23, Association for Computing Ma- chinery, New York, NY, USA (2024). https://doi.org/10.1145/3595916.3626437, https://doi.org/10.1145/3595916.3626437

  6. [6]

    In: Proceedings of the 39th International Conference on Machine Learning (ICML)

    Ilyas, A., Engstrom, L., Santurkar, S., Tsipras, D., Madry, A.: Data models: Understanding data through models. In: Proceedings of the 39th International Conference on Machine Learning (ICML). pp. 9525–9589. PMLR (2022),https: //proceedings.mlr.press/v162/ilyas22a.html

  7. [7]

    In: Burstein, J., Doran, C., Solorio, T

    Jain, S., Wallace, B.C.: Attention is not Explanation. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long and Short Papers). pp. 3543–3556. As- sociation for Computational Linguistics, Minneapolis,...

  8. [8]

    In: Precup, D., Teh, Y.W

    Koh, P.W., Liang, P.: Understanding black-box predictions via influence functions. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference onMachineLearning.ProceedingsofMachineLearningResearch,vol.70,pp.1885–

Show all 27 references
  1. [9]

    Transactions of the Association for Compu- tational Linguistics 7, 453–466 (2019)

    Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al.: Natural questions: a bench- mark for question answering research. Transactions of the Association for Compu- tational Linguistics 7, 453...

  2. [10]

    In: Camps-Valls, G., Ruiz, F.J.R., Valera, I

    Kwon, Y., Zou, J.: Beta shapley: a unified and noise-reduced data valuation frame- work for machine learning. In: Camps-Valls, G., Ruiz, F.J.R., Valera, I. (eds.) AIS- TATS. Proceedings of Machine Learning Research, vol. 151, pp. 8780–8802. PMLR (2022)

  3. [11]

    In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017),h...

  4. [12]

    In: Bollegala, D., Huang, R., Ritter, A

    Sarti, G., Feldhus, N., Sickert, L., van der Wal, O.: Inseq: An interpretabil- ity toolkit for sequence generation models. In: Bollegala, D., Huang, R., Ritter, A. (eds.) Proceedings of the 61st Annual Meeting of the Associa- 16 Nematov et al. tion for Computational Linguistic...

  5. [13]

    Shapley, L.S., et al.: A value for n-person games (1953)

  6. [14]

    arXiv preprint arXiv:1312.6034 (2013)

    Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional net- works: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)

  7. [15]

    arXiv preprint arXiv:1706.03825 (2017)

    Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: re- moving noise by adding noise. arXiv preprint arXiv:1706.03825 (2017)

  8. [16]

    In: Precup, D., Teh, Y.W

    Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Confer- ence on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 3319–3328. PMLR (06–11 Aug 2017),https:...

  9. [17]

    BMC bioinformatics16, 1–28 (2015)

    Tsatsaronis, G., Balikas, G., Malakasiotis, P., Partalas, I., Zschunke, M., Alvers, M.R., Weissenborn, D., Krithara, A., Petridis, S., Polychronopoulos, D., et al.: An overview of the bioasq large-scale biomedical semantic indexing and question answering competition. BMC bioin...

  10. [18]

    In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processi...

  11. [19]

    In: Inui, K., Jiang, J., Ng, V., Wan, X

    Wiegreffe, S., Pinter, Y.: Attention is not not explanation. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empiri- cal Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNL...

  12. [20]

    Yin, K., Neubig, G.: Interpreting language models with contrastive explanations Source Attribution in Retrieval-Augmented Generation 17 A Model Evaluation We employ two evaluation metrics to evaluate the quality of responses by LLMs:

  13. [22]

    For the BioASQ dataset, we useMedEmbed-large-v0.1; for all other datasets, we useall-MiniLM-L6-v2, both from the Sentence Trans- formers library

    Cosine similaritybetween the embeddings of the generated and ground- truth answers. For the BioASQ dataset, we useMedEmbed-large-v0.1; for all other datasets, we useall-MiniLM-L6-v2, both from the Sentence Trans- formers library

  14. [23]

    The judge returns a binary output, True or False

    LLM-as-a-Judge - Given the question, generated and ground-truth an- swers, we use a fine-tuned LLM to decide whether the generated response entails the ground truth. The judge returns a binary output, True or False. Evaluation of the models for all datasets is demonstrated in ...

  15. [24]

    Focus on semantic equivalence, not exact wording

  16. [25]

    Minor grammatical differences don’t affect equivalence

  17. [26]

    The generated answer must capture all key information from the ground truth

  18. [27]

    evaluation

    Additional relevant information in the generated answer is acceptable Output Format (strictly follow this JSON format): {"evaluation": "yes"/"no", "explanation": "short explanation about the provided evaluation"} Examples: Query: "What is photosynthesis?" Ground Truth: "Photos...

  19. [1894]

    PMLR (06–11 Aug 2017), https://proceedings.mlr.press/v70/koh17a. html

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.