acceptodds
Under review as a conference paper at ICLR 2027

From Fragments to Functions: Evidence-Linked Trajectories for Tracing Scientific Recombination

Abstract

Scientific designs often combine functional relations first reported in separate papers, yet literature extraction rarely distinguishes a new local mechanism from a new arrangement of established ones. We ask how such recombination appears in the scientific record and whether historically supported but unreported paths precede later complete reports. To study these questions, we construct source-linked Material–Sensor–Signal–System trajectories from full-text conductive-fibre flexible-sensor papers. A paper-disjoint evaluation on 200 gold papers assesses source fidelity and record coverage. Of 1,302 complete trajectories first observed in 2022–2025, 78.4% combine three relations already observed before 2022. In four rolling-origin tests, a source-linked relation-support score places 29–36 future complete reports among its top 100 candidates, versus 19–29 for term popularity and 2.5–3.0 expected under random ranking. These temporal analyses make functional recombination observable and test whether local precedent anticipates complete-path reports.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific literature mining, evidence-grounded information extraction, scientific discovery, temporal backtesting, materials informatics, flexible sensors

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JGNrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper develops an LLM-based scientific knowledge discovery approach based on the proposed MS3MS^3 schema. It first extracts evidence-linked functional tuples from scientific literature and then uses the tuples to build scientific knowledge recombination and rank unreported combinations that appear in the future literature. The idea itself sounds plausible, but the methodological justification, novelty, and validation present a few concerns from my perspective.

Critical strengths
  1. This paper looks at scientific knowledge recombination from a functional perspective, which is interesting and has the potential to help researchers identify combination so established relationships even when experimental evidence is not appearing.
  2. The study adopts a temporal evaluation strategy by freezing historical features at each cutoff to investigate future literature outcomes. This is a practical and functional design.
Critical weaknesses
  1. The general idea of extracting structured relations from literature and combing them to identify knowledge gaps is not entirely new. The paper does not sufficiently differentiate its contributions from the existing approaches or establish the novelty of the proposed MS3MS^3 representation and evidence-linked tuples.
  2. The main experiments primarily just compare direct LLM generation, single-stage fine-tuning, and relatively simple ranking baselines. The comparison with more recent approaches, particularly knowledge graph link prediction and path-ranking approaches, are needed. Also, more baseline and ablation studies should be in place.
  3. The reliability of terminology normalization and trajectory assembly is not fully validated, which may lead to error propagation in downstream tasks.
Questions for authors
  1. Why is the functional schema specifically defined as MS3? Is this choice made based on an established theoretical framework or and empirical analysis of this domain? The authors should explain why these four roles are appropriate and ordered like the current form.
  2. The proposed framework adopts three supervised adaptors, but individual performance and contributions are not evaluated, reporting this is important for the initial extraction and normalization stages, as errors introduced early can propagate into following analyses.
  3. How were the 100 papers used for initial annotation and 200 evaluation papers samples? The authors also need to provide inter-annotator agreement stats and additional robustness analysis. Also, although I understand that annotation is a time and resource consuming task, the small number of samples may not be significant enough to justify and generalize the findings.
  4. Given that future publication is used as the outcome for evaluating candidate trajectories, how do the authors distinguish literature evidence from real scientific feasibility in practical use?
Reviewer_UxR6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors start from an observation about how engineering knowledge accumulates. A working device rarely appears in one paper. One group shows a material can sense strain, another shows strain sensing can be read out as a change in capacitance, a third shows capacitance readouts work for health monitoring. The complete design, material through application, may go unreported for years even though every link in the chain already has evidence. The literature holds the pieces but no index of which end-to-end designs have been assembled and which have not. The aspiration is to build that index from full text, with every link tied back to the sentence or figure that supports it, and then use it to ask two questions. 1) How much of what looks like new science is actually a rearrangement of established links? And 2) can we spot the assembled-but-unreported designs before someone reports them?

The author(s) then built the index for one field, conductive-fibre flexible sensors, using a language model split into three stages so that nothing enters the record without a citable source. They showed the staged pipeline recovers more of what papers actually report than a single fine-tuned model. Applying the index across time, they found that most newly reported complete designs reuse links that were all already known, which supports the view that integration, not mechanism discovery, is the dominant mode of progress in this corpus. They then froze the record at past years and asked whether a simple evidence-weighted score could have pointed to designs that were later reported. It did, well above chance and somewhat above counting edges. Finally, they built one device suggested by the method, as proof that the workflow can reach the bench. The result is a careful descriptive study of recombination with a modest predictive add-on, not a new learning method.

Critical strengths
  • The temporal discipline associated with the research design and analysis feels strong here, involving frozen relation sets per year, paper-disjoint evaluation, the base frozen for the whole observation year, and separation of exact and system-validated endpoints.

  • The relation-novelty versus path-novelty distinction is conceptually clean, and the records are auditable (but for a small and specific cast). The authors label the validation-cue reranker as retrospective rather than claiming it (which felt honest).

Critical weaknesses
  • The ranking result is small. Against the three-edge-count control, the method adds 0–8 closures at top-100 and loses in 2018. The score’s weights (0.30, 0.15, 0.40, 0.10, 0.05) and the reranker constants in Eq. 5 (120, 20, 0.20) arrive with no tuning protocol. With four cutoffs as the only test, tuning on the evaluation windows is the obvious worry and the paper does not rule it out.

  • There is no learned baseline. The cited Marwitz et al. concept-graph predictor and any link-prediction model on the typed relation graph seem like the natural comparators. The three-edge count is the method minus its features, but not an external baseline.

  • The worked example (CNT composite -> strain -> capacitance -> health monitoring) sits at rank 151, outside the top-100 review budget the paper uses as its headline metric. The showcase undercuts the ranking story.

  • The 78.4% recombination figure depends entirely on term normalization. Coarse vocabularies make nearly everything a recombination. Vocabulary size per role and sensitivity to granularity are absent.

  • Extraction metrics are 0–4 expert rubrics with no inter-annotator agreement, no precision/recall against the gold records, and no definition sharp enough to reproduce.

  • Device study: n=3 per architecture, representative traces only, no statistics. It is an anecdote, which is fine if presented as one (but its not currently).

  • Single (narrow but apparently important) domain; the generalization claim in the discussion is asserted, not tested.

Questions for authors
  • How were the score weights and Eq. 5 constants chosen, and on what data?

  • What was Inter-annotator agreement on the 200 gold papers?

  • What was the vocabulary size per role and recombination rate under coarser and finer normalization?

  • Why was there no learned ranker?

  • Is the relation-support versus three-edge gap significant under paired bootstrap (or another, stricter statistical evaluation)?

Reviewer_6V9LOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies functional recombination in the scientific literature by extracting MS³ trajectories from full-text papers on conductive-fibre flexible sensors, using a factorized three-adapter pipeline over Qwen2.5-72B-Instruct. It shows that many complete MS³ paths recombine previously observed relations, that the pipeline outperforms single-stage fine-tuning and direct generation on expert-annotated gold papers, and that ranking unreported but fully supported paths recovers later published designs above simple baselines, with a small fabrication case study illustrating the workflow.

Critical strengths
  • Separating local relation novelty from complete arrangement novelty is a useful lens, and the source-linked MS³ schema that leaves incomplete paths incomplete is clean and auditable.
  • The paper uses a paper-disjoint gold split with dual expert annotation, leakage-aware temporal validation, and candid reporting of retrospective components and limitations.
Critical weaknesses
  • The ML contribution is mainly three LoRA adapters and post-generation checks around an off-the-shelf backbone, with no evidence that the framework transfers beyond one narrow subfield.
  • The ranking evaluation is retrospective, the reranker’s incremental value is unclear, and the baselines lack embedding-similarity, link-prediction, learned-ranking, or LLM-ranker comparisons, with no uncertainty estimates or significance tests.
  • Error propagation and term-normalization sensitivity are not quantified, gold-evaluation details are incomplete, and the physical case study is illustrative rather than confirmatory, while the paper’s three loosely integrated narratives blur the central contribution.
Questions for authors
  • How sensitive are the recombination share and ranking results to extraction errors and term normalization, and can you recompute them on manually audited or high-confidence paths under an alternative normalization scheme?
  • How does relation support compare against stronger baselines with the same frozen evidence, such as embedding-similarity scorers, link-prediction, learned rankers, or LLM rankers?
  • Can you add a lead-time or survival analysis and a pseudo-prospective variant with validation-cue weights frozen on pre-2018 windows, and report mean±SD for the case-study devices?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.