Pith. sign in

REVIEW 4 major objections 6 minor 44 references

How Reliable are LLMs for Reasoning on the Re-ranking task?

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Training method decides whether LLM reranking explanations can be trusted.

desk verdict A modest new dataset and a clear DPO-over-SFT result sit under a headline explainability claim the paper never actually tests. read the letter →

arxiv 2508.18444 v2 pith:VUXRJDAD submitted 2025-08-25 cs.CL cs.AI

Nafis Tanveer Islam , Zhiming Zhao This is my paper
classification cs.CLcs.AI
keywords LLMre-rankingExplainabilitySHAPattributionDirectPreferenceOptimizationRetrievalTrainingmethodsEarthsciencesearchTransparency
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

The paper asks whether LLMs can do transparent re-ranking when ranking data is scarce. It trains small open-weight LLMs on expert-annotated relevance scores from an Earth-science knowledge base, comparing supervised fine-tuning, reward modeling, PPO, and DPO. It reports that DPO gives the best ranking accuracy, that larger 7B models beat smaller ones, and that all LLMs beat the BM25 baseline. It also reports that explanations generated from SHAP token-attribution scores are more coherent than explanations generated without those scores. The paper's point is to show a low-cost pipeline—DPO for reranking, SHAP for token-level attribution, and a general-purpose LLM for prose—that gives users a concrete reason they can check.

What carries the argument

The central mechanism is a three-stage pipeline. First, a training objective (SFT, reward modeling, PPO, or DPO) adapts an LLM whose last layer is collapsed to a single neuron outputting a ranking score. Second, SHAP assigns an attribution score to each token of the selected response, and the top ten scores are kept. Third, a general-purpose LLM turns those scores, the query, and the response text into a natural-language explanation. The SHAP scores are the load-bearing link: they determine which tokens the final explanation can cite.

What would settle it

Re-annotate the same queries with independent domain experts and compute inter-annotator agreement, or build a test set where the true relevant item is deliberately placed outside the BM25 top-10. If expert labels disagree substantially, or if DPO-trained rerankers never surface the outside item, the paper's ranking and reliability conclusions would not survive.

Watch

Extended reading notes

Core claim

On the paper's terms: different training methods do not just change ranking accuracy; they change whether a reranker has learned genuine semantic relevance or only task-optimizing abstract patterns. The evidence is a comparison of training objectives on a small dataset where DPO consistently yields the lowest pairwise loss and highest top-1 accuracy. SHAP attribution scores for DPO-trained models concentrate on semantically meaningful tokens—for example, 'concentration' for an aerosol dataset query. Feeding those scores to a general LLM produces explanations that cite the contributing tokens; withholding the scores produces generic restatements of the document. The paper treats that contrast

Load-bearing premise

The expert scores from 94 hackathon participants, given on a 0–9 scale for BM25 top-10 results, are treated as reliable ground truth even though agreement between raters is never checked; every ranking and explanation conclusion inherits that assumption.

Editorial extensions

If this is right

  • If DPO-trained rerankers beat PPO, SFT, and reward modeling on small data, preference-optimization methods become the default choice for data-limited reranking.
  • If LLMs beat BM25 on NDCG@5/10 and similarity metrics in this domain, retrieval pipelines can replace or augment lexical first-stage ranking with semantic reranking.
  • If explanations built on SHAP scores are more coherent than attribution-free explanations, token-level interpretability tools can feed LLM-generated justifications in ranking systems.
  • The single-neuron last-layer trick suggests a practical way to turn a generative LLM into a scorer with minimal architectural change.

Reading between the lines

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

  • Because the dataset only contains BM25 top-10 candidates, the reported gains may not transfer to settings where relevant items fall outside the initial candidate list; testing with a deeper or different first-stage retriever would show how much of the improvement is reranking versus recovery.
  • The claim that some training methods learn abstract knowledge rather than semantic understanding is inferred from loss, token patterns, and explanation quality, not tested directly; adversarial or deliberately misleading queries could probe whether DPO-trained models cite genuinely relevant tokens or merely salient ones.
  • The pipeline should be portable: replacing the explanation-generating LLM with a smaller open model would test whether the SHAP signal alone drives explanation quality, or whether the generator's strength matters.
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

4 major / 6 minor

Summary. The paper proposes and evaluates an LLM-based re-ranking pipeline for a small environmental-science dataset collected during the ENVRI-Hub-Next hackathon. It compares SFT, PPO, DPO, and reward modeling on LLaMa, Mistral, and Phi-3, then adopts DPO to train five LLMs and compares them to BM25. It adds a post-hoc explanation module: SHAP attribution scores over the top-10 tokens of the selected response are fed to GPT-4o to generate natural-language justifications. The authors claim that DPO is the best training method, that larger LLMs improve re-ranking, and that explanations generated with SHAP scores are 'markedly more coherent and relevant' than those without.

Significance. If fully supported, the paper would provide a practical, low-cost recipe for transparent LLM re-ranking in data-scarce scientific domains, and it would be one of the first to connect training-method choice with explainability. The dataset, based on 94 hackathon participants' judgments, is a useful resource. However, the current evidence does not support the central explainability claim, and the ranking comparisons lack statistical grounding. The paper's strengths are its concrete setup, the comparison across four training paradigms, and the promise of released code; these do not compensate for the missing evaluation.

major comments (4)
  1. [§5.4, Table 1] RQ1 concludes DPO is the best training method from loss and top-1 accuracy on a single 80:10:10 split. No confidence intervals, error bars, significance tests, or held-out NDCG are reported for the four training methods. Since DPO is then used for all subsequent experiments (RQ2) and the conclusion, the selection is load-bearing. At minimum, report variance over multiple seeds/splits and NDCG on the test set for each training method.
  2. [§5.4, Tables 3–4, §6] The central claim (RQ3, abstract, conclusion) that SHAP-attribution-based explanations are 'markedly more coherent and relevant' is not tested. Only two (or three) hand-picked examples are shown, with no blinding, no human annotators, no inter-annotator agreement, and no automated explanation-quality metric. Section 6 explicitly defers 'human evaluations and automated explainability metrics' to future work. Moreover, the 'with As' vs 'without As' prompts differ not only in attribution values but in the instruction text (Figure 3 vs. the implied prompt), so any perceived improvement is confounded. This is not a minor gap: the paper's headline claim rests on it.
  3. [§5.2, §5.4] All ranking and explanation conclusions depend on expert-generated labels from 94 hackathon participants on a 0–9 scale over BM25 top-10 candidates. No inter-annotator agreement is reported, and relevant items outside the BM25 top-10 are never recovered, biasing NDCG and similarity metrics. The single 80:10:10 split further means that no uncertainty estimate is available. These issues affect every quantitative result in Tables 1–2.
  4. [§5.3, §5.4] The metric definitions are inconsistent with their use. BERTScore, BLEU, and Rouge-L are defined as text-overlap/similarity metrics (Section 5.3), but Section 5.4 says they 'compare the ground-truth ranking of items with the rankings produced by different models.' Table 2 reports values for BM25, a retriever that outputs a ranked list, not generated text. It is unclear what reference text is used and how these metrics are computed for a ranked list; without this, Table 2's numbers are uninterpretable.
minor comments (6)
  1. [§3.1] Notation errors: the set U contains duplicate 'u1', and Q is written as 'q1,2, q3' with a typo; 'between then users' should read 'between the users'.
  2. [§5.4] The text refers to 'Model 1' and 'Model 3 (our 7-billion-parameter configuration)', but Table 1 lists model families (LLaMa, Mistral, Phi-3) without parameter counts; these references are undefined.
  3. [§5.4] The sentence 'Mistral-7B ... underperforms by approximately 0.03 higher in NDCG' is self-contradictory; 'higher' should likely be 'lower'.
  4. [§4.2, §5.1] The trained model is described as outputting a single ranking score from a single neuron, yet Section 5.1 specifies a beam size and temperature for 'sequence generation tasks.' Clarify whether the model is used as a scorer or a generator during inference.
  5. [§5.1, §5.4] The choice of 'top 10 tokens with the highest attribution scores' is stated to be based on experimental analysis, but no sensitivity analysis or evidence is provided; report how this cutoff was selected.
  6. [Figure 3, Tables 3–4] The 'without As' prompt is not shown; include both prompt templates in an appendix so that the comparison is reproducible.

Circularity Check

1 steps flagged · score 2.0 of 10

Ranking results are externally validated; the SHAP-explanation claim is mildly self-confirming because the prompt itself supplies the top tokens and instructs the LLM to use them.

  1. self definitional [Section 4.2 (LLM Explanation Generation Prompt, Figure 3); Section 5.4 Answering RQ3; Tables 3-4]
    "Using the attribution values of the top tokens, generate a concise, clear, and coherent explanation describing why this response was selected to answer the query. Focus on how the most important tokens contribute to the relevance and ranking of this response."

    The 'with As' explanations are generated from a prompt that explicitly supplies the top SHAP tokens and instructs the model to focus on them. The paper's RQ3 evidence is that with-As explanations 'explicitly reference the most influential tokens that contributed to the ranking decision' (Section 5.4). That observation is constructed by the prompt, not an independent demonstration that SHAP improves reasoning. The qualitative claim of greater coherence is also confounded because the with-As prompt itself asks for a 'concise, clear, and coherent explanation,' while the without-As prompt is not shown. Thus the central explainability result reduces to prompt-following rather than a validated property of attribution scores.

full rationale

The ranking evaluations are self-contained and externally grounded: expert-generated rankings from the ENVRI-Hub-Next hackathon are used as training and test labels, and models are evaluated on held-out data with NDCG, BERTScore, BLEU, ROUGE-L, and cosine similarity. No fitted parameter is renamed as a prediction, and no load-bearing self-citation or imported uniqueness theorem was found. The main circularity concern is limited to the RQ3 explainability comparison: the prompt for the 'with As' condition supplies the top SHAP tokens and instructs the LLM to base its explanation on them, so the observation that those explanations mention the tokens is partly self-fulfilling. The paper's own conclusion acknowledges that human evaluation and automated explainability metrics remain future work, which further indicates this claim is not yet rigorously validated. Overall, the core ranking contribution is independent, and the circularity is mild and localized.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities, forces, or conserved quantities. The proposed 'single neuron' ranking head is an architectural modification of an existing LLM, not a new entity. The load-bearing assumptions are about data quality, the BM25 candidate set, and the validity of SHAP as an explanation mechanism.

free parameters (2)
  • top-10 attribution token cutoff = 10 top tokens
    Chosen post hoc; Section 5.1 says 'we found' the top 10 tokens have the highest impact, but no quantitative comparison is shown.
  • training hyperparameters = lr=2e-5, batch=4, epochs=8, beam=4, temperature=0.5
    Chosen by hand for all experiments; the ranking conclusions are not shown to be robust to these choices.
assumptions (4)
  • domain assumption ENVRI-Hub expert 0-9 rankings are a valid gold standard for query-item relevance
    Used as supervision and evaluation in Sections 5.2-5.4; no inter-annotator agreement or label validation is provided.
  • domain assumption BM25 top-10 candidate lists contain all relevant documents
    The reranking pipeline only reorders these 10 items; relevant items outside the initial list cannot be recovered, see Section 5.2.
  • standard math SHAP token attributions on the LLM output are meaningful explanations of ranking decisions
    SHAP is applied to a single-neuron ranking head on summary text (Section 4.2), but no validation links SHAP values to actual ranking behavior.
  • ad hoc to paper Top-10 tokens with highest SHAP values carry the causal explanation
    Section 5.1 claims 'we found' without showing the comparison; the entire reasoning module relies on this cutoff.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Reliable are LLMs for Reasoning on the Re-ranking task?." pith.science (2026). https://pith.science/paper/VUXRJDAD

@misc{pith2026250818444,
  author       = {Pith},
  title        = {Pith review of: How Reliable are LLMs for Reasoning on the Re-ranking task?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VUXRJDAD}},
  note         = {Machine review of arXiv:2508.18444}
}
read the original abstract

With the improving semantic understanding capability of Large Language Models (LLMs), they exhibit a greater awareness and alignment with human values, but this comes at the cost of transparency. Although promising results are achieved via experimental analysis, an in-depth understanding of the LLM's internal workings is unavoidable to comprehend the reasoning behind the re-ranking, which provides end users with an explanation that enables them to make an informed decision. Moreover, in newly developed systems with limited user engagement and insufficient ranking data, accurately re-ranking content remains a significant challenge. While various training methods affect the training of LLMs and generate inference, our analysis has found that some training methods exhibit better explainability than others, implying that an accurate semantic understanding has not been learned through all training methods; instead, abstract knowledge has been gained to optimize evaluation, which raises questions about the true reliability of LLMs. Therefore, in this work, we analyze how different training methods affect the semantic understanding of the re-ranking task in LLMs and investigate whether these models can generate more informed textual reasoning to overcome the challenges of transparency or LLMs and limited training data. To analyze the LLMs for re-ranking tasks, we utilize a relatively small ranking dataset from the environment and the Earth science domain to re-rank retrieved content. Furthermore, we also analyze the explainable information to see if the re-ranking can be reasoned using explainability.

Figures

Figures reproduced from arXiv: 2508.18444 by the authors.

Figure 1
Figure 1. Contribution of our proposed work In summary, our contributions are as follows: – We analyze and compare a set of re-ranking techniques using large language models. – We propose a SHAP explanation-based method powered by LLM to generate reasoning behind each recommendation. – We also provide qualitative studies to analyze and support the textual rea￾soning of the ranked items. 1 https://envri.eu/envri-hubnext/ [PIT… view at source ↗
Figure 2
Figure 2. Overall Architecture of our Proposed System [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Prompt template used for generating LLM explanations [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Top attribution scores of the selected item based on a query [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 28 canonical work pages

  1. [1]

    In: SC22: International Conference for High Performance Computing, Networking, Storage and Analysis

    Aminabadi, R.Y., Rajbhandari, S., Awan, A.A., Li, C., Li, D., Zheng, E., Ruwase, O., Smith, S., Zhang, M., Rasley, J., et al.: Deepspeed-inference: enabling efficient inference of transformer models at unprecedented scale. In: SC22: International Conference for High Performance Computing, Networking, Storage and Analysis. pp. 1–15. IEEE (2022)

    work page 2022

  2. [2]

    In: Learn- ing

    Burges, C.: From ranknet to lambdarank to lambdamart: An overview. In: Learn- ing. vol. 11, p. 81 (2010)

    work page 2010

  3. [3]

    In: Proceedings of the 22nd International Conference on Machine Learning (ICML)

    Burges, C.J., Shaked, T., Renshaw, E., Lazier, M., Deeds, A., Hamilton, N., Hullender, G.: Learning to rank using gradient descent. In: Proceedings of the 22nd International Conference on Machine Learning (ICML). pp. 89–96 (2005). https://doi.org/10.1145/1102351.1102363

    arXiv 2005

  4. [4]

    ACM Transactions on Recommender Systems (2024)

    Carraro, D., Bridge, D.: Enhancing recommendation diversity by re-ranking with large language models. ACM Transactions on Recommender Systems (2024)

    work page 2024

  5. [5]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2017)

    Christiano, P.F., Leike, J., Brown, T.B., Martic, M., Legg, S., Amodei, D.: Deep re- inforcement learning from human preferences. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)

    work page 2017

  6. [6]

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

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. 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 Paper...

    doi:10.18653/v1/n19-1423 2019

  7. [7]

    https://www.elastic.co/elasticsearch/ (2025), version 8.x

    Elastic: Elasticsearch: Distributed, restful search and analytics engine. https://www.elastic.co/elasticsearch/ (2025), version 8.x

    work page 2025

  8. [8]

    https://eur- lex.europa.eu/eli/reg/2016/679/oj (2016), regulation (EU) 2016/679 of the Eu- ropean Parliament and of the Council

    European Union: General Data Protection Regulation (GDPR). https://eur- lex.europa.eu/eli/reg/2016/679/oj (2016), regulation (EU) 2016/679 of the Eu- ropean Parliament and of the Council

    work page 2016

Show all 44 references
  1. [9]

    Open Research Europe1, 68 (2023)

    Farshidi, S., Liao, X., Li, N., Goldfarb, D., Magagna, B., Stocker, M., Jeffery, K., Thijsse, P., Pichot, C., Petzold, A., et al.: Knowledge sharing and discovery across heterogeneous research infrastructures. Open Research Europe1, 68 (2023)

    2023

  2. [10]

    arXiv preprint arXiv:2406.12433 (2024)

    Gao, J., Chen, B., Zhao, X., Liu, W., Li, X., Wang, Y., Zhang, Z., Wang, W., Ye, Y., Lin, S., et al.: Llm-enhanced reranking in recommender systems. arXiv preprint arXiv:2406.12433 (2024)

    2024 arXiv

  3. [11]

    ACM Transactions on Information Systems42(3), 1–23 (2024) 16 Islam et al

    Gao, J., Zhao, X., Li, M., Zhao, M., Wu, R., Guo, R., Liu, Y., Yin, D.: Smlp4rec: an efficient all-mlp architecture for sequential recommendations. ACM Transactions on Information Systems42(3), 1–23 (2024) 16 Islam et al

    2024

  4. [12]

    In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management

    Guo, J., Fan, Y., Ai, Q., Croft, W.B.: A deep relevance matching model for ad- hoc retrieval. In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. pp. 55–64 (2016)

    2016

  5. [13]

    In: Proceedings of the 24th ACM international on conference on information and knowledge management

    He, X., Chen, T., Kan, M.Y., Chen, X.: Trirank: Review-aware explainable recom- mendation by modeling aspects. In: Proceedings of the 24th ACM international on conference on information and knowledge management. pp. 1661–1670 (2015)

    2015

  6. [14]

    In: Proceedings of the 2000 ACM conference on Computer supported cooperative work

    Herlocker, J.L., Konstan, J.A., Riedl, J.: Explaining collaborative filtering recom- mendations. In: Proceedings of the 2000 ACM conference on Computer supported cooperative work. pp. 241–250 (2000)

    2000

  7. [15]

    In: Proceedings of the 22nd ACM International Conference on Information & Knowledge Manage- ment

    Huang, P.S., He, X., Gao, J., Deng, L., Acero, A., Heck, L.: Learning deep struc- tured semantic models for web search using clickthrough data. In: Proceedings of the 22nd ACM International Conference on Information & Knowledge Manage- ment. pp. 2333–2338 (2013)

    2013

  8. [16]

    In: Proceedings of the 44th International ACM SIGIR conference on Research and Development in Information Retrieval

    Li, L., Zhang, Y., Chen, L.: Extra: Explanation ranking datasets for explainable recommendation. In: Proceedings of the 44th International ACM SIGIR conference on Research and Development in Information Retrieval. pp. 2463–2469 (2021)

    2021

  9. [17]

    arXiv preprint arXiv:2105.11601 (2021)

    Li, L., Zhang, Y., Chen, L.: Personalized transformer for explainable recommen- dation. arXiv preprint arXiv:2105.11601 (2021)

    2021 arXiv

  10. [18]

    In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management

    Li, X., Yan, F., Zhao, X., Wang, Y., Chen, B., Guo, H., Tang, R.: Hamur: Hy- per adapter for multi-domain recommendation. In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. pp. 1268– 1277 (2023)

    2023

  11. [19]

    arXiv preprint arXiv:2311.12399 (2023)

    Li, Y., Li, Z., Wang, P., Li, J., Sun, X., Cheng, H., Yu, J.X.: A survey of graph meets large language model: Progress and future directions. arXiv preprint arXiv:2311.12399 (2023)

    2023 arXiv

  12. [20]

    In: Proceedings of the ACM on Web Conference 2025

    Li, Y., Zhang, X., Luo, L., Chang, H., Ren, Y., King, I., Li, J.: G-refer: Graph retrieval-augmented large language model for explainable recommendation. In: Proceedings of the ACM on Web Conference 2025. pp. 240–251 (2025)

    2025

  13. [21]

    In: Text Summarization Branches Out

    Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74–81. Association for Computational Linguis- tics, Barcelona, Spain (Jul 2004), https://aclanthology.org/W04-1013/

    2004

  14. [22]

    ACM Trans

    Lin, J., Dai, X., Xi, Y., Liu, W., Chen, B., Zhang, H., Liu, Y., Wu, C., Li, X., Zhu, C., Guo, H., Yu, Y., Tang, R., Zhang, W.: How can recommender systems benefit from large language models: A survey. ACM Trans. Inf. Syst.43(2) (Jan 2025). https://doi.org/10.1145/3678004, htt...

    2025 doi

  15. [23]

    arXiv preprint arXiv:2504.07439 (2025)

    Liu, Q., Duan, H., Chen, Y., Lu, Q., Sun, W., Mao, J.: Llm4ranking: An easy-to- use framework of utilizing large language models for document reranking. arXiv preprint arXiv:2504.07439 (2025)

    2025 arXiv

  16. [24]

    In: Proceedings of the 29th ACM Interna- tional Conference on Information & Knowledge Management

    Liu, W., Liu, Q., Tang, R., Chen, J., He, X., Heng, P.A.: Personalized re-ranking with item relationships for e-commerce. In: Proceedings of the 29th ACM Interna- tional Conference on Information & Knowledge Management. pp. 925–934 (2020)

    2020

  17. [25]

    In: Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS)

    Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS). pp. 4765–4774. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/fi...

    2017

  18. [26]

    arXiv preprint arXiv:2406.02377 (2024)

    Ma, Q., Ren, X., Huang, C.: Xrec: Large language models for explainable recom- mendation. arXiv preprint arXiv:2406.02377 (2024)

    2024 arXiv

  19. [27]

    In: arXiv preprint arXiv:1901.04085 (2019) How Reliable are LLMs for Reasoning on the Re-ranking task? 17

    Nogueira, R., Cho, K.: Passage re-ranking with bert. In: arXiv preprint arXiv:1901.04085 (2019) How Reliable are LLMs for Reasoning on the Re-ranking task? 17

    1901 arXiv

  20. [28]

    In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics

    Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311–318 (2002)

    2002

  21. [29]

    arXiv preprint arXiv:2412.13337 (2024)

    Pareja, A., Nayak, N.S., Wang, H., Killamsetty, K., Sudalairaj, S., Zhao, W., Han, S., Bhandwaldar, A., Xu, G., Xu, K., et al.: Unveiling the secret recipe: A guide for supervised fine-tuning small llms. arXiv preprint arXiv:2412.13337 (2024)

    2024 arXiv

  22. [30]

    In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

    Peake, G., Wang, J.: Explanation mining: Post hoc interpretability of latent factor models for recommendation systems. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 2060–2069 (2018)

  23. [31]

    In: Interna- tional Conference on Learning Representations (ICLR) (2023)

    Rafailov, R.P., Wieting, J., Mitchell, E., Ermon, S., Manning, C.D.: Direct pref- erence optimization: Your language model is secretly a reward model. In: Interna- tional Conference on Learning Representations (ICLR) (2023)

    2023

  24. [32]

    why should i trust you?

    Ribeiro, M.T., Singh, S., Guestrin, C.: "why should i trust you?": Explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD Interna- tional Conference on Knowledge Discovery and Data Mining. pp. 1135–1144 (2016)

    2016

  25. [33]

    In: ACM SIGIR Conference

    Robertson, S., Walker, S.: Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In: ACM SIGIR Conference. pp. 232–241 (1994). https://doi.org/10.1007/BFb0026825

    1994 doi

  26. [34]

    Foundations and Trends in Information Retrieval3(4), 333–389 (2009)

    Robertson, S., Zaragoza, H.: The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval3(4), 333–389 (2009)

    2009

  27. [35]

    In: Proceedings of the 10th international conference on World Wide Web

    Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Item-based collaborative filtering recommendation algorithms. In: Proceedings of the 10th international conference on World Wide Web. pp. 285–295 (2001)

    2001

  28. [36]

    In: The adaptive web: methods and strategies of web personal- ization, pp

    Schafer, J.B., Frankowski, D., Herlocker, J., Sen, S.: Collaborative filtering recom- mender systems. In: The adaptive web: methods and strategies of web personal- ization, pp. 291–324. Springer (2007)

    2007

  29. [37]

    In: arXiv preprint arXiv:1707.06347 (2017)

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. In: arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  30. [38]

    In: Proceedings of the 2018 world wide web conference

    Wang, X., He, X., Feng, F., Nie, L., Chua, T.S.: Tem: Tree-enhanced embedding model for explainable recommendation. In: Proceedings of the 2018 world wide web conference. pp. 1543–1552 (2018)

    2018

  31. [39]

    In: Proceedings of the 40th International ACM SIGIR Con- ference on Research and Development in Information Retrieval

    Xiong, C., Dai, Z., Callan, J., Liu, Z., Power, R.: End-to-end neural ad-hoc ranking with kernel pooling. In: Proceedings of the 40th International ACM SIGIR Con- ference on Research and Development in Information Retrieval. pp. 55–64 (2017)

    2017

  32. [40]

    arXiv preprint arXiv:2408.09865 (2024)

    Yang, C.W., Feng, Z.Q., Lin, Y.J., Chen, C.W., Wu, K.d., Xu, H., Yao, J.F., Kao, H.Y.: Maple: Enhancing review generation with multi-aspect prompt learning in explainable recommendation. arXiv preprint arXiv:2408.09865 (2024)

    2024 arXiv

  33. [41]

    In: First Conference on Language Modeling (2024)

    Zhang, T., Patil, S.G., Jain, N., Shen, S., Zaharia, M., Stoica, I., Gonzalez, J.E.: Raft: Adapting language model to domain specific rag. In: First Conference on Language Modeling (2024)

    2024

  34. [42]

    arXiv preprint arXiv:1904.09675 (2019)

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019)

    1904 arXiv

  35. [43]

    In: Proceedings of the 37th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Zhang, Y., Lai, G., Zhang, M., Zhang, Y., Liu, Y., Ma, S.: Explicit factor mod- els for explainable recommendation based on phrase-level sentiment analysis. In: Proceedings of the 37th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp....

    2014

  36. [44]

    arXiv preprint arXiv:2308.07107 (2023)

    Zhu, Y., Yuan, H., Wang, S., Liu, J., Liu, W., Deng, C., Chen, H., Liu, Z., Dou, Z., Wen, J.R.: Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107 (2023)

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.