Evaluating a 4B open-weights local LLM for agentic DFT workflows: a literature reproducibility audit

Evaluating a 4B open-weights local LLM for agentic DFT workflows: a literature reproducibility audit Shambhu Bhandari Sharma∗1

arXiv:2608.29665v1 [cond-mat.mtrl-sci] 30 Aug 2026

1

University College London, Department of Earth Sciences, London WC1E 6BT, United Kingdom September 1, 2026

Abstract Agentic workflows in materials science relying on hosted commercial models face severe reproducibility, economic, and data-privacy constraints. To explore fully local agentic science, this work evaluates an open-weights qwen3:4b model executing an autonomous scientific pipeline across varying hardware constraints. Applied to pentagonal two-dimensional materials, the system extracts parameters from unstructured text, translates them into density functional theory (DFT) inputs, and drives simulations to convergence under a strict neurosymbolic architecture where agents propose and deterministic code disposes. The workflow is guarded by verbatim text grounding and multi-pass inference unions to counteract hardware-induced structural collapse. Evaluated against 201 expert judgements, the extractor achieves 95.7 % precision (95% CI 90.3–98.1 %) and 67.3 % recall (59.8–74.0 %), ensuring extracted parameters are strictly factual. However, precision identifying absent parameters does not exceed 47.0 %, establishing that the measured omission rate constitutes a loose upper bound on true literature incompleteness. Across three hardware configurations, complete GPU residency governs extraction quality more fundamentally than weight or cache precision, raising Matthews’ correlation from 0.414 to 0.530 at fixed quantisation, and to 0.560 with an unquantised cache. A corpus-scale audit indicates only 19 (33.3 %) of the 57 studies are reproducible in principle, reporting every method parameter needed to re-initialise the calculation. Driven to convergence, the workflow reproduces published lattice constants with a mean absolute relative error of 2.3 % where the relaxed structure retains its prototype, establishing that lightweight open-weights models can reliably drive autonomous agentic workflows when bounded by deterministic code gates.

Keywords: agentic workflows; open-weights LLM; local inference; neurosymbolic architecture; reproducibility auditing; density functional theory; information extraction; penta-monolayers; two-dimensional materials

1

Introduction

point math during inference [9]. However, almost all of these agentic systems share a critical dependency: they are driven by hosted, closed-weight frontier models. This reliance on commercial APIs introduces significant operational and epistemological vulnerabilities. A workflow tied to a commercial key ceases to run when the endpoint is repriced, the model is reversioned, or the key expires, meaning long-term reproducibility is never guaranteed. Furthermore, metered inference is billed per token, making iterative debugging prohibitively expensive. Frontier inference also carries a measurable energy cost per query [10]. Crucially, much computational chemistry operates on isolated networks where proprietary data cannot be sent to third-party endpoints, rendering cloud-hosted models unusable. Fortunately, capability on agentic tasks appears to depend more on interface design than raw param-

Large language model (LLM) agents are already driving autonomous workflows across computational chemistry and materials science. Recent pipelines span the entire research lifecycle, from coupling expert tools for chemical planning and executing robotic synthesis [1–3], to orchestrating density functional theory (DFT) calculations through hierarchical multi-agent frameworks [4, 5]. Within materials informatics, transformer-based architectures have firmly established the viability of extracting composition-property records at corpus scale [6, 7]. These systems depend on grammar-constrained decoding to ensure parseable output and strict grounding techniques to verify generated statements against source texts [8], and must confront known reproducibility issues with non-deterministic floating1

eter count. Recent studies demonstrate that tool- cache are quantised. Weights are statically quanaugmented smaller models can outperform larger tised to Q4_K_M, and, on the 4 GiB GPU, the KV unconstrained baselines [11]. A 4B retrosynthesis cache is additionally quantised to 4-bit (q4_0) to planner surpasses a 7B reasoning baseline given a accommodate the sequence length. Residency is not constrained tool interface [12], and 7-8B chemistry guaranteed on a GPU of this capacity: the placeagents on a consumer GPU have matched datacentre- ment of individual layers between device and host is hosted equivalents [13]. determined by the inference server at load time and While the ∆-project established that independent is not fixed by configuration. To decouple the two DFT codes agree closely given identical inputs [14], quantisation axes from this uncontrolled residency it presupposed that methodological parameters are condition, the pipeline is additionally evaluated on a completely and transparently reported in the source Tesla T4 (15 GiB), which accommodates the model text. Whether computational literature, spanning in full and permits weight and cache precision to be both peer-reviewed articles and preprints, provides varied independently (Section 3.2). sufficient methodological transparency to enable autonomous simulation recomputation across an entire Inference parameters and grammar concorpus remains completely untested in density func- straints. Operating within these physical contional theory. straints, all agent generations use greedy decoding This work applies a fully open-source, local agen- (T = 0) with a generation limit of 2048 tokens, a tic workflow to address this critical gap: how of- cap on output; internal reasoning (‘thinking’) is exten does a first-principles study report sufficient pa- plicitly disabled to prevent models from expending rameters to permit independent replication? Every token budgets on chain-of-thought sequences that component runs locally and unmetered, on an open- the grammar discards. The context window dynamiweights qwen3:4b model [15] on a single local GPU cally accommodates the document length; truncated via Ollama [16] and the SIESTA electronic structure prompts are retried at double capacity, and those package [17]. Bounding the model’s autonomy is exceeding the 32,768-token ceiling are recorded as the central architectural constraint: agents select failures. Output is strictly constrained at the logits among named options rather than free-form argu- level using a grammar compiled directly from the ments, and every proposal passes a deterministic target JSON schema (GBNF). Two schema propcode gate before reaching a calculation or tally [18]. erties are lost during compilation and require exThe contributions are threefold: a neurosymbolic plicit programmatic compensation. First, semantic architecture spanning extraction, input translation, field descriptions are stripped, necessitating their reand autonomous execution; a taxonomy of silent injection into the zero-shot prompts. Second, fields failure modes including head-truncated prompts carrying default values are not inherently marked and structural collapse under incomplete GPU res- as required. All schema fields are explicitly declared idency [19]; and a case study auditing 57 first- as required prior to compilation; for the extraction principles studies of pentagonal lattice-based [20] schema, where this forced the model to emit a null two-dimensional materials, recomputing published object for every parameter, only the fields determinlattice constants against fresh SIESTA calculations. ing reproducibility are required. String length and pattern constraints, which the grammar compiler rejects, are removed from the compiled schema and 2 Methods enforced on the returned object instead. This section details the system architecture and hardNeurosymbolic architecture: agents propose, ware constraints governing an automated workflow code disposes. Because syntactic schema comthat spans four sequential phases, from document pliance does not guarantee factual correctness, the acquisition to physical simulation (Figure 1). pipeline enforces a strict neurosymbolic hierarchy: agents propose, and deterministic code disposes. To 2.1 System architecture and con- prevent arbitrary code execution, each agent is restricted to a named allow-list of read-only tools, and straints tool arguments are programmatically bound. The Hardware constraints and model quantisa- agent selects a routine by name, arguments are poption. All extraction and translation agents run ulated from the current verification state, and tool qwen3:4b locally via the Ollama inference server [16] execution rounds are strictly capped. Every returned (backed by llama.cpp [21]) on a NVIDIA RTX 3050 object passes a deterministic gate: Pydantic [22] valiLaptop (4 GiB), the constrained deployment target dation of the schema and length constraints withheld the method is designed for. To execute the model from the grammar, followed by explicit range and within the strict VRAM capacity of this hardware vocabulary checks in Python. Agent proposals failclass, both weights and the attention key/value (KV) ing these checks are corrected or refused, and the 2

Retrieval

Extraction

Translation

LLM Parser

Prototype

Full-text Fetch

Multi-Pass

Translator Agent

Format Gate

Grounding Gate

Format Gate

OpenAlex

Recomputation Convergence Agent Parity Check

Critic Agent

Diagnostician Agent

Figure 1: The autonomous audit pipeline. Retrieval fetches and screens OpenAlex full texts. Extraction parses and deterministically grounds parameters across three passes. Translation maps extracted claims and topological prototypes to SIESTA inputs. Recomputation drives physical simulations to convergence, monitored by diagnostician and critic agents. The workflow concludes with a parity check against the literature. Yellow boxes denote LLM agents; every other box is a deterministic code step. Red crosses indicate discarded records. intervention is logged; values extracted from a paper are instead flagged and recorded unaltered, since an out-of-range value may be an extraction error or the paper’s reported choice, and the two are separate findings. Each agent call is written to a JSONL audit log recording what the agent proposed, what was finally applied, and which gate fired on any difference, from which override counts per agent are tallied.

functional”, or a named DFT code), and match a pentagonal structural pattern. Surviving records are deduplicated (grouped by arXiv identifier, otherwise by normalised title), preferring non-arXiv DOIs and higher citation counts. The published DOI is retained as the primary key, with the OpenAlex identifier as fallback; arXiv identifiers and open-access URIs are preserved to facilitate access to paywalled full text.

2.2

Full-text acquisition and lexical validation. Full-text acquisition prioritises institutional repositories and arXiv. Documents are parsed natively (PDFs via PyMuPDF [24], HTML via BeautifulSoup [25]) into raw text arrays. To ensure only verified, full-length manuscripts proceed, a document is accepted if at least four of nine structural markers (e.g., “abstract”, “conclusion”) are present; a document shorter than 12,000 characters must additionally contain at least three methodological markers (e.g., “cutoff”, “pseudopotential”), and documents yielding fewer than 2,000 characters are discarded. Supplementary information is never fetched; a parameter documented solely in the supplementary material is recorded as absent (Section 4).

Phase 1: Literature retrieval

Metadata retrieval. The corpus is constructed via a three-stage pipeline, beginning with the OpenAlex API [23], selected over proprietary databases such as Scopus for its keyless, freely redistributable access and its indexing of preprints alongside publisher records. A curated set of 16 seed terms is applied to titles and abstracts: nine structural terms targeting topological families (e.g., “Cairo tiling”) and seven heavily studied compositions (e.g., penta-SiC2), retrieving records whose abstracts designate solely the chemical formula.

Filtration and deduplication. Each record passes four deterministic filters. It must be an article or preprint, and its DOI must not match patterns indicating supplementary information. Furthermore, Deterministic subsampling. Where a subset is it must contain a first-principles indicator in its necessary for evaluation, it is drawn deterministititle or abstract (e.g., “first-principles”, “density cally via a SHA-256 hash of each paper’s identifier. 3

This mechanism yields a strictly reproducible sample with K = 3. This monotonic union ensures a coluncorrelated with year, venue, or citation count. lapsed pass (Vk = ∅) withholds evidence rather than falsely voting for absence, and ensures widening coverage cannot admit ungrounded claims. What the 2.3 Phase 2: Parameter extraction union cannot repair is misattribution, since a corFor each retrieved manuscript, the extraction agent rectly quoted sentence assigned to the wrong field evaluates a schema of twenty distinct computational satisfies G in every pass. method parameters, returning for each a boolean indicating presence, the extracted numerical value The deterministic grounding gate. Extraction or string, and the verbatim sentence supporting the is governed by a strict, one-directional evidence gate: extraction. A paper is defined as reproducible in asserted parameters must supply verbatim textual principle if it reports every parameter required to evidence; absent claims are never challenged. Prior re-initialise the calculation. Because the required to verification, quotes are normalised (N ) by stripparameter set is code-dependent (e.g., a plane-wave ping model framing, lowercasing, and converting cutoff and a real-space mesh cutoff are not inter- hyphens to spaces. Writing w for the words of the changeable), these specific requirements are defined normalised quote and n = 6, evidence e is accepted deterministically in the pipeline’s configuration. For- against document d when: mally, for a paper p reporting code c(p), the scored  set is   1, N (e) ⊆ N (d),
S(p) = Runiv ∪ R2D ∪ B c(p) ∪ {code}, (1) G(e, d) = 1, |w| ≥ n ∧ ∃ i : wi:i+n−1 ⊆ N (d),   0, otherwise. where Runiv denotes universal DFT settings (functional, pseudopotential, k_mesh, (4) force_threshold, energy_threshold), R2D If G(e, d) = 0, the boolean is reset (reps ← 0) and denotes two-dimensional monolayer con- the override logged. straints (vacuum_spacing), B(c) = RPW (plane_wave_cutoff) for plane-wave codes, Unit-aware value support. Where the groundRNAO (mesh_cutoff, basis_size) for numerical- ing gate verifies the existence of the quoted senatomic-orbital codes, and ∅ where no code is stated. tence, a secondary value-support diagnostic evaluThe reported fraction is ates whether any reported numerical value mathematically follows from that text within a relative { s ∈ S(p) : reps (p) } , (2) tolerance. Operating strictly as an observational f (p) = |S(p)| diagnostic, this unit-aware check does not override where reps (p) ∈ {0, 1} indicates whether parameter the primary boolean extraction; the value-grounding s is reported in paper p, and p is reproducible in gate described below does. To correctly evaluate standardized numerical values, the algorithm applies principle exactly when f (p) = 1. deterministic, unit-aware conversions (e.g., Ry → eV) Full-context information extraction. To pre- via a predefined unit conversion mapping, decoupling serve global context for parameter disambiguation, semantic accuracy from syntactic formatting. the full text (excluding the bibliography and truncated to a 120,000-character limit) is supplied to the Numeric value grounding. A second gate govlanguage model in a single prompt context, strictly erns the numerical content of a claim. Where the circumventing retrieval-augmented generation. evidence gate establishes that a quoted sentence occurs in the source, the value gate requires that Multi-pass extraction and grounded union. the asserted magnitude itself occur verbatim within Given the variance of greedy decoding, full-context the document. A value that cannot be so located is inference is executed in three independent passes. cleared while the surrounding claim is retained, so Outputs are combined via a grounded union rather that an ungrounded magnitude is removed without than a majority vote to mitigate whole-paper col- discarding the evidence that the parameter was dislapse. Writing Sk for fields asserted in pass k, ex cussed. The gate is deterministic, and consequently for the supporting textual evidence quoted for field reproducible without re-invoking the model; it is x, d for the source document text, and G for the applied at extraction time and, for corpora predatevidence gate, the validated set of a pass (Vk ) and ing its introduction, retrospectively under identical criteria. the shipped record (E) are: Vk = { x ∈ Sk : G(ex , d) = 1 },

E=

K [ k=1

Quantifying systematic error: precision versus recall. Finally, to evaluate the extractor’s (3) baseline recall rate against these automated gates,

Vk ,

4

an expert annotator produced hand-labelled ground truth for a targeted subset of core schema fields across a subset drawn from the hash-sampled evaluation set (retaining only viable manuscripts and applicable judgements). The annotator labelled from the published PDFs rather than from the converted text, so that a parameter lost in conversion could be separated from one the authors never reported; the labels are a single expert’s judgements, and no second-rater agreement statistic is claimed. A second batch of papers was labelled from the retrieved text itself, recording only the reported/not-reported decision, which is the sole input to precision and recall.

from topological prototype geometries parameterized with published lattice constants, and executed with standard SIESTA configurations.

Automated convergence and diagnosis. Convergence testing is an autonomous loop whose length is decided at runtime. Following each variable-cell −1 relaxation (target force tolerance 0.02 eV Å ), the Convergence agent reads the output and selects from four actions (increasing the mesh cutoff, expanding the basis set, densifying the in-plane k-grid, or terminating execution), advancing monotonically along predefined parameter progressions (e.g., mesh cutoffs step along 200–900 Ry and bases along SZ, SZP, DZ, DZP, TZP). Termination requires successive completed steps i − 1 and i of a cell of N atoms with 2.4 Phase 3: Input translation total energy Ei and in-plane lattice constant ai to Initialisation and topological prototypes. Be- satisfy: cause author-provided coordinates are universally absent, structures are generated exclusively from |ai − ai−1 | Ei − Ei−1 ≤ εE , ≤ εa , (5) recorded topological prototypes and scaled via emN ai−1 pirical covalent radii. Compositions lacking a known pentagonal prototype are skipped. These unrelaxed where εE = 5 meV and εa = 0.5 %. Convergence is geometries are passed directly to SIESTA 5.4.2, en- never the agent’s to declare: it is evaluated arithforcing a 20 Å out-of-plane vacuum. The calcula- metically in code before the agent is consulted, and tions employ scalar-relativistic norm-conserving ON- any proposal that would make the calculation comCVPSP pseudopotentials in PSML format along- putationally cheaper is programmatically refused. side a baseline double-ζ polarised (DZP) numerical The relaxation loop is constrained to a maximum of six iterations and a four-hour wall-clock budget per atomic orbital basis. target. Non-converged calculations trigger an automated Diagnostician agent that parses error logs The Translator agent and override guardrails. and extracts supporting failure lines; in the present The DFT calculation is configured by the Translacampaign, all simulated targets achieved SCF contor agent, which maps extracted parameters onto vergence nominally without activating diagnostic SIESTA input directives. Consistent with the neurecovery. rosymbolic architecture, the agent’s proposals are strictly governed by thirteen deterministic gates. Rather than halting execution or prompting the The Critic agent and epistemic verdicts. Beagent, these gates programmatically override mal- fore a final verdict is issued, the Critic agent selects formed proposals with physically consistent fallback verification checks to run from a vocabulary of eight parameters and log the intervention. For example, options (symmetry, expected symmetry, calibration, the gates force the out-of-plane k-grid component to convergence, units, structure file, literature claims, unity for a 2D monolayer, enforce a minimum 9×9×1 and literature methods). The agent chooses which mesh for sub-unity grids, balance in-plane grid com- checks are warranted; the underlying code strictly ponents differing by more than a factor of two, and controls what data they run on. Compiling the reset mesh cutoffs outside the physical bounds of valid check names directly into the GBNF decod50–2000 Ry to a default of 350 Ry. Furthermore, ing grammar structurally prevents the model from the gates enforce semantic boundaries on simulation hallucinating non-existent tests. The pipeline issues no categorical verdicts. Becapability: parameters that SIESTA cannot natively reproduce (such as hybrid functionals (e.g., HSE, cause starting from prototype geometries and usPBE0), plane-wave cutoffs, or specific dispersion ing independent basis sets conflates methodological corrections) are intercepted and explicitly recorded differences with reporting divergence, the system reports only raw physical deviations. When calas unmappable methodological divergences. culations fail convergence or leave their prototype symmetry, the Critic marks them as inconclusive 2.5 Phase 4: Recomputation and withholds the comparison. Candidate materials are prioritized by reporting completeness. Because relaxed coordinates are Axis-matched comparison. An in-plane cell is rarely deposited, initial structures are constructed unordered, and publications do not agree on which 5

axis is a: comparing label to label rather than magnitude to magnitude manufactured deviations of several per cent where the axis-matched agreement was sub-percent. Published and recomputed cells are therefore both sorted before comparison, and a comparison is formed only where the paper reports both in-plane constants. A single reported constant cannot be assigned to an axis; pairing it with the nearer recomputed axis would select the smaller of two residuals by construction, so such targets are reported as raw deviations in the text and never plotted.

Table 1: Corpus-scale effect of combining passes on the 4 GiB GPU over 69 documents. Papers Fields Papers reported reporting none Pass 1 Pass 2 Pass 3

69 69 69

435 440 439

18 20 20

Grounded union

69

475

17

caution applies to the extraction channel, since both proportions are measured through a pipeline whose misses are not known to be distributed evenly across the two publication routes.

3

Results

3.1

3.2 Local inference stability Retrieval yield and corpus comTo quantify the stability of locally quantised inferposition

ence, the pipeline was executed multiple times over the corpus. Field-level decisions remained highly stable across passes (Figure 2). This instability manifests as an abrupt structural collapse rather than gradual semantic drift (Figure 2): a flipped boolean token forces the constrained grammar down an empty generation path. Instability results from incomplete GPU residency rather than cache precision. On a 4 GiB GPU, the server placed between 23 and 37 layers on the GPU. This split varied between loads due to the uncontrolled residency condition (Section 2.1). On a GPU that holds the model, three passes with an identical q4_0 cache are byte-identical. Consequently, a collapsed pass records zero parameters, falsely inflating the non-reporting rate. To mitigate this hardware-induced variance, the final dataset was constructed via a grounded union of three independent passes (Table 1). This approach leverages the monotonic union defined in Section 2.3 to inherently recover false negatives without adjudication. Across three passes, 21 of 69 papers suffered total extraction collapse at least once. The union recovered parameters for 4 of them, while the remaining 17 failed in every pass (Table 1), which the pipeline records as complete non-reporting; the labelled subset does not permit these to be separated from consistent extraction failure. Overall, the union yielded 475 supported fields, recovering 35 fields over the best individual pass (440 fields). The merge is not a plain superset: resolving plane-wave and real-space mesh cutoff collisions removes nine of the 484 fields a naive union would carry (Section 2.3). At a 24,576-token context window, an f16 cache requires 3.38 GiB of VRAM, forcing severe memory offloading on a 4 GiB GPU. On hardware that holds both (the 15 GiB Tesla T4), the f16 configuration is 27 % faster than q4_0. However, because f16 cannot fit on the 4 GiB GPU, the q4_0 baseline was maintained for consumer-hardware metrics.

Execution of the retrieval pipeline (16 August 2026) returned an initial candidate set, yielding a baseline corpus of 298 unique papers. Of the 135 manuscripts with an open-access location, automated retrieval returned 403 Forbidden errors for 36 records, and validation stripped the remainder: 21 failed the minimum length threshold (Section 2.2), 7 were non-article-like landing pages, and 2 returned other HTTP errors. This yielded 69 full-text manuscripts, of which 41 were sourced via arXiv. During Phase 2, 62 of these full texts were successfully processed by the extraction agent. However, the agent subsequently classified 5 records as lacking a genuine pentagonal structural motif, representing a model-dependent classification rather than a deterministic filter. Excluding these non-pentagonal structures defines the final audited population of 57 papers. Because a paper’s arXiv copy and its journal version may report differently, the audited set was partitioned by publication status (17 preprints with arXiv DOIs versus 40 journal articles with publisher DOIs) and compared against the strict reproducibility criterion. The analysis demonstrates significantly higher reporting completeness in preprints: 10 of 17 (59 %) preprints state every required parameter, compared to 9 of 40 (22 %) publisher versions (difference of 36 percentage points, 95% CI 10–63 pp, Fisher’s exact test, p = 0.013). The mean parameter reporting rate is 84 % for preprints versus 62 % for journal articles. This divergence is consistent with structural publishing constraints rather than disparities in scientific diligence. Journal formatting limits routinely displace methodological parameters into Supporting Information files, which the retrieval architecture deliberately excludes (Section 2.2). Consequently, the measured disparity likely reflects where parameters are physically located rather than whether they are entirely omitted by the authors, and the same 6

6

x

Difference (pass A − pass B)

16 14

y=

Parameters reported (pass B)

18

12 10 8 6 4 n = 69

2

identical: 56/69

0 0

2

4

6

8

4 2

Parameters reported (pass A)

bias (-0.07)

0 −2 −4 −6

10 12 14 16 18

+1.96 SD (+5.1)

−1.96 SD (-5.2) 0

2

4

6

8

10 12 14 16 18

Mean parameters reported

Figure 2: Repeatability of extraction over two identical passes on the 4 GiB GPU. Left: paired per-paper counts against the identity line. Right: Bland–Altman plot showing disagreement is independent of parameter count. would wrongly charge to the extractor. Scored conventionally (presence as positive), every configuration exhibits high nominal performance. When evaluated against parameter absence (the primary quantity measured in this audit), precision does not exceed 47.0 %, indicating that the majority of parameters flagged as unreported are present in VRAM Time GPU W / KV (GiB) (s) Layers the source text. Full GPU residency, not cache precision, produces RTX 3050 (4 GiB) Q4_K_M/q4_0 3.61 120 23–37 Tesla T4 (15 GiB) Q4_K_M/q4_0 3.61 58 37 the larger effect: holding the cache at q4_0 on a Tesla T4 (15 GiB) Q4_K_M/f16 6.44 43 37 fully-resident GPU raises recall and MCC (Table 3). Tesla T4 (15 GiB) Q8_0/f16 8.10 50 37 Unquantising the cache then trades presence recall for conservatism, rendering it the configuration least Hardware holding the full model separates the prone to asserting ungrounded parameters. The pipeline’s baseline performance (precision, two quantisation axes (Table 2). Doubling weight recall, and F1 ) is detailed in Figure 3. The precision from Q4_K_M to Q8_0 costs 1.7 GiB and near-perfect precision confirms the structural ef17 % more time. (The effect of cache quantisation ficacy of the grounding gate. In the final union on extraction accuracy is evaluated in Section 3.3.) dataset, 473 of 475 asserted fields carry correctly matching textual evidence. The five false positives 3.3 Extraction accuracy and report- behind the precision score were tightly localized (pseudopotential_type 2, energy_threshold 2, ing completeness k_mesh 1). Precision, recall, and grounding taxonomy Mathematically, the measured non-reporting rate is defined as the sum of genuine author omissions An expert annotated 25 papers against the nine reand the extractor’s false negative rate. Writing T quired parameters, yielding 201 independent judgeand O for the parameters genuinely reported and ments blind to the pipeline’s output. 2 abstract-only genuinely omitted across the labelled judgements, records were excluded. Basis size applies only to and ρ and π for recall and precision, the measured numerical-orbital codes; recall over the eight univernon-reporting count is sal fields is 69.8 %, versus 67.3 % across all nine. Two limits of presence-based metrics emerge. 1−π M = O + (1 − ρ) T − ρT , (6) First, precision records presence, not correctness: | {z } | π {z } missed of 10 labelled packages, the extractor named 8 corfalse positives rectly. Remaining errors are consequential, as package identity governs the physical meaning of ambigu- so M ≥ O whenever 1 − ρ ≥ ρ (1 − π)/π, which ous numeric cutoffs (discussed below). Second, 6 holds here by a factor of eleven: recall error domparameters (6.7 %) found in the PDF were destroyed inates precision error by an order of magnitude. in PDF-to-text conversion, which PDF-based recall Extrapolated from the labelled subset to the full

Table 2: Hardware constraints and execution profiles. Minimum VRAM includes the weight buffer, the key/value cache at the corpus’s widest window (28,672 tokens), and the compute buffer. “Layers” records how many of the 37 model layers fit on the GPU.

7

Table 3: Extraction accuracy scored against expert hand labels. Accuracy is reported under both decision orientations due to extreme label skew (only 18 % of judgements are not reported) and because parameter absence is the primary quantity this work measures. Matthews’ correlation coefficient [26] is included because a constant “reported” baseline scores zero. Reported Configuration RTX 3050 4 GiB, q4_0 Tesla T4 15 GiB, q4_0 Tesla T4 15 GiB, f16

RTX 4 GB, q4_0

Not reported

P

R

F1

P

R

F1

MCC

95.7 96.3 99.2

67.3 78.8 73.9

79.0 86.7 84.7

36.5 47.0 44.9

86.1 86.1 97.2

51.2 60.8 61.4

0.414 0.530 0.560

T4, q4_0

T4, f16

vacuum spacing (n=17) pseudopotential (n=19) stated

k-mesh (n=22)

in paper

130

122

54

35

43

5

5

1

31

31

35

Hand label

XC functional (n=25)

111

code (n=25) cutoff (n=19) energy threshold (n=15)

not stated in paper

force threshold (n=17) basis size (n=6) 20

40

60

Recall (%)

80

100

reported

not reported

Pipeline prediction

Figure 3: Extractor accuracy against hand labels. Recall is shown with its 95% CI; basis size is marked because it lacks sufficient positive support to state a rate. Dashed vertical lines mark overall recall for each configuration. The confusion matrix gives the counts from which the rates derive. 20-field schema, this performance is of the same order as figures reported for established literature extractors such as ComProScanner [7], Cactus [13], MiniCheck [8], and ChemDataExtractor [6], though on different corpora and extraction schemas. Per-field recall varies significantly: notably, basis_size was never recovered (Figure 3). As a strictly required field for that methodology, this represents a structural omission in extraction capability rather than an isolated variance. Prior to the grounding gate, 93.3 % of supporting sentences across the multi-pass union were accurately located in the raw text. Across the 169 extracted numeric values in the union dataset, 95.3 % were correctly derived from their quotes. The pipeline actively intercepts a specific taxonomy of five silent failure modes inherent to open-weights LLM extraction:

placeholder literals (e.g., 0.0 or 1.0) when parameters are unstated. 3. Hardware-induced structural collapse: Partial GPU residency perturbs logits under greedy decoding, forcing the constrained grammar to silently collapse into empty JSON outputs (Section 3.2). 4. Self-negating evidence: The model sets reported=true while generating an evidence quote explicitly stating that the authors did not report the value (observed in 30 instances, 5.7 %). 5. Memorised-value substitution and plausible sweeps: The model bypasses document retrieval by emitting canonical literature constants memorised during pretraining (e.g., substituting the ideal penta-graphene lattice constant a = 3.64 Å or penta-PdSe2 constant b = 5.92 Å into manuscripts describing unrelated compositions), or synthesising plausible smooth numerical sweeps (e.g., stepping cohesive energies from −7.1 eV to −7.7 eV).

1. Head-truncated prompts: The inference server silently discards the prompt head to retain the tail under constrained context (observed in 9 of 69 manuscripts).

2. Default-value schema hallucinations: The model satisfies the forced schema type con- The secondary value-grounding gate (Section 2.3) straints (Section 2.1) by inserting canonical successfully intercepted this behaviour, automat8

ically stripping 51 ungrounded values (16.4 % of testable valued claims) across the corpus and preforce threshold venting ungrounded physical parameters from propenergy threshold agating into downstream simulation workflows. The vacuum spacing two fields lacking evidence strings entirely represent a distinct, uncaught failure mode. pseudopotential Testing a deterministic localiser against known evcode idence sentences confirms why full-context inference k-mesh is required (Section 2.3): keyword retrieval recovXC functional ered only 58.2 % of the necessary context; doubling RTX 3050 4 GB, q4_0 the retrieval window recovered 71.1 %, yet every paplane-wave cutoff Tesla T4, q4_0 rameter field continued to fail to capture critical Tesla T4, f16 mesh cutoff context. The failure is fundamental to the physics 0 10 20 30 40 50 text: parameters such as “40 Ry” ambiguously map Papers not reporting the parameter (%) to specific cutoffs depending entirely on the underlying software code. Supplying a localized context window structurally severs this dependency. Figure 4: Parameters absent from the retrieved full Across the 57 papers, only 19 (33.3 %) are reprotext, n = 57. These rates are upper bounds on ducible in principle in the main text, with a mean non-reporting: at 67.3 % recall a fraction of these parameter reporting rate of 68.3 %. These omissions absences are extraction failures rather than omisare heavily skewed toward convergence controls (Figsions by the authors. ure 4); notably, 21.1 % of the audited papers are recorded as not naming the specific software packtical configurations nine times; the sole discrepancy age used for the calculations. In accordance with (a mesh cutoff) was corrected deterministically by Equation 6, this reflects both author omission and the convergence routine. extractor failure: on the hand-labelled subset evConvergence. Across 24 invocations, the agent ery paper names both its software package and its uniformly selected increase_mesh; because it never exchange–correlation functional, each of which the exercised alternative discretion, it did not alter any extractor records as missing in roughly one paper physical outcomes. Because loop termination is dein seven. A third channel is methodological: ten of terministically orchestrated (Section 2.1), agent colthe twelve records naming no package also report lapse cannot falsely declare convergence; it can only no functional, pseudopotential or cutoff, a pattern halt execution. Convergence typically required two absent from every paper that does name one. Such to four steps, never reaching the imposed iteration records are consistent with studies that are compuor wall-clock limits. tational but not first-principles, and are therefore Diagnostician. The agent was not invoked; no scored against parameters their methodology never calculation suffered self-consistent field (SCF) or defines. Paper-level completeness represents the structural failure. most statistically volatile metric reported: the strict Critic. It raised 18 concerns across seven physif (p) = 1 criterion (Equation 2) is acutely sensitive cal categories (six blocking errors, 12 warnings) over to single-field errors, since a single false negative nine verification records, downgrading 4 verdicts records a paper as incomplete and a false positive to inconclusive and identifying geometric symmetry as complete, so the two error channels partly cancel. breaks. In one instance, failing to select checks trigOn the labelled subset the pipeline and the annotagered a deterministic override that substituted the tor disagreed on five of twenty-five papers, in both complete set. The Critic occasionally failed to return directions, with a net difference of one. a schema-valid object, leaving those verdicts unreviewed. Because the verification rule set enforces an 3.4 Agent activity and the effect of asymmetric hierarchy (Section 2.1), it cannot override its own contrary evidence, and a collapse safely cache quantisation defaults the pipeline to the unadjusted numerical The activity and stability of the four execution conclusion. agents (Table 4) were evaluated to track cache quanInstability scales with context injection, not tisation instability across roles. output length. Collapse correlates with injected Translator. The agent mapped parameters from tool context rather than generation volume. The 12 verification records into 11 distinct sets. The code- Translator safely emits a median of ∼2400 characgate intercepted all 17 invocations because the agent ters; the Critic emits ∼1400. However, the Critic emitted standard literature notation (e.g., GGA/PBE) is uniquely injected with the full JSON output of rather than strict parser syntax. Across a single executed checks before issuing its verdict. Repeatinput over ten passes, the Translator produced iden- ing the campaign on fully-resident hardware isolates 9

Table 4: Agent activity across the recomputation campaign. The multi-agent system was evaluated in three configurations differing only in GPU and KV-cache precision: the original RTX 3050 4 GiB with q4_0 cache (12 targets), and two fully-resident Tesla T4 15 GiB runs over 9 targets: one with a q4_0 cache, and one with an unquantised f16 cache (pooling 17 total invocations across a repeated run). For a given target, the extraction and prompts remain identical. Only the Critic operates over multiple tool-calling rounds, and it alone collapsed in every configuration. RTX 3050 (q4_0) Agent

Calls

Tesla T4 (q4_0)

Tesla T4 (f16)

Collapsed

Tool calls

Calls

Collapsed

Tool calls

Calls

Collapsed

Tool calls

Translator Convergence Diagnostician Critic

17 24 0 13

0 0 — 3

— — — 48

9 15 0 9

0 0 — 3

1 — — 16

17 26 0 17

0 1 — 1

1 — — 52

All agents

54

3

—

33

3

—

60

2

—

BN2 C C2Si NiP2

6

PdSe2 a axis b axis

Critic held (not scored)

5

4

y

=

x

This work, SIESTA/DZP (Å)

the cause (Table 4). The Critic collapsed in 3 of 13 and 3 of 9 invocations on a quantised cache (evaluated across two GPUs at indistinguishable rates, p = 0.66), compared to 1 of 17 on an unquantised cache. Unlike the extraction phase, the Critic’s collapse thus survives full residency. It falls roughly fivefold under unquantised cache precision, though this lacks significance either on identical targets (3 of 9 vs 1 of 9, p = 0.58) or pooled (6 of 22 vs 1 of 17, p = 0.11). The effect remains non-uniform: Convergence collapsed once in 26 f16 invocations and never in 39 q4_0 invocations, while the Translator never collapsed. Consistent with context-injection fragility, the Critic issued 1.8 tool calls per invocation on the quantised run versus 3.1 unquantised, failing early in its tool loop rather than after extended generation. Cache precision, operating on dynamically injected tool context, remains the primary explanation for Critic instability, despite lacking statistical significance when pooled across roles. Its effect is confined to the agents: across the 8 targets run under both cache precisions, the recomputed cells differ by at most 0.0052 Å, so the electronic structure is untouched.

4

5

6

Published lattice constant (Å) Figure 5: Published versus recomputed in-plane lattice constants (y = x dashed line; shaded band ±2 %) across 8 comparisons. Circles and triangles denote the shorter lattice parameter a and longer lattice parameter b, respectively; open markers indicate unverified targets withheld by the Critic agent due to symmetry departure [27–29].

ally shows, faded, the targets whose relaxed cell 3.5 Recomputation of published lat- left its prototype. The mean absolute error (MAE) is 0.110 Å (mean absolute relative error (MARE) tice constants 2.3 %), median absolute deviation is 1.1 %, and 6 Complete reporting is a necessary condition for re- of the 8 comparisons fall within 2 %. The signed production, not a sufficient one: a methods section bias is 0.090 Å: the recomputation is systematically may state every parameter and still not determine larger than the published cell rather than centred the calculation, because the relaxed coordinates are on it. All four reference calculations are plane-wave almost never deposited. Recomputation directly (VASP, Quantum ESPRESSO), against which a DZP assesses what textual auditing cannot: whether numerical-orbital basis is expected to overestimate, complete methodological reporting enables success- so the offset is a basis-set difference rather than a disful physical simulation execution. The execution agreement about the material. Physical agreement pipeline ran the top-ranked materials possessing the nonetheless varies by composition: penta-PdSe2 most complete sets of reported computational fields. matches closely, spanning 0.08–1.73 % across four Standard SIESTA input configurations, alongside comparisons, penta-NiP2 yields sub-percent deviathe published lattice parameters and atomic coordi- tion (0.46–0.61 % across two comparisons), whereas nates, were used to start the calculations. penta-C2 Si deviates by ∼6.9 %. Two further targets Figure 5 reports 8 matched comparisons from 4 are withheld because their papers report only one papers [30–33] across 3 compositions, and addition- in-plane constant [34] (Section 2.5). One of them 10

exposed a compositional misattribution (substituting elemental Si for a p-SiS2 claim), satisfying the textual evidence gate while specifying an incorrect material. The Critic correctly withheld verdicts when recomputed structures left their prototype, excluding them from every deviation reported above; Figure 5 retains them as open markers. For penta-BN2 , the cell broke symmetry to b/a = 1.058 against a tetragonal prototype; for a carbon paper, it reached b/a = 1.1857 (a 18.6 % departure). Both triggered symmetry concerns, placing the structures on hold rather than plotting invalid comparisons. The dominant residual error is structural rather than electronic. Because relaxed CIF (Crystallographic Information File) coordinates are rarely deposited, structures are rebuilt from topological prototypes, and a relaxation that settles into a different local minimum measures the starting geometry rather than the published physics; the two symmetry holds above exemplify this exact failure mode (Figure 6). A perfectly reported methods section can therefore remain irreproducible when the geometry is withheld. Calibration against public reference geometries yielded only four usable points across two distinct pentagonal structures, falling short of the required threshold of six. No categorical verdict is therefore issued (Section 2.5). Because recomputations employ independent basis sets and pseudopotentials, the raw physical deviations shown in Figure 5, rather than categorical pass/fail verdicts, constitute the primary reported measurement.

4

Discussion

The capability audit demonstrates a 4 GiB qwen3:4b possesses usable semantic competence but severe structural fragility, necessitating rigid programmatic gates for transcription and data provenance. Extraction recall dominates systematic error: since the pipeline measures reporting absence, extraction failures (false negatives) artificially inflate apparent omission rates. The 67.3% recall ensures reported per-parameter omission rates are conservative upper bounds, measured on nine parameters at 95.7 % precision (95% CI 90.3–98.1 %; five false positives per 116 asserted presences). Incomplete GPU residency compounds this, precluding single-pass reproducibility on 4 GiB hardware and mandating the three-pass grounded union as a methodological necessity.

conventions displace parameters into inaccessible supplementary files, meaning the gap likely measures parameter localisation rather than absolute presence. Resolving this requires dedicated perpublisher supplementary retrieval. Methodological limitations Evidence grounding bounds fabrication but permits misattribution. For example, a p-MS2 (M = Si, Ge, Pb) paper’s lattice claim was stored solely under the metal, so a p-SiS2 constant was attributed to elemental Si. That target was excluded from Figure 5 because the source reported only a single in-plane constant; consequently, the misattribution was identified solely through manual inspection rather than automated filtering. The pipeline cannot detect such composition mismatches; they remain identifiable solely through expert manual annotation. These labels cover nine fields across 25 papers (201 judgements) by a single expert annotator, yielding a wide 59.8– 74.0 % recall interval. Additionally, the basis_size parameter’s extreme rarity precluded statistically meaningful evaluation of its extraction recall. A related limitation attaches to corpus definition rather than to extraction: the screening flag distinguishes computational from non-computational work, not density functional theory from other computational methods. Studies employing tight-binding, classical dynamics or continuum models therefore remain in the audited population, where their absent fields are recorded as omissions rather than as evidence of a different methodology. Implications for materials science reporting Methodological completeness does not guarantee structural reproducibility: where relaxed geometries are withheld, the rebuilt prototype rather than the published structure determines the result (Section 3.5). With mature deposition infrastructure (AiiDA [35], NOMAD [36]), this barrier remains strictly conventional. Threshold-based calibration encounters similar data scarcity: the calibration set yielded only two of the six distinct structures required before a threshold is defined, so no comparison in this campaign rests on a calibrated tolerance and no categorical verdict is issued. Moreover, independent basis sets and pseudopotentials conflate methodological divergence with basis-set incompleteness and starting-geometry entrapment. Decoupling these requires the authors’ original structures. Consequently, the pipeline issues no refuted verdicts, reporting only the raw physical deviation between published claims and independent recomputations.

Defining omission within the corpus The pipeline strictly targets direct arXiv or open-access text, excluding separately indexed DOIs and supplementary files. Consequently, all reported rates Provenance of the reported campaign Two reflect full-text presence alone. This contextualises artefacts remain intentionally unreconciled: 3 unthe preprint–publisher gap (Section 3.1): journal reviewed verdicts from Critic collapses, and 12 ver11

b/a = 1.025

Δz = 1.46 Å

b/a = 1.047

Δz = 1.63 Å

b/a = 1.186

Δz = 0.05 Å

Figure 6: Visualizing conformational entrapment during variable-cell relaxation. (Left) The ideal pentagonal starting prototype. (Middle) Relaxed penta-PdSe2 retains its pentagonal symmetry and proper out-of-plane buckling (∆z = 1.63 Å). (Right) The carbon target exemplifies a failure mode: the structure is trapped in a different local minimum where the b/a ratio expands to 1.186, flattening the lattice (∆z = 0.05 Å) and stretching the bonds. ification records spanning three distinct configuration fingerprints. These surface solely because outputs are stamped with their generative configuration. The capacity to detect its own systemic drift is a precondition for evaluating external literature reproducibility. The reported campaign is itself one of ten: each defect the gates exposed invalidated every extraction that preceded its correction, and the retained logs record at least 509 extractions, approximately 4.7 million input tokens, before the corpus reported here was produced. It is that iteration, rather than the final run, which unmetered local execution makes practicable.

discovery is absent structural data, not omitted parameters. With relaxed coordinates rarely deposited, workflows must rely on topological prototypes. This forces conformational entrapment, conflating true methodological divergence with starting-geometry sensitivity. Definitively adjudicating published claims requires addressing these systemic data gaps. Future pipelines must parse publisher-specific supplementary materials housing siloed structural files and build community-driven reference geometry datasets. Finally, while this implementation orchestrates density functional theory relaxations, the underlying framework remains property-agnostic. Expanding verification workflows to derived physical proper5 Conclusions ties (such as elastic constants, formation energies, and electronic band gaps) will establish local openThis work presents an autonomous, local framework weights agentic systems as viable instruments for for reproducibility auditing in computational macontinuous verification of the scientific record. terials science. It demonstrates an open-weights qwen3:4b model on consumer hardware can drive an end-to-end first-principles workflow. Despite severe structural fragility, such models exhibit suffi- Data and code availability cient semantic competence for complex extraction when strictly constrained by deterministic code gates. The complete pipeline source code and configSuch a model can be trusted in what it asserts, but uration files are available at https://github. not in what it fails to find; and that reliability de- com/shambhubhandari/Agentic-Workflow (MIT pends on the conditions of its deployment, complete license). Raw full-text manuscripts are omitted residency on the accelerator above all. Trust is for copyright compliance; the retrieval module proconferred not by the model but by the constraints vides the OpenAlex queries and repository-fallback imposed upon it: every assertion must quote its routines to regenerate the corpus from public endsource verbatim, and every magnitude must be lo- points. The repository provides the agent-extracted cated within the document before it may inform records, the hand-labelled ground-truth dataset, and a calculation. Neither constraint is negotiable by all SIESTA input/output files. Every output file is the agent, and neither depends upon a network, a cryptographically stamped with its generative consubscription, or a service that may be withdrawn. figuration to guarantee auditability. The automated audit reveals incomplete textual reporting is not the binding reproducibility constraint. Successful parameter extraction does not Conflicts of interest guarantee structural reproducibility. Physical recomputations indicate the primary barrier to data-driven There are no conflicts to declare. 12

CRediT authorship contribution [13] A. D. McNaughton, G. K. Sankar Ramalaxmi, A. Kruel, C. R. Knutson, R. A. Varikoti and statement N. Kumar, ACS omega, 2024, 9, 46563–46573. Shambhu Bhandari Sharma (S.B.S): Concep- [14] K. Lejaeghere, G. Bihlmayer, T. Björkman, P. Blaha, S. Blügel, V. Blum, D. Caliste, I. E. tualization, Methodology, Software, Formal analysis, Castelli, S. J. Clark, A. Dal Corso et al., SciInvestigation, Data Curation, Validation, Visualizaence, 2016, 351, aad3000. tion, Writing – original draft, Writing – review & A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, [15] editing. B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv et al., arXiv preprint arXiv:2505.09388, 2025. Acknowledgements [16] Ollama contributors, Ollama: get up and running with large language models locally, SoftThe author acknowledges Alibaba Cloud (Qwen), ware, 2023, https://github.com/ollama/ llama.cpp, and Ollama for enabling local, grammarollama, Version 0.32.11 used in this work. constrained execution. Bibliographic retrieval used [17] J. M. Soler, E. Artacho, J. D. Gale, A. García, OpenAlex (OurResearch); recomputations used J. Junquera, P. Ordejón and D. Sánchez-Portal, SIESTA. The author thanks the developers of Journal of physics: Condensed matter, 2002, PyMuPDF, BeautifulSoup, and the Python ecosys14, 2745–2779. tem. [18] A. Doshi, Y. Hong, C. Xu, E. Kang, A. Kapravelos and C. Kästner, Proceedings of the IEEE/ACM 48th International Conference on References Software Engineering, 2026, pp. 201–205. [19] S. Barnett, S. Kurniawan, S. Thudumu, [1] A. M. Bran, S. Cox, O. Schilter, C. Baldassari, Z. Brannelly and M. Abdelrazek, Proceedings A. D. White and P. Schwaller, Nature machine of the IEEE/ACM 3rd International Conferintelligence, 2024, 6, 525–535. ence on AI Engineering-Software Engineering [2] D. A. Boiko, R. MacKnight, B. Kline and for AI, 2024, pp. 194–199. G. Gomes, Nature, 2023, 624, 570–578. [3] N. J. Szymanski, B. Rendy, Y. Fei, R. E. Kumar, [20] S. Zhang, J. Zhou, Q. Wang, X. Chen, Y. Kawazoe and P. Jena, Proceedings of the National T. He, D. Milsted, M. J. McDermott, M. GalAcademy of Sciences, 2015, 112, 2372–2377. lant, E. D. Cubuk, A. Merchant et al., Nature, [21] G. Gerganov and llama.cpp contributors, 2023, 624, 86–91. llama.cpp: LLM inference in C/C++, Software, [4] Z. Wang, H. Huang, H. Zhao, C. Xu, S. Zhu, 2023, https://github.com/ggml-org/llama. J. Janssen and V. Viswanathan, arXiv preprint cpp. arXiv:2507.14267, 2025. [22] S. Colvin and Pydantic contributors, Pydantic: [5] P. Yang, Z. Zhang, Y. Li, X. Wang, Y. Deng, data validation using Python type hints, SoftY. Lu, B. Tang, Z. Liu and B. An, arXiv ware, 2017, https://github.com/pydantic/ preprint arXiv:2605.26179, 2026. pydantic. [6] M. C. Swain and J. M. Cole, Journal of chemical information and modeling, 2016, 56, 1894– [23] J. Priem, H. Piwowar and R. Orr, arXiv preprint arXiv:2205.01833, 2022. 1904. [7] A. Roy, E. Grisan, J. Buckeridge and C. Gatti- [24] Artifex Software, PyMuPDF: a Python binding for MuPDF, Software, 2024, https://github. noni, Digital Discovery, 2026, 5, 1794–1808. com/pymupdf/PyMuPDF. [8] L. Tang, P. Laban and G. Durrett, Proceedings of the 2024 Conference on Empirical Methods [25] L. Richardson, Beautiful Soup: a Python library for pulling data out of HTML and XML in Natural Language Processing, 2024, pp. 8818– files, Software, 2004, https://www.crummy. 8847. com/software/BeautifulSoup/. [9] J. Yuan, H. Li, X. Ding, W. Xie, Y.-J. Li, B. W. Matthews, Biochimica et Biophysica Acta [26] W. Zhao, K. Wan, J. Shi, X. Hu and Z. Liu, (BBA)-Protein Structure, 1975, 405, 442–451. arXiv preprint arXiv:2506.09501, 2025. [10] N. Jegham, M. Abdelatti, C. Y. Koh, L. El- [27] J. Li, X. Fan, Y. Wei and G. Chen, Scientific reports, 2016, 6, 31840. moubarki and A. Hendawi, arXiv preprint arXiv:2505.09598, 2025. [28] Y. Xu, Z. Ning, H. Zhang, G. Ni, H. Shao, B. Peng, X. Zhang, X. He, Y. Zhu and H. Zhu, [11] A. Zywot, X. Chen and M. de Rijke, arXiv arXiv preprint arXiv:1701.03715, 2017. preprint arXiv:2601.11327, 2026. [12] Y. Zhu, J. Gan, X. Sun, F. Sun, Y. Shi, M. M. [29] C. Zhou, R. Huo, B. Situ, Z. Yan, Y. Tu and Islam, C. Shang, W. Gao, C. W. Coley, Y. Sun Z. Zhang, Physical Chemistry Chemical Physics, et al., arXiv preprint arXiv:2607.14512, 2026. 2025, 27, 11056–11065. 13

[30] A. Lopez-Bezanilla and P. B. Littlewood, The Journal of Physical Chemistry C, 2015, 119, 19469–19474. [31] X. Shao, X. Liu, X. Zhao, J. Wang, X. Zhang and M. Zhao, Physical Review B, 2018, 98, 085437. [32] K. Nandan, A. Agarwal, S. Bhowmick and Y. S. Chauhan, IEEE Transactions on Electron Devices, 2021, 68, 6551–6557. [33] S. Bravo, M. Pacheco, J. Correa and L. Chico, Physical Chemistry Chemical Physics, 2022, 24, 15749–15755. [34] L. Li, Z. Huang, J. Xu and H. Huang, Frontiers in Chemistry, 2022, 10, 1061703. [35] S. P. Huber, S. Zoupanos, M. Uhrin, L. Talirz, L. Kahle, R. Häuselmann, D. Gresch, T. Müller, A. V. Yakutovich, C. W. Andersen et al., Scientific Data, 2020, 7, 300. [36] L. M. Ghiringhelli, C. Baldauf, T. Bereau, S. Brockhauser, C. Carbogno, J. Chamanara, S. Cozzini, S. Curtarolo, C. Draxl, S. Dwaraknath et al., Scientific Data, 2023, 10, 626.

14

Related documents

• Typed Federated Artifacts for the Agentic Web:Sharing Tool-Routing Knowledge Across Frozen,Heterogeneous LLM Agents#974902
• SENTINEL-RL: Offloading Topological Reasoning from LLM Agents in the Security Operations Center#927025
• Scepsy: Serving Agentic Workflows Using Aggregate LLM Pipelines#1105012
• AgentR A Stateful and Recovery-Aware Software Architecture for LLM-based Auditable Workflows#615242
• AgentR A Stateful and Recovery-Aware Software Architecture for LLM-based Auditable Workflows#614275