Pith. sign in

REVIEW 4 major objections 5 minor 13 references

The paper finds that a 4-billion-parameter local text model is the strongest LEED document verifier, that deterministic numeric checking fixes threshold errors, and that the full neuro-symbolic pipeline and low-resolution images both reduce

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 22:40 UTC pith:6NZ6NOZF

load-bearing objection A useful first benchmark for LEED document screening, but the headline 4B-vs-8B result is confounded by unmatched prompt versions, and the corpus is too small to support anything beyond descriptive claims. the 4 major comments →

arxiv 2607.15647 v1 pith:6NZ6NOZF submitted 2026-07-17 cs.AI

Neuro-Symbolic AI for LEED compliance: Document-Centric Benchmarking, Deterministic Numeric Checking, and When Multimodal Hurts

Aritro De (1) , Juliana Felkner (1) ((1) The University of Texas at Austin) This is my paper
classification cs.AI
keywords LEED v4.1 BD+Cautomated compliance checkingneuro-symbolic AIlarge language modelsdeterministic numeric checkingretrieval-augmented generationmultimodal ablationsustainable buildings
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish a practical design rule for automating LEED v4.1 BD+C compliance checks from raw project documents: a small, locally deployed text model should be the default verifier, deterministic numeric checking should be applied selectively to quantitative credits, and extraction of structured values is the frontier that limits everything. The evidence comes from four university buildings, 484 PDFs, and 153 credit-level decisions. The paper reports that a 4-billion-parameter model, gemma3:4b, reaches 67.3% accuracy and beats an 8-billion-parameter model, and that a deterministic numeric checker fixes threshold errors on key quantitative credits, moving EA-p2 from 50% to 100%. The full neuro-symbolic pipeline trails the text-only baseline at 61.6%, because extraction failures and conservative behavior on qualitative credits outweigh the numerical gains. The paper also finds that adding low-resolution drawing images consistently reduces accuracy, and that prompt choice interacts with the project's PASS/INSUFFICIENT_DATA distribution.

Core claim

The central claim is that within the narrow scope of LEED v4.1 BD+C compliance verification over raw project documentation, a locally deployed 4-billion-parameter text model (gemma3:4b) is the strongest verifier at 67.3% credit-level accuracy, outperforming a larger 8-billion-parameter model (llama3.1:8b, 60.5%). The deterministic numeric checker corrects threshold/arithmetic errors on quantitative credits, raising EA-p2 from 50% to 100% accuracy and improving several other credits when the required values are reliably extracted. The full neuro-symbolic configuration achieves only 61.6% because extraction failures and conservative behavior on qualitative categories offset those gains. The pa

What carries the argument

The load-bearing mechanism is the separation of concerns between a language model and a deterministic rule engine. The LLM handles contextual interpretation, evidence retrieval, and value extraction; the numeric checker applies exact LEED threshold logic and overrides the LLM only when it can produce a determinate PASS or FAIL. Supporting machinery includes credit-section-aware chunking, a three-tier lexical-semantic retrieval system, and a confidence-based promotion step that converts high-confidence INSUFFICIENT_DATA verdicts to PASS on documentation-rich projects.

Load-bearing premise

The reported promotion gains and the neuro-symbolic 61.6% figure assume the confidence threshold of 0.6 was chosen without peeking at the final test labels; if it was tuned on those labels, the gains are an optimistic in-sample bound and the comparison to text-only is not a fair deployment setting.

What would settle it

A leave-one-building-out recalibration of the promotion threshold on development credits only: if the neuro-symbolic accuracy then falls further below the text-only 67.3%, or if the EA-p2 gains vanish when the threshold is chosen out-of-sample, the paper's hybrid-deployment recommendation would require re-evaluation. Alternatively, adding verified FAIL cases to the ground truth and re-running the full pipeline would test whether the accuracy numbers and design rule hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A small, local text model can serve as a privacy-preserving default for LEED pre-review, avoiding the need to send proprietary project documents to cloud APIs.
  • Deterministic numeric checking guarantees correct threshold decisions on quantitative credits such as the EA-p2 prerequisite, provided the required values can be extracted.
  • Adding low-resolution drawing images to the pipeline is counterproductive at current rendering resolutions; text-only processing is the safer default for technical documents.
  • Prompt selection should be treated as a calibration decision: rubric prompts for documentation-rich projects, chain-of-thought prompts for documentation-lean projects, baseline prompts for balanced ones.
  • The main bottleneck is not model choice or threshold logic but reliable extraction of structured numeric values from heterogeneous construction documents.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If extraction improves via table-aware parsing and cross-document reconciliation, the neuro-symbolic configuration could plausibly surpass the text-only baseline; the current 61.6% result may be a snapshot of an immature extraction stage rather than a ceiling for the approach.
  • The finding that multimodal inputs hurt at 150–300 dpi suggests a targeted path forward: isolate and upscale evidence-heavy schedule tables or train a vision model on construction drawings, rather than passing full pages at standard resolutions.
  • Because the ground truth contains no verified FAIL labels, the system's ability to detect genuine non-compliance is unmeasured; a benchmark with known failing credits could materially change the design recommendation.
  • The observed prompt-prior interaction could be exploited by estimating documentation completeness in advance and automatically selecting the prompt type per project, though the paper only shows the interaction, not a deployment mechanism.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces a neuro-symbolic pipeline for LEED v4.1 BD+C compliance verification over raw project PDFs, combining credit-section-aware chunking, keyword/semantic retrieval, a locally hosted 4B-parameter LLM (gemma3:4b), a confidence-based promotion rule, and a deterministic numeric checker that applies externally specified LEED thresholds. On a corpus of four university buildings (484 PDFs, 153 credit-level decisions), the paper reports that text-only gemma3:4b achieves 67.3% accuracy, outperforming llama3.1:8b (60.5%); that the deterministic checker improves EA-p2 from 50% to 100%; that the full neuro-symbolic pipeline reaches only 61.6% overall; that adding drawing images at 150–300 DPI consistently reduces accuracy; and that prompt effectiveness interacts with the building-level PASS rate. The paper is carefully hedged as descriptive and identifies extraction as the main bottleneck, but the headline comparisons rely on prompt versions, a potentially post-hoc promotion threshold, and small samples.

Significance. If the empirical claims survive the confounds identified below, the paper would make a useful contribution as an initial, transparent benchmark for document-centric LEED compliance checking and as a demonstration of a principled division of labor between an LLM and deterministic rules. Its strengths include the use of externally specified LEED thresholds in the numeric checker (avoiding circular rule induction), the explicit audit trail for verdicts, the systematic ablation of image inputs, and the honest reporting of failure modes and limitations. The design rule — use a strong local text model by default, add symbolic checking selectively, and treat extraction as the frontier — is actionable and falsifiable. However, the headline model comparison and the 'multimodal hurts' result are currently not fully controlled, and the promotion-based gains rest on a threshold whose selection protocol is ambiguous.

major comments (4)
  1. [Table 4; §4.6] The central claim that gemma3:4b outperforms llama3.1:8b is confounded by prompt version. Table 4 labels the two models as 'llama3.1:8b (V11)' and 'gemma3:4b (V17)', but §4.6 reports prompt tuning only for gemma3:4b; no equivalent sweep or even a statement that V11 and V17 are identical except trivially is given for llama3.1:8b. Thus the 6.8 pp gap may be a prompt artifact. To support the practical recommendation 'use a strong local text-only model as the default', the authors must either evaluate llama3.1:8b under the same V17 prompt (and ideally gemma under V11) or explicitly document why V11 and V17 are equivalent. Without this, the model comparison should be withdrawn or relabeled as a comparison of two model-prompt configurations.
  2. [Table 8] The multimodal ablation contains an internal inconsistency that undermines the 'multimodal hurts' claim. Table 8 reports Building 3 with Images/Credit = 0.0, yet the +Images condition is shown as 45.7% versus 60.0% text-only (−14.3 pp). If no images were actually attached for Building 3, the +Images condition is identical to text-only and should produce the same accuracy; if images were attached for some credits, the column is mislabeled and the aggregate −16.4 pp result is not clearly attributable to image inputs. The paper must clarify what Images/Credit means and reconcile this row. At minimum, the claim that adding images 'consistently reduced accuracy' across all evaluated buildings is not supported by the current table for Building 3.
  3. [§2.7, Eq. (19); Table 7; §5.5] The promotion-threshold τ=0.6 is load-bearing for the neuro-symbolic aggregate result (61.6%) and for the Table 7 gains, but the paper itself concedes the ambiguity in §2.7 and §5.5: the threshold 'was not tuned separately per building using held-out data' and 'should be described as either a development-set optimum or a global fixed threshold, depending on the actual experimental protocol.' If τ was selected after observing performance on the full four-building set, the Table 7 improvements and the neuro-symbolic accuracy are in-sample fits. The authors must report the actual selection protocol: if τ was chosen post hoc, they should present the no-promotion result as primary and treat promotion as exploratory, or run a proper leave-one-building-out calibration of τ and report the resulting (possibly lower) gains.
  4. [§3.3; Table 4] The headline performance differences are reported without any uncertainty quantification. With 153 credit-level decisions and high building-to-building variation — llama3.1:8b beats gemma3:4b on Building 2 by 8.9 pp while gemma wins on the other buildings — the 6.8 pp overall gap may be within sampling noise. The paper says the analysis is descriptive, but it also draws practical recommendations from these comparisons. At minimum, report per-building and overall exact binomial confidence intervals, and avoid unqualified 'outperforms' language unless the comparison is shown to be robust to the prompt control and sampling uncertainty.
minor comments (5)
  1. [§4.3; Table 7] The text says the numeric checker was invoked on '41 of 141 applicable credit evaluations,' and the denominators in Table 7 sum to 141, yet the corpus contains 153 credit-level decisions. The paper should define the 141-credit subset or correct the numbers, since the current discrepancy is confusing.
  2. [§5.4] The prompt-sensitivity discussion switches from Building 1–4 to unexplained names GLT, HMA, RRH, and SEA. Map these to the building identifiers in Table 3 or use a single naming convention throughout.
  3. [§4.7] The runtime text mentions 'an evaluation of 154 credit instances,' but the corpus has 153 credit decisions. Correct the typo.
  4. [Table 7] The 'Before Promotion' and 'After Promotion' columns are not fully defined. Since promotion occurs before numeric checking (§2.7), the reader cannot tell whether the numeric checker is active in both columns; if it is, the delta does not isolate promotion. State explicitly which pipeline stages are included in each column.
  5. [Figure 6] Figure 6 reports additional Building 1 tests at 300 DPI, but no corresponding table with exact numbers is provided. Add a small table or state that the figure is illustrative.

Circularity Check

1 steps flagged

Self-flagged τ=0.6 fit risk plus an uncontrolled prompt-version confound; no equation reduces the main derivation to its inputs by construction.

specific steps
  1. fitted input called prediction [§2.7 (Eqs. 18–19), §4.4 (Table 7), §5.5]
    "The reported value of τ=0.6 should therefore be described as either a development-set optimum or a global fixed threshold, depending on the actual experimental protocol. If it was selected after observing performance on all four buildings, it should be treated as exploratory rather than confirmatory. ... the promotion threshold was chosen as a single global value. It was not tuned separately per building using held-out data. The reported gains therefore represent an optimistic bound rather than a fully validated deployment setting."

    Eq. 19 defines τ* as the argmax of accuracy on a development set. If τ=0.6 was chosen after observing the four final buildings, then the Table 7 'After Promotion' gains (+10.8, +5.9, +2.9, +2.9 pp) and the resulting 61.6% neuro-symbolic total are in-sample fits rather than out-of-sample predictions: the promoted verdicts are exactly the ones the threshold was selected to recover. The paper's own §5.5 calls this an 'optimistic bound,' so it is a self-flagged partial circularity, not a hidden one. It is not load-bearing for the text-only-default recommendation, which is why the score is 3 rather than higher.

full rationale

Most of the pipeline is not circular. The deterministic numeric checker (Eqs. 15–17, Table 2) applies external LEED v4.1 thresholds and point tables (e.g., ≥5% improvement for EA-p2, ≥20% for WE-p2), so EA-p2's 50→100% movement is an external-rule computation, not a fit; the demonstrated difficulty is extraction. The text-only model comparison rests on independent runs, but it is threatened by an uncontrolled prompt-version difference: Table 4 labels llama3.1:8b as 'V11' and gemma3:4b as 'V17', while §4.6 reports prompt engineering sweeps only for gemma3:4b. If V17 is a more engineered prompt, the 6.8 pp gap may be a prompt artifact; that is a confound, not a definitional circularity, so it does not by itself raise the circularity score. The one genuine circularity risk is the promotion threshold τ=0.6: Eq. 19 fits τ by maximizing accuracy, and if the final labels were used to select it, the Table 7 promotion gains and the neuro-symbolic 61.6% figure are in-sample. The paper explicitly flags this in §2.7 and §5.5 and labels the gains 'an optimistic bound.' There is no load-bearing self-citation chain in the references, no imported uniqueness theorem, and no ansatz smuggled via citation. Overall, the central text-only-default conclusion has independent content; the circularity score is 3 because of the self-flagged τ fit risk and the additional prompt confound.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or theoretical entities. Its 'credit-aware chunks', 'keyword signatures', and 'numeric checker' are data structures and rule modules, not postulates requiring independent evidence. All symbolic rules are drawn from the external LEED standard, so the main ledger burden is on hand-chosen hyperparameters and the trustworthiness of the ground-truth labels.

free parameters (5)
  • promotion threshold τ = 0.6
    Confidence threshold above which INSUFFICIENT_DATA is promoted to PASS. The paper states it may be a development-set optimum or a global fixed threshold; if selected after seeing all four buildings, the Table 7 gains are in-sample and not a validated prediction.
  • sampling temperature = 0.85
    Hand-chosen to reduce excessive abstention while retaining structured output constraints; affects stochastic variation and the self-consistency vote.
  • retrieval top-K / evidence budget = K=8, 8000 chars
    Hand-chosen constants limiting the evidence set; influence which chunks reach the verifier but are not fitted to the evaluation labels.
  • maximum chunk length L_max = 3000 chars
    Hand-chosen split threshold for credit-section-aware chunking; a configuration choice, not a parameter fitted to the results.
  • self-consistency run count = 3
    Number of LLM inference runs used for majority voting; selected for reliability, but changes the variance of the reported accuracy.
axioms (4)
  • domain assumption LEED v4.1 BD+C quantitative thresholds and point tables encoded in the numeric checker (e.g., EA-p2 requires ≥5% improvement; WE-p2 ≥20% reduction) are the correct official requirements.
    The entire numeric checker (§2.6, Table 2) depends on these rules; if any threshold is outdated or misread, the checker's 'corrections' would be wrong. No LEED reference standard is cited in the paper.
  • domain assumption Expert ground-truth labels (LEED AP + GBCI info) for 153 credit-level decisions are accurate.
    Section 3.1 states labels were verified by a LEED AP using project documentation and official GBCI review; no inter-annotator agreement or release of labels is provided.
  • domain assumption PDF text layers are intact for all evaluated documents, so text-only retrieval captures the relevant evidence.
    Sections 2.2 and 5.5 state the pipeline assumes reliable PDF text extraction; scanned/image-only documents are excluded, limiting generalizability.
  • domain assumption Model-reported confidence scores are informative enough to drive the promotion rule.
    Promotion (§2.7) uses the LLM's confidence γ; if confidence is miscalibrated, promoting INSUFFICIENT_DATA above τ=0.6 will be wrong. The paper acknowledges miscalibration risk in §5.1.

pith-pipeline@v1.3.0-alltime-deepseek · 24334 in / 17520 out tokens · 152074 ms · 2026-08-01T22:40:56.272695+00:00 · methodology

0 comments
read the original abstract

LEED v4.1 BD+C certification remains a document-intensive process that requires reviewers to read hundreds of pages of project evidence and apply credit-specific threshold logic by hand. This paper investigates whether small, locally deployed language models can perform meaningful screening of LEED documentation and how deterministic symbolic components should share that work. A neuro-symbolic pipeline is introduced that aligns project PDFs to LEED credit sections, retrieves evidence with credit-aware keyword signatures, verifies compliance with a locally hosted 4-billion-parameter language model, and applies a LEED-specific numeric checker to quantitative thresholds. Experiments on four university buildings (484 PDFs, 153 credit-level decisions) show that a 4-billion-parameter model (gemma3:4b) is the strongest text-only core verifier, achieving 67.3% accuracy and outperforming a larger 8-billion-parameter model (llama3.1:8b) in this task. The deterministic numeric checker corrects arithmetic errors on key quantitative credits, moving EA-p2 from 50% to 100% accuracy and improving several other credits when required values are reliably extracted. At the same time, the full neuro-symbolic configuration achieves 61.6% overall accuracy, trailing the best text-only baseline due to extraction failures and conservative behavior on qualitative categories. Systematic ablations show that adding low-resolution drawing images (150-300 dpi) consistently reduces accuracy, and that prompt effectiveness depends on the building's ground-truth PASS rate: rubric prompts perform best on documentation-rich projects, while chain-of-thought prompts perform best on documentation-lean projects. Within the specific scope of LEED v4.1 BD+C compliance verification over raw project documentation, this pipeline and its baselines provide an initial reproducible reference point for both accuracy and failure modes.

Figures

Figures reproduced from arXiv: 2607.15647 by Aritro De (1), Juliana Felkner (1) ((1) The University of Texas at Austin).

Figure 1
Figure 1. Figure 1: Neuro-symbolic LEED verification pipeline. The pipeline transforms heterogeneous project documentation into credit-level verdicts. The language model performs retrieval-guided evidence interpretation and value extraction. The symbolic checker applies deterministic credit-specific threshold logic. The checker overrides the language-model result only when it returns a determinate PASS or FAIL verdict. Aritro… view at source ↗
Figure 2
Figure 2. Figure 2: Text-only model accuracy across four buildings. gemma3:4b achieves the highest overall accuracy (67.3%), outperforming llama3.1:8b by 6.8 pp despite having half the parameters. smollm3:3b performs poorly on PASS-heavy buildings due to conservative bias. Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 25 of 23 [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Neuro-symbolic accuracy by LEED category across all four buildings. EA achieves the highest accuracy (76.7%) due to numeric checker intervention. SS shows the largest degradation (−45.1 pp vs. text-only), revealing that Sustainable Sites evidence is poorly captured by text-based retrieval [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-credit accuracy delta (neuro-symbolic minus V17 text-only) across 12 quantitative credits. EA-p2 shows the largest improvement (+50 pp, 50%→100%). Regressions on SS-c5, SS-c3, EQ-c8, and WE-c2 are caused by extraction failures, not checker logic errors. Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 26 of 23 [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Multimodal ablation comparing text-only gemma3:4b with +Images at 150 DPI. Adding images reduces accuracy by 14–21 pp across all three buildings. Delta annotations show the per-building change in percentage points. Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 27 of 23 [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Vision model comparison on Building 1 at 300 DPI. No vision model exceeds 50% accuracy. text-only gemma3:4b remains the strongest configuration. llava:7b reads clean documents but fails on technical drawings; moondream hallucinates drawing content. Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 28 of 23 [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prompt engineering sensitivity to class imbalance. Rubric prompts perform best on PASS-heavy buildings (Building 1, 81.6% PASS); CoT prompts perform best on INSUFFICIENT_DATA-heavy buildings (Building 3, 42.9% PASS). Few-shot and CoT degrade performance on balanced corpora (Building 4). Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 29 of 23 [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt effectiveness vs. ground-truth PASS rate per building. Rubric (red) shows positive accuracy deltas on PASS-heavy buildings (PASS% > 80); CoT (blue) shows positive deltas on INSUFFICIENT_DATA-heavy buildings (PASS% < 50). The crossover occurs near 50% PASS [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Inference time (left) and token usage (right) per credit. Neuro-symbolic adds approximately 30 s overhead over text-only gemma3:4b. Multimodal adds 20 s without accuracy benefit. smollm3:3b is the slowest despite fewer parameters. Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 30 of 23 [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Confusion matrices (PASS vs. INSUFFICIENT_DATA) for V17 text-only (top row) and neuro-symbolic (bottom row) across all four buildings. Neuro-symbolic reduces false negatives on PASS-heavy buildings (Building 1, Building 2) via promotion logic, but introduces additional errors on Building 3 and Building 4 where ground truth is more balanced. Aritro De; Juliana Felkner: Preprint submitted to Elsevier Page 3… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 6 canonical work pages · 3 internal anchors

  1. [2]

    A Survey of Confidence Estimation and Calibration in Large Language Models, in: Duh, K., Gomez, H., Bethard, S. (Eds.), Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Association for Computational Linguistics, Mexico City, Mexico. pp. 6...

    doi:10.18653/v1/2024.naacl-long.366 2024

  2. [4]

    (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, Association for Computational Linguistics, Miami, Florida, USA

    Mirror-Consistency: Harnessing Inconsistency in Majority Voting, in: Al- Onaizan, Y., Bansal, M., Chen, Y.N. (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, Association for Computational Linguistics, Miami, Florida, USA. pp. 2408–2420. URL:https://aclanthology.org/2024.findings-emnlp.135/, doi:10.18653/v1/2024.findings-emnlp...

    work page doi:10.18653/v1/2024.findings-emnlp.135 2024

  3. [6]

    An Integrated Platform for LEED Certification Automation Using Computer Vision and LLM-RAG

    An Integrated Platform for LEED Certification Automation Using Computer Vision and LLM-RAG. URL:http://arxiv.org/ abs/2506.00888, doi:10.48550/arXiv.2506.00888. arXiv:2506.00888 [cs.SE]. Lee, J.H., Ji, S.Y., Ostwald, M.J.,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.00888

  4. [8]

    URL:http://arxiv.org/abs/2506.20551, doi:10.48550/arXiv.2506.20551

    Large Language Model-Driven Code Compliance Checking in Building Information Modeling. URL:http://arxiv.org/abs/2506.20551, doi:10.48550/arXiv.2506.20551. arXiv:2506.20551 [cs.SE]. Michael, N., BenShushan, D., Bien, J., Moore, D.A.,

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.20551

  5. [9]

    Confidence Calibration in Large Language Models

    Confidence Calibration in Large Language Models. URL:http://arxiv.org/abs/ 2605.23909, doi:10.48550/arXiv.2605.23909. arXiv:2605.23909 [cs.AI] version:

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.23909

  6. [10]

    Building Research & Information 0, 1–24

    A systematic review of methods for interpreting building code regulations in automated compliance systems. Building Research & Information 0, 1–24. URL:https://doi.org/10.1080/09613218.2026.2637965, doi:10.1080/ 09613218.2026.2637965. _eprint: https://doi.org/10.1080/09613218.2026.2637965. Oche,A.J.,Folashade,A.G.,Ghosal,T.,Biswas,A.,2025. ASystematicRevi...

    arXiv 2026

  7. [11]

    20090–20111

    Confidence Improves Self-Consistency in LLMs, in: Findings of the Association for Computational Linguistics: ACL 2025, pp. 20090–20111. URL:http://arxiv.org/abs/2502.06233, doi:10.18653/v1/2025.findings-acl.1030. arXiv:2502.06233 [cs.CL]. Wan, H., Zhang, J., Chen, Y., Xu, W., Feng, F.,

    arXiv 2025

  8. [12]

    Building Simulation 18, 1251–1273

    Exploring Gen-AI applications in building research and industry: A review. Building Simulation 18, 1251–1273. URL:http://arxiv.org/abs/2410.01098, doi:10.1007/s12273-025-1279-x. arXiv:2410.01098 [cs.AI]. Zhang, J., El-Gohary, N.M.,

    Pith/arXiv arXiv

  9. [170]

    Huang, S., Ma, Z., Du, J., Meng, C., Wang, W., Lin, Z.,

    URL:https://www.nature.com/articles/s41597-024-04320-x, doi:10.1038/s41597-024-04320-x. Huang, S., Ma, Z., Du, J., Meng, C., Wang, W., Lin, Z.,

    work page doi:10.1038/s41597-024-04320-x

  10. [2016]

    Journal of Computing in Civil Engineering 30, 04015014

    Semantic NLP-Based Information Extraction from Construction Regulatory Documents for Automated Compliance Checking. Journal of Computing in Civil Engineering 30, 04015014. URL:https://ascelibrary.org/doi/10.1061/ %28ASCE%29CP.1943-5487.0000346, doi:10.1061/(ASCE)CP.1943-5487.0000346. Aritro De; Juliana Felkner:Preprint submitted to ElsevierPage 23 of 23 N...

    arXiv 1943

  11. [2024]

    Journal of Building Engineering 96, 110401

    Comparative analysis of sustainable building certification processes. Journal of Building Engineering 96, 110401. URL:https://www.sciencedirect.com/science/article/pii/S2352710224019697, doi:10.1016/j.jobe.2024. 110401. Geng, J., Cai, F., Wang, Y., Koeppl, H., Nakov, P., Gurevych, I.,

    work page doi:10.1016/j.jobe.2024 2024

  12. [2025]

    Enhancing Domain-Specific Retrieval-Augmented Generation: Synthetic Data Generation and Evaluation using Reasoning Models

    Enhancing Domain-Specific Retrieval-Augmented Generation: Synthetic Data Generation and Evaluation using Reasoning Models. URL:http://arxiv.org/abs/2502.15854, doi:10.48550/arXiv.2502.15854. arXiv:2502.15854 [cs.LG]. Kalai, A.T., Nachum, O., Vempala, S.S., Zhang, E.,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.15854

  13. [2026]

    Automation in Construction 185, 106859

    Automated compliance checking across the building lifecycle: Systematic and semantic review integrating PRISMA and deep search. Automation in Construction 185, 106859. URL:https://www.sciencedirect.com/science/article/pii/ S0926580526001007, doi:10.1016/j.autcon.2026.106859. Madireddy, S., Gao, L., Din, Z., Kim, K., Senouci, A., Han, Z., Zhang, Y.,

    arXiv 2026