Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

RMIT-ADM+S at the SIGIR 2025 LiveRAG Challenge

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper reports a first-place finish: the G-RAG pipeline—hypothetical answer generation before retrieval, hybrid retrieval, pointwise LLM re-ranking, and a simple answer prompt—won the SIGIR 2025 LiveRAG Challenge on aggregated manual…

desk verdict A transparent, externally-anchored systems report from the LiveRAG winner; the win is real, but the causal claim that G-RAG's design choices caused it is unsupported by the paper's own small-sample comparison. read the letter →

arxiv 2506.14516 v2 pith:MHFQZQYM submitted 2025-06-17 cs.IR

classification cs.IR
keywords Retrieval-AugmentedGenerationLiveRAGChallengehypotheticalanswerHyDEpointwiseLLMre-rankingGridofPointsN-wayANOVAjudging
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

At the first SIGIR LiveRAG Challenge, the authors' G-RAG system took first place on the official manual evaluation, which aggregates Coverage, Relatedness, and Quality scores into a Borda score. The paper's central claim is that the win came from inserting one extra step before retrieval: a fixed 10B generation model writes a short hypothetical answer that is then used as a search query alongside the original question, followed by hybrid sparse-dense retrieval, pointwise LLM re-ranking, and a deliberately simple answer prompt. The authors also claim that this configuration choice was not arbitrary: a Grid-of-Points sweep over 96 configurations and an N-way ANOVA identified question augmentation as by far the most influential component, with $\omega^2_p = 0.7571$. If these claims hold, they suggest that cheap, zero-shot generate-then-retrieve can improve RAG outcomes even when the final answer generator is frozen, and that factorial component analysis is a practical tuning tool under contest time limits. The evidence base has a narrow point: the decisive head-to-head used only 100 questions, where the winning variant beat the champion configuration on 8 questions and lost on 7 for Relevance, with LLM-judge scores rather than the official manual protocol.

What carries the argument

The carrying object is G-RAG, a RAG pipeline with an added generation-before-retrieval stage: the fixed Falcon3-10B-Instruct model writes a hypothetical answer, and that hypothetical answer is used as a search query alongside the original question. Three supporting mechanisms do the work: (1) hybrid retrieval merging BM25 and dense cosine-similarity rankings with reciprocal rank fusion at $k = 60$; (2) a pointwise LLM re-ranker, defined as a classifier that scores each document by the likelihood of a 'Yes' token and drops documents scoring below 0.5; and (3) a simple answer-generation prompt. The hypothetical answer is the load-bearing novelty: the ANOVA attributes most explained variance to question augmentation ($\omega^2_p = 0.7571$), while the re-ranker and fusion method provide the ranked context the generator uses.

What would settle it

Re-run the submitted G-RAG and the best non-G-RAG configuration on the official LiveRAG test questions using the organizers' manual protocol; if G-RAG does not beat the champion, or the 8-versus-7 relevance split flips on a larger sample, the claim that hypothetical-answer generation caused the win is falsified. Separately, compute per-question agreement between the internal LLM judge's Relevance and Faithfulness scores and the official manual scores; low agreement would mean the 96-configuration selection was tuned to a proxy rather than to the target.

Watch

Extended reading notes

Core claim

The paper's central claim is that G-RAG won the first SIGIR 2025 LiveRAG Challenge. The submitted run used the fixed Falcon3-10B-Instruct model in every LLM role: it first writes a short hypothetical answer that could be true to the question, and that generated text is added to the original question as an extra retrieval query. Retrieval is hybrid: BM25 over OpenSearch plus dense embeddings on Pinecone, merged by reciprocal rank fusion. A pointwise LLM re-ranker then scores each candidate document by the logit-likelihood of generating 'Yes' when asked whether the document helps answer the question, discarding documents scoring below 0.5. The final answer is produced by a deliberately simple 'answer based on the provided documents' prompt. On the official manual evaluation, the aggregated Borda score placed the system first, with top individual scores of Coverage 1.61, Relatedness 1.88, and Quality 1.67 on a 0–2 Likert scale. The paper attributes this result to the design and evaluation strategy: a 96-configuration Grid-of-Points sweep and an N-way ANOVA pointed to question augmentation as the dominant component, and a follow-up comparison motivated replacing the champion's no-augmentation setting with hypothetical-answer generation.

Load-bearing premise

The whole configuration story depends on the internal LLM judge being a faithful stand-in for the hidden official manual evaluation, and on a 100-question comparison with a near-tie head-to-head being large enough to tell which component change actually helped.

Editorial extensions

If this is right

  • If G-RAG's win is attributable to its components, future LiveRAG-style tasks can improve RAG by generating a hypothetical answer and using it as an extra query, without changing the fixed answer generator.
  • The 96-configuration Grid-of-Points sweep and ANOVA show that in-house component analysis can pick a winning configuration under contest deadlines: question augmentation alone accounted for the largest share of explained variance.
  • The pointwise LLM re-ranker, based on the logit probability of generating 'Yes', is compatible with hybrid BM25 plus dense retrieval and can be applied with a 10B open-weight model.
  • Reducing input context from 15k to 10k tokens had no significant measured effect on Relevance or Faithfulness, so latency can be traded for context length without sacrificing these metrics.

Reading between the lines

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

  • Beyond the paper: the decisive head-to-head is a 100-question sample where the G-RAG variant wins Relevance on 8 questions and loses on 7, and loses Faithfulness 14 to 12; those counts are too close to establish by themselves that hypothetical-answer generation is the operative improvement, so the win may partly come from hybrid retrieval or re-ranking.
  • Beyond the paper: a clean test is to ablate only the hypothetical-answer query from the submitted run and re-run the official manual evaluation; if the gap disappears, the 'G' step is essential, and if not, the paper's evaluation narrative would need revision.
  • Beyond the paper: the internal LLM judge is a proxy for a hidden official protocol, so the ANOVA component rankings should be re-validated against official scores on a shared sample; low judge-to-official agreement would make the factorial conclusions a property of the proxy rather than the task.
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 describes the RMIT–ADM+S system, G-RAG, submitted to the SIGIR 2025 LiveRAG Challenge. G-RAG generates a hypothetical answer that is used as an additional search query, combines sparse and dense retrieval with reciprocal rank fusion, applies a pointwise LLM-based re-ranker, and produces the final answer with Falcon3-10B-Instruct using a simple prompt. The authors report an in-house evaluation based on DataMorgana-generated data, a Grid of Points search over 96 configurations, an N-way ANOVA analysis, and a focused post-ANOVA comparison against a hypothetical-answer variant. The paper claims that the submitted system won the challenge by achieving the highest Borda score on the official manual evaluation, and it interprets this outcome as validating the effectiveness of the proposed design and evaluation strategy.

Significance. If the official win is confirmed by the challenge report, the paper provides a useful, reproducible system description: the architecture is simple, the code is publicly available, and the use of Grid of Points and ANOVA is a sensible approach to configuration selection under time pressure. The paper also honestly acknowledges several limitations of LLM-based evaluation in Section 5. However, the internal evidence supporting the final selection of G-RAG is weak: the 100-question comparison in Table 3 shows essentially no statistical advantage over the champion configuration, and the in-house LLM judge is not validated against the official manual evaluation. The official victory therefore does not, by itself, substantiate the claim that the specific G-RAG design or the GoP/ANOVA strategy caused the win. The paper's central factual claim is sound, but the causal/validating inference needs significant moderation or additional evidence.

major comments (3)
  1. [§3.4, Table 3] The selection of G-RAG over the champion configuration rests on a 100-question comparison in which the average Relevance scores are identical (1.75 vs. 1.75), G-RAG wins on Relevance in 8 questions and loses in 7, and it scores higher on Faithfulness in 12 questions and lower in 14. With 15 and 26 discordant pairs respectively, a sign test or McNemar test would return large p-values (approximately 1.0 and 0.85), indicating that the observed win counts are fully compatible with chance. Furthermore, the 100-question sample was constructed by combining 15 deliberately selected 'tricky' questions with 85 questions from a separate dataset (Section 3.1.1), so it is not a random sample of the challenge distribution. The paper should provide a significance test and a more representative sample, or explicitly refrain from claiming that this comparison validates G-RAG.
  2. [§3.1.3 and §3.4] The in-house evaluation uses Claude 3.5 Sonnet as a proxy for the official LiveRAG evaluation, but the exact official prompts and procedures are not public and no validation of the proxy is provided. The discrepancy between the in-house average scores (Relevance 1.75, Faithfulness 0.59) and the official private-leaderboard scores reported in Section 4 (Relevance 1.199, Faithfulness 0.477) suggests that the proxy is not calibrated to the official protocol. Without agreement statistics between the proxy and the official evaluation on a common sample, the internal ranking of configurations may not transfer to the challenge setting. The authors should either supply such validation or temper the conclusion that the official win validates their evaluation strategy.
  3. [§3.3 and §3.4] The final submitted configuration was never evaluated inside the GoP/ANOVA framework. The ANOVA indicated that Question Augmentation was the most influential component, but the GoP champion used the 'None' augmentation setting; G-RAG introduces a new augmentation level (hypothetical answer generation), and the paper explicitly states that the full GoP and ANOVA were not re-run for this change. Consequently, the conclusion in Section 4 that the GoP-ANOVA approach 'enabled' the selection of the winning configuration overstates what the systematic evaluation actually supports. The systematic evaluation selected the champion, whereas the final winning run was chosen by a separate, underpowered ad-hoc comparison. This distinction should be acknowledged and the conclusion reworded.
minor comments (4)
  1. [§4] The relationship between the private-leaderboard rank (third) and the final manual-evaluation win is not explained; the text should clarify that these are distinct evaluation stages (automatic relevance/faithfulness scores versus the manual Borda aggregation).
  2. [Table 3] The table headings 'Relevance Score' and 'Faithfulness Score' are misleading because the entries are counts of per-question wins, not scores; use 'Relevance wins' and 'Faithfulness wins' or similar.
  3. [§3.1.1] The description of the test set construction is ambiguous: it says 15 'tricky' questions were combined with 85 questions from a separate dataset, but it is unclear whether the 15 'challenging' questions are a subset of the 179 'tricky' questions and whether the 85 questions overlap with other datasets used later.
  4. [§3.2] The paper states that 96 configurations were evaluated while holding the number of query variants at eight and always applying the re-ranker, but it does not list the exact factor levels for each component in the grid; including a table or appendix of the full configuration space would improve reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the winning result is anchored to the external LiveRAG challenge report; in-house tuning and self-citations are not load-bearing reductions, though the selection step is statistically fragile.

full rationale

The paper's central factual claim, that the submitted G-RAG system ranked first in the SIGIR 2025 LiveRAG Challenge, is anchored to the external challenge report [4], not to the paper's own in-house evaluation. The in-house chain (GoP, ANOVA, post-ANOVA comparison) is a configuration-selection procedure, not a derivation of the official outcome. Section 3.4 reports that G-RAG and the champion tie on average Relevance (1.75 vs 1.75) with an 8-7 per-question advantage, and Section 3.1.3 states that the exact official prompts and procedures are not publicly disclosed. These are threats to the causal claim that the design choices caused the win, but they are statistical-validity concerns, not circular reductions: the official manual evaluation is independent of the Claude 3.5 Sonnet proxy used for selection. Self-citations ([16], [17], [24]) supply prompts or future-work directions; none is invoked as a uniqueness theorem or as the justification for the winning configuration. The closest sentence to circularity, 'G-RAG was ultimately announced as the winning system, validating the effectiveness of our design and evaluation strategy' (Section 4), over-attributes success to the evaluation strategy, but it does not fold the official result into an input assumption. Section 5 appropriately acknowledges LLM-judge limitations and the need for human validation. Overall, no load-bearing circular step is present; the paper is externally benchmarked, so the circularity score is low.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No theoretical derivation is attempted. The reported improvement rests on a synthetic dev set, an LLM-judge proxy, a 96-configuration sweep with fixed re-ranker and variant count, and a small ad-hoc comparison for the final G-RAG decision. Four numeric choices are effectively free parameters tuned on development data.

free parameters (4)
  • Reciprocal Rank Fusion k = 60
    Fusion constant used to merge sparse and dense rankings; the paper states k=60 but does not report tuning over values.
  • Re-ranker likelihood threshold = 0.5
    Documents with a 'Yes' likelihood below 0.5 are discarded; the threshold appears fixed, with no ablation reported.
  • Number of query variants = 8
    Question augmentation produced 8 variants in the GoP sweep; the paper says this was set to 8 to manage computational constraints.
  • Context word limit = 10,000 tokens
    Chosen over the 15k context that gave the best relevance, to lower latency; the ANOVA showed no significant effect of context length.
assumptions (3)
  • domain assumption The DataMorgana synthetic question-answer sets are representative of the LiveRAG test corpus
    The entire in-house optimization uses synthetic data; Section 3.1.1 describes the configuration and test set construction.
  • domain assumption Claude 3.5 Sonnet relevance and faithfulness scores approximate the official LiveRAG manual evaluation
    Section 3.1.3 states the official prompts are undisclosed and the team 'aimed to replicate the setup as closely as possible.'
  • standard math Standard ANOVA assumptions (independent observations, additivity, normality) hold for the 96-configuration comparison
    Section 3.3 applies N-way ANOVA with partial omega squared without checking residual assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RMIT-ADM+S at the SIGIR 2025 LiveRAG Challenge." pith.science (2026). https://pith.science/paper/MHFQZQYM

@misc{pith2026250614516,
  author       = {Pith},
  title        = {Pith review of: RMIT-ADM+S at the SIGIR 2025 LiveRAG Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHFQZQYM}},
  note         = {Machine review of arXiv:2506.14516}
}
read the original abstract

This paper presents the RMIT--ADM+S winning system in the SIGIR 2025 LiveRAG Challenge. Our Generation-Retrieval-Augmented Generation (G-RAG) approach generates a hypothetical answer that is used during the retrieval phase, alongside the original question. G-RAG also incorporates a pointwise large language model (LLM)-based re-ranking step prior to final answer generation. We describe the system architecture and the rationale behind our design choices. In particular, a systematic evaluation using the Grid of Points approach and N-way ANOVA enabled a controlled comparison of multiple configurations, including query variant generation, question decomposition, rank fusion strategies, and prompting techniques for answer generation. The submitted system achieved the highest Borda score based on the aggregation of Coverage, Relatedness, and Quality scores from manual evaluations, ranking first in the SIGIR 2025 LiveRAG Challenge.

Figures

Figures reproduced from arXiv: 2506.14516 by the authors.

Figure 1
Figure 1. Pipeline of our G-RAG approach. Components used in the final selected run are shown with solid borders; components [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SIGIR 2025 -- LiveRAG Challenge Report

    cs.CL 2025-07 conditional novelty 3.0 of 10

    In the SIGIR 2025 LiveRAG Challenge, all 25 active RAG teams beat the no-RAG baseline on LLM-judged correctness, and LLM scores correlated with human scores at r=0.88.

Reference graph

Works this paper leans on

32 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [4]

    David Carmel, Simone Filice, Guy Horowitz, Yoelle Maarek, Oren Somekh, and Ran Tavory. 2025. SIGIR 2025–LiveRAG Challenge Report.arXiv preprint arXiv:2507.04942(2025)

  2. [1]

    Marwah Alaofi, Negar Arabzadeh, Charles L. A. Clarke, and Mark Sanderson. 2025.Generative Information Retrieval Evaluation. Springer Nature Switzerland, Cham, 135–159. doi:10.1007/978-3-031-73147-1_6

  3. [2]

    Marwah Alaofi, Luke Gallagher, Mark Sanderson, Falk Scholer, and Paul Thomas

  4. [3]

    Ameer Albahem, Damiano Spina, Falk Scholer, and Lawrence Cavedon. 2021. Component-based Analysis of Dynamic Search Performance.ACM Trans. Inf. Syst.40, 3 (Nov. 2021), 61:1–61:47. doi:10.1145/3483237

  5. [5]

    Cormack, Charles L A Clarke, and Stefan Buettcher

    Gordon V. Cormack, Charles L A Clarke, and Stefan Buettcher. 2009. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. In Proc. SIGIR. 758–759. doi:10.1145/1571941.1572114

  6. [6]

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. The Power of Noise: Redefining Retrieval for RAG Systems. InProc. SIGIR. 719–729. doi:10.1145/3626772.3657834

  7. [7]

    Laura Dietz, Oleg Zendel, Peter Bailey, Charles Clarke, Ellese Cotterill, Jeff Dalton, Faegheh Hasibi, Mark Sanderson, and Nick Craswell. 2025. Principles and Guidelines for the Use of LLM Judges. InProc. ICTIR. doi:10.1145/3731120.3744588

  8. [8]

    Guglielmo Faggioli, Oleg Zendel, J Shane Culpepper, Nicola Ferro, and Falk Scholer. 2022. sMARE: A New Paradigm to Evaluate and Understand Query Performance Prediction Methods.Information Retrieval Journal25, 2 (2022), 94–122. doi:10.1007/s10791-022-09407-w

Show all 32 references
  1. [9]

    Nicola Ferro and Gianmaria Silvello. 2016. A General Linear Mixed Models Approach to Study System Component Effects. InProc. SIGIR. 25–34. doi:10. 1145/2911451.2911530

  2. [10]

    Simone Filice, Guy Horowitz, David Carmel, Zohar Karnin, Liane Lewin-Eytan, and Yoelle Maarek. 2025. Generating Diverse Q&A Benchmarks for RAG Evalua- tion with DataMorgana. arXiv:2501.12789 [cs.CL]

  3. [11]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise Zero-Shot Dense Retrieval without Relevance Labels. InProc. ACL. 1762–1777. doi:10.18653/ v1/2023.acl-long.99

  4. [12]

    Antoinette DA Kroes and Jason R Finley. 2023. Demystifying Omega Squared: Practical Guidance for Effect Size in Common Analysis of Variance Designs. Psychological Methods(2023). doi:10.1037/met0000581

  5. [13]

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michi- hiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D Manning, Christopher Re, Diana Acosta-Nava...

  6. [14]

    2017.Designing Experiments and Analyzing Data: A Model Comparison Perspective

    Scott E Maxwell, Harold D Delaney, and Ken Kelley. 2017.Designing Experiments and Analyzing Data: A Model Comparison Perspective. Routledge. doi:10.4324/ 9781315642956

  7. [15]

    Erxue Min, Hsiu-Yuan Huang, Min Yang, Xihong Yang, Xin Jia, Yunfang Wu, Hengyi Cai, Junfeng Wang, Shuaiqiang Wang, and Dawei Yin. 2025. From Prompting to Alignment: A Generative Framework for Query Recommendation. arXiv:2504.10208 [cs.IR]

  8. [16]

    Abushaqra, Angel Felipe Mag- nossão de Paula, Kaixin Ji, Halil Ali, Danula Hettiachchi, Johanne R

    Sachin Pathiyan Cherumanal, Lin Tian, Futoon M. Abushaqra, Angel Felipe Mag- nossão de Paula, Kaixin Ji, Halil Ali, Danula Hettiachchi, Johanne R. Trippas, Falk Scholer, and Damiano Spina. 2024. Walert: Putting Conversational Information Seeking Knowledge into Action by Buildi...

  9. [17]

    Kun Ran, Marwah Alaofi, Mark Sanderson, and Damiano Spina. 2025. Two Heads Are Better Than One: Improving Search Effectiveness Through LLM-Generated Query Variants. InProc. CHIIR. 333–341. doi:10.1145/3698204.3716468

  10. [18]

    SE Robertson, S Walker, S Jones, M Hancock-Beaulieu, and M Gatford. 1994. Okapi at TREC-3. InProc. of the Third Text REtrieval Conference (TREC 1994). 18 pages. https://trec.nist.gov/pubs/trec3/papers/city.ps.gz

  11. [19]

    2011.ANOV A and ANCOV A: A GLM Approach(1 ed.)

    Andrew Rutherford. 2011.ANOV A and ANCOV A: A GLM Approach(1 ed.). Wiley. doi:10.1002/9781118491683

  12. [20]

    Shuoqi Sun, Shengyao Zhuang, Shuai Wang, and Guido Zuccon. 2025. An Investigation of Prompt Variations for Zero-shot LLM-based Rankers. InECIR. 185–201. doi:10.1007/978-3-031-88711-6_12

  13. [21]

    Manveer Singh Tamber, Forrest Sheng Bao, Chenyu Xu, Ge Luo, Suleman Kazi, Minseok Bae, Miaoran Li, Ofer Mendelevitch, Renyi Qu, and Jimmy Lin

  14. [22]

    TII Team. 2024. The Falcon 3 family of Open Models

  15. [23]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2024. Text Embeddings by Weakly-Supervised Contrastive Pre-training. arXiv:2212.03533 [cs.CL]

  16. [24]

    Shane Culpepper, and Falk Scholer

    Oleg Zendel, J. Shane Culpepper, and Falk Scholer. 2021. Is Query Performance Prediction With Multiple Query Variations Harder Than Topic Performance Prediction?. InProc. SIGIR. 1713–1717. doi:10.1145/3404835.3463039 A Prompts A.1 Query Generation A.1.1 Naïve Prompt. System Pr...

  17. [27]

    Each query should focus on different aspects or interpretations of the original question

  18. [28]

    Use synonyms and related terms where appropriate

  19. [29]

    Include both broad and specific variations

  20. [30]

    Maintain the core meaning while varying the expression

  21. [31]

    Write each query on a new line

  22. [32]

    No-RES” with “I don’t know

    Do not include any additional text or formatting The original question is correctly spelled. 6 RMIT–ADM+S at the SIGIR 2025 LiveRAG Challenge User Prompt: Question to analyze: {question} Please generate diverse query variations that capture different aspects of this question: ...

  23. [2023]

    Can Generative LLMs Create Query Variants for Test Collections? An Exploratory Study. InProc. SIGIR. 1869–1873. doi:10.1145/3539618.3591960

  24. [2025]

    arXiv:2505.04847 [cs.CL]

    Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards. arXiv:2505.04847 [cs.CL]

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.