Many engineering fields hit the same physical limit: a device heats quickly but cools slowly, because heat leaves a solid only by diffusion. Researchers in soft robotics, phase-change thermal storage, hydrogen storage beds, electrolysis stacks and electric motors all study this "thermal cycle latency," but each field names it differently and cites mostly its own papers. Searching by keyword misses work that uses another field's terms. Following citations, also called reference chaining or snowballing (reading a paper's bibliography and the later papers that cite it), misses work outside the starting field, because citation graphs cluster tightly around each community.
We propose a retrieval method for scoping reviews (structured surveys that map what a heterogeneous literature contains, rather than pooling effect sizes) that gets across these silos by matching the physics instead of words or citations. When citation traversal stops finding new papers, an LLM agent reads the full texts it has already accepted and extracts a physical signature: what is heated, its characteristic thickness, its thermal diffusivity, and the reported heating and reset times. It then screens a broad candidate pool for other papers whose reported numbers show the same diffusion-limited reset, regardless of what they call it. Papers found this way become new seeds, so chaining restarts inside the next silo. The same extracted numbers also give scoping reviews something they currently lack: an automatic physical-plausibility check on the values entered into the review's data table.
Core Claim
Our central claim is that, in fragmented engineering literatures, multi-hop citation chaining stops early. It reaches a closed set of papers inside the first disciplinary cluster it enters, and that closure is mistaken for completeness. A relevance test based on dimensionless physics breaks the closure and raises recall without the precision collapse of broad keyword search.
The diagnostic quantity is the Fourier number of the reset,
$$\mathrm{Fo} = \frac{\alpha\, t_{\text{reset}}}{L^2}, \qquad \alpha = \frac{k}{\rho c_p},$$
where $t_{\text{reset}}$ is the reported cooling or recovery time, $L$ is the characteristic thickness through which heat must leave, and $\alpha$ is the thermal diffusivity. The diffusivity is built from conductivity $k$, density $\rho$ and specific heat $c_p$. The standard diffusion scaling says a reset limited by conduction has Fo of order 0.1 to 1. A paper reporting a much smaller Fo is limited by something else, or it contains an error.
Two examples from different fields show why this test works where vocabulary fails. The 3.2-second origami switch reset in Origami-based integration of robots that sense, decide, and respond and the melted-layer barrier in phase-change storage described by Kailiang Huang et al. share no keywords and probably no citations. Both are nonetheless conduction-limited resets that the signature can recognize.
Positioning Against Prior Work
The closest work motivates the proposal and also sets its targets.
Automated citation searching in systematic review production: A simulation study (also indexed as Rajit et al.) ran one-hop API chaining from seed papers across 27 biomedical, environmental and social-policy reviews. It found higher precision than conventional search but recall of only 35.79%. That study used no semantic filtering, made only one hop, and covered no engineering domain. We add multi-hop traversal, full-text screening and the physics criterion, and we test on thermal engineering.
Two agentic systems couple citation traversal with LLM judgment, but only on computer science abstracts. Evaluating and Improving LitLLMs with Deep Research combines bidirectional graph expansion with LLM ranking and reports more than 5x coverage over keyword search. PaSa trains a crawler agent that follows citations alongside a selector agent that filters candidates. Neither system extracts quantitative variables, and both judge relevance by topical similarity of abstracts. Our relevance criterion is a physical regime computed from full-text numbers. This is the specific difference, and it is also the thing most likely to fail.
Two papers on citation-graph structure, Dornberger and Kanakia et al., report that bidirectional traversal converges into closure graphs that reduce bias. We treat convergence as a hypothesis to test, not a guarantee. If those closures are local to one discipline, the vocabulary divides documented by Ghanbarzadeh et al. should reappear as divides in the citation graph.
On the methodology side, two reviews define the setting and the data-quality problem. Annual Operation Energy Efficiency Benchmarking of Chilled Water Plants is a thermal-domain systematic review built from keyword search over two databases. A Scoping Review of Publicly Available Language Tasks in Clinical NLP documents poor reporting of exactly the parameters a data table needs.
Research Program
We will build a pipeline that alternates two modes. In the chaining mode, an LLM screener runs forward and backward citation traversal from seeds over OpenAlex and Semantic Scholar until a round adds no relevant papers. In the bridging mode, the agent first extracts the signature from each accepted full text. The signature records the heated body, $L$, the material and its $k$, $\rho$, $c_p$ (or $\alpha$ directly), the heating time, the reset time, and whether cooling is active. It then queries a deliberately broad candidate pool built from generic physical terms such as "cooling time," "cycle," "conduction" and "transient." Candidates are screened by whether their full texts report a heating–cooling cycle whose computed Fo falls in the conduction-limited range and whose reset is the stated performance limit. Accepted candidates that land in an unvisited citation community, detected by graph clustering, become new seeds.
Extracted values then go into a data-charting matrix, meaning the review's standardized extraction table. Any row whose Fo is physically implausible is flagged for human audit. This targets the gap the file identifies: scoping reviews extract numbers with no automatic sanity check.
The application regime spans the thermal-latency literature profiled in the file, which we expect the gold standard to reach across several silos:
- soft and liquid-crystal-elastomer actuators (Hao et al., Song et al.);
- shape-memory alloy muscles (SMA artificial muscle with active and passive heat dissipation);
- thermal logic (Yan et al.);
- hydride beds (Dornheim et al.);
- electrolysis stacks (Jiang et al.);
- motor windings (Van Damme et al.).
The gold standard must not be built with our method. We will assemble it from independent subdomain experts. Each expert runs an exhaustive multi-database search in their own field's vocabulary. The pooled union is screened at full text by two reviewers, with Cohen's kappa reported for agreement. We will estimate how many relevant papers remain unfound with capture–recapture across the independent teams, so recall is not capped at what any single search could see. Seeds will come only from soft robotics, which is the hard case: every other silo must be reached by crossing. The output is a PRISMA-ScR-compliant scoping review of thermal cycle latency, together with a released gold set.
Baselines
The naive reference is conventional multi-database Boolean search with one field's vocabulary, the setting where Rajit et al. give the reference standard 100% recall by construction and 0.83% precision. Here we run it honestly from the soft-robotics vocabulary alone, which shows what a single-silo reviewer actually retrieves.
The direct prior-art baseline is one-hop API chaining from the same seeds. On non-engineering reviews it reported 35.79% recall, 2.57% precision, F1 0.048 and F3 0.028, against F3 0.077 for conventional search (Rajit et al.). We will rerun it on our gold set rather than assume those numbers transfer.
The strongest existing method is LLM-screened bidirectional multi-hop chaining in the style of Deep Research LitLLMs, run to convergence. We will also include PaSa as an agentic search competitor. PaSa reports a 37.78% relative Recall@20 gain over Google with GPT-4o paraphrasing and 85% selector F1 on computer science queries. No reported result was found on Lacuna for either system on engineering or thermal literature. All LLM baselines will use the same backbone model, the same seed set, the same candidate APIs, a matched token budget and the same full-text access. Abstract-only variants will be reported separately so that full-text access alone does not get credited to the physics criterion.
The key ablation keeps everything in our pipeline fixed, including the convergence-triggered bridging step and the reseeding. It replaces only the relevance test: the Fo-based physical screen is swapped for LLM topical judgment of the same full texts, given an instruction describing thermal cycle latency. Our method beats the field only if it exceeds converged multi-hop chaining by a clear margin in recall at matched screening workload, measured as papers a human must check per relevant paper found. It must also do better on the ablation, specifically in how many gold-set citation communities it reaches.
Why not just…?
Why not just build a cross-disciplinary synonym list and search with it? This is a real threat. A thesaurus containing "cooling lag," "thermal relaxation time," "Kapitza resistance" and similar terms might recover most silos cheaply. The difficulty is that the list can only be written by someone who already knows the silos exist. The independent-teams gold standard is effectively the result of that expert thesaurus. We will therefore run a thesaurus baseline compiled by one non-expert reviewer, with an LLM's help, from the seed papers only. If it matches our recall at similar precision, the physics criterion is unnecessary and we will report that.
Why not just use dense embedding retrieval over abstracts? Embeddings capture meaning beyond exact words and may bridge vocabulary on their own. However, abstracts often omit the reset time, thickness and material constants that define the regime. That is the abstract-only bottleneck the file notes for PaSa and Deep Research. We include an embedding-retrieval-plus-chaining baseline at the same budget. If it reaches as many communities as our method, the claim fails.
Why not just chain for more hops? One-hop chaining clearly underperforms, and multi-hop convergence might suffice. The converged multi-hop baseline tests this directly. Our hypothesis predicts that it stalls inside the seed community. Measuring how its recall changes with each added hop, broken down by community, will show whether that happens.
Proposed Experiments and Decisive Checks
The primary experiment is the thermal-latency gold set. We report recall, precision, F1, F3 and screening workload, plus the fraction of gold-set citation communities reached from soft-robotics-only seeds, for every baseline and the ablation.
Two external checks follow. First, we try to rediscover the 124 included publications of the chilled-water review (Annual Operation Energy Efficiency Benchmarking) from a handful of its seeds. This tests generality within thermal engineering on a gold set we did not construct. Second, we run on the clinical NLP review (A Scoping Review of Publicly Available Language Tasks in Clinical NLP) as a negative control. That literature has no physical signature, so bridging should add nothing there. If it helps, the gain comes from somewhere other than the physics.
The charting audit has humans verify a random sample of extracted values together with every Fo-flagged row. The flag must show markedly higher error rates than unflagged rows to count as a useful sanity check.
The decisive result is the ablation. If the physics screen and topical LLM judgment reach the same communities, then the contribution is simply "restart from new seeds after chaining converges." That is a smaller but still publishable finding, and we would state it as such.
References
- Origami-based integration of robots that sense, decide, and respond
- Macro-Encapsulated PCM Cylinder Module Based on Paraffin and Float Stones
- Automated citation searching in systematic review production: A simulation study
- Evaluating and Improving LitLLMs with Deep Research
- PaSa: An LLM Agent for Comprehensive Academic Paper Search
- What Are the Determinants of China’s Export Growth?―A Systematic Review of Chinese Literature (1997-2016)
- Mitigating Biases in CORD-19 for Analyzing COVID-19 Literature
- Application areas of multi-user virtual environments in the healthcare context
- Annual Operation Energy Efficiency Benchmarking of Chilled Water Plants: A Systematic Review of Global Cases (2014–2025)
- A scoping review of publicly available language tasks in clinical natural language processing
- A Review of Smart Materials for the Boost of Soft Actuators, Soft Sensors, and Robotics Applications
- 3D printing of liquid crystal elastomers-based actuator for an inchworm-inspired crawling soft robot
- Research on a Novel Shape-Memory Alloy Artificial Muscle with Active and Passive Heat Dissipation
- Research and development of hydrogen carrier based solutions for hydrogen compression and storage
- Holistic Dynamic Modeling and Simulation of Alkaline Water Electrolysis Systems Based on Heat Current Method
- Performance comparison of Axial Flux PM machine with Anodised Aluminium Foil and Round Copper Wire