PaperlayerView exact revision on arXiv

Source-linked AI summary

PathAnchor: Path-Structured Evidence for Scientific Agents

Qiuhui Chen, Jiafan Lu, Shuaimin Tang, Tao Dai, Suyuan Wang, Chenrui Ji, Zhenglei Zhou, Weimin Zhong

arXiv:2609.38766v1cs.AI

TL;DR

Scientific agents can retrieve relevant passages yet lose functional order, mix sources, or state conclusions beyond the record. PathAnchor organizes evidence as source-linked MS3 paths and uses a bounded controller to open supporting records. It scores 82.6% on FS EvidenceQA and, in a matched ablation, improves source recall and citation provenance while reducing tool use.

  • Problem

    Retrieved passages may lose functional order, combine incompatible sources, or let plausible completions be presented as reported evidence.

    Source: §1 · p. 1

  • Method

    PathAnchor uses source-linked MS3 trajectories and a bounded controller that searches paths, opens exact evidence, and marks an explicit evidence boundary.

    Source: §1 · p. 1, §5 · p. 4

  • Results

    PathAnchor scores 82.6% (3.304/4), the highest observed score among six tested systems; path records improve source recall and citation-provenance completeness by 21.6 and 20.8 points while reducing tool use.

    Source: §1 · p. 1, §4.1 · p. 3, §5 · p. 4

  • Takeaways & Limitations

    Under a limited call budget, source-local functional order supports evidence selection and cross-paper synthesis, while the explicit boundary makes unsupported requests visible.

    Source: §4.4 · p. 4, §5 · p. 4

    The representation ablation bundles changes to role labels, relation direction, path order, and mechanism statements, so it does not isolate their individual effects.

    Source: §4.3 · p. 4

Abstract

from arXiv · show

Scientific agents can retrieve relevant passages yet still lose functional order, mix evidence across sources, or state conclusions that exceed the retrieved record. We introduce PathAnchor, a bounded scientific reasoning system built on path-structured evidence workspaces. Instead of treating passages or extracted concepts as independent units, the system retrieves source-linked Material-Sensor-Signal-System trajectories that preserve role, direction, and the evidence supporting each transition. A controller uses three read-only tools to search paper-specific trajectories, trace paths across candidate sources, and open exact evidence before producing a claim-cited answer and an explicit evidence boundary. On 120 single- and cross-paper flexible-sensor questions, PathAnchor scores 82.6% and leads six evaluated systems. Under a matched controller, corpus, and six-call budget, replacing unordered concept graphs with path-structured records raises source recall from 61.3% to 82.9%, increases answers whose claims all cite opened evidence from 69.2% to 90.0%, and reduces tool calls. These results show that evidence organization affects retrieval and citation completeness under fixed agent resources.

1. INTRODUCTION

Scientific agents may retrieve relevant passages yet lose functional order, combine incompatible sources, or present plausible completions as reported evidence. PathAnchor organizes source-linked functional trajectories and, under matched conditions, improves source recall and citation-provenance completeness while reducing tool use.

1. INTRODUCTION

  • Isolated passages can leave agents to reconstruct linked facts’ roles and order, risking incompatible-source combinations or unsupported completions.The motivating sequence runs from a material to a sensing element, measurable signal, and system-level function.Source: §1 · p. 1
  • Prior work covers agentic retrieval, citation-aware synthesis, benchmarks, structured extraction, and scientific knowledge graphs; the paper asks whether organization changes faithfulness and efficiency under matched access.The related-work passage names PaperQA, LitQA2, SciRAG, OpenScholar, AstaBench, and SciCUEval.Source: §1 · p. 1
  • PathAnchor uses ordered, source-linked Material–Sensor–Signal–System trajectories and separates supported conclusions from an explicit evidence boundary.The agent uses three read-only tools and at most six calls.Source: §1 · p. 1
  • Path records improve source recall and citation-provenance completeness by 21.6 and 20.8 points while reducing tool use under a matched ablation.The ablation fixes the controller, corpus, evidence access, prompts, and call budget.Source: §1 · p. 1

2. PATH-STRUCTURED SCIENTIFIC AGENT

PathAnchor represents evidence as source-linked functional paths, gives an agent a bounded read-only workspace, and compares path records with an otherwise matched concept-graph control.

2.1. Evidence-linked functional trajectories

  • An MS3 record contains typed Material, Sensor, Signal, and System nodes, supported directed transitions, and source-evidence links; experiments retrieve complete four-role paths.The frozen corpus contains 13,689 full-text papers and 26,648 trajectory records linked to 131,083 evidence items; separate indexes retrieve candidate papers and rank paths within them.Source: §2.1 · pp. 1–2

2.2. Bounded evidence workspace

  • The controller uses read-only search, path-tracing, and evidence-inspection actions, stopping when support is sufficient or the six-call budget is reached.A draft pairs each substantive claim with opened evidence and returns a separate boundary statement; citations are checked against opened records.Source: §2.2 · pp. 2–3

2.3. Matched representation control

  • Role Graph-Agent keeps the controller, corpus, prompts, evidence passages, and six-call budget fixed while withholding MS3 roles, directed relations, path order, and mechanism statements.This changes the structure shown to the agent without changing its underlying evidence access.Source: §2.3 · p. 3

3. EXPERIMENTAL PROTOCOL

The protocol combines three public evaluations with a domain-specific flexible-sensor benchmark, then compares systems and evidence-retrieval performance using task-specific scoring procedures.

3.1. Benchmarks

  • The public benchmarks use distinct endpoints: exact-label accuracy for LitQA2-FullText, an official four-component average for ScholarQA-CS2, and an eight-stratum macro-average for SciCUEval Materials.SciCUEval’s 200 questions span two input modalities and four competencies; the other public evaluations include 75 and 100 test cases, respectively.Source: §3.1 · p. 3, §3.2 · p. 3
  • FS EvidenceQA contains 120 evidence-derived questions—60 single-paper and 60 cross-paper—balanced over six categories and scored for correctness, grounding, evidence completeness, and boundary control.Questions were generated from anonymized excerpts, with source papers disjoint across questions; the authors characterize the benchmark as a controlled diagnostic of source-grounded answering.Source: §3.1 · p. 3

3.2. Systems and evaluation

  • Six configurations are compared under controlled source access, with failed or truncated outputs scored incorrect; FS EvidenceQA uses paired question-level bootstrap intervals.The matched representation test measures hidden-gold source recall, fully cited answers, uncited claims, and tool calls.Source: §3.2 · p. 3
  • A separate retrieval diagnostic compares full-text evidence search with MS3 mechanism-field search on 180 frozen questions, using proxy-graded relevance and nDCG@10.Paired bootstrap intervals are computed at the question level.Source: §3.2 · p. 3

4. RESULTS AND ANALYSIS

PathAnchor leads the reported flexible-sensor and public evaluations, while matched tests find that path structure improves evidence retrieval and citation completeness with fewer tool calls. The discussion notes that the representation components remain bundled and identifies broader tests as future work.

4.1. Multi-benchmark performance

  • On FS EvidenceQA, PathAnchor achieves 82.6% (3.304/4) and leads both question types; task-adapted MS3 workflows rank first on three public evaluations, with margins of 10.7, 12.0, and 6.5 points.The evaluations are LitQA2-FullText, ScholarQA-CS2, and SciCUEval Materials, respectively.Source: §4.1 · p. 3
  • Cross-paper scores fall for every system, but PathAnchor remains highest; unsupported claims rise from 0.583 to 1.950 per PathAnchor answer.Future work will test partial paths and individual representation components across additional controllers and expert-authored questions.Source: §4.1 · p. 3, §4.4 · p. 4

4.2. Claim-level quality and uncertainty

  • PathAnchor’s correctness, grounding, evidence-completeness, and boundary-control scores are 3.325, 3.300, 3.308, and 3.283, respectively; it produces the fewest unsupported claims (1.267 per question), with Kimi K3 the only interval crossing zero.Source: §4.2 · p. 4

4.3. What path structure changes

  • 21.6-point source-recall gain and opened-evidence citation completeness rising from 69.2% to 90.0% accompany 0.91 fewer calls under matched access.Both agents can open the same passages; claims without opened-evidence citations fall from 0.450 to 0.133 per answer.Source: §4.3 · p. 4
  • PathAnchor exposes 11.81 papers versus Role Graph’s 13.70, with 99.90% versus 96.39% citation validity and 6.7% versus 37.5% six-call-limit rates.The authors describe the lower limit-hit rate as consistent with earlier stopping once source-specific support is found.Source: §4.3 · p. 4
  • The representation ablation changes role labels, relation direction, path order, and mechanism statements as a bundle.Future work will test partial paths and individual components across additional controllers and expert-authored questions.Source: §4.3 · p. 4, §4.4 · p. 4

4.4. Discussion

  • The retrieval and agent tests locate benefits at two stages: ranking on covered questions and finding more gold sources despite a smaller workspace.The discussion also connects source-local functional order to cross-paper synthesis under a limited call budget.Source: §4.4 · p. 4

5. CONCLUSION

PathAnchor uses ordered, source-linked MS3 trajectories under a six-call budget. The conclusion reports the highest observed FS EvidenceQA score among six tested systems and improved recall and citation provenance in the matched ablation.

  • PathAnchor organizes scientific evidence as ordered, source-linked MS3 trajectories under a six-call budget.Source: §5 · p. 4
  • On FS EvidenceQA, PathAnchor achieves the highest observed score among six tested systems.Source: §5 · p. 4
  • The matched ablation finds improved source recall and opened-evidence citation provenance alongside fewer tool calls.An explicit evidence boundary makes unsupported requests visible, while final citations remain tied to opened article records.Source: §5 · p. 4
Loading 2609.38766v1…