Paper deep dive
From Papers to Mechanisms: An Evidence-Grounded Knowledge Substrate for Scientific Language Models
Qiuhui Chen, Yibo Liu, Tao Dai, Jiafan Lu, Zhenglei Zhou, Weimin Zhong
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Scientific language models often access literature through untyped text chunks, which fragment the functional and evidential structure required for mechanism-rich questions. We introduce an evidence-grounded mechanism knowledge substrate that organizes scientific literature into provenance-linked evidence units, role-typed entities, and directed mechanism paths. We instantiate it as MS$^3$, a Material-Sensor-Signal-System schema for conductive-fiber flexible sensors, over 13,689 papers, 131,083 evidence items, and 26,648 mechanism objects. On in-domain and coverage-shift question-answering benchmarks, we compare closed-book generation, Web search, Raw-PDF RAG, and MS$^3$ retrieval across ten language models. MS$^3$ improves macro-averaged scientific correctness. It also improves citation entailment and answer completeness. These results support mechanism substrates as a reliable representation layer for scientific language models and motivate a source-repair workflow in which insufficient MS$^3$ evidence triggers targeted retrieval from its linked papers rather than assuming that a user has already supplied the correct PDFs.
Tags
Links
- Source: https://arxiv.org/abs/2610.06248v1
- Canonical: https://arxiv.org/abs/2610.06248v1
Trouble viewing inline? Open PDF directly →
Full Text
36,357 characters extracted from source content.
Expand or collapse full text
From Papers to Mechanisms: An Evidence-Grounded Knowledge Substrate for Scientific Language Models Qiuhui Chen 1 , Yibo Liu 2 , Tao Dai 1 , Jiafan Lu 1 , Zhenglei Zhou 3 , Weimin Zhong 1,* 1 School of Information Science and Engineering, East China University of Science and Technology 2 School of Computer Science, Shanghai Jiao Tong University 3 Tencent Abstract Scientific language models often access literature through un- typed text chunks, which fragment the functional and eviden- tial structure required for mechanism-rich questions. We intro- duce an evidence-grounded mechanism knowledge substrate that organizes scientific literature into provenance-linked ev- idence units, role-typed entities, and directed mechanism paths. We instantiate it as MS 3 , a Material–Sensor–Signal– System schema for conductive-fiber flexible sensors, over 13,689 papers, 131,083 evidence items, and 26,648 mech- anism objects. On in-domain and coverage-shift question- answering benchmarks, we compare closed-book generation, Web search, Raw-PDF RAG, and MS 3 retrieval across ten language models. MS 3 improves macro-averaged scientific correctness. It also improves citation entailment and answer completeness. These results support mechanism substrates as a reliable representation layer for scientific language models and motivate a source-repair workflow in which insufficient MS 3 evidence triggers targeted retrieval from its linked pa- pers rather than assuming that a user has already supplied the correct PDFs. Introduction Large language models are increasingly used to search, sum- marize, and reason over scientific literature (Asai et al. 2026; Skarlinski et al. 2024). Because their internal knowledge is incomplete and difficult to verify, retrieval-augmented gen- eration has become a common interface between language models and scientific corpora (Lewis et al. 2020; Lála et al. 2024). Most systems, however, represent papers as indepen- dently indexed text chunks. This representation is effective for document lookup and local factual questions, but is poorly matched to mechanism-oriented reasoning. Explaining how a material enables a system-level function may require con- necting its composition or fabrication, the resulting device, its measurable response, and the downstream application, with supporting evidence distributed across text, figures, cap- tions, and tables. Structured knowledge representations and agentic retrieval improve how scientific information is organized and accessed (Auer et al. 2018; Jaradeh et al. 2019; Edge et al. 2024; Lála et al. 2024; Skarlinski et al. 2024; Ding et al. 2026). Never- theless, existing systems often either abstract away the source ∗ Corresponding author. Figure 1: Evidence-grounded construction of an MS 3 mech- anism object. evidence supporting individual relations or continue to oper- ate over untyped document chunks. As a result, the language model must reconstruct functional structure at inference time while also determining which claims are supported by the re- trieved evidence. We introduce an evidence-grounded mechanism knowl- edge substrate as an intermediate representation between scientific documents and language models. It organizes liter- ature into provenance-linked evidence units, role-typed en- tities, directed relations, and complete or partial mechanism paths. The framework is schema-configurable: each domain defines its entity types, admissible relations, and structural arXiv:2610.06248v1 [cs.AI] 5 Oct 2026 constraints, while evidence extraction, provenance preserva- tion, mechanism construction, retrieval, and evidence pack- ing remain shared. Unsupported nodes and relations are left absent rather than completed from model priors, and every retained element is linked to its supporting source evidence. We instantiate the framework for conductive-fiber flexi- ble sensors using the Material–Sensor–Signal–System (MS 3 ) schema, Material→ Sensor→ Signal→ System. Material represents the constituent material or fabrication- related implementation, Sensor the functional sensing el- ement, Signal the measurable response, and System the demonstrated or proposed downstream use. To distinguish experimental evidence from envisioned applications, System records the demonstrated system, proposed application, val- idation setting, and deployment status separately. An MS 3 object may contain a complete four-layer path or only the subpath supported by the paper, with provenance linked to source passages and, when available, pages, sections, figures, captions, tables, and quantitative evidence. Using this framework, we construct an MS 3 substrate from 13,689 papers, containing 131,083 source-grounded evidence items and 26,648 complete or partial mechanism objects. At inference time, the system retrieves candidate pa- pers and selects evidence-linked mechanisms relevant to the query. We evaluate MS 3 across 10 language models. MS 3 substantially improves scientific correctness for nine of the ten models in blinded cross-family evaluation. Our main contributions are: • Evidence-grounded mechanism representation. We formulate a schema-configurable substrate that jointly represents source evidence, functional roles, directed re- lations, partial or complete mechanism paths, and prove- nance. • Large-scale MS 3 instantiation. We construct an evidence-linked substrate for conductive-fiber flexible sensors covering 13,689 papers and 26,648 mechanism objects. • Mechanism-aware retrieval interface. We develop de- terministic lexical retrieval and provenance-aware top- k evidence selection that expose structured, source- verifiable mechanism contexts to heterogeneous language models. • Cross-model evaluation. We evaluate the substrate across 10 language models using paired uncertainty esti- mates and blinded cross-family judging. Related Work Scientific literature retrieval and synthesis. Retrieval- augmented generation (RAG) augments language models with non-parametric corpora and has become a standard approach for knowledge-intensive generation (Lewis et al. 2020). Scientific variants increasingly combine retrieval with multi-stage language-model workflows. PaperQA retrieves and evaluates passages from full-text papers before gener- ating cited answers (Lála et al. 2024), while PaperQA2 ex- tends this paradigm to literature search, synthesis, and con- tradiction detection (Skarlinski et al. 2024). OpenScholar couples large-scale passage retrieval with citation-backed, multi-paper synthesis (Asai et al. 2026), and SciRAG in- troduces adaptive retrieval, citation-aware reasoning, and outline-guided generation (Ding et al. 2026). These sys- tems substantially improve access to scientific literature, but their retrieval units typically remain documents, passages, or citation-linked neighborhoods. Consequently, the language model must still infer the functional roles of different evi- dence pieces and determine whether they form a coherent mechanism. Our work is complementary to advances in re- trieval and agent orchestration: we focus on the representation layer and construct a persistent, provenance-linked mecha- nism substrate that can be queried by heterogeneous language models. Scientific information extraction and knowledge graphs. Scientific information extraction seeks to convert papers into entities and relations. SciERC jointly models enti- ties, relations, and coreference in scientific abstracts (Luan et al. 2018), and SciREX extends scientific extraction to document-level salient entities and n-ary relations (Jain et al. 2020). More recent schema-driven approaches such as SPIRES use language models to populate user-defined knowledge structures (Caufield et al. 2024). In parallel, scien- tific knowledge graph efforts, including the Science Knowl- edge Graph and the Open Research Knowledge Graph, rep- resent scholarly entities, contributions, and comparisons in machine-actionable form (Auer et al. 2018; Jaradeh et al. 2019). Domain-specific resources further demonstrate the value of structured representations; for example, MatKG or- ganizes materials, properties, synthesis methods, character- ization methods, and applications extracted from materials- science literature (Venugopal and Olivetti 2024). Graph- based RAG methods such as GraphRAG and G-Retriever retrieve graph communities or query-relevant subgraphs to support summarization and question answering (Edge et al. 2024; He et al. 2024). Existing lines of work, however, do not generally provide all of source-level evidence, domain- specific functional roles, directed mechanism relations, ex- plicit partial-path uncertainty, and a common interface for downstream language models. Our substrate combines these elements by treating evidence-grounded mechanism paths, rather than isolated triples or document chunks, as the pri- mary knowledge unit. Evidence-grounded scientific question answering and verification. Evidence attribution is central to reliable sci- entific reasoning. QASPER pairs full-paper questions with answers and supporting evidence spans (Dasigi et al. 2021). SciFact formulates scientific claim verification as retrieving evidence that supports or refutes a claim and identifying the corresponding rationales (Wadden et al. 2020); SciFact- Open extends this setting to large-corpus retrieval and ex- poses challenges such as evidence specificity and conflicting findings (Wadden et al. 2022). SciVer further evaluates claim verification over multimodal scientific evidence, including text, tables, and figures (Wang et al. 2025). These bench- marks provide important task-level supervision and evalua- tion, but evidence is typically attached to individual questions or claims rather than organized into a reusable scientific rep- resentation. In contrast, our framework incorporates prove- nance during substrate construction, links evidence to both mechanism nodes and directed relations, and retains missing mechanism layers as explicit evidence gaps. This represen- tation supports not only question answering and verification, but also mechanism reconstruction and cross-paper compari- son. Thus, our work lies at the intersection of scientific RAG, knowledge-graph construction, and evidence-grounded veri- fication, while making the organization of scientific evidence itself the primary object of study. Methodology Problem Formulation Let D = d i N i=1 be a scientific corpus and let Σ δ = (T δ ,R δ ,C δ ) denote the mechanism schema for domain δ. Here, T δ contains functional entity types, R δ contains ad- missible directed relations, and C δ contains structural con- straints. The framework is schema-configurable: construc- tion and retrieval are shared across domains, whereas Σ δ supplies domain-specific semantics. For flexible sensors, the schema is instantiated as Material → Sensor → Signal → System. We transformD into a substrate K δ = (D,E,V,R,P,A),(1) where E contains source-grounded evidence units, V and R are typed mechanism nodes and relations, P contains complete or partial mechanism paths, andA maps retained nodes and relations to their supporting evidence. The sub- strate is subsequently accessed through mechanism-aware retrieval rather than model fine-tuning. Figure 1 illustrates an evidence-grounded MS 3 object, and Figure 3 summarizes the decomposed construction pipeline. Constructing the Mechanism Knowledge Substrate Evidence extraction and provenance grounding. Each paper is parsed into ordered source blocks with page and section metadata and links to nearby figures, captions, tables, and quantitative content when available. Consecutive blocks are grouped into context-preserving windows for extraction while retaining their original locations. A task-specialized extractor produces evidence units e = (x e ,ρ e ,π e ), where x e is source-derived text, ρ e denotes its evidential role (e.g., experiment, property, validation, or reference), and π e stores document-level and local provenance. We accept an evidence unit only when its text, after deterministic normalization, can be aligned to the originating source block. Unaligned candidates are rejected, and near duplicates are merged while preserving the union of their provenance links. This stage produces compact, traceable evidence records rather than free-form summaries. Decomposed mechanism construction. We decompose substrate construction into three structured generation stages. All three stages use independently trained QLoRA adapters over Qwen2.5-72B-Instruct. The evidence extractor (E- LoRA) maps a focused PDF chunk to a small set of local evidence items. The terminology model (T-LoRA) performs evidence-grounded MS 3 terminology normalization, recov- ering Material, Sensor, Signal, and System terms from the accepted evidence rather than linking them to an external ontology. The mechanism model (M-LoRA) receives the ev- idence, grounded MS 3 terms, and source-grounded figure captions, and produces ordered mechanism objects with ev- idence chains. Mechanism induction is constrained to the ordered Material→Sensor→Signal→System schema. Com- plete paths and evidence-supported partial paths are retained, while unsupported layers are left absent. A mechanism object is p = (V p ,R p ,A p ),(2) where V p contains typed nodes, R p contains directed rela- tions allowed by Σ δ , and A p links each retained node and relation to evidence identifiers. A complete MS 3 object con- tains all four layers, while a partial object contains only the supported subpath. Missing layers and relations are left ab- sent rather than completed using model priors, making evi- dence gaps explicit. Validation and assembly. We apply deterministic gates for node-type validity, evidence-reference consistency, source alignment, provenance recovery, and duplicate removal. At final inference time, an evidence excerpt is retained only when its normalized text can be exactly aligned to a parsed source chunk; otherwise it is removed. Invalid evidence refer- ences are removed from mechanism chains, and a mechanism is discarded when no valid supporting evidence remains. Duplicate evidence and mechanism objects are removed using normalized exact matching. Token-overlap rules are used only to reconcile complementary partial mechanism paths. The accepted records are assembled into linked evi- dence and mechanism views that share paper-level evidence identifiers. Mechanism-Aware Access for Language Models Query-conditioned retrieval. The evaluated system uses a deterministic lexical query parser rather than a learned query interpreter. It normalizes and tokenizes the query, removes fixed stopwords and short tokens, and uses the remaining terms to search separate SQLite FTS5 indices for mecha- nisms and evidence. Quoted paper titles are resolved by exact title matching when available. Candidate papers are ranked by summing their mechanism-index and evidence-index BM25 retrieval scores with equal weights: S(p| q) = X m∈R m (p,q) − BM25 m (m,q) + X e∈R e (p,q) − BM25 e (e,q). (3) Within each candidate paper, mechanism objects are ranked by query-token overlap with their statements and MS 3 fields. Evidence referenced by the selected mechanisms is priori- tized, followed by other evidence ranked by query overlap. This procedure expands mechanism-linked evidence identi- fiers; it does not perform multi-hop traversal over database graph edges. a Corpus scale and temporal coverage Figure 2: Overview of the MS 3 knowledge substrate, including corpus scale, domain coverage, evidence composition, structural coverage, and multimodal provenance. Provenance-aware evidence selection and generation. For the in-domain evaluation, the system retains at most four papers, five mechanism objects per paper, and twelve evidence items per paper. The selected records are serialized directly as mechanism cards containing typed paths, missing layers, evidence excerpts, identifiers, and source locations. The evaluated implementation uses fixed top-k limits rather than an explicit input-token budget, minimum-cover solver, or redundancy-aware packing objective. The resulting context C q is supplied to the answer model: ˆy = f θ (q,C q ).(4) The prompt requires evidence identifiers for supported claims and permits an insufficient evidence response when the requested information is not present. Evidence sufficiency and source repair. The target sys- tem treats an MS 3 card as a verified index into source evi- dence rather than as an immutable summary. A claim-level verifier first checks whether the retrieved cards support every requested role, relation, number, unit, validation setting, and deployment statement. If support is missing, conflicting, or too coarse, the system follows the selected paper identifiers back to their stored full text, retrieves claim-directed raw passages, and re-audits the affected fields. If no relevant pa- per is covered by the substrate, scholarly Web search is used for source discovery before the same evidence gates are ap- plied. Thus, raw-PDF retrieval is an internal repair operator reached through retrieval and verification, not a user-facing condition that is handed a gold paper. Repaired cards re- main request-local unless a versioned update receives expert approval. The experiments below evaluate the single-pass MS 3 retrieval component. The ongoing open-domain study evaluates coverage routing, while the full source-repair loop requires a separate controlled evaluation. Experiments Research questions. We ask whether MS 3 retrieval (i) im- proves scientific correctness over a language model without retrieval, (i) improves over scholarly Web evidence available to a realistic search agent, and (i) yields consistent benefits across model families rather than for a single generator. Benchmark. We use a 300-question in-domain benchmark frozen before answer generation. It contains 100 mechanism- path reconstruction questions, 60 quantitative evidence questions, 60 cross-paper comparisons, and 80 evidence- boundary questions that require separating a demonstrated system from a proposal or unsupported deployment claim. The questions cover 362 target papers and are disjoint from the earlier development benchmark at the paper level. Several questions include a paper title; the benchmark therefore mea- sures grounded evidence use after a source can be resolved, rather than fully open-domain literature discovery. We additionally construct a 200-question out-of-domain benchmark. The covered stratum includes 70 single-topic corpus-discovery questions and 30 cross-paper synthesis questions. The out-of-domain stratum contains 50 near- domain and 50 far-domain questions. We remove paper titles, DOIs, paper and evidence IDs, and exact source-case iden- tifiers from every question. Hidden source-paper IDs and ❄ ❄ ❄ ❄ Figure 3: Expert-guided supervision, paper-disjoint bench- mark construction, and decomposed extraction of evidence- grounded MS 3 mechanism objects. answer outlines are retained only for post-retrieval auditing and are never exposed to the retriever or answer generator. Compared systems. We compare four conditions in the in- domain evaluation and three single-source or hybrid retrieval conditions in the coverage-shift evaluation. Closed-book Di- rect receives only the question and is instructed to provide a best-effort answer from parametric knowledge without fab- ricating paper-specific details. Web Search receives evidence frozen before answer generation from at most three queries and four scholarly sources discovered through OpenAlex and Crossref. Verified MS 3 receives the audited mechanism cards and their source excerpts. For the in-domain evaluation, Raw- PDF RAG receives the complete text of the known target paper and serves as an oracle-input comparison. It is not evaluated in the coverage-shift setting, where no single tar- get paper is assumed to be known in advance. Hybrid com- bines frozen Web evidence with Verified-MS 3 mechanism cards. The two sources are retrieved independently, dedu- plicated while preserving provenance, and jointly provided to the same generator. MS 3 supplies structured mechanism paths and evidence boundaries, whereas Web retrieval con- tributes evidence beyond the corpus coverage. The generator must cite the corresponding source and abstain when neither source supports a claim. Models and generation. The completed evaluation covers Qwen2.5-14B, Qwen2.5-72B, Qwen3-30B-A3B, Mistral- Small-24B, DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi- K3, Kimi-K2.6, GLM-5.2, and GLM-5-Turbo. Open models are run from frozen local checkpoints and the DeepSeek, Kimi and GLM models through their versioned API end- points. Each model receives the same question and method- specific frozen context. We use deterministic decoding where supported and a common structured answer schema. The Di- rect prompt permits the use of parametric knowledge and does not force refusal merely because citations are unavail- able; its citation field is empty by design. Blind evaluation and statistics. For each question, candidate identities, method names, model names, and representation-specific citation identifiers are hidden from the judge. The judge receives the union of verbatim evidence needed to check all candidates and independently scores sci- entific correctness, citation entailment, and completeness on a 0–4 scale. Absolute calibration can therefore differ between judge groups; primary inference uses within-model paired comparisons. Primary comparisons are conducted within each generator using paired question-level scores. Citation entailment is not applicable to Closed-book Direct and is not used to penalize its scientific correctness. Hardware and implementation. All local experiments were conducted on a server equipped with three NVIDIA A100 80 GB GPUs. Training used all three GPUs, whereas each local generator was evaluated on a single GPU. Our implementation uses Python 3.11.13, PyTorch 2.5.1 with CUDA 12.4, Transformers 4.57.6, PEFT 0.18.1, bitsand- bytes 0.49.2, DeepSpeed 0.19.1, and FlashAttention 2.7.4. We fix the training seed to 42 and freeze all benchmark questions, retrieved contexts, and source documents before answer generation. MS 3 extraction models. The evidence extractor (E), ter- minology grounder (T), and mechanism inducer (M) are task-specific adapters of Qwen2.5-72B-Instruct. We use 4- bit QLoRA with NF4 quantization and double quantiza- tion. LoRA adapters are applied to the q_proj, k_proj, v_proj, and o_proj modules with rank 16, scaling fac- tor 16, and dropout 0.05. The maximum input length is 14,000 tokens, and sequence packing is disabled. Training uses BF16, gradient checkpointing, and DeepSpeed ZeRO-3. The per-device batch size is 1 with eight gradient accumu- lation steps, giving an effective batch size of 24 over three GPUs. We use a cosine learning-rate schedule, zero weight decay, and a maximum gradient norm of 0.3. Main results. Table 1 shows a consistent retrieval benefit. MS 3 improves scientific correctness over Closed-book Direct for all ten generators, with a macro-average gain of 54.4 points. Compared with frozen Web Search, MS 3 gains 26.9 points on average and is better for 9 of 10 generators. The exception is Qwen3-30B-A3B, for which Web exceeds MS 3 by 5.0 points; Thus, structured context is not equally usable by every generator. Table 2 shows that the correctness gain is accompanied by stronger grounding rather than only longer answers. Com- pared with Web Search, MS 3 improves citation entailment Figure 4: Mechanism-centered MS 3 retrieval and grounding. A scientific query is answered either from parametric knowledge alone or through dual-index retrieval, provenance expansion, and evidence-aware context packing over the MS 3 substrate, enabling mechanism-grounded responses with claim-level citations and explicit support boundaries. In-domainCoverage-shift GeneratorParams.Inference Direct Web Raw-PDF Verified-MS 3 ∆ V−R Web Verified-MS 3 Hybrid∆ H−B Qwen2.5-14B14BLocal-4bit 37.5 58.380.880.80.0 51.952.1 54.2 +2.1 Qwen2.5-72B72BLocal-4bit 34.2 60.081.786.7 +5.0 52.254.6 57.8 +3.2 Qwen3-30B-A3B 30B/3B active Local-4bit 23.3 53.346.748.3 +1.7 48.951.1 53.8 +2.7 Mistral-Small-24B 24BLocal-4bit 25.0 54.265.070.0 +5.0 53.352.1 55.6 +2.3 DeepSeek-V4-Flash Undisc.API25.8 51.780.089.2 +9.2 56.258.3 68.8 +10.4 DeepSeek-V4-Pro Undisc.API28.3 46.782.590.0 +7.5 56.364.6 70.9 +6.3 Kimi-K3Undisc.API26.7 56.782.583.3 +0.8 49.241.7 41.7-7.5 Kimi-K2.6Undisc.API26.7 56.780.090.8 +10.8 61.364.6 71.3 +6.7 GLM-5.2Undisc.API29.2 60.883.394.2 +10.8 63.466.7 68.8 +2.1 GLM-5-TurboUndisc.API24.2 57.584.291.7 +7.5 43.550.0 58.3 +8.3 Macro average–28.1 55.676.782.5 +5.8 53.655.6 60.1 +3.7 Table 1: Scientific correctness (%). V , R, and H denote Verified-MS 3 , Raw-PDF RAG, and Hybrid retrieval, respectively. The Raw-PDF condition is a source-resolved oracle supplied with the benchmark-matched paper, rather than a user-facing retrieval baseline. ∆ V−R = V − R measures the improvement of Verified-MS 3 over Raw-PDF RAG. For the coverage-shift evaluation, B = max(Web,V ) denotes the strongest single-source retrieval condition, and ∆ H−B = H − B measures the additional benefit of combining Web and Verified-MS 3 . Bold indicates the best result within each evaluation setting. Scientific correctnessCitation entailmentCompleteness GeneratorDirect Web Raw-PDF Verified-MS 3 Web Raw-PDF Verified-MS 3 Direct Web Raw-PDF Verified-MS 3 Qwen2.5-14B37.5 58.380.880.8 58.387.585.0 20.8 42.569.273.3 Qwen2.5-72B34.2 60.081.786.7 61.786.790.0 20.0 45.075.879.2 Qwen3-30B-A3B23.3 53.346.748.3 49.250.049.2 24.2 48.345.849.2 Mistral-Small-24B 25.0 54.265.070.0 50.870.075.0 20.0 40.059.264.2 DeepSeek-V4-Flash 25.8 51.780.089.2 68.390.897.5 18.3 36.777.587.5 DeepSeek-V4-Pro28.3 46.782.590.0 62.592.597.5 23.3 34.285.093.3 Kimi-K326.7 56.782.583.3 56.785.084.2 20.8 50.083.383.3 Kimi-K2.626.7 56.780.090.8 55.083.391.7 20.8 47.580.090.0 GLM-5.229.2 60.883.394.2 60.088.396.7 15.8 46.782.592.5 GLM-5-Turbo24.2 57.584.291.7 56.789.294.2 14.2 43.375.885.8 Macro average28.1 55.676.782.5 57.982.386.1 19.8 43.473.479.8 Table 2: Per-generator results on the in-domain evaluation (%). Citation entailment is not applicable to Closed-book Direct. Bold indicates the best method within each metric group for a given generator. 88% 80% 65% 58% 80% 86% 74% 43% 59% 89% 84% 71% 72% 82% 67% 59% Figure 5: Ablation of the MS 3 substrate construction pipeline. Error bars denote 95% confidence intervals. by 28.2 points and completeness by 36.4 points. This pat- tern is consistent with the benefit of same-paper evidence links, functional roles, and explicit evidence boundaries. It does not establish the contribution of each structural com- ponent individually; those effects require evidence-matched ablations. In the coverage-shift evaluation, Hybrid retrieval achieves a macro-average scientific correctness of 60.1%, exceeding the strongest single-source condition for each generator by 3.7 points on average. Hybrid improves over the better of Web Search and Verified-MS 3 for nine of the ten gener- ators, with particularly large gains for DeepSeek-V4-Flash (+10.4 points), GLM-5-Turbo (+8.3 points), and Kimi-K2.6 (+6.7 points). The exception is Kimi-K3, for which Hybrid underperforms the strongest single-source condition by 7.5 points. These results indicate that Web evidence and MS 3 mechanism cards are generally complementary under cov- erage shift, but that combining them does not automatically improve generation. The failure case suggests that some gen- erators may be sensitive to conflicting evidence, increased context length, or the need to integrate information with het- erogeneous structures and provenance. Ablation Studies We compare direct generation, one-shot LoRA, decomposed extraction, and the complete E/T/M pipeline with verification. As shown in Figure 5, the full pipeline achieves the best semantic full-path accuracy (74%) and effective quality (72%), while maintaining strong MS 3 term accuracy (88%) and mechanism support (84%). No- tably, decomposed extraction without verification performs poorly on full-path reconstruction (43%), suggesting that the verifier is important for correcting inconsistencies and as- sembling reliable mechanism chains. The results also show that decomposition alone is insuf- ficient: although it preserves term extraction performance, errors accumulate during full-path assembly. Adding verifi- cation recovers 31 points in semantic full-path accuracy and 22 points in effective quality, indicating that validation is particularly important for relation-level consistency. Limitations The current instantiation is restricted to conductive-fiber flexible sensors, and schema transfer to other scientific domains remains unevaluated. The bench- mark primarily measures evidence use after candidate sources can be identified, rather than unrestricted litera- ture discovery. In addition, the present experiments evaluate single-pass MS 3 retrieval rather than the complete source- repair workflow. Conclusion We presented an evidence-grounded mechanism knowledge substrate that organizes scientific literature into provenance- linked, role-typed mechanism paths. Instantiated as MS 3 for conductive-fiber flexible sensors, the substrate supports grounded mechanism reconstruction, quantitative lookup, cross-paper comparison, and evidence-boundary reasoning. On the in-domain benchmark, Verified MS 3 improves macro- averaged scientific correctness by 26.9 percentage points over Web search and by 5.8 points over a source-resolved Raw- PDF oracle. Under coverage shift, Hybrid retrieval improves over the stronger single-source condition by 3.7 points on average. These findings demonstrate that knowledge repre- sentation and evidence routing materially affect grounded scientific question answering. References Asai, A.; He, J.; Shao, R.; Shi, W.; Singh, A.; Chang, J. C.; Lo, K.; Soldaini, L.; Feldman, S.; D’Arcy, M.; Wadden, D.; Latzke, M.; Sparks, J.; Hwang, J. D.; Kishore, V.; Tian, M.; Ji, P.; Liu, S.; Tong, H.; Wu, B.; Xiong, Y.; Zettlemoyer, L.; Neubig, G.; Weld, D. S.; Downey, D.; Yih, W.-t.; Koh, P. W.; and Hajishirzi, H. 2026. Synthesizing Scientific Literature with Retrieval-Augmented Language Models. Nature, 650: 857–863. Auer, S.; Kovtun, V.; Prinz, M.; Kasprzik, A.; Stocker, M.; and Vidal, M.-E. 2018. Towards a Knowledge Graph for Science. In Proceedings of the 8th International Conference on Web Intelligence, Mining and Semantics, WIMS ’18, 1–6. New York, NY, USA: Association for Computing Machinery. Caufield, J. H.; Hegde, H.; Emonet, V.; Harris, N. L.; Joachimiak, M. P.; Matentzoglu, N.; Kim, H.; Moxon, S. A. T.; Reese, J. T.; Haendel, M. A.; Robinson, P. N.; and Mungall, C. J. 2024. Structured Prompt Interrogation and Recursive Extraction of Semantics (SPIRES): A Method for Populating Knowledge Bases Using Zero-Shot Learning. Bioinformatics, 40(3): btae104. Dasigi, P.; Lo, K.; Beltagy, I.; Cohan, A.; Smith, N. A.; and Gardner, M. 2021. A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4599–4610. Online: Asso- ciation for Computational Linguistics. Ding, H.; Zhao, Y.; Hu, T.; Wang, Z.; Patwardhan, M.; and Cohan, A. 2026. SciRAG: Adaptive, Citation-Aware, and Outline-Guided Retrieval and Synthesis for Scientific Litera- ture. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 6440–6460. Rabat, Morocco: As- sociation for Computational Linguistics. Edge, D.; Trinh, H.; Cheng, N.; Bradley, J.; Chao, A.; Mody, A.; Truitt, S.; Metropolitansky, D.; Ness, R. O.; and Lar- son, J. 2024. From Local to Global: A Graph RAG Ap- proach to Query-Focused Summarization. arXiv preprint arXiv:2404.16130. He, X.; Tian, Y.; Sun, Y.; Chawla, N. V.; Laurent, T.; LeCun, Y.; Bresson, X.; and Hooi, B. 2024. G-Retriever: Retrieval- Augmented Generation for Textual Graph Understanding and Question Answering. In Advances in Neural Information Processing Systems, volume 37, 132876–132907. Curran As- sociates, Inc. Jain, S.; van Zuylen, M.; Hajishirzi, H.; and Beltagy, I. 2020. SciREX: A Challenge Dataset for Document-Level Informa- tion Extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 7506–7516. Online: Association for Computational Linguistics. Jaradeh, M. Y.; Oelen, A.; Farfar, K. E.; Prinz, M.; D’Souza, J.; Kismihók, G.; Stocker, M.; and Auer, S. 2019. Open Research Knowledge Graph: Next Generation Infrastructure for Semantic Scholarly Knowledge. In Proceedings of the 10th International Conference on Knowledge Capture, K- CAP ’19, 243–246. New York, NY, USA: Association for Computing Machinery. Lála, J.; O’Donoghue, O.; Shtedritski, A.; Cox, S.; Ro- driques, S. G.; and White, A. D. 2024. PaperQA: Retrieval- Augmented Generative Agent for Scientific Research. In The Twelfth International Conference on Learning Representa- tions. Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; Küttler, H.; Lewis, M.; Yih, W.-t.; Rocktäschel, T.; Riedel, S.; and Kiela, D. 2020. Retrieval-Augmented Gen- eration for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems, volume 33, 9459– 9474. Curran Associates, Inc. Luan, Y.; He, L.; Ostendorf, M.; and Hajishirzi, H. 2018. Multi-Task Identification of Entities, Relations, and Corefer- ence for Scientific Knowledge Graph Construction. In Pro- ceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 3219–3232. Brussels, Bel- gium: Association for Computational Linguistics. Skarlinski, M. D.; Cox, S.; Laurent, J. M.; Braza, J. D.; Hinks, M.; Hammerling, M. J.; Ponnapati, M.; Rodriques, S. G.; and White, A. D. 2024. Language Agents Achieve Su- perhuman Synthesis of Scientific Knowledge. arXiv preprint arXiv:2409.13740. Venugopal, V.; and Olivetti, E. 2024. MatKG: An Au- tonomously Generated Knowledge Graph in Material Sci- ence. Scientific Data, 11: 217. Wadden, D.; Lin, S.; Lo, K.; Wang, L. L.; van Zuylen, M.; Cohan, A.; and Hajishirzi, H. 2020. Fact or Fiction: Verify- ing Scientific Claims. In Proceedings of the 2020 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), 7534–7550. Online: Association for Computa- tional Linguistics. Wadden, D.; Lo, K.; Kuehl, B.; Cohan, A.; Beltagy, I.; Wang, L. L.; and Hajishirzi, H. 2022. SciFact-Open: Towards Open- Domain Scientific Claim Verification. In Findings of the Association for Computational Linguistics: EMNLP 2022, 4719–4734. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics. Wang, C.; Shen, Y.; Kuang, Z.; Cohan, A.; and Zhao, Y. 2025. SciVer: Evaluating Foundation Models for Multi- modal Scientific Claim Verification. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8562–8579. Vienna, Austria: Association for Computational Linguistics.