Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval
Open accessThe DOI matches its Crossref record; 88 of the 88 references it cites were found in a registry or an index; 7 of the 10 sources its claims cite were checked for support.
Abstract
Generative information retrieval (GenIR) is a promising neural retrieval paradigm that formulates document retrieval as a document identifier (docid) generation task, allowing for end-to-end optimization toward a unified global retrieval objective. To address token-level misalignment, we propose direct document relevance optimization (DDRO), which aligns token-level docid generation with document-level relevance estimation through direct optimization via pairwise ranking, eliminating the need for explicit reward modeling and reinforcement learning. Experimental results show DDRO achieves a 7.4% improvement in MRR@10 for MS MARCO and a 19.9% improvement for Natural Questions.
Overview
CoChat Summary
TL;DR
DDRO aligns token-level docid generation with document-level relevance by adding pairwise ranking after supervised fine-tuning, avoiding reward modeling and reinforcement learning. On MS MARCO and NQ320k it improves early ranking over GenRRL and established baselines, including 7.4% and 19.9% MRR@10 gains over RL-based methods.
Key findings
- DDRO outperforms RLRF/GenRRL: on MS MARCO, DDRO (TU) beats GenRRL (Sum) by 15.06% in R@1 and 7.4% in MRR@10; on NQ, DDRO (PQ) beats GenRRL (Sum) by 34.69% in R@1 and 19.87% in MRR@10.
- Against established baselines: on MS MARCO, DDRO (TU) improves over the best baseline ROGER-Ultron by 15.63% in R@1 and 8.03% in MRR@10; on NQ, DDRO (PQ) improves over ROGER-Ultron (TU) by 36.27% in R@1 and 23.58% in MRR@10.
- Ablation shows pairwise ranking matters: removing it lowers NQ DDRO (PQ) from 48.92 to 44.19 R@1 and from 55.51 to 50.48 MRR@10; on MS MARCO, DDRO (TU) drops from 66.46 to 64.60 R@5 and from 74.01 to 72.90 R@10.
- Docid design interacts with dataset: DDRO (TU) is strongest on MS MARCO (38.24 R@1, 50.07 MRR@10), while DDRO (PQ) is strongest on NQ (48.92 R@1, 55.51 MRR@10).
Methods
- Two docid types are constructed: TU (reverse URL if semantically rich, otherwise title + domain) and PQ (product quantization codes from a T5-based encoder).
- Supervised fine-tuning uses T5-base with indexing and retrieval tasks, passage-to-docid pairs, tf-idf-selected term-to-docid pairs, and pseudo-queries generated by DocT5Query (10 per document).
- DDRO then directly optimizes pairwise ranking with a reference-policy constraint, initialized from the SFT model and trained with a modified Hugging Face TRL DPOTrainer; experiments use MS MARCO document ranking (320k documents, 360k query-document pairs) and NQ320k (320k query-document pairs), evaluated by R@1/5/10 and MRR@10 with paired t-tests at p < 0.05.
Limitations
- Stated: evaluation primarily focuses on a subset of the MS MARCO document dataset, leaving scalability to larger, domain-specific, and multilingual corpora unexplored.
- Stated: future work includes listwise ranking, advanced hard negative mining, and multi-objective optimization; a comprehensive comparison with scalable GenIR baselines such as RIPOR and PAG remains future work.
- Evident from the text: only two benchmark datasets are used, and document summaries are excluded as docids because they depend on external summarization models, which limits direct comparison with GenRRL (Sum).
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures

![Figure 3 Architecture of the DDRO model, which fine-tunes the retrieval model through direct learning-to-rank (L2R) optimization using relevance feedback. Unlike GenRRL [68], DDRO directly optimizes with relevance judgment data, avoiding reinforcement learning, explicit reward modeling, and extensive hyperparameter tuning. For clarity , the model 𝜋 ref 𝜃 from the SFT phase is referred to as 𝜋 ref , with its parameters frozen during this phase.](https://figures.cochat.ai/paper-figures/05526d7477994286982bb807fa34759b/lightweight-and-direct-document-relevance-optimization-for-generative-figure-3.png)



Verification
Claims the paper makes
2 Valid · 3 Not valid · 1 Not checked · 50 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“However, existing GenIR models suffer from token-level misalignment, where models trained to predict the next token often fail to capture document-level relevance effectively.”
- Also stated in: § 3.2 Learning to Rank (L2R) · plain
“A fundamental challenge in GenIR stems from the inherent misalignment between the optimization objectives of autoregressive models and the overarching objectives of document ranking tasks.”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > Abstract
What the sources say
- Could not verifycites [28] 28. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (Eds.). Association for Computational Linguistics, Online, 7871-7880. https://doi.org/10.18653/v1/2020.acl-main.703 sourceExperimental — accuracy not yet measured
Section 2 (Model) describes BART’s autoregressive decoder, while Section 4.3 (Results) compares pre-training methods across tasks; page numbers are not provided in our copy. Neither section addresses GenIR, docid generation, token-level misalignment with document relevance, or its effect on ranking. Searches for those terms and paraphrases found no passage that does. (LLM, conf=0.78)
- Could not verifycites [68] 68. Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval. https://doi.org/10. 48550/ARXIV.2007.00808 sourceExperimental — accuracy not yet measured
The Related Work section (p. 8) describes dense retrieval through embedding-space search, and the training analysis (p. 3) discusses the informativeness of negative samples. Searches for autoregressive docid generation, next-token prediction, and token-level misalignment found no passage addressing the claimed failure of GenIR models. (LLM, conf=0.76)
- Could not verifycites [69] 69. Andrew Yates, Rodrigo Nogueira, and Jimmy Lin. 2021. Pretrained Transformers for Text Ranking: BERT and Beyond. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Tutorials, Greg Kondrak, Kalina Bontcheva, and Dan Gillick (Eds.). Association for Computational Linguistics, Online, 1-4. sourceExperimental — accuracy not yet measured
Our copy of the 4-page NAACL 2021 tutorial covers only two categories of transformer ranking work — multi-stage ranking/reranking (Sec. 2) and learned dense representations (Sec. 3) — plus a short 'Looking Ahead' (Sec. 4). Generative retrieval with autoregressive document-identifier (docid) generation is mentioned only in passing as 'adapting pretrained sequence-to-sequence models (Nogueira et al., 2020; dos Santos et al., 2020)' among efforts 'beyond BERT'; the tutorial makes no statement about token-level misalignment, next-token training objectives failing to capture document-level relevance, or resulting suboptimal ranking performance. Searches for 'token-level misalignment', 'docid/identifier generation', 'autoregressive/generative retrieval', and 'next token prediction vs ranking loss' returned no passage on the point, so the claim can be neither confirmed nor contradicted from our copy. (LLM, conf=0.72)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 3.2 Learning to Rank (L2R) · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The performance of dense retrieval baselines is generally better than that of sparse retrieval baselines, likely because the former uses dense vectors to capture richer semantic information, which is also consistent with the findings [31,32].”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 6.2 Comparison with Established Baselines
What the sources say
- Partly supportedcites [31] 31. Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 6636-6648. https://doi.org/10. 18653/v1/2023.acl-long.366 sourceExperimental — accuracy not yet measured
Section 4.3 and Table 2 compare dense and sparse retrieval baselines on MS MARCO; several dense models outperform the sparse baselines. The paper does not report results on MS300K, so it supports the general comparison but not the claim for that dataset. (LLM, conf=0.85)
- Not checked yetcites [32] 32. Zihao Li, Zhuoran Yang, and Mengdi Wang. 2023. Reinforcement Learning with Human Feedback: Learning Dynamic Choices via Pessimism. https: //doi.org/10.48550/ARXIV.2305.18438 source
Checked before — its earlier verdict is held back until a new check runs.
hedgedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“(ii) Dense retrieval baselines, including DPR and ANCE, improve early ranking metrics over term-based methods.”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > Comparison to baseline retrieval models.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“On the MS MARCO dataset, removing pairwise ranking optimization significantly reduces intermediate and broader recall metrics (R@5 and R@10) for both DDRO variants, highlighting its critical role in improving retrieval performance.”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 6.3 Ablation Study
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, at broader recall levels (R@10), ROGER-Ultron (TU) shows a slight advantage over DDRO (TU).”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 6.5 Analysis of Relevance Distribution
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
18 more claims
Create a free account to see all 56 claims and to verify the 2 references still open.
- It may state two separate claims as one. 25
- Probably not a substantial claim. 11
- Borderline: one of the tests for a substantial claim is unclear. 2
- The quote only partly states this claim. 2
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 4.3 Direct L2R Optimization Using Relevance Feedback
“The gradient with respect to the model parameters 𝜃 is defined as in Eq. 7, where”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 4.2 Supervised Fine-tuning
“passage : {𝑤 𝑖 , 𝑤 𝑖+1 , . . . , 𝑤 𝑖+𝑚-1 } → 𝑑𝑜𝑐𝑖𝑑, (7”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 6.1 Comparison with Reinforcement Learning-Based Methods
“To address RQ1, we compare DDRO with GenRRL [68] on both datasets.Results are presented in Table 1 and Table 2.”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > References
“68. Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval. https://doi.org/10. 48550/ARXIV.2007.00808”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > 5.1 Datasets and Evaluation Metrics
“the Natural Questions (NQ) datasetfoot_1 [23].”
§ Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval > References
“23. Varsha Kishore, Chao Wan, Justin Lovelace, Yoav Artzi, and Kilian Q. Weinberger. 2023. IncDSI: incrementally updatable document retrieval.”
Citations
Cites
Proceedings of the 24th international conference on Machine learning
(2007) · doi:10.1145/1273496 · Cited by 11,630
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
Proximal Policy Optimization Algorithms
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017) · arXiv · doi:10.48550/arxiv.1707.06347 · Cited by 10,625
Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., & Zettlemoyer, L. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.703 · Cited by 8,821 · cited 1×
“GenIR models build on large pre-trained language models, such as T5 [41] and BART [27], and are designed to take a query string and generate a ranked list of document identifiers (docids) based on their generation probabilities, ordered in descending sequence.”
Dense Passage Retrieval for Open-Domain Question Answering
Karpukhin, V., Oğuz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W.-T. (2020) · arXiv · doi:10.48550/arxiv.2004.04906 · Cited by 5,311
LLaMA: Open and Efficient Foundation Language Models
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023) · arXiv · doi:10.48550/arxiv.2302.13971 · Cited by 4,602
A Learning Algorithm for Continually Running Fully Recurrent Neural Networks
Williams, R. J., & Zipser, D. (1989) · Neural Computation · doi:10.1162/neco.1989.1.2.270 · Cited by 4,369
Dense passage retrieval for open-domain question answering
Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W.-T. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.550 · Cited by 3,575 · cited 4×
“A widely recognized approach [39, 47, 73] optimizes the reward function using Proximal Policy Optimization (PPO) [46], and defines the reward function as: GenRRL [68] trains a reward model using relevance-annotated data derived from BM25 [44], DPR [21], and LLaMA-13b [53].”
InfluentialThe probabilistic relevance framework: BM25 and beyond
Robertson, S., & Zaragoza, H. (2009) · Foundations and Trends® in Information Retrieval · doi:10.1561/1500000019 · Cited by 3,374 · cited 5×
“Positive samples were selected based on qrels relevance judgments, while negatives were drawn from the top 1000 BM25-retrieved documents, stratified into top (1–100), mid (101–500), and lower (501–1000) ranks.”
InfluentialTraining Language Models to Follow Instructions with Human Feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-2011 · Cited by 3,117
Research works printed but not linked: 12
- Christopher JC Burges. 2010. From RankNet to LambdaRank to LambdaMart: An Overview. Learning, Vol. 11, 23--581 (2010), 81.
- Yanyan Lan, Yadong Zhu, Jiafeng Guo, Shuzi Niu, and Xueqi Cheng. 2014. Position-aware ListMLE: A Sequential Learning Process for Ranking. In Proceedings of the Thirtieth Conference on Uncertainty in Artificial Intelligence (Quebec City, Quebec, Canada) (UAI'14). AUAI Press, Arlington, Virginia, USA, 449--458.
- 10.5555/3600270.3602281 · doi:10.5555/3600270.3602281
- 10.5555/3666122.3666574 · doi:10.5555/3666122.3666574
- Stephen E. Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford. 1994. Okapi at TREC-3. In Proceedings of the Third Text REtrieval Conference (TREC-3). National Institute of Standards and Technology (NIST), Gaithersburg, Maryland, USA, 109-126. https://www.microsoft.com/en-us/research/publication/okapi-at-trec-3/
- 10.5555/3495724.3495977 · doi:10.5555/3495724.3495977
- 10.5555/3666122.3668132 · doi:10.5555/3666122.3668132
- Yubao Tang, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Wei Chen, and Xueqi Cheng. 2024a. Generative Retrieval Meets Multi-Graded Relevance. In Advances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37. Curran Associates, Inc., 72790-72817. https://proceedings.neurips.cc/paper_files/paper/2024/file/853e781cb2af58956ed5c89aa59da3fc-Paper-Conference.pdf
- Yubao Tang, Ruqing Zhang, Zhaochun Ren, Jiafeng Guo, and Maarten de Rijke. 2024 e. Recent Advances in Generative Information Retrieval. In Advances in Information Retrieval, Nazli Goharian, Nicola Tonellotto, Yulan He, Aldo Lipani, Graham McDonald, Craig Macdonald, and Iadh Ounis (Eds.). Springer Nature Switzerland, Cham, 363-368.
- 10.1145/3589335.3641239 · DOI
- 10.5555/3600270.3601857 · doi:10.5555/3600270.3601857
- 10.5555/3600270.3602126 · doi:10.5555/3600270.3602126
Cited by
Investigating LLM Variability in Personalized Conversational Information Retrieval
Lupart, S., Dijk, D. V., Langezaal, E., van Dort, I., & Aliannejadi, M. (2025) · Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP) · doi:10.1145/3767695.3769502 · Cited by 8
Multi-Step Semantic Reasoning in Generative Retrieval
Dong, S., Tang, Y., & de Rijke, M. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-21300-6_17 · Cited by 4
Self-citationGenerative Retrieval with Few-Shot Indexing
Askari, A., Meng, C., Aliannejadi, M., Ren, Z., Kanoulas, E., & Verberne, S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-21300-6_52 · Cited by 16
Metadata
- Authors
- Mekonnen, K. A., Tang, Y., & de Rijke, M.
- Affiliations
- University of Amsterdam (NL)
- University of Amsterdam, Amsterdam, Netherlands
- Venue
- Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval
- Publisher
- ACM
- Published
- 2025-07-13
- Pages
- 1327-1338
- OpenAlex
- W4412376958
- Semantic Scholar
- ea9f2b40e15cf26943e16dcc3982b6168a31ec0b
- Cited by
- 3
- References
- 72
- Citations by year
- 2026: 2 · 2025: 1
- FWCI
- 3.43
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) (publishedVersion, cc-by)
- UvA-DARE (University of Amsterdam) (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Funders
- Funder: Dutch Research Council (Identifiers: 024.004.022, NWA.1389.20.\-183, KICH3.LTP.20.006)
- Funder: Ahold Delhaize (Identifier: AIRLab Amsterdam)
- Funder: EU (Identifier: 101070212.)
- Topics
- Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Information Systems › Recommender Systems and Techniques
- Keywords
- pairwise ranking, reward modeling, MS MARCO, Natural Questions
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval. https://app.cochat.ai/papers/10.1145/3726302.3730023
- The paper
- Mekonnen, K. A., Tang, Y., & de Rijke, M. (2025). Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval. Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. https://doi.org/10.1145/3726302.3730023
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.