Paper deep dive
RECIPER: A Dual-View Retrieval Pipeline for Procedure-Oriented Materials Question Answering
Zhuoyu Wu, Wenhui Ou, Pei-Sze Tan, Wenqi Fang, Sailaja Rajanala, Raphaël C. -W. Phan
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 99%
Last extracted: 4/14/2026, 2:32:11 AM
Summary
RECIPER is a dual-view retrieval pipeline designed for procedure-oriented materials science question answering. It improves retrieval performance by indexing both paragraph-level context and LLM-extracted procedural summaries, integrating them through a lightweight lexical reranking mechanism. The framework consistently outperforms paragraph-only dense retrieval across various backbones and LLMs, enhancing evidence selection and downstream answer quality.
Entities (4)
Relation Signals (3)
RECIPER → improves → Materials Science Question Answering
confidence 100% · RECIPER consistently improves early-rank retrieval over paragraph-only dense retrieval
RECIPER → uses → DeepSeek-R1 Distill Qwen-32B
confidence 100% · we construct a procedure-centric representation for each paper using an instruction-following LLM, DeepSeek-R1-Distill-Qwen-32B.
RECIPER → utilizes → BGE-large-en-v1.5
confidence 100% · With BGE-large-en-v1.5, it reaches 86.82%, 97.07%, and 97.85% on Recall@1, Recall@5, and Recall@10
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Retrieving procedure-oriented evidence from materials science papers is difficult because key synthesis details are often scattered across long, context-heavy documents and are not well captured by paragraph-only dense retrieval. We present RECIPER, a dual-view retrieval pipeline that indexes both paragraph-level context and compact large language model-extracted procedural summaries, then combines the two candidate streams with lightweight lexical reranking. Across four dense retrieval backbones, RECIPER consistently improves early-rank retrieval over paragraph-only dense retrieval, achieving average gains of +3.73 in Recall@1, +2.85 in nDCG@10, and +3.13 in MRR. With BGE-large-en-v1.5, it reaches 86.82%, 97.07%, and 97.85% on Recall@1, Recall@5, and Recall@10, respectively. We further observe improved downstream question answering under automatic metrics, suggesting that procedural summaries can serve as a useful complementary retrieval signal for procedure-oriented materials question answering. Code and data are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2604.11229v1
- Canonical: https://arxiv.org/abs/2604.11229v1
Trouble viewing inline? Open PDF directly →
Full Text
21,834 characters extracted from source content.
Expand or collapse full text
RECIPER: A Dual-View Retrieval Pipeline for Procedure-Oriented Materials Question Answering Abstract Retrieving procedure-oriented evidence from materials science papers is difficult because key synthesis details are often scattered across long, context-heavy documents and are not well captured by paragraph-only dense retrieval. We present RECIPER, a dual-view retrieval pipeline that indexes both paragraph-level context and compact LLM-extracted procedural summaries, then combines the two candidate streams with lightweight lexical reranking. Across four dense retrieval backbones, RECIPER consistently improves early-rank retrieval over paragraph-only dense retrieval, achieving average gains of +3.73 in Recall@1, +2.85 in nDCG@10, and +3.13 in MRR. With BGE-large-en-v1.5, it reaches 86.82%, 97.07%, and 97.85% on Recall@1, Recall@5, and Recall@10 respectively. We further observe improved downstream QA under automatic metrics, suggesting that procedural summaries can serve as a useful complementary retrieval signal for procedure-oriented materials QA. Code and data are available at https://github.com/ReaganWu/RECIPER. Index Terms— Materials science retrieval, Scientific question answering, Retrieval-augmented generation Fig. 1: Overview of RECIPER Framework. In this framework, the user query is first transformed into feature vectors, which are then used to retrieve the most relevant Paragraph and Recipe embedding vectors from the Chunk Vector Database. Recipe vectors pass through a Threshold Screener to filter out highly similar entries and increase content diversity. The filtered Recipes are then combined with the Paragraphs in a Rule-based Re-rank module and merged with the original query before being fed into the LLM, producing a precise, well-grounded answer. 1 Introduction Large-scale scientific literature contains rich domain-specific knowledge, including experimental procedures, synthesis workflows, and contextual descriptions of materials [22]. However, locating such information remains labor-intensive, as key procedural details are often buried in context-heavy documents. Although large language models (LLMs) enable interactive question answering (QA) [15], they remain unreliable for fine-grained scientific queries and may produce hallucinated responses due to their static training data [21]. Retrieval-Augmented Generation (RAG) improves reliability by retrieving supporting evidence before answer generation. Nevertheless, retrieval in materials science remains challenging. First, many questions require precise synthesis steps or material properties, whereas standard dense retrieval mainly returns unstructured text chunks, making such details difficult to identify [2, 1]. Second, procedural knowledge is often distributed across multiple interdependent sections, while chunk-level retrieval breaks these connections [12]. Third, existing approaches have explored expert priors [1], structured representations [12], and summary-based signals [7]. However, it remains unclear whether compact procedural abstractions can serve as an effective auxiliary retrieval view for procedure-oriented materials QA. To address these limitations, we propose RECIPER, a recipe-enhanced dual-view retrieval framework that treats procedural knowledge as a complementary retrieval signal. RECIPER represents each paper using two views: a Recipe view, which encodes compact step-level procedural summaries, and a paragraph view, which preserves broader contextual evidence. These two views are jointly retrieved and integrated within a unified ranking pipeline, enabling more effective combination of procedural and contextual signals. • We introduce a dual-view retrieval pipeline that combines paragraph-level context with LLM-extracted procedural summaries for materials literature retrieval. • We show empirically that procedural summaries are weak as standalone retrieval units but provide complementary signals when combined with paragraph retrieval. • We demonstrate consistent gains across four dense backbones, with average improvements of +3.73 in Recall@1, +2.85 in nDCG@10, and +3.13 in MRR over paragraph-only dense retrieval, indicating that RECIPER provides a robust and backbone-agnostic improvement for scientific retrieval. 2 Methodology We propose RECIPER, a procedure-aware dual-view retrieval framework for materials science question answering. The central idea is to represent each paper from two complementary views: (1) a contextual view composed of paragraph-level text chunks, and (2) a procedural view composed of compact LLM-extracted procedural summaries. Given a user query, RECIPER retrieves candidates from both views, merges them into a unified candidate pool, and applies a lightweight query-aware reranking step to prioritize evidence that is both semantically relevant and lexically aligned with the query. Figure 1 shows the overall framework. 2.1 Procedure-Centric Knowledge Extraction Scientific papers in materials science often describe synthesis workflows in long, dispersed, and context-heavy paragraphs, making direct retrieval inefficient for procedure-oriented questions. To expose this procedural signal more explicitly, we construct a procedure-centric representation for each paper using an instruction-following LLM, DeepSeek-R1-Distill-Qwen-32B. For each document, the LLM generates a compact procedural summary from the full text. Each summary is formatted as a compact step-oriented description covering materials, operations, and conditions. Compared with raw paragraphs, these summaries compress long-form methodological text into a more retrieval-friendly form while preserving the major procedural cues needed for downstream question answering. We denote the set of paragraph chunks as P and the set of procedure-centric summaries as ℛR. These summaries are not intended to replace paragraph evidence, but to expose procedural cues in a more retrieval-friendly form. 2.2 Dual-View Candidate Retrieval Given a query q, RECIPER retrieves candidates independently from the two views. The paragraph view provides broad contextual evidence, while the procedural view emphasizes condensed synthesis logic and experimentally relevant operations. Let q denote the query embedding, and let ie_i denote the embedding of a candidate item from either view. We compute the base retrieval score as si=11+d(,i),s_i= 11+d(q,e_i), (1) where d(⋅,⋅)d(·,·) is the embedding-space distance. Using this scoring function, we retrieve the top-KcK_c paragraph candidates q=TopKKc(q,),P_q=TopK_K_c(q,P), (2) and the top-KcK_c procedural candidates ℛq=TopKKc(q,ℛ).R_q=TopK_K_c(q,R). (3) The two candidate sets provide complementary evidence: paragraph candidates tend to preserve narrative and descriptive context, whereas procedural candidates more directly capture synthesis-oriented information. 2.3 Candidate Merging and Stream-Aware Deduplication After dual-view retrieval, we merge the two candidate lists into a single pool q(0)=q∪ℛq.C_q^(0)=P_q _q. (4) Since multiple candidates may originate from the same paper and the same retrieval stream, we perform stream-aware deduplication to reduce redundant evidence while preserving cross-view complementarity. Specifically, for each (paper_id,stream)(paper\_id,stream) pair, we keep only the highest-ranked candidate. This gives the filtered candidate pool q=Dedupe(q(0)).C_q=Dedupe(C_q^(0)). (5) This design intentionally preserves cross-view complementarity while preventing within-stream redundancy from dominating the final candidate pool. 2.4 Query-Aware Lexical Reranking Dense retrieval is effective for coarse semantic matching, but within a high-quality candidate pool, semantically similar candidates may still differ in how directly they address the query. To refine the ranking, we introduce a lightweight query-aware lexical reranking step. Let Q be the token set of the query, and let DiD_i denote the lexical evidence of candidate cic_i, constructed from both its title and body text, where the title provides a compact topic cue, and the body text provides local content evidence: Q=Tokenize(q),Di=Tokenize(titlei⊕texti),Q=Tokenize(q), D_i=Tokenize(title_i _i), (6) where ⊕ denotes string concatenation. We define the query-coverage score of candidate cic_i as oi=|Q∩Di|max(|Q|,1).o_i= |Q∩ D_i| (|Q|,1). (7) The final reranked score is then computed as s^i=si+λoi, s_i=s_i+λ o_i, (8) where a small constant λ (= 0.1) is controlling the strength of lexical adjustment. This reranking step is intentionally lightweight. It preserves the main semantic ordering induced by dense retrieval, while promoting candidates that explicitly cover a larger fraction of the query terms. 2.5 Evidence Selection for Downstream QA Finally, all candidates in qC_q are sorted by s^i s_i, and the top-K items are selected as the evidence context ℰq=c1,c2,…,cK.E_q=\c_1,c_2,…,c_K\. (9) The selected evidence is then passed to a downstream large language model for answer generation. This setup allows us to examine whether the proposed retrieval pipeline improves evidence selection and downstream answer quality under automatic metrics. 3 Experiments 3.1 Experimental Setup We evaluate RECIPER on a materials-science QA benchmark built from 300+ research articles collected from public sources (e.g., arXiv and Semantic Scholar). Each paper is paired with GPT-5.3-generated question-answer instances and linked to its source document, yielding 1,024 query-document pairs for retrieval evaluation. The benchmark emphasizes synthesis-oriented questions involving procedures, material properties, and characteristic behaviors. For retrieval, we index both paragraph chunks and procedure-centric summaries using dense embeddings. Unless otherwise stated, the main results use BGE-large-en-v1.5; we further test all-MiniLM-L6-v2, Contriever, and E5-large-v2 to assess backbone robustness. We report Recall@K (K=1,5,10K=1,5,10), nDCG@10, and MRR. For downstream QA, retrieved evidence is fed into multiple LLMs ranging from 0.5B to 40B parameters, and we report BERTScore-F1, ROUGE-L, cosine similarity, and BLEURT. Group System R@1 R@5 R@10 nDCG@10 MRR External Paragraph Baselines BM25 [10] 0.6172 0.8066 0.8477 0.7335 0.6967 all-MiniLM-L6-v2 [17] 0.7432 0.9102 0.9307 0.8438 0.8150 Contriever [5] 0.7793 0.9131 0.9375 0.8615 0.8367 BGE-large-en-v1.5 [18] 0.8408 0.9512 0.9619 0.9061 0.8875 E5-large-v2 [16] 0.8477 0.9561 0.9717 0.9136 0.8945 BM25 + BGE-large-en-v1.5 0.7549 0.9443 0.9629 0.8665 0.8345 Recipe / Fusion Ablations (BGE backbone) Dense (Paragraph) [6] 0.8408 0.9512 0.9619 0.9060 0.8875 Rerank (Paragraph) [11] 0.8604 0.9570 0.9619 0.9161 0.9007 Dense (Recipe) [6] 0.5107 0.6299 0.6533 0.5837 0.5610 Hybrid (Recipe+Paragraph) [13] 0.8486 0.9658 0.9795 0.9181 0.8979 Hybrid + RRF (Recipe+Paragraph) [9] 0.7754 0.9619 0.9707 0.8815 0.8517 Rerank (Recipe + Paragraph) [11] 0.5703 0.8887 0.9521 0.7634 0.7024 RECIPER (Ours) 0.8682 0.9707 0.9785 0.9283 0.9116 Table 1: Retrieval performance comparison and ablation study. The upper block reports paragraph-based baselines, while the lower block analyzes recipe-based and dual-view variants under the BGE backbone. Recipe-only retrieval is weak, but combining it with paragraph retrieval improves performance, showing that procedural and contextual signals are complementary. RECIPER achieves the best overall results, especially on early-rank metrics, indicating more effective integration of procedural and contextual evidence. Backbone vs Paragraph vs Hybrid Δ @1 Δ @10 Δ @1 Δ @10 MiniLM +0.0468 +0.0344 +0.0273 +0.0151 Contriever +0.0361 +0.0332 +0.0263 +0.0186 BGE +0.0274 +0.0223 +0.0196 +0.0102 E5 +0.0390 +0.0242 +0.0293 +0.0138 Average +0.0373 +0.0285 +0.0256 +0.0144 Table 2: Cross-backbone improvement of RECIPER over paragraph-only dense retrieval and naive hybrid fusion. RECIPER consistently improves early-rank accuracy across all embedding backbones. Model M BERT-F1 R-L Cos BLT GPT-5 [8] ◆ 0.8612 0.2387 0.7745 0.3787 ∙ 0.8552 0.2137 0.7592 0.3747 ⋆ 0.8465 0.2035 0.7350 0.3794 Deepseek-32B [4] ◆ 0.8662 0.2517 0.7379 0.3604 ∙ 0.8646 0.2428 0.7430 0.3555 ⋆ 0.8520 0.1945 0.7030 0.3366 Llama-3.1-8B [3] ◆ 0.8627 0.2552 0.7385 0.3615 ∙ 0.8601 0.2365 0.7395 0.3598 ⋆ 0.8418 0.1997 0.6668 0.3602 Qwen-2.5-7B [14] ◆ 0.8677 0.2752 0.7671 0.3799 ∙ 0.8655 0.2579 0.7656 0.3703 ⋆ 0.8349 0.1799 0.6377 0.3925 Qwen-3-4B [14] ◆ 0.8569 0.2232 0.7312 0.3688 ∙ 0.8553 0.2230 0.7359 0.3707 ⋆ 0.8339 0.1721 0.6570 0.3532 Llama-3.2-3B [3] ◆ 0.8487 0.2278 0.7168 0.3719 ∙ 0.8476 0.2121 0.7197 0.3651 ⋆ 0.8409 0.2001 0.6828 0.3701 Qwen-3-1.7B [20] ◆ 0.8408 0.2236 0.6802 0.4362 ∙ 0.8341 0.2175 0.6606 0.4139 ⋆ 0.8249 0.1818 0.6283 0.4109 Vibe-1.5B [19] ◆ 0.7626 0.0930 0.3361 0.2725 ∙ 0.7540 0.0809 0.2999 0.2536 ⋆ 0.7010 0.0095 0.0016 0.1986 Qwen2.5-0.5B [14] ◆ 0.8524 0.2124 0.7532 0.3920 ∙ 0.8491 0.1966 0.7420 0.3865 ⋆ 0.8328 0.1600 0.6772 0.3844 Table 3: Overall QA performance across models and retrieval modes. Symbols indicate retrieval mode: ◆ RECIPER, ∙ Paragraph-Dense RAG, ⋆ NoRAG. Metrics are BERT-F1 (F1 score of BERTScore), R-L (ROUGE-L), Cos (Cosine similarity), and BLT (BLEURT). The best value per metric is highlighted with a yellow block and in bold font. 3.2 Retrieval Results Tables 1 and 2 summarize the retrieval results. Table 1 reports the main comparison and ablations under the BGE backbone, while Table 2 shows cross-backbone gains. Three findings are clear from Table 1. First, the procedural view alone is much weaker than paragraph-only dense retrieval, indicating that compact recipe-style summaries are insufficient as a standalone retrieval space. Second, naive dual-view fusion already improves over paragraph-only retrieval, confirming that contextual and procedural signals are complementary. Third, RECIPER further improves over naive hybrid fusion, showing that the gain comes from more effective integration of the two views. With the BGE backbone, RECIPER improves Recall@1 from 0.8408 to 0.8682 over paragraph-only retrieval and from 0.8486 to 0.8682 over naive hybrid fusion, while also achieving the best nDCG@10 (0.9283) and MRR (0.9116). The improvement is most pronounced on early-rank metrics, indicating better top-evidence selection. Table 2 shows that this trend is consistent across all four dense encoders. On average, RECIPER improves over paragraph-only retrieval by +3.73 points in Recall@1, +2.85 in nDCG@10, and +3.13 in MRR; compared with naive hybrid fusion, it still gains +2.56, +1.44, and +1.87 points, respectively. These results suggest that RECIPER is not tied to a specific embedding model, but offers a generally effective way to integrate contextual and recipe-based procedural retrieval signals. 3.3 Transfer to Downstream QA We further test whether improved retrieval quality translates into better answer generation. Across LLMs from 0.5B to 40B parameters, RECIPER consistently outperforms both with NoRAG and Paragraph-Dense RAG on most metrics. The improvement is most visible on ROUGE-L and BERT-F1, suggesting that better evidence selection leads to more grounded answers. The effect is especially clear for smaller models, indicating that stronger retrieval can partially compensate for limited parametric knowledge. Using Qwen-2.5-7B as an example, RECIPER improves ROUGE-L from 0.2579 to 0.2752 and BERT-F1 from 0.8655 to 0.8677 over Paragraph-Dense RAG. Similar trends are observed across the model spectrum, supporting that the retrieval design is architecture-agnostic and mainly benefits evidence quality rather than any specific generator. 4 Conclusion In this work, we introduced RECIPER, a dual-view retrieval framework that integrates structured procedural knowledge with paragraph-level evidence for materials-science QA. Across eight LLMs ranging from 0.5B to 40B parameters, RECIPER consistently outperforms both No-RAG and paragraph-only baselines, achieving higher BERTScore, ROUGE-L, BLEURT, and semantic similarity. Our results show that recipe-based procedural representations complement dense retrieval by providing property- and step-level signals, with particularly strong benefits for smaller models. These findings indicate that RECIPER offers a robust, architecture-agnostic retrieval improvement and provides a scalable foundation for scientific QA and knowledge extraction from complex materials literature. References [1] M. Ansari and S. M. Moosavi (2024) Agent-based learning of materials datasets from the scientific literature. Digital Discovery 3 (12), p. 2607–2617. Cited by: §1. [2] M. J. Buehler (2024) Generative retrieval-augmented ontologic graph and multiagent strategies for interpretive large language model-based materials design. ACS Engineering Au 4 (2), p. 241–277. Cited by: §1. [3] A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al. (2024) The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Cited by: Table 3, Table 3. [4] D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. (2025) Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Cited by: Table 3. [5] G. Izacard, M. Caron, L. Hosseini, S. Riedel, P. Bojanowski, A. Joulin, and E. Grave (2021) Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118. Cited by: Table 1. [6] P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, et al. (2020) Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, p. 9459–9474. Cited by: Table 1, Table 1. [7] R. Mostafa, M. N. Baig, M. T. Ehsan, and J. Hasan (2024) G-rag: knowledge expansion in material science. arXiv preprint arXiv:2411.14592. Cited by: §1. [8] OpenAI (2025) GPT-5. Note: https://platform.openai.comAccessed: 2025-01-10 Cited by: Table 3. [9] Z. Rackauckas (2024) Rag-fusion: a new take on retrieval-augmented generation. arXiv preprint arXiv:2402.03367. Cited by: Table 1. [10] S. Robertson and H. Zaragoza (2009) The probabilistic relevance framework: bm25 and beyond. Vol. 4, Now Publishers Inc. Cited by: Table 1. [11] D. Sachan, M. Lewis, M. Joshi, A. Aghajanyan, W. Yih, J. Pineau, and L. Zettlemoyer (2022) Improving passage retrieval with zero-shot question generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, p. 3781–3797. Cited by: Table 1, Table 1. [12] M. Scheidgen, L. Himanen, A. N. Ladines, D. Sikter, M. Nakhaee, Á. Fekete, T. Chang, A. Golparvar, J. A. Márquez, S. Brockhauser, et al. (2023) NOMAD: a distributed web-based platform for managing materials science research data. Journal of Open Source Software 8 (90), p. 5388. Cited by: §1. [13] W. Takahara, Y. Yamaguchi, M. Ogano, F. Kakami, Y. Harashima, T. Takayama, S. Takasuka, A. Kudo, and M. Fujii (2025) Materials dual-source knowledge retrieval-augmented generation for local large language models in photocatalysts. Journal of Chemical Information and Modeling 65 (24), p. 13098–13114. Cited by: Table 1. [14] Q. Team et al. (2025) Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. Cited by: Table 3, Table 3, Table 3. [15] M. Van, P. Verma, C. Zhao, and X. Wu (2025) A survey of ai for materials science: foundation models, llm agents, datasets, and tools. arXiv preprint arXiv:2506.20743. Cited by: §1. [16] L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, D. Jiang, R. Majumder, and F. Wei (2022) Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533. Cited by: Table 1. [17] W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou (2020) Minilm: deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in neural information processing systems 33, p. 5776–5788. Cited by: Table 1. [18] S. Xiao, Z. Liu, P. Zhang, and N. Muennighoff (2023) C-pack: packaged resources to advance general chinese embedding. External Links: 2309.07597 Cited by: Table 1. [19] S. Xu, Y. Zhou, W. Wang, J. Min, Z. Yin, Y. Dai, S. Liu, L. Pang, Y. Chen, and J. Zhang (2025) Tiny model, big logic: diversity-driven optimization elicits large-model reasoning ability in vibethinker-1.5 b. arXiv preprint arXiv:2511.06221. Cited by: Table 3. [20] A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, others, Q. Team, et al. (2025) Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: Table 3. [21] M. Zaki, N. Krishnan, et al. (2023) Mascqa: a question answering dataset for investigating materials science knowledge of large language models. arXiv preprint arXiv:2308.09115. Cited by: §1. [22] Y. Zimmermann, A. Bazgir, A. Al-Feghali, M. Ansari, J. Bocarsly, L. C. Brinson, Y. Chiang, D. Circi, M. Chiu, N. Daelman, et al. (2025) 34 examples of llm applications in materials science and chemistry: towards automation, assistants, agents, and accelerated scientific discovery. arXiv preprint arXiv:2505.03049. Cited by: §1.