FindLink: Learning to Assess Scientific Connections from Source-Grounded Premises

Jingxuan Wei; Song Tan; Xiaohan Liu; Chao Sang; Changwei Wang; Linzhuang Sun; Bihui Yu; Fuqiang Wang

cs.CLNov 2, 2026

FindLink: Learning to Assess Scientific Connections from Source-Grounded Premises

Authors: Jingxuan Wei, Song Tan, Xiaohan Liu, Chao Sang, Changwei Wang, Linzhuang Sun, Bihui Yu, Fuqiang Wang,

Organizations: Qilu University of Technology · Shandong Computer Science Center · University of the Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences

Abstract

Scientific discovery from the literature requires assessing connections among findings obtained under different study conditions. Source attribution makes individual findings traceable, but does not determine what they jointly support. We introduce GI-TraceBench, a benchmark of 10,282 questions in gastrointestinal oncology that unifies provenance question answering, knowledge discovery, and hypothesis evaluation through shared Paper, Claim, and Evidence Span records. We further propose FindLink, a two-stage framework for provenance-grounded scientific reasoning. Premise Grounding constructs source-bound premises with explicit study context and support status. Structured Reasoning organizes these premises into support plans that represent both supported relations and unresolved conditions, then generates evidence-conditioned answers using deterministic compiler feedback. The compiler verifies structural and source-binding consistency, but does not assess scientific entailment. On the 1,545-question fixed-evidence test set, FindLink improves the balanced family macro score from 67.95 to 87.25 over direct supervised fine-tuning of the same backbone. A retrained variant without Structured Reasoning achieves slightly higher premise-identifier F1, but substantially lower relation and evidence-sufficiency F1. These results indicate that scientific reasoning requires not only identifying grounded premises, but also assessing what connections those premises support.

Figures & tables

Figure 1: Overview of Source-Grounded Scientific Connection Assessment. Findings from multiple papers may be individually supported but fail to establish the connecting claim, leaving an evidence gap that provenance-aware reasoning must identify.

Figure 2: GI-TraceBench construction. Source papers are converted into linked Paper–Claim–Evidence Span records, expanded into candidate relations and hypotheses, and transformed into QA instances after reference, deduplication, and split checks.

Figure 3: GI-TraceBench composition. Task families are shown in the inner disk and their primary subtypes in the outer ring.

Figure 4: FindLink framework. An answer-level SFT checkpoint is sequentially adapted through Premise Grounding and Structured Reasoning. The compiler checks the support plan before generating an evidence-grounded answer with explicit gaps and conditional assessment fields.

Model	PGS $\uparrow$	DRS $\uparrow$	HSS $\uparrow$	BFM $\uparrow$	R-L $\uparrow$	MET $\uparrow$	BERT $\uparrow$	Mover $\uparrow$	LLM Judge $\uparrow$
Open-source models
GLM-4-9B-0414 ( Z.ai, 2025 )	10.29	34.94	24.01	23.08	20.11	20.67	62.94	42.93	2.75
Gemma-3-12B-IT ( Google DeepMind, 2025 )	12.55	21.42	24.99	19.65	16.94	24.26	54.50	36.76	2.42
Granite-3.3-8B ( IBM, 2025 )	0.97	11.91	12.76	8.55	6.46	9.78	23.82	15.84	0.91
Hunyuan-7B ( Tencent Hunyuan Team, 2025 )	12.41	23.18	23.46	19.68	25.18	26.42	57.78	41.12	2.17
Llama-3.1-8B-Instruct ( Meta, 2024 )	11.08	8.83	1.90	7.27	5.77	6.73	14.86	10.16	0.61

Table 1: Main results and ablations on GI-TraceBench. Scores use 0–100, except LLM Judge (0–5). Bold marks column maxima; green marks the best scores excluding FindLink. A/B denote Premise Grounding/Structured Reasoning.

Figure 5: Fine-grained knowledge-discovery performance. Columns show seven open-weight models, four closed-source models, four adapted methods, and FindLink; rows show the four F1 components of DRS. The shared color scale is 0–100. Default SFT and ablations remain in Table 11 .

Figure 6: Consistency between LLM-as-Judge and human evaluation.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Split	Clusters	Provenance	Discovery	Hypothesis	Total
Train	42	5,009	1,624	2,104	8,737
Test	8	898	286	361	1,545
Total	50	5,907	1,910	2,465	10,282

Appendix

Table 2: Train/Test composition. Counts refer to retained QA instances and thematic clusters.

Domain	Train	Test	Total
Biliary	300	31	331
Colorectal	2,157	560	2,717
Cross-GI	2,380	596	2,976
Esophageal	973	221	1,194
Gastric	1,380	83	1,463
Pancreatic	1,547	54	1,601

Appendix

Table 3: Domain distribution. Cross-GI denotes shared gastrointestinal oncology topics, not a separate cancer site.

Object or construction route	Count	Counting unit / scope
Papers	1,614	Source-document records.
Claims	5,475	Scientific-statement records.
Evidence Spans	4,968	Source-passage records.
Associations	733	Candidate-association records.
Hypothesis cards: association-based	733	Cards derived from association records.
Hypothesis cards: typed-chain	838	Cards derived from typed discovery chains.

Appendix

Table 4: Upstream construction inventory. These counts precede formal QA filtering and are not split-level entity counts.

	Evidence grounding					Scientific reasoning			Further testing
Benchmark	SA	EL	MD	CV	UI	SC	SR	HA	VP	RC
PubMedQA ( Jin et al., 2019 )	✗	✗	✗	✗	✓	✗	✗	✗	✗	✗
BioASQ-QA ( Krithara et al., 2023 )	✓	✓	✓	✗	✗	✗	✗	✗	✗	✗
QASPER ( Dasigi et al., 2021 )	✓	✓	✗	✗	✓	✗	✗	✗	✗	✗
SciFact ( Wadden et al., 2020 )	✓	✓	✓	✓	✓	✗	✗	✗	✗	✗
ScholarQABench ( Asai et al., 2026 )	✓	✓	✓	✓	✗	✗	✗	✗	✗	✗

Appendix

Table 5: Capability coverage. ✓: explicitly covered; ✗: not explicitly covered. Coverage is benchmark-level, not required of every instance.

Field	Notation	Meaning
id	$i$	Local premise identifier, such as P0 .
statement	$s_{i}$	Atomic statement to be assessed or reused.
context	$c_{i}$	Study setting and applicable qualifications.
polarity	$\pi_{i}$	Positive or negative orientation.
origin	$o_{i}$	Observation, author claim, or candidate origin.
sources	$I_{i}$	References to canonical evidence pointers.

Appendix

Table 6: Premise representation. Support labels are model predictions, not independently verified truth labels.

Interface	Role and interpretation
from_a	Selects a generated premise by local identifier and retains its source bindings.
from_source	Binds a proposition directly to an original evidence pointer.
all_of	Combines existing support expressions without introducing a new relation.
Context metadata	Preserves study setting, polarity, and scope.
Target requirements	Records requested propositions and concrete verification conditions.
Validity judgment	Records whether a candidate relation is admissible as a represented assessment.

Appendix

Table 7: Support-plan interface. The operators organize a represented assessment and do not constitute a general scientific proof system.

Setting	Premise adaptation	Reasoning adaptation
Initialization	$S$	$SA$
Targets	Premises	Plans and answers
Adaptation	LoRA	LoRA
LoRA rank / scaling alpha	16 / 32	16 / 32
Epochs	1	1
Learning rate	$10^{-4}$	$10^{-4}$

Appendix

Table 8: Configuration of the two adaptation stages.

Configuration	Training route	Inference interface
Default SFT	Reuse $S$	Question and evidence directly produce the answer.
w/o A	$S\rightarrow SB$	Plan and answer directly from original evidence; no premise inventory or from_a leaves.
w/o B	$SA\rightarrow D$	Frozen premise output plus question and evidence; no support plan or compiler-derived fields.
FindLink	$S\rightarrow SA\rightarrow SAB$	Premises, support plan, deterministic compilation, and evidence-conditioned answer.

Appendix

Table 9: Core comparison design.

Criterion	Assessment
Answer quality	Correctness, relevance, completeness, and coverage of the requested content.
Evidence entailment	Whether substantive claims follow from the supplied evidence without contradiction or unsupported extrapolation.
Citation correctness	Whether valid source references support the claims to which they are attached.
Citation completeness	Whether important evidence-dependent claims have the required source support.
Boundary control	Correct polarity, study scope, uncertainty, causal strength, and distinction between candidates, hypotheses, and facts.

Appendix

Table 10: Shared evaluation criteria. The LLM judge assigns one holistic score; human reviewers score each applicable dimension. Citation criteria apply only when citations are required.

System	Premise	Relation	Boundary	Sufficiency
Default SFT ( Qwen Team, 2026 )	59.93	58.69	61.63	62.06
w/o A	86.01	87.54	89.25	65.79
w/o B	88.10	76.44	89.74	60.71
FindLink	87.69	88.98	92.36	78.40

Appendix

Table 11: Knowledge Discovery QA components for the in-domain configurations. Values use the 0–100 scale. Bold marks the column maximum among these four configurations. Results for the other 15 comparison systems are displayed in Figure 5 .

System	LLM Judge	Human
Proposed method
FindLink (Ours)	4.19	4.40
Open-weight general models
GLM-4-9B-0414 ( Z.ai, 2025 )	2.50	1.92
Gemma-3-12B-IT ( Google DeepMind, 2025 )	2.85	3.25
Granite-3.3-8B ( IBM, 2025 )	0.79	0.73

Appendix

Table 12: Aligned human and LLM-judge means. Each row reports a system mean on the common 24-question subset and the 0–5 scale. The table includes the 11 comparison models and FindLink; the judge values are subset-specific for Human–LLM agreement, whereas the main table reports the separate 120-question judge evaluation.

Figure 7: Candidate–verification mismatch. The evidence supports one candidate, whereas the generated verification plan targets a different intervention.

Figure 8: Study-context boundary diagnostic. The appropriate assessment preserves the original population and treatment setting instead of making an unsupported generalization.

Figure 9: Bounded-assessment diagnostic. The response supports a candidate-level association while preserving uncertainty and avoiding an unsupported causal conclusion.

Figure 10: Provenance-localization diagnostic. The claim is incorrectly linked to an irrelevant safety passage, whereas the appropriate source directly reports the claimed biomarker result.

Scientific Reasoning in Language Models

Scientific reasoning in language models studies how AI systems use evidence, domain knowledge, and tools to solve problems and support discovery.

See all 161 papers

LitTraceQA: A Benchmark for Multi-Stage Grounding and Verification in Scientific Question Answering
Aug 7, 2026Xuye Liu, Yimu Wang, Peng Shi +7Evidence-Grounded Generation Scientific QA
University of Waterloo · 2Amazon · 3City University of Hong Kong +1
PathAnchor: Path-Structured Evidence for Scientific Agents
Sep 30, 2026Qiuhui Chen, Jiafan Lu, Shuaimin Tang +5Evidence-Grounded Reasoning Scientific QA
School of Information Science and Engineering, East China University of Science and Technology · College of Intelligent Robotics and Advanced Manufacturing, Fudan University · Tencent
VisProve: Learning Typed Visual Provenance for Scientific Document Question Answering
Nov 7, 2026Jingxuan Wei, Zhengyang Wang, Zheng Guo +6Evidence-Grounded QA Scientific QA
University of Chinese Academy of Sciences · Qilu University of Technology · Shandong Computer Science Center +2

Aug 7, 2026J/K move · Enter open · S save

LitTraceQA: A Benchmark for Multi-Stage Grounding and Verification in Scientific Question Answering

Xuye Liu, Yimu Wang, Peng Shi, Bo Xue+6

University of Waterloo · 2Amazon · 3City University of Hong Kong+1

Evidence-Grounded Generation Scientific QA Evidence Retrieval Evidence-Grounded QA Factual QA Multi-Document QA+1

Open paper AskPDF arXiv

Scientific literature is increasingly used as a knowledge source for language models, retrieval-augmented generation systems, and research assistants, but answering research questions from papers requires more than fluent generation. A reliable system must identify the relevant papers, locate the concrete evidence that supports the answer, and produce a response that is faithful to that evidence. We present LitTraceQA, a benchmark for literature-grounded question answering over scientific papers. Given a research question and a metadata pool of papers, a system must return three connected outputs: canonical paper identifiers, supporting evidence locations, and answers in one or more requested formats, including free-form text, multiple-choice answers, and structured tables. LitTraceQA targets evidence types common in scientific reading: tables, figures, text spans, equations or algorithms, and citation contexts. The public development split contains 55 examples, including 26 hidden-source single-paper questions and 29 multi-paper questions, and provides gold papers, evidence annotations, and answers for local validation. We also analyze a larger final annotation collection with 4,978 unique-question records over 4,859 unique gold papers. By evaluating paper retrieval, evidence grounding, and answer accuracy separately, LitTraceQA provides a testbed for scientific QA systems that produce verifiable answers rather than unsupported summaries.

Abstract

Scientific Reasoning in Language Models

Explore similar work

LitTraceQA: A Benchmark for Multi-Stage Grounding and Verification in Scientific Question Answering

PathAnchor: Path-Structured Evidence for Scientific Agents

VisProve: Learning Typed Visual Provenance for Scientific Document Question Answering