acceptodds
Under review as a conference paper at ICLR 2027

Revela+: Adapting Multi-Stage Language Modeling Training for Reasoning Retrieval

Abstract

Reasoning-intensive retrieval is difficult because useful documents may support an answer through multi-step inference rather than resemble the query. Dense retrievers are mostly trained contrastively on annotated or synthetic query–document pairs, which reduces each document to a binary label on a similarity score: it records which document is relevant, but not how, or how much, it helps. A language model's next-token loss offers a finer signal, and Revela uses it to train a standalone dense retriever by letting the retriever weight the in-batch sequences that the model attends to. Revela, however, applies this only to pre-training, whereas language models are trained through a pipeline of pre-training, supervised fine-tuning (SFT), and preference tuning. Can a dense retriever be trained through this full pipeline, and what does each stage contribute? We introduce Revela+, which carries the retriever and the language model through all three stages: while attending to retriever-weighted sequences in its batch, the model predicts document text, then answers to queries, and then compares answers grounded in relevant and mined negative documents. No stage adds a document-ranking loss. Across five encoders from two model families, mean retrieval performance improves at both stage transitions on BRIGHT and BEIR. For Llama-3.1-8B, mean BRIGHT NDCG@10 rises from 25.1% after pre-training to 29.8% after SFT and 32.1% after preference tuning, surpassing contrastively trained ReasonIR-8B (24.4%) and RaDeR-7B (26.0%) with 6.6–7.7× less training compute than ReasonIR-8B.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
dense retrieval, reasoning-intensive retrieval, retriever training through language modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hfceOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Revela+, extending Revela to train dense retrievers through pre-training, supervised fine-tuning, and preference tuning. It uses document prediction, answer prediction, and answer preference objectives to train the retriever through LM attention weights. Experiments on BRIGHT and BEIR show stage-wise retrieval improvements across five encoder backbones.

Critical strengths
  • S1: The paper presents a promising extension of Revela by incorporating SFT and preference tuning into retriever training. Experiments show improvements over Revela and consistent stage-wise gains across multiple encoder backbones on BRIGHT and BEIR.
Critical weaknesses
  • W1: The motivation for extending Revela to a three-stage pipeline is insufficiently developed. While answer prediction and preference comparison provide plausible supervision, the paper does not clearly explain what limitation of the preceding stage each addresses or why this progression is preferable to directly exploiting relevance supervision.
  • W2: The empirical contribution relative to existing methods is insufficiently established. ReasonEmbed [1], also initialized from pretrained LLMs, achieves stronger BRIGHT results with contrastive learning but is omitted from the comparisons. It remains unclear what additional advantage the proposed training pipeline offers over existing approaches.

[1] Jianlyu Chen, Junwei Lan, Chaofan Li, Defu Lian, and Zheng Liu. 2026. ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1203–1221, San Diego, California, United States. Association for Computational Linguistics.

Questions for authors

None

4 comments

We thank the reviewer for their careful reading and for recognising Revela+ as a promising extension of Revela with consistent stage-wise gains across encoders and benchmarks. We address each weakness below.

Motivation for the three-stage pipeline

We agree the motivation should be stated more explicitly, and we have revised the introduction accordingly.

What each stage addresses

Pre-training needs no labels, but its retriever never scores documents from a query. In a blinded LLM-judged analysis of 70 BRIGHT-Pro queries, a quarter of its top-10 documents are off-topic. SFT addresses this by training the retriever on the documents that help the language model answer a query: the off-topic share halves and supporting documents rise from 39% to 47%. SFT, however, rewards any document that makes the answer more likely, and it never contrasts the relevant document with a plausible alternative. Preference tuning adds this contrast by comparing answers grounded in the relevant and in a mined negative document. Supporting documents rise further, to 50%, and the comparison adds 1.9 points on BRIGHT beyond continued answer prediction on the same data. Each stage also depends on the previous one: SFT from the base models reaches 20.1 NDCG@10, against 29.8 from the pre-trained pair.

Why not direct relevance supervision

We use the same relevance pairs, but through the language model’s loss, which measures how much a document helps produce the answer. With identical data and updates, encoder-only InfoNCE reaches 25.7 against 29.8 NDCG@10, and it falls below the pre-trained starting point on Biology and Earth Science, where it loses the pre-trained preference for complete passages that the language-model objective preserves (Appendix F.2). The three stages follow the same reasoning as language-model training, where each stage supplies a signal the previous one lacks (Ouyang et al., 2022).

We have stated this motivation in the introduction and added the analysis to Appendix H.

Comparison with stronger baselines (ReasonEmbed)

We have evaluated ReasonEmbed under our harness and have added this comparison (Section 4.2, Appendix D.2).

Our main contribution is a training signal applied across three training stages, and we control the training data. ReasonEmbed’s gains come largely from data. It synthesizes 81.7K training queries from BRIGHT’s own corpora as said in their paper (“we directly leverage the 12 datasets included in BRIGHT”), with BRIGHT’s task instructions, covering all 12 evaluation domains, and its authors acknowledge that their study is “currently limited to domains covered by existing benchmarks, particularly BRIGHT.” Its own ablation reflects this: its proposed loss adds only +1.3 points on BRIGHT over standard InfoNCE on the same data. We instead train on ReasonIR’s (Shao et al., 2025) varied-length (VL) split, whose queries and documents are LLM-generated from brainstormed task types and not drawn from BRIGHT’s domains, and we deliberately exclude ReasonIR’s HQ split, which is seeded from BRIGHT documents. Even so, without any training data (queries or documents) synthesized from BRIGHT’s corpora, Revela+ trails ReasonEmbed on the same Llama-3.1-8B backbone by only 4.0 points (32.1 vs. 36.1) and outperforms it on Earth Science (56.4 vs. 56.0), LeetCode (32.1 vs. 28.8) and AoPS (13.4 vs. 9.1). The fair comparisons for our signal are on shared data. With matched data, batches and updates, our signal outperforms InfoNCE by +4.1 points (29.8 vs. 25.7), and Revela+-8B outperforms ReasonIR and RaDeR (Das et al., 2025), the contrastive methods whose data we use, by +7.7 and +6.1 points (32.1 vs. 24.4 and 26.0).

We hope these clarifications and the additional analyses address the reviewer’s concerns, and we would be glad to answer any further questions.

References

Jianlyu Chen, Junwei Lan, Chaofan Li, Defu Lian, and Zheng Liu. 2026. ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1203–1221.

Debrup Das, Sam O’Nuallain, and Razieh Rahimi. 2025. RaDeR: Reasoning-aware Dense Retrieval Models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 19970–19997.

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training Language Models to Follow Instructions with Human Feedback. In Advances in Neural Information Processing Systems 35 (NeurIPS 2022).

Rulin Shao, Rui Qiao, Varsha Kishore, Niklas Muennighoff, Xi Victoria Lin, Daniela Rus, Bryan Kian Hsiang Low, Sewon Min, Wen-tau Yih, Pang Wei Koh, and Luke Zettlemoyer. 2025. ReasonIR: Training Retrievers for Reasoning Tasks. In Second Conference on Language Modeling (COLM 2025).

Thank you for the clarification and for the ReasonEmbed comparison. I agree that training data quality may explain much of the gap, but ReasonEmbed reports a higher BRIGHT score than Revela+ even with standard InfoNCE. How would Revela+ perform if trained on the same ReasonEmbed relevance pairs? This comparison would help clarify whether the gains over InfoNCE observed on your training data also hold on ReasonEmbed’s data.

We thank the reviewer for the follow-up and the suggestion. Training Revela+ on ReasonEmbed’s relevance pairs is not a drop-in change: our signal is trained on sequences rather than query–document pairs, so the data must first be adapted, which involves design choices that deserve careful study. This is not feasible to do properly within the remaining rebuttal period.

We note that Revela+ already trains on data curated by other groups specifically for contrastive training: ReasonIR and RaDeR were each trained on their own authors’ data and recipes, and Revela+ trained on these datasets outperforms both on BRIGHT (+7.7 and +6.1 points). ReasonEmbed builds on the same recipe as these baselines, LLM-synthesized reasoning-intensive queries with contrastive training, with more refined data curation. We will investigate training Revela+ on ReasonEmbed’s data and discuss it in the camera-ready version.

Thank you for the clarification. Whether the observed advantage over standard InfoNCE carries over to ReasonEmbed’s data remains an open question. My concern in W2 therefore remains, and I will keep my original rating. I view the paper as borderline and would not object to acceptance.

Reviewer_tGuLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Revela+. It extends Revela’s approach to training dense retrievers through language modeling. The training pipeline has three stages: pre-training, supervised fine-tuning, and preference tuning. The method uses weighted in-batch cross-sequence attention. Gradients from the language model’s training objective reach the retriever through the context weights. The authors evaluate five encoders from two model families. They report consistent gains in average retrieval performance across stages on BRIGHT and BEIR.

Critical strengths
  • Strong empirical results. The method achieves substantial retrieval gains. Its smaller models outperform the proprietary embedding models evaluated on BRIGHT.

  • Targeted ablation studies. The ablation studies demonstrate the benefits of each training stage.

  • A unified framework with practical deployment. All three stages use the same retriever–language model coupling mechanism for supervision. Only the retriever is needed at inference time. Document embeddings can be precomputed. The method therefore retains the standard dense retrieval workflow.

Critical weaknesses
  • The paper’s presentation needs improvement. Figure 1 caption lack sufficient detail. This makes the figures difficult to interpret and the method harder to follow.

  • The contrastive learning baseline is not sufficiently strong. The current comparison provides limited evidence of the method’s advantage over well-tuned contrastive learning. I suggest expanding the comparison to include stronger, well-tuned contrastive baselines.

  • Performance on Pony decreases after SFT for all five encoders. Performance on AoPS decreases after preference tuning for all five encoders. These consistent declines warrant a more systematic discussion.

Questions for authors

See weakness

6 comments

We thank the reviewer for the thoughtful review and for highlighting the strong empirical results, the targeted ablations, and the unified framework that keeps the standard dense-retrieval workflow at inference. We address each weakness below.

Presentation and Figure 1

We agree. We have rewritten the caption of Figure 1 tied the method section to it (Section 3) to explain each panel and its notation: the contrastive setup, the three Revela+ stages with the retriever’s weights over in-batch sequences, the shared learning mechanism, and the BRIGHT results.

Stronger contrastive baselines

Following this comment and related ones from other reviewers, we have added ReasonEmbed (Chen et al., 2026), a recent contrastively trained reasoning retriever that reports stronger BRIGHT results than the baselines in our paper, and evaluated it under our harness (Section 4.2, Appendix D.2).

Our main contribution is a training signal applied across three training stages, and we control the training data. ReasonEmbed’s gains come largely from data. It trains on documents from BRIGHT’s own corpora, with BRIGHT’s relevant documents filtered out, and synthesizes 81.7K training queries from them (“we directly leverage the 12 datasets included in BRIGHT”), with BRIGHT’s task instructions, covering all 12 evaluation domains, and its authors acknowledge that their study is “currently limited to domains covered by existing benchmarks, particularly BRIGHT.” Its own ablation reflects this: its proposed loss adds only +1.3 points on BRIGHT over standard InfoNCE on the same data. We instead train on ReasonIR’s (Shao et al., 2025) varied-length (VL) split, whose queries and documents are LLM-generated from brainstormed task types, and we deliberately exclude ReasonIR’s HQ split, which is seeded from BRIGHT documents. Even so, without any training data (queries or documents) synthesized from BRIGHT’s corpora, Revela+ trails ReasonEmbed on the same Llama-3.1-8B backbone by only 4.0 points (32.1 vs. 36.1) and outperforms it on Earth Science (56.4 vs. 56.0), LeetCode (32.1 vs. 28.8) and AoPS (13.4 vs. 9.1). The fair comparisons for our signal are on shared data. With matched data, batches and updates, our signal outperforms InfoNCE by +4.1 points (29.8 vs. 25.7), and Revela+-8B outperforms ReasonIR and RaDeR (Das et al., 2025), the contrastive methods whose data we use, by +7.7 and +6.1 points (32.1 vs. 24.4 and 26.0).

Performance on Pony and AoPS

These consistent declines prompted a per-query analysis of both tasks. The two declines have different causes.

Pony’s notion of relevance differs from the other coding tasks. Its corpus has 7,894 documents, yet all 112 queries share just 43 gold documents (19.8 per query), and one documentation page is gold for 109 of the 112 queries. Each query asks for a program in an unfamiliar language, so the gold documents are the language’s general syntax pages, needed whatever the problem is. In LeetCode, relevance is problem-specific: each of the 142 queries has 1.8 gold documents on average, drawn from a 414k-document corpus, and no document is gold for more than three queries. SFT teaches problem-specific relevance of the LeetCode kind: it improves the other eleven tasks, but on Pony it replaces the pre-trained encoders’ general preference for tutorial pages with topic-matched source-code chunks.

The AoPS decline is produced by the comparative term of preference tuning. On Llama-3.1-8B, removing it (control (E), β = 0) reduces the AoPS drop from −6.7 to −1.9 points, and a larger β enlarges it. The term has the pairwise form of a contrastive loss with one mined negative, mediated by the reference LM, so it strengthens whatever distinguishes d+ from d− in the preference data. In our data every pair contrasts theorem or definition statements, and the comparative term raises TheoremQA-Questions from 32.5 to 37.2 over continued answer prediction. AoPS relevance instead links problems that share a solution technique, and these become outranked by elementary problems on the same topic. To test this, we trained a variant whose preference data mixes the theorem-level pairs (one third) with general-domain pairs. AoPS recovers to 18.1 (from 13.4), while TheoremQA-Questions keeps most of its gain (36.0).

We have added these analyses and ablations to Appendix I and summarize them in Section 5.

We hope these clarifications and additional analyses address the reviewer’s concerns, and we are happy to answer any further questions.

References

Jianlyu Chen, Junwei Lan, Chaofan Li, Defu Lian, and Zheng Liu. 2026. ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1203–1221.

Rulin Shao, Rui Qiao, Varsha Kishore, Niklas Muennighoff, Xi Victoria Lin, Daniela Rus, Bryan Kian Hsiang Low, Sewon Min, Wen-tau Yih, Pang Wei Koh, and Luke Zettlemoyer. 2025. ReasonIR: Training Retrievers for Reasoning Tasks. In Second Conference on Language Modeling (COLM).

Debrup Das, Sam O’Nuallain, and Razieh Rahimi. 2025. RaDeR: Reasoning-aware Dense Retrieval Models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 19970–19997.

Thank you for the response. I remain concerned about the contribution of this work over Revela. The main extension appears to be applying Revela’s approach to the SFT stage. Could you briefly clarify what additional insights this work offers beyond Revela, apart from the performance gains achieved by extending its approach to two additional training stages? I am considering maintaining my borderline rating.

Reviewer_cjDAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Revela+, which jointly trains a retriever and a language model across all three stages: pre-training, supervised fine-tuning (SFT), and preference tuning. Using in-batch sequences weighted by the retriever, the model first predicts document text, then predicts answers to queries, and finally compares answers conditioned on relevant documents with those conditioned on mined negative documents. In contrast, Revela applies this approach only during pre-training.

Across five encoders from two model families, Revela+ generally improves average retrieval performance on BRIGHT and BEIR through SFT and preference tuning. For Llama-3.1-8B, the average NDCG@10 on BRIGHT increases from 25.1% after pre-training to 29.8% after SFT and 32.1% after preference tuning, outperforming the contrastively trained ReasonIR-8B.

Critical strengths
  1. Revela+ trains a dense retriever throughout the full language-model training pipeline: pre-training, SFT, and preference tuning, using only the language model’s objective at each stage.

  2. The authors provide ablation studies showing that each of the three training stages contributes performance gains. The final 8B retriever outperforms a dedicated, contrastively trained reasoning retriever while using approximately one-seventh of the training compute.

Critical weaknesses
  1. Results on CoIR are missing. In addition, Revela+-1B underperforms Revela-1B on BEIR, but the authors do not explain this discrepancy.

  2. The baselines are relatively weak, and comparisons with methods from 2026, such as ReasonEmbed and llama-nv-embed-reasoning-3b, are missing.

  3. InfoNCE reduces performance on Biology from 51.4% to 33.2% and on Earth Science from 48.6% to 30.7%, both below the pre-training baseline. In contrast, answer next-token prediction (NTP) achieves 53.2% and 56.0%, respectively. This is the paper’s strongest evidence that generative signals outperform contrastive signals, yet the authors discuss it in only one sentence. Why does the contrastive objective degrade previously acquired retrieval capabilities in knowledge-intensive domains?

  4. The authors should specify the exact models and versions used for the Cohere, OpenAI, Voyage, and Google baselines.

Questions for authors
  1. Could you include comparative results on CoIR? Revela was evaluated on this benchmark, but Revela+ is not.
  2. Could you add comparisons with recent methods, such as ReasonEmbed and llama-nv-embed-reasoning-3b?
  • Chen, J., Lan, J., Li, C., et al. “ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval.” Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026, pp. 1203–1221.
  • https://huggingface.co/nvidia/llama-nv-embed-reasoning-3b
4 comments

We thank the reviewer for the careful assessment and for recognising that Revela+ trains a dense retriever through the full language-model pipeline, with ablations showing that each stage contributes and an 8B retriever that outperforms a dedicated reasoning retriever at about one-seventh of the training compute. We address each comment below.

CoIR results

Revela+ is designed for reasoning-intensive retrieval, so we evaluate it on benchmarks built for that setting, as prior work in this space does [Shao et al., 2025; Sun et al., 2025; Chen et al., 2026]. CoIR primarily tests code search and code matching. That said, BRIGHT’s LeetCode and StackOverflow domains show how the model handles coding: every encoder improves from pre-training to preference tuning, by up to 13.6 NDCG@10 points on LeetCode and 8.5 on StackOverflow (+5.9 to +9.3 and +3.3 to +8.5 for the main 0.6B–8B encoders). A fair CoIR evaluation would also require curating a coding-focused training set and training Revela+ on it, which we could not do within the rebuttal period given the compute required. We will add these results to the camera-ready version.

BEIR at 1B

The gap reflects task overfitting and does not indicate a general weakness of Revela+. Revela+-1B trails Revela-1B by 1.6 BEIR points (41.1 vs. 42.7), and the gap comes from specific datasets such as HotpotQA (−17.2), which retrieves Wikipedia passages, the domain of Revela’s entire training corpus. Excluding HotpotQA alone, the two models are within 0.3 points on average, and Revela+ outperforms Revela on BEIR at the other sizes (0.27B, 0.6B, 3B).

InfoNCE on Biology and Earth Science

We analysed the saved rankings of the pre-trained, InfoNCE and Answer-NTP encoders on BRIGHT. InfoNCE’s drop on Biology and Earth Science comes from short fragments entering the top ranks. BRIGHT splits web pages into chunks, and in these two domains almost all chunks (89–98%) are fragments under 600 characters, while relevant passages are long. InfoNCE still ranks the relevant passage above other long passages, just as the pre-trained model does. However, it also places irrelevant fragments ahead of it: a median of two fragments in Biology and four in Earth Science are ranked above the first relevant passage, whereas the pre-trained model ranks none above it. As a result, the relevant passage drops out of the top 10 for 35–43% of queries, against 15–22% after pre-training. The effect holds when we compare queries whose rankings changed equally from pre-training: there, InfoNCE still admits 1.7 more fragments per query than Answer-NTP. With the same batches and updates, Answer-NTP largely keeps the pre-trained preference for complete passages, while InfoNCE loses it. We hypothesise that this reflects how contrastive training can suppress “previously well represented features” [Robinson et al., 2021; Chen et al., 2021], in line with dense retrievers’ known preference for “shorter documents” [Fayyaz et al., 2025]. We have added this analysis to Appendix F.2.

Stronger baselines

Thanks for the references. We have evaluated ReasonEmbed [Chen et al., 2026] under our harness and have added this comparison (Section 4.2, Appendix D.2).

Our main contribution is a training signal applied across three training stages, and we control the training data. ReasonEmbed’s gains come largely from data. It trains on documents from BRIGHT’s own corpora, with BRIGHT’s pos documents filtered out, and synthesizes 81.7K training queries from them (“we directly leverage the 12 datasets included in BRIGHT”), with BRIGHT’s task instructions, covering all 12 evaluation domains, and its authors acknowledge that their study is “currently limited to domains covered by existing benchmarks, particularly BRIGHT.” Its own ablation reflects this: its proposed loss adds only +1.3 points on BRIGHT over standard InfoNCE on the same data. We instead train on the varied-length (VL) split of ReasonIR [Shao et al., 2025], whose queries and documents are LLM-generated from brainstormed task types, and we deliberately exclude ReasonIR’s HQ split, which is seeded from BRIGHT documents. Even so, without any training documents or queries synthesized from BRIGHT’s corpora, Revela+ trails ReasonEmbed on the same Llama-3.1-8B backbone by only 4.0 points (32.1 vs. 36.1) and outperforms it on Earth Science (56.4 vs. 56.0), LeetCode (32.1 vs. 28.8) and AoPS (13.4 vs. 9.1). llama-nv-embed-reasoning-3b scales ReasonEmbed’s approach further: it synthesizes additional queries over BRIGHT’s corpora and also trains on the ReasonEmbed, ReasonRank, and ReasonAug datasets. The fair comparisons for our signal are on shared data. With matched data, batches and updates, our signal outperforms InfoNCE by +4.1 points (29.8 vs. 25.7), and Revela+-8B outperforms ReasonIR and RaDeR [Das et al., 2025], the contrastive methods whose data we use, by +7.7 and +6.1 points (32.1 vs. 24.4 and 26.0).

Versions of the proprietary baselines

The proprietary baseline scores are those reported by BRIGHT [Su et al., 2025]: Cohere embed-english-v3.0, OpenAI text-embedding-3-large, Voyage voyage-large-2-instruct, and Google text-embedding-preview-0409 (768 dimensions). We now list these versions in Table 4 (Appendix D.1).

We hope these clarifications and additional analyses address the reviewer’s concerns, and we are happy to answer any further questions.

References

Jianlyu Chen, Junwei Lan, Chaofan Li, Defu Lian, and Zheng Liu. 2026. ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1203–1221.

Ting Chen, Calvin Luo, and Lala Li. 2021. Intriguing Properties of Contrastive Losses. In Advances in Neural Information Processing Systems 34 (NeurIPS 2021).

Debrup Das, Sam O’Nuallain, and Razieh Rahimi. 2025. RaDeR: Reasoning-aware Dense Retrieval Models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 19970–19997.

Mohsen Fayyaz, Ali Modarressi, Hinrich Schuetze, and Nanyun Peng. 2025. Collapse of Dense Retrievers: Short, Early, and Literal Biases Outranking Factual Evidence. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025).

Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. 2021. Can Contrastive Learning Avoid Shortcut Solutions? In Advances in Neural Information Processing Systems 34 (NeurIPS 2021).

Rulin Shao, Rui Qiao, Varsha Kishore, Niklas Muennighoff, Xi Victoria Lin, Daniela Rus, Bryan Kian Hsiang Low, Sewon Min, Wen-tau Yih, Pang Wei Koh, and Luke Zettlemoyer. 2025. ReasonIR: Training Retrievers for Reasoning Tasks. In Second Conference on Language Modeling (COLM 2025).

Hongjin Su, Howard Yen, Mengzhou Xia, Weijia Shi, Niklas Muennighoff, Han-yu Wang, Haisu Liu, Quan Shi, Zachary S. Siegel, Michael Tang, Ruoxi Sun, Jinsung Yoon, Sercan O. Arik, Danqi Chen, and Tao Yu. 2025. BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval. In The Thirteenth International Conference on Learning Representations (ICLR 2025).

Duolin Sun, Meixiu Long, Dan Yang, Junjie Wang, Yecheng Luo, Yue Shen, Jian Wang, Hualei Zhou, Chunxiao Guo, Peng Wei, Jiahai Wang, and Jinjie Gu. 2025. DIVER: A Multi-Stage Approach for Reasoning-intensive Information Retrieval. arXiv preprint arXiv:2508.07995.

Thank you to the authors for their response. I will maintain my original score.

Thank you for your response. Please let us know if there are any outstanding concerns that we can help clarify.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.