A property-registry contract for retrieve-or-refuse thermal-mechanical lattice search

Shaoliang Yang; Henry Chu; Zu Yashengjiang; Jun Wang

cs.CESep 13, 2026

A property-registry contract for retrieve-or-refuse thermal-mechanical lattice search

Authors: Shaoliang Yang, Henry Chu, Zu Yashengjiang, Jun Wang

Organizations: Department of Mechanical Engineering, Santa Clara University, Santa Clara, CA 95053, USA. · Santa Clara University, Santa Clara, CA 95053, USA.

Abstract

Early thermal-mechanical lattice requirements are knowledge-intensive and often jointly unsatisfiable: an engineer asks for a cell that is light, stiff, laterally conducting and cheap, and no cell in the library satisfies it. A design system should say so, and say which requirement to loosen and by how much, rather than return the nearest row. A generative model can return a candidate even when the library holds none. This work applies established conflict diagnosis to a catalogue of homogenised properties. Search over 1,397 homogenised unit cells and 19 base materials (26,543 combinations) returns a catalogue row that a second solver can rebuild, or, for an empty feasible set of at most eight constraints, inclusion-minimal unsatisfiable subsets (MUS) and the slack of a repair, not a neighbour. One registry declaration generates the prompt and the evaluator, so undeclared parsed keys cannot reach search. On a frozen suite of 64 typed queries, retrieve-or-refuse matches min-repair on every feasible query (48/48) and refuses every empty one (16/16) with MUS and slack. Constraint-ignoring nearest-neighbour and penalty search violate stated cost or density limits. On 216 empty queries, repairs printed at three significant figures and rounded outward stay feasible in every case (216/216). When density and cost must be kept, a list of minimum-cardinality repairs keeps them on 209 of 211 queries; a protection-first repair keeps them on all 211, the same as full diagnosis. A 308-request parse benchmark is a template-text check, not unconstrained engineer prose.

Figures & tables

System	Knowledge	Mode	Refuse	Headline evidence
Kartashov & Vlassis [ 4 ]	commands	generate	no	2D cases
TrussGPT [ 5 ]	mech. targets	generate	no	experiment
Khanghah et al. [ 7 ]	graph tokens a	generate	no	reconstruction
MetaGen [ 6 ]	DSL + database	generate	no	case-study re-sim.
MetaSymbO [ 34 ]	language + latent	retrieve+evolve	no	validity / novelty
AutoMS [ 35 ]	NL + FEA loop	evolve	no	17 cross-physics

Table 1: Where neighbouring design systems represent engineering knowledge, generate or retrieve, and refuse. Headline evidence is the authors’ own, not a ranking.

Figure 1: Retrieve or refuse. The registry is the contract between language and search. Only the parse is learned. An empty feasible set of at most eight remaining constraints returns a minimal conflict and slack; a larger empty query is refused with that cap named. A non-empty set returns a catalogue row, re-homogenised on return. The engineer revises the requirements; the system does not invent a neighbour. D prints one worked tool transcript.

Key	Kind	Unit	Provenance	Operators
rho	geometry	—	solved (isovalue bisection)	$\leq,\geq$
k_aniso	geometry	—	solved ( $k_{33}/k_{11}$ , smaller is more directional)	$\leq,\geq$
D_11	geometry	—	solved ( $D^{\ast}/D_{0}$ , complete periodic pore)	$\leq,\geq$
symmetry	geometry	—	solved (point group of $f$ )	$=$
cost_per_kg	material	USD/kg	handbook	$\leq,\geq$
printable	material	—	handbook	$=$

Table 2: Excerpt of the property-registry contract. Provenance is solved (homogenisation), handbook (nominal metal table), or estimated (not claimed). The full registry has 26 queryable keys; 24 of them are searchable numeric columns.

Field	Role
$\mathcal{O}$	weighted objectives
$\mathcal{C}$	constraint atoms
$F$	material filter
UNDERSTOOD AS	compiled $q$ , printed before retrieval
CANNOT EXPRESS	schema cannot name the ask
DISCARDED	undeclared or vacuous keys

Table 3: Parse output. Only

q

reaches search; residue is reported, not scored.

Factor	In this catalogue	Solved?
Solid phase	19 handbook metals/ceramics as $k_{s}$ , $E_{s}$	no
Void phase	empty pore, $k_{\mathrm{void}}=0$	no
Volume fraction	$\rho\in[0.108,0.514]$ by isovalue bisection	yes
Porosity layout	family, mode, $f$ , isovalue; voxel FE	yes

Table 4: What enters an effective property. Only the last two rows are solved. The void is the exact empty-pore model, not a soft inclusion.

Solid	$\Delta k^{\ast}$
copper	$+0.02$ to $+0.04\%$
aluminium 6061	$+0.06$ to $+0.11\%$
AlSi10Mg	$+0.07$ to $+0.14\%$
stainless 316L	$+0.6$ to $+1.1\%$
Ti–6Al–4V	$+1.4$ to $+2.7\%$

Table 5: Change in

k^{\ast}

when the empty pore is filled with still air (

0.026

W/mK). Three cells re-solved per metal.

Target $\rho$	Failure rate
$0.10$ – $0.15$	$36.9\%$
$0.15$ – $0.20$	$19.8\%$
above $0.35$	none

Table 6: Unsolvable enumerated cells (

1{,}536

planned rows;

139

fail).

Scheme	Fit to this geometry
Voxel FE (used)	One routine, both physics; deleting void elements is the zero-contrast treatment.
FFT [ 60 ]	Slow at infinite contrast, which is the regime here.
Mori–Tanaka [ 61 ]	Dilute ellipsoids; TPMS networks are percolating and not dilute.

Table 7: Solver choice for a true-zero void and a TPMS mask.

	$\nu\in[0.29,0.34]$	$[0.15,0.45]$
Selection regret	$0$	$0.40\%$ at $\nu=0.45$
Top-three set	unchanged	unchanged
Max rank shift	—	$9$ of $16$
$E_{11}$ value drift	$1.6\%$	$11.8\%$
$E_{33}$ value drift	$3.9\%$	$33.5\%$
$E_{33}/E_{11}$ drift	$3.5\%$	$29.0\%$

Table 8: Single-

\nu

catalogue (

\nu=0.30

) versus re-solves. Selection is which cell is returned; value drift is the reported

E

Axis	Count
Families $\times$ modes $\times$ $f$ $\times$ target $\rho$	$8\times 2\times 8\times 12$
Target $\rho$ interval	$[0.12,0.50]$
Enumerated cells	$1{,}536$
Usable rows	$1{,}397$ ( $90.9\%$ )
Metals / product	$19$ / $26{,}543$

Table 9: Enumerated catalogue before and after percolation discard.

Cell size	Median $t$	Fraction $<0.4$ mm	Fraction $<0.2$ mm
$10$ mm	$0.91$ mm	$22.5\%$	$2.7\%$
$20$ mm	—	—	none below $0.3$ mm

Table 10: Estimated min_feature on the

1{,}397

feasible rows. Micro LPBF has printed TPMS walls of

100\,\mu

m [ 62 ] ; the

0.2

mm floor is a chosen screening threshold at a

10

mm cell, not a universal machine limit.

Figure 2: Four catalogue cells, rebuilt from the stored parameters. Left to right: gyroid network $f{=}111$ ; diamond network $f{=}113$ (tetragonal, the steering mechanism); Schwarz P sheet $f{=}111$ ; IWP network $f{=}111$ . The triad on the first panel fixes the coordinate convention used throughout: axes $1$ and $2$ are in-plane, axis $3$ is through-thickness, and all four cells are drawn in the same orientation. Every directional quantity in this paper ( $k_{33}/k_{11}$ , $E_{33}/E_{11}$ , the steering result) is referred to these axes.

$n$	backend	independent	worst rel. err.	all within $10^{-6}$
60	GPU, matrix-free	yes	$3.02\times 10^{-14}$	$60/60$

Table 11: Sampled GPU versus CPU re-homogenisation of catalogue rows (seed 0). Agreement is on the discrete problem, not the physical cell.

Slice	Residual
Median $k^{\ast}$ / $E^{\ast}$ / $k_{33}/k_{11}$	$1.6\%$ / $2.4\%$ / $0.56\%$
Median $k^{\ast}$ at $\rho=0.25$ / $0.35$	$2.5\%$ / $1.3\%$
Worst $k^{\ast}$ at $\rho=0.25$ / $0.35$	$21.1\%$ / $2.9\%$
Median $k^{\ast}$ at $f=(1,1,1)$ , $(1,1,3)$ , $(1,2,3)$	$0.9\%$ , $2.1\%$ , $2.0\%$
Cells with $k^{\ast}>3\%$	$13/60$ , all at $\rho=0.25$

Table 12: Working-grid residual against

n=64

(sixty-cell factorial).

cell	$k_{33}/k_{11}$	voxel $E_{11}$	scikit-fem hex	hex vs voxel
gyroid network, $\rho=0.30$	$1.000$	$0.07004$	$0.07006$	$0.03\%$
Schwarz P sheet, $\rho=0.30$	$1.000$	$0.08786$	$0.08786$	$2\times 10^{-6}$
IWP network, $\rho=0.40$	$1.000$	$0.21106$	$0.21106$	$7\times 10^{-7}$
gyroid network, $f{=}(1,1,2)$	$0.489$	$0.08897$	$0.08897$	$1.3\times 10^{-6}$
diamond network, $f{=}(1,1,3)$	$0.178$	$0.16464$	$0.16464$	$8.3\times 10^{-7}$
diamond network, $f{=}(1,2,3)$	$0.312$	$0.24204$	$0.24204$	$1.4\times 10^{-6}$

Table 13: Catalogue

E_{11}

(periodic,

E=1

\nu=0.30

). Voxel is the stored number. Hex is scikit-fem on the same occupancy and the same element family. Absolute relative difference is

\lvert

hex

-

voxel

\rvert/\lvert

voxel

\rvert

. STL cube checks are in the text, not this table.

cell	our $k_{33}/k_{11}$	scikit-fem	worst component
gyroid network, $f{=}(1,1,1)$	$1.0000$	$1.0000$	$1.5\times 10^{-14}$
Schwarz P sheet, $f{=}(1,1,1)$	$1.0002$	$1.0002$	$1.2\times 10^{-14}$
IWP network, $f{=}(1,1,1)$	$1.0000$	$1.0000$	$1.6\times 10^{-14}$
diamond network, $f{=}(1,1,3)$	$0.1780$	$0.1780$	$1.4\times 10^{-13}$

Table 14: Periodic

k^{\ast}

, scikit-fem against the catalogue solver on the same occupancy,

k_{s}=1

. The last row is the cell behind the steering result: an independent library returns the same

k_{33}/k_{11}

Panel of Figure 3 — Figure 3: (a) Effective conductivity against effective stiffness for all $26{,}543$ combinations. Clouds use colour and marker shape by material class (circles: high-conductivity metals; squares: structural metals; triangles: ceramics); bulk metals are larger markers with a dark edge. Labels are a subset, to avoid overlap. (b) Spread of the ratio $k/E$ . Material choice over 19 dense solids: $94\times$ . Geometry at one metal (aluminium 6061), over the whole density range: $7\times$ globally, $2.1$ – $3.3\times$ inside density bands of width $0.04$ . The two levers are not interchangeable.

	Flash-Lite	Flash	Keyword table	TF-IDF 1-NN
Overall	$0.897\pm 0.007$	$0.997$	$0.606$	—
Literal	$1.000\pm 0.000$	$1.000$	—	—
Paraphrased	$0.793\pm 0.014$	$0.994$	$0.49$	$0.57$

Table 15: Concept F1 on 308 template-generated requests. Flash-Lite is mean

\pm

sample standard deviation over five runs. Flash is two identical runs; we do not write a five-run error bar. Keyword and TF-IDF are registry-derived baselines, not learned parsers.

Figure 5: Frozen typed suite ( $n=64$ ; 48 feasible, 16 empty) against a mask-intersection oracle computed before search. (a) Fraction of feasible queries whose returned row meets every original constraint. (b) Empty queries: comparable yes/no outcomes at the same printed precision, columns 1–5: states infeasibility, returns a row, that row violates an original bound, names a selected correction set, enumerates all minimal conflicts. Retrieve-or-refuse agrees with the oracle on all 64 decisions and matches an independent MUS enumerator on the 16 empty queries. Min-repair satisfies every kept constraint ( $64/64$ ). Median search $2.7$ ms ( $|\mathcal{C}|\leq 3$ ).

	count	rate
First repair executable at nearest-even 3 s.f.	$107/216$	$0.50$
First repair executable, outward 3 s.f. (same queries)	$216/216$	$1.00$
Best-objective repair executable at nearest-even 3 s.f.	$94/216$	$0.44$
Best-objective repair executable, outward 3 s.f.	$216/216$	$1.00$
Disagree on dropped set	$119/216$	$0.55$
of which minimum cardinality one	$117/119$	$0.98$

Table 16: Printed-precision repair check (

n=216

empty queries). The nearest-even three-significant-figure rows are the pre-registered metric. The outward rows are a follow-up display rule on the same frozen queries, not a new draw. Retained original bounds are kept exactly; only repaired atoms are printed. Executability is a non-empty catalogue search after substituting the displayed repair values. Disagreement is over queries where both single-answer methods return a repair. Policy names are those of Sec. 5.3 .

	count	rate
Eligible queries	$211/216$	$0.98$
First repair (protection unused) preserves protected	$133/211$	$0.63$
Smallest-first list with protection	$209/211$	$0.99$
Protection-first repair (no MUS)	$211/211$	$1.00$
Full diagnosis (all inclusion-minimal MCS)	$211/211$	$1.00$
Smallest-first list fails; others succeed	$2/211$	$0.01$

Table 17: Requirement-revision check on the same 216 queries. The task was specified before its results. Density and unit cost are protected; stiffness, conductivity, anisotropy, and part density may be loosened.

211

queries have at least one protected and one allowable constraint. Smallest-first list with protection ranks globally smallest correction sets, then rejects those that touch a protected bound. Protection-first repair is a post-result control on the same frozen queries, not a new draw: it enforces protection before minimising allowable relaxations, without MUS or MCS enumeration. Preserves means the printed revised query is executable and every protected original bound is still present.

Brief	Retrieve-or-refuse	NN / min-repair
Heat spreader $\max k_{11}$ , $k_{33}/k_{11}{\leq}0.7$ , $\rho{\leq}0.40$ , cost $\leq 5$	Al 6061 diamond sheet, $k_{11}{=}59.8$ ( $n{=}3168$ )	NN: silver (violates cost). Min-repair: same as retrieve
Cheap conductor $\max k_{11}$ , cost $\leq 3$	Al 6061 split-P, $k_{11}{=}78.2$ , $2.5$ USD/kg ( $n{=}4191$ )	NN: silver, $900$ USD/kg (violates cost). Min-repair: same as retrieve
Stiff and light $\max E/\rho$ , $\rho{\leq}0.25$ , cost $\leq 10$	AlSi10Mg Schwarz P sheet, $\rho{=}0.227$ ( $n{=}3336$ )	NN: SiC (violates cost). Min-repair: same as retrieve
Light and stiff $\rho{\leq}0.15$ , $E_{11}{\geq}100$	refuse; MUS $\{\rho,E_{11}\}$ ; slacks $0.326$ / $36.5$ GPa	NN: W Schwarz P, $\rho{=}0.257$ , $E_{11}{=}81$ (violates both). Min-repair: W diamond, $\rho{=}0.325$ , $E_{11}{=}103$ (keeps $E_{11}$ )

Table 18: Worked typed briefs. NN ignores constraints. Min-repair returns the retrieve row when the query is feasible, and the jointly attainable MCS repair when it is not.

Figure 6: Gold-query retrieval for maximum in-plane conductivity at bulk price at most $3$ USD/kg. Left: the returned cell. Centre: its location in the $(E_{11},k_{11})$ cloud; the lighter points exceed the price cap. The diamond is unconstrained nearest-neighbour (silver, $900$ USD/kg); the square is penalty search (copper, $9$ USD/kg). Right: stored properties of the top row.

Figure 7: Gold-query refusal of $\rho\leq 0.15$ and $E_{11}\geq 100$ GPa. Left: the catalogue in $(\rho,E_{11})$ . Teal points meet the density bound ( $n=1{,}900$ ); gold points meet the stiffness bound ( $n=1{,}301$ ). Their intersection (shaded) is empty. Arrows are catalogue extrema, not the printed slack: the lightest cell at $E_{11}\geq 100$ GPa has $\rho=0.325$ , and the stiffest cell at $\rho\leq 0.15$ has $E_{11}=36.6$ GPa. The deployed printer reports those repair bounds as $\rho\leq 0.326$ and $E_{11}\geq 36.5$ GPa (Table 18 ). Right: rows remaining after dropping each singleton MCS. Each constraint is feasible alone, so both are minimum-cardinality repairs.

Slice	$n$	refuse	exact	set-F1
All	80	56	$80/80$	$1.00$
Single-constraint boundary	56	32	$56/56$	$1.00$
Jointly infeasible pairs	24	24	$24/24$	$1.00$

Table 19: Union of minimum-cardinality MCS on the 80-item constructed boundary bench. Items are typed from the boundary set; the language model is out of the loop. This set is disjoint from the two worked gold queries in Figs. 6 – 7 .

Query	LOO	MUS	Pair $n$
$\rho{\leq}0.15$ , $E_{11}{\geq}100$ , $k_{11}{\geq}40$	$\rho,E_{11}$	$\{\rho,E_{11}\}$	extra $k_{11}$ not in MUS
$\rho{\leq}0.18$ , $E_{11}{\geq}50$ , $k_{11}{\geq}60$	$\rho$	$\{\rho,E_{11}\}$ , $\{\rho,k_{11}\}$	two pairwise conflicts
$\rho{\leq}0.20$ , $E_{11}{\geq}40$ , $k_{11}{\geq}40$	$\rho,E_{11},k_{11}$	$\{\rho,E_{11},k_{11}\}$	$67$ , $287$ , $2272$

Table 20: Leave-one-out binding versus deployed inclusion-minimal unsatisfiable subsets on three-constraint queries that are not in the 80-item constructed set. Pair counts are feasible rows for each two-constraint projection. The middle row is why MUS is not decoration: leave-one-out names the unique minimum MCS and hides the two pairwise conflicts.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Key	Kind	Unit	Provenance	Operators
rho	geometry	—	solved (isovalue bisection)	$\leq,\geq$
porosity	geometry	—	solved ( $1-\rho$ )	$\leq,\geq$
k_aniso	geometry	—	solved ( $k_{33}/k_{11}$ , smaller is more directional)	$\leq,\geq$
E_aniso	geometry	—	solved ( $E_{33}/E_{11}$ )	$\leq,\geq$
D_11	geometry	—	solved ( $D^{\ast}/D_{0}$ , complete periodic pore)	$\leq,\geq$
D_33	geometry	—	solved ( $D^{\ast}/D_{0}$ , complete periodic pore)	$\leq,\geq$

Appendix

Table 21: Full property-registry contract. Table 2 in the main text is an excerpt.


Query id	Feas.	R	NN	Pen	MR	LD
s_rho_<=_0.12	Y	Y	Y	Y	Y	Y
s_rho_<=_0.15	Y	Y	Y	Y	Y	Y
s_rho_<=_0.18	Y	Y	Y	Y	Y	Y
s_rho_<=_0.2	Y	Y	Y	Y	Y	Y
s_rho_<=_0.25	Y	Y	Y	Y	Y	Y

Appendix

Table 22: Frozen suite of 64 typed queries. Feas. is the mask-intersection label. R = retrieve-or-refuse (answered and satisfies); NN / Pen / MR / LD are whether that policy’s returned row satisfies the original constraints. Y/N. Retrieve-or-refuse leaves empty queries unanswered.

Constraint Satisfaction

Constraint satisfaction studies how to find or optimize assignments that obey interacting rules across discrete and structured problems.

See all 86 papers

Aug 28, 2026

Operator Packages, Proposer Strength, and Construction-Family Plateaus in Office-Scale Verified Search

Verified search, in which a language model proposes programs, a hard evaluator scores them, and selection keeps the best, has recently moved mathematical records; controlled ablations of the proposer-side components remain rare. We instrument a minimal FunSearch-style loop at office scale (a 30B local model on a laptop, 120-600 verified samples per run) with three operator packages: a schematic notebook the model writes and carries instead of verbatim elites, a named obstacle, and behavioural repulsion from constructions already found. On nine construction problems from a public repository, the complete 2^3 factorial with two replicates favours the primary contrast in a nominal two-stage analysis: the composition closes more of the seed-to-record gap (+0.196; nominal pooled p=0.023, stage-combination p~0.08; median per-problem effect +0.045). Repulsion raises construction-hash diversity in eight of nine problems (one-sided p=0.0039; partly a manipulation check). The factorial finds no positive memory-by-repulsion interaction (bounded to about +/-0.04); the gain decomposes additively, and memory+repulsion is the only arm that never collapses (0 of 18 runs), within 0.025 of the full composition. A frontier proposer under the identical loop reaches in tens of samples what the local model does not in hundreds; in single scoping runs its gains arrive without the operators. The searches stall after closing 85-92% of the gap on the flagship problem, and a later analysis reads that plateau as a resolution limit: the reference value is a continuum limit and the unaided long-run finals already carry its family's structure. In the registered family-hint test the family named in words is adopted and loses; handed as code, it is optimized, but our best finite-grid encoding stays below the plateau reached unaided. We release the harness, every candidate, and the dated pre-registrations.

Roberto I. Ono Filho

Independent researcher

LLMs for Evolutionary Optimization LLM-Based Program Synthesis Inference-Time Search Automated Algorithm Discovery Evolutionary Optimization Automated Heuristic Design

Ask PDF

Mar 25, 2026

ThermoQA: A Three-Tier Benchmark for Evaluating Thermodynamic Reasoning in Large Language Models

We present ThermoQA, a benchmark of 293 open-ended engineering thermodynamics problems in three tiers: property lookups (110 Q), component analysis (101 Q), and full cycle analysis (82 Q). Ground truth is computed programmatically from CoolProp 7.2.0, covering water, R-134a, and variable-cp air. Six frontier LLMs are evaluated across three independent runs each. The composite leaderboard is led by Claude Opus 4.6 (94.1%), GPT-5.4 (93.1%), and Gemini 3.1 Pro (92.5%). Cross-tier degradation ranges from 2.8 pp (Opus) to 32.5 pp (MiniMax), confirming that property memorization does not imply thermodynamic reasoning. Supercritical water, R-134a refrigerant, and combined-cycle gas turbine analysis serve as natural discriminators with 40-60 pp performance spreads. Multi-run sigma ranges from +/-0.1% to +/-2.5%, quantifying reasoning consistency as a distinct evaluation axis. Dataset and code are open-source at https://huggingface.co/datasets/olivenet/thermoqa

Kemal Düzkar

Olivenet Kyrenia, Cyprus

LLM Evaluation Scientific Reasoning in Language Models Numerical Reasoning in Language Models Reasoning Consistency in Language Models

Ask PDF

May 28, 2026

From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals

As candidate generation and high-throughput experimentation advance, the primary bottleneck in materials discovery is shifting from property prediction to making reliable evaluations among massive candidate sets. We propose a Knowledge-Augmented Preference Signals Framework, MaterEval, that automatically produces, for the same candidate, two evaluations: an informed judgment that follows expert rules and provides supporting evidence, and a rule-removed blind guess. By pairing the two evaluations as preference data, we guide general-purpose large language models (LLMs), originally lacking materials-specific criteria, from intuitive judgment toward reliable evaluation supported by explicit evidence. To balance throughput, cost, and reliability, we further introduce a fast-slow reasoning scheme that decouples large-scale rapid screening from in-depth review on a small subset. Using high-entropy alloy (HEA) assessment as a case study, we show that, without external retrieval and relying solely on internalized capabilities, small open-source LLMs achieve substantial gains in accuracy, conclusion consistency, and evidence discrimination, approaching the performance of rule-based closed-source LLMs. These results demonstrate that expert rules can be systematically transformed into learnable preference signals, enabling a low-cost and deployable evaluation module for autonomous materials discovery loops.

Yeyong Yu, Wenya Hu, Xing Wu +1

School of Computer Engineering & Science, Shanghai University, Shanghai, 200444, China. · Center of Materials Informatics and Data Science, Materials Genome2026 Institute, Shanghai University, Shanghai, 200444, China. · Key Laboratory of Silicate Cultural Relics Conservation (Shanghai University), Ministry of Education, China. +1

LLM Evaluation Pairwise Preference Evaluation Materials Science Large Language Model-Guided Scientific Discovery Preference Learning

Ask PDF

Abstract

Constraint Satisfaction

Explore similar work

Operator Packages, Proposer Strength, and Construction-Family Plateaus in Office-Scale Verified Search

ThermoQA: A Three-Tier Benchmark for Evaluating Thermodynamic Reasoning in Large Language Models

From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals