Paper deep dive
Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies
Dalton Raphael Harmsen, Swier Garst, Thomas van Osch, ZarĂš Palanciyan, Joaquin Vanschoren
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cross-lingual transfer describes how knowledge in a source language benefits a target language. Measuring it quantitatively requires broad multilingual pre-training, as prior work has done with cross-lingual transfer matrices. We ask whether transfer is predictable from freely available typological features, and whether the prominence of high-resource source languages reflects typology or data quality and quantity. We show that typological databases contain cheap and dense signals about cross-lingual transfer. Our typology-only random forest on a 24-language prior-work transfer matrix scores leave-one-language-out $\rho{=}0.705$ and $R^2{=}0.49$, beating a non-typological control at $\rho{=}0.62$, which verifies the ability of typology-only predictions to reconstruct costly measured cross-lingual transfer. The signal survives leave-one-script-out and leave-one-family-out protocols, so script and family confounding do not explain the effect. By decomposing the transfer into a typology term and a resource-and-script bias term, we find the best-source ranking sensitive to this bias. In contrast, typology is not affected by this bias, which makes it a zero-compute screening tool that replaces hundreds of training runs with a model fit. Our code is available \href{this https URL}{here}.
Tags
Links
- Source: https://arxiv.org/abs/2609.39640v1
- Canonical: https://arxiv.org/abs/2609.39640v1
Trouble viewing inline? Open PDF directly â
Full Text
95,236 characters extracted from source content.
Expand or collapse full text
Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies Dalton Raphael Harmsen *1 , Swier Garst 2 , Thomas van Osch 2 , ZarĂš Palanciyan 2 , and Joaquin Vanschoren 1 1 AMOR/e Lab, Eindhoven University of Technology, Eindhoven, The Netherlands 2 SURF, Amsterdam, The Netherlands Abstract Cross-lingual transfer describes how knowledge in a source language benefits a target language. Measuring it quantitatively requires broad multilingual pre- training, as prior work has done with cross-lingual transfer matrices. We ask whether transfer is predictable from freely available typological features, and whether the prominence of high-resource source languages reflects typology or data quality and quantity. We show that typological databases contain cheap and dense signals about cross-lingual transfer. Our typology-only random forest on a 24- language prior-work transfer matrix scores leave-one-language-outÏ=0.705and R 2 =0.49, beating a non-typological control atÏ=0.62, which verifies the ability of typology-only predictions to reconstruct costly measured cross-lingual transfer. The signal survives leave-one-script-out and leave-one-family-out protocols, so script and family confounding do not explain the effect. By decomposing the transfer into a typology term and a resource-and-script bias term, we find the best-source ranking sensitive to this bias. In contrast, typology is not affected by this bias, which makes it a zero-compute screening tool that replaces hundreds of training runs with a model fit. Our code is available here. 1 Introduction Cross-lingual transfer describes how knowledge in a source language benefits a target language. Longpre et al. introduced ATLAS, which measures this through multilingual pre-training, finding that a few high-resource languages dominate in transfer due to their abundance of training data. However, methods such as ATLAS are prohibitively expensive, requiring hundreds of pre-training runs [1,6]. On the other hand, Grambank [14,15] and WALS [3] are readily-available typological databases holding hundreds of structural properties per language. Linguistic similarity has previously been shown to correlate with transfer [4,9,12]. We thus hypothesize that these databases contain valuable transfer signal, which can rank candidate source languages before any compute is spent. We therefore pose two questions; RQ1: can typological features reconstruct the transfer scores of the ATLAS matrix? RQ2: is the prominence of high-resource languages such as English driven by typological features, or by data quality and quantity? Contributions First, we introduce a typology-only random forest that predicts transfer on a 24- language ATLAS subset, verifying the ability of typology-only predictions to reconstruct the cross- lingual transfer measured by ATLAS. Our inexpensive model replaces hundreds of LLM training runs, complementing forward-pass estimators such as LEEP [10] and LogME [16]. Second, we decompose the ATLAS scores into a typology term and a resource-and-script bias term, and show that best-source ranking is sensitive to this bias. 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Workshop: Linguistic Principles for Foundation Models. arXiv:2609.39640v1 [cs.CL] 30 Sep 2026 2 Method and Experiments Data We use typological features from both the Grambank [14,15] and WALS [3] databases. In total, 387 Glottocode-keyed [5] features were used, of which 195 originated from Grambank, and 192 from WALS. The exact list of features is given in Appendix A, Table 7. We use these features to predict the bilingual transfer scores (BTS) and fine-tuning adaptation score (FAS) introduced by ATLAS [8]. Keeping languages with at least 70% Grambank feature coverage leaves 24 languages and 552 language pairs; a set we refer to as ATLAS-24. Model trainingOur main model is a 200-tree random forest swept per feature source. We compare four tiers: two random forest models, one using all features and one using the features determined by a feature selection and hyperparameter sweep, and two ridge regression models as baselines: a similarity proxy on typological similarity alone, and a generalized linear model on typological similarity, script, genealogical, and geographical distances. We also investigate a control of equal model capacity which uses five non-typological descriptors: writing script, WALS family, WALS genus, geographic region, and Wikipedia resource bin. We use leave-one-language-out (LOLO) and leave-k-languages-out cross-validation, reporting held-out PearsonÏfor ranking performance and R 2 for regression performance. In addition, we use leave-one-family-out (LOFO) and leave-one- script-out (LOSO) as more general holdouts. A fold-respecting permutation null, cluster bootstrap intervals by source language, and paired bootstrap tests ofâÏare employed for statistical tests. Refer to Appendix A for details. Resource-bias decomposition To separate typology from resources, we modelBTS = T + b, composed of typological signalTand resource-and-script biasb, which we model as Wikipedia article counts (source and target language separate), script difference, and genealogical and geographic distance. We fit a ridge regression to estimate Ë band then a random forest that fits the typology term Ë T on the residual BTSâ Ë b. Experiments Experiments are grouped as follows: E1âE4 probe RQ1 and E5 probes RQ2. E1 examines whether typological features reconstruct measured transfer, by comparing held-out perfor- mance across the four model tiers and the matched-capacity metadata control against the permutation null. E2 examines whether the reconstruction is carried by shared writing script rather than by typology, comparing cross-script, same-script, and Latin-only held-out pairs against a script-only directed baseline. E3 examines the genealogical confound through grouped holdouts that leave an entire script or an entire family out of training at once. E4 examines where the signal lives, sweeping the number of features from a single one to the full set and measuring the stability of the importance ranking across folds. E5 examines which source language transfers best, and how a sourceâs data availability shapes this result, by ranking candidate sources under four operationalizations of transfer that vary in how much resource bias they admit. Full experimental configurations are in Appendix A. 3 Results Reconstruction performance across model capacity (E1) The reconstruction performance for different model setups is given in Table 1. The predictive performance of the similarity proxy and ridge GLM baselines in both protocols does not meet the simple baseline thresholdÏ=0.225(Appendix B, Table 16). The untuned random forest passes this threshold and is able to empirically reconstruct the ATLAS transfer matrix. Performance is further increased with random forest tuned using a hyperparameter sweep, highlighting the value of the typological features for transfer prediction. Ablations In harder protocols the results degrade but do not collapse, reachingÏ=0.41in the worst case where both languages of each held-out pair are unseen (Appendix B, Table 15). In a matched- capacity setting, where metadata is compared against typology with 5 features each, the performance difference disappears, suggesting that the strength of typology lies in the breadth of the features (Appendix B, Table 12). Statistical inference supports the performance of the tuned model: the cluster bootstrap 95% CI of[0.51, 0.82]excludes the permutation null (meanâ0.007) atp=0.005. The ordering of feature sources survives metric and language-set changes (Tables 8 and 9). Prediction heatmaps of all tiers are in Appendix C. Interestingly, Grambank features alone outperform the Grambank + WALS union:Ï=0.78vs.0.705(Appendix B, Table 8). The unionâs degradation comes from the model fitting WALS patterns that hold only within the small training set and fail on held-out 2 languages, rather than from column dilution or WALS feature quality (Appendix B, Table 10). For the final results, the Grambank + WALS union is adopted anyway due to post-hoc selection bias. Table 1: LOLO and leave-2-languages-out predictive performance on ATLAS-24 (E1). ModelLOLO Ï R 2 L2LO Ï similarity proxy0.037 â0.0100.029 ridge GLM (sim/script/gene/geo)0.1620.0140.154 random forest (all features)0.6150.3500.648 tuned RF (typology)0.7050.4920.702 metadata RF (matched-capacity, 5 fields)0.5810.2640.611 metadata RF (own-sweep tuned)0.6200.3750.621 single-feature ridges (script/family/wiki) â€0.17 â0 â€0.17 Generalization across scripts and families (E2, E3) Cross-script held-out pairs are predicted at least as well as same-script, atÏ=0.728vs.0.600(Table 2), and the model stays well above the script-only baseline even on Latin-Latin pairs, where script carries no information. Grouped holdout does not decrease performance either: LOSO and LOFO macro-Ïexceed the matched LOLO baseline on the same held-out pairs, and the pair-weightedÏis unchanged (Appendix B, Table 14). A symmetric cross-check agrees: language pairs that share a script transfer more on average, but a script-only model cannot predict the directed scores and stays near chance (Appendix B, Tables 15 and 13), highlighting that indeed script strongly correlates with transfer, but has no incremental directed power. To conclude, holding out an entire group does not collapse performance (p<0.01 for paired bootstrap test), indicating that the signal is typological and not confounded by script or language family. Table 2: Cross-script robustness and grouped holdouts (E2, E3) on ATLAS-24. SplitÏ cross-script pairs0.728 same-script pairs0.600 LatinâLatin only0.509 script-only model0.138 LOLO (leave-one-language-out)0.705 LOSO (leave-one-script-out)0.776 matched LOLO (LOSO pairs)0.667 LOFO (leave-one-family-out)0.729 matched LOLO (LOFO pairs)0.671 Where the signal lives (E4) Typological signal for cross-lingual transfer is highly redundant rather than concentrated in a few decisive features (Table 3). The top-ksweep is flat up tok=100 (Ï=0.57â0.65), peaks atk=200(0.742), and settles at0.705for the full 387 feature set. The top-25 importance ranking is almost completely unstable across folds, yet the single best feature (k=1) already reachesÏ=0.622, because typological features are correlated (Appendix B, Table 15). This suggests that hundreds of typological features encode one coarse signal, and any one of them recovers most of it. Table 3: Top-kfeature sweep for Experiment E4 under Grambank+WALS; the leak-free row selects its top-k within each fold. top-k15102550100200387 global-rank Ï0.6220.6130.5730.5700.6470.6270.7420.705 leak-free Ï0.6220.6270.5730.6790.6570.7160.7100.705 3 The best pre-training source (E5) Table 4 lists the top-scoring languages per operationalization, along with their transfer scores. On every BTS-scale operationalization, English ranks first in average transfer score across all languages. Debiasing the residual reverses this: Marathi and Filipino top the ranking instead. This reversal shows that raw rankings are highly sensitive to the resource-and-script bias term, which is exactly what makes typology an attractive alternative, since it never observes this metadata bias in the first place. The flip is statistically robust: Marathiâs mean residual exceeds Englishâs by0.24(p<0.001), holds under a leak-free refit of the bias layer, and clears a permutation null (Appendix B, Table 15). Across 200 reruns, the specific top debiased language varies, so the robust finding is the flip away from English, rather than any particular ranking (Appendix B). Table 4: âBest pre-training sourceâ over a fixed candidate set (Experiment E5); rows are not comparable across scales. Operationalization#1 source#2 source empirical (ATLAS, raw BTS mean)English (+0.022)Hebrew (â0.074) on-scale reweighted RFEnglish (â0.005)Hebrew (â0.070) OOD-generalizer (cross-family)English (â0.011)Hebrew (â0.065) debiased residual (mean-zero, off-scale)Marathi (+0.36)Filipino (+0.36) 4 Conclusion, Discussion and Limitations 4.1 Conclusion and Discussion We introduce a typology-only predictor that beats a non-typological control on ATLAS transfer under leave-one-language, leave-one-script, and leave-one-family holdouts. The signal it uses is redundant rather than localized: hundreds of correlated features encode one coarse dimension, and a single feature recovers most of it. Furthermore, we show that best-source rankings built from resource statistics inherit a resource-and-script bias; typological features never observe these resources, so they offer an independent lens on the same question. Our model requires no new LLM training runs. Given an existing transfer matrix to fit on, adding a language costs only a prediction of a random forest, which makes typology a promising candidate for prioritizing source languages, or for proposing data mixtures to test on low compute budget. Further assessing the quality of such data mixtures requires more in-depth downstream analyses, which we leave for future work. Beyond our results, the distributed signal also calls on linguistic theory: transfer emerges from many small structural alignments rather than any single feature family. The bias decomposition warns that a source can look excellent merely because of resource abundance, raising the question whether current high-resource languages are the theoretical optimal training sources, especially in light of future dataset collection efforts. 4.2 Limitations Ground truth and compute Our results rest entirely on the ATLAS matrix [8], which mixes measured BTS with FAS-estimated scores, so we model a noisy target whose confounds propagate to our findings. Compute constraints confine us to ATLASâs published results, and strong claims about pre-training require verification at scale. Scale and samplingAll source rankings are 24-language extrapolations on the non-random ATLASâ Grambank overlap, which drops strong European sources such as Spanish and German, and the unseen-language Ïâ0.41 is our generalization estimate. Residual debiasing is a diagnosticThe bias layerâs weights describe ATLAS-24 rather than a stable law, so the off-scale residualTonly exposes the rankingâs bias-sensitivity (Appendix D). Residuals are mean-zero by construction and unitless with respect to BTS, so they support only relative comparisons. A script-by-genus reweighted random forest on raw BTS scores keeps predictions on-scale and comparable. 4 Acknowledgments and Disclosure of Funding This research was supported by the OpenEuroLLM project, co-funded by the Digital Europe Pro- gramme under GA no. 101195233. References [1]X. Cao, Q. Liu, C. Xiao, Y. Oda, J. Wang, P. Stenetorp, D. Kawahara, M. Onizuka, S. Kurohashi, and S. Zheng. Shapleylaw: A game-theoretic approach to multilingual scaling laws. arXiv preprint arXiv:2603.17945, 2026. [2] W. B. Cavnar, J. M. Trenkle, et al. N-gram-based text categorization. In Proceedings of SDAIR-94, 3rd annual symposium on document analysis and information retrieval, volume 161175, page 14. Ann Arbor, Michigan, 1994. [3]M. S. Dryer and M. Haspelmath, editors. WALS Online (v2020.4). Zenodo, 2013. doi: 10.5281/zenodo.13950591. URL https://doi.org/10.5281/zenodo.13950591. [4]J. Eronen, M. Ptaszynski, T. Wicherkiewicz, R. Borges, K. Janic, Z. Liu, T. Mahmud, and F. Masui. Language models are polyglots: Language similarity predicts cross-lingual transfer learning performance. Machine Learning and Knowledge Extraction, 8(3), 2026. ISSN 2504- 4990. doi: 10.3390/make8030065. URL https://w.mdpi.com/2504-4990/8/3/65. [5]R. Forkel and H. Hammarström. Glottocodes: Identifiers linking families, languages and dialects to comprehensive reference information. Semantic Web, 13(6):917â924, 2022. doi: 10.3233/ SW-212843. URL https://journals.sagepub.com/doi/abs/10.3233/SW-212843. [6] Y. He, A. Benhaim, B. Patra, P. Vaddamanu, S. Ahuja, P. Chopra, V. Chaudhary, H. Zhao, and X. Song. Scaling laws for multilingual language models. In W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, editors, Findings of the Association for Computational Linguistics: ACL 2025, pages 4257â4273, Vienna, Austria, July 2025. Association for Computational Linguistics. ISBN 979-8-89176-256-5. doi: 10.18653/v1/2025.findings-acl.221. URLhttps://aclanthology. org/2025.findings-acl.221/. [7] P. Littell, D. R. Mortensen, K. Lin, K. Kairis, C. Turner, and L. Levin. URIEL and lang2vec: Representing languages as typological, geographical, and phylogenetic vectors. In M. Lapata, P. Blunsom, and A. Koller, editors, Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 8â14, Valencia, Spain, Apr. 2017. Association for Computational Linguistics. URLhttps: //aclanthology.org/E17-2002/. [8]S. Longpre, S. Kudugunta, N. Muennighoff, I.-H. Hsu, I. Caswell, A. Pentland, S. Arik, C.-Y. Lee, and S. Ebrahimi. Atlas: Adaptive transfer scaling laws for multilingual pretraining, finetun- ing, and decoding the curse of multilinguality. In C. Vondrick, B. Hariharan, C. Raffel, L. Pinto, D. Yang, and A. Faust, editors, International Conference on Learning Representations, volume 2026, pages 31839â31867, 2026. URLhttps://proceedings.iclr.c/paper_files/ paper/2026/file/35c26a810471039b3427e524565cbef9-Paper-Conference.pdf. [9]B. Muller, D. Gupta, J.-P. Fauconnier, S. Patwardhan, D. Vandyke, and S. Agarwal. Languages you know influence those you learn: Impact of language characteristics on multi-lingual text-to- text transfer. In A. Albalak, C. Zhou, C. Raffel, D. Ramachandran, S. Ruder, and X. Ma, editors, Proceedings of The 1st Transfer Learning for Natural Language Processing Workshop, volume 203 of Proceedings of Machine Learning Research, pages 88â102. PMLR, 03 Dec 2023. URL https://proceedings.mlr.press/v203/muller23a.html. [10] C. Nguyen, T. Hassner, M. Seeger, and C. Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In H. D. I and A. Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7294â7305. PMLR, 13â18 Jul 2020. URLhttps://proceedings. mlr.press/v119/nguyen20b.html. 5 [11]A. Petrov, E. La Malfa, P. Torr, and A. Bibi.Language model tokenizers introduce unfairness between languages.In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Sys- tems, volume 36, pages 36963â36990. Curran Associates, Inc., 2023.doi: 10.52202/ 075280-1608.URLhttps://proceedings.neurips.c/paper_files/paper/2023/ file/74b24dca8334adce292883b4b651eda-Paper-Conference.pdf. [12]E. Rice, A. Marashian, H. Haynie, K. von der Wense, and A. Palmer. Untangling the influence of typology, data, and model architecture on ranking transfer languages for cross-lingual POS tagging. In S. Truong, R. A. Putri, D. Nguyen, A. Wang, D. Ho, A. Oh, and S. Koyejo, editors, Proceedings of the 1st Workshop on Language Models for Underserved Communities (LM4UC 2025), pages 22â31, Albuquerque, New Mexico, May 2025. Association for Computational Linguistics. ISBN 979-8-89176-242-8. doi: 10.18653/v1/2025.lm4uc-1.4. URLhttps: //aclanthology.org/2025.lm4uc-1.4/. [13] R. Sennrich, B. Haddow, and A. Birch. Neural machine translation of rare words with subword units. In K. Erk and N. A. Smith, editors, Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715â1725, Berlin, Germany, Aug. 2016. Association for Computational Linguistics. doi: 10.18653/v1/P16-1162. URL https://aclanthology.org/P16-1162/. [14] H. SkirgĂ„rd, H. J. Haynie, D. E. Blasi, H. Hammarström, J. Collins, J. J. Latarche, J. Lesage, T. Weber, A. Witzlack-Makarevich, S. Passmore, A. Chira, L. Maurits, R. Dinnage, M. Dunn, G. Reesink, R. Singer, C. Bowern, P. Epps, J. Hill, O. Vesakoski, M. Robbeets, N. K. Abbas, D. Auer, N. A. Bakker, G. Barbos, R. D. Borges, S. Danielsen, L. Dorenbusch, E. Dorn, J. Elliott, G. Falcone, J. Fischer, Y. Ghanggo Ate, H. Gibson, H.-P. Göbel, J. A. Goodall, V. Gruner, A. Harvey, R. Hayes, L. Heer, R. E. Herrera Miranda, N. HĂŒbler, B. Huntington-Rainey, J. K. Ivani, M. Johns, E. Just, E. Kashima, C. Kipf, J. V. Klingenberg, N. König, A. Koti, R. G. A. Kowalik, O. Krasnoukhova, N. L. Lindvall, M. Lorenzen, H. Lutzenberger, T. R. Martins, C. Mata German, S. van der Meer, J. Montoya SamamĂ©, M. MĂŒller, S. Muradoglu, K. Neely, J. Nickel, M. Norvik, C. A. Oluoch, J. Peacock, I. O. Pearey, N. Peck, S. Petit, S. Pieper, M. Poblete, D. Prestipino, L. Raabe, A. Raja, J. Reimringer, S. C. Rey, J. Rizaew, E. Ruppert, K. K. Salmon, J. Sammet, R. Schembri, L. Schlabbach, F. W. Schmidt, A. Skilton, W. D. Smith, H. de Sousa, K. Sverredal, D. Valle, J. Vera, J. VoĂ, T. Witte, H. Wu, S. Yam, J. Ye, M. Yong, T. Yuditha, R. Zariquiey, R. Forkel, N. Evans, S. C. Levinson, M. Haspelmath, S. J. Greenhill, Q. D. Atkinson, and R. D. Gray. Grambank reveals the importance of genealogical constraints on linguistic diversity and highlights the impact of language loss. Science Advances, 9(16), 2023. doi: 10.1126/sciadv.adg6175. [15]H. SkirgĂ„rd, H. J. Haynie, H. Hammarström, D. E. Blasi, J. Collins, J. Latarche, J. Lesage, T. Weber, A. Witzlack-Makarevich, M. Dunn, G. Reesink, R. Singer, C. Bowern, P. Epps, J. Hill, O. Vesakoski, N. K. Abbas, S. Ananth, D. Auer, N. A. Bakker, G. Barbos, A. Bolls, R. D. Borges, M. Browen, L. Chevallier, S. Danielsen, S. Dohlen, L. Dorenbusch, E. Dorn, M. Duhamel, F. E. H. Ali, J. Elliott, G. Falcone, A.-M. Fehn, J. Fischer, Y. G. Ate, H. Gibson, H.-P. Göbel, J. A. Goodall, V. Gruner, A. Harvey, R. Hayes, L. Heer, R. E. H. Miranda, N. HĂŒbler, B. H. Huntington-Rainey, G. Inglese, J. K. Ivani, M. Johns, E. Just, I. Kapitonov, E. Kashima, C. Kipf, J. V. Klingenberg, N. König, A. Koti, R. G. A. Kowalik, O. Krasnoukhova, K. L. Lindsey, N. L. M. Lindvall, M. Lorenzen, H. Lutzenberger, A. Marley, T. R. A. Martins, C. M. German, S. van der Meer, J. Montoya, M. MĂŒller, S. Murado Ìglu, HunterGatherer, D. Nash, K. Neely, J. Nickel, M. Norvik, B. Olsson, C. A. Oluoch, D. Osgarby, J. Peacock, I. O. Pearey, N. Peck, J. Peter, S. Petit, S. Pieper, M. Poblete, D. Prestipino, L. Raabe, A. Raja, J. Reimringer, S. C. Rey, J. Rizaew, E. Ruppert, K. K. Salmon, J. Sammet, R. Schembri, L. Schlabbach, F. W. P. Schmidt, D. Schokkin, J. Siegel, A. Skilton, H. de Sousa, K. Sverredal, D. Valle, J. Vera, J. VoĂ, D. W. Smith, T. Witte, H. Wu, S. Yam, J. Ye, M. Yong, T. Yuditha, R. Zariquiey, R. Forkel, N. Evans, S. C. Levinson, M. Haspelmath, S. J. Greenhill, Q. D. Atkinson, and R. D. Gray. Grambank v1.0, Mar. 2023. URL https://doi.org/10.5281/zenodo.7740140. Dataset. [16]K. You, Y. Liu, J. Wang, and M. Long. Logme: Practical assessment of pre-trained models for transfer learning. In M. Meila and T. Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, 6 pages 12133â12143. PMLR, 18â24 Jul 2021. URLhttps://proceedings.mlr.press/ v139/you21b.html. A Experimental configuration Notation Formally, the input is[M s ,M t , â]withM s ,M t â 0, 1 n codes the one-hot per-feature profiles (n codes = P fâF |C f |, withC f the codes of featuref) andâ â 0, 1 n feat a per-feature disagreement block (n feat =|F|). Protocols, metrics, and controls in plain language Every experiment trains a model on some language pairs and then scores it on pairs it never trained on, and the protocols below differ only in how the pairs are split. In leave-one-language-out, every pair touching one language is hidden, the model trains on the remaining pairs and predicts the hidden ones, and each of the 24 languages takes a turn as the hidden language. Leave-k-languages-out hidesklanguages at once, so a test pair can involve languages that never appear in training, which is the harder setting. PearsonÏchecks whether the model orders pairs as ATLAS does:1means a perfect ordering,0means the predicted ordering carries no information about the true one, and negative values mean systematically reversed orderings.R 2 asks the stronger question of whether the predicted values are numerically close: it compares the modelâs squared error against that of a baseline that always predicts the average BTS, so 1is perfect,0matches the baseline, and negative is worse than it. This is the quantity also known as explained variance, the share of the variation in BTS that the predictions account for. The permutation null shuffles the training scores within each fold, reruns the procedure, and reports how often such chance data score as well as the real data. Because pairs that share a language are not independent, the cluster bootstrap resamples whole languages with replacement, and each drawn language brings all of its pairs. Repeating the evaluation on each resample expresses how much the reported score could plausibly vary. The matched-capacity control is a random forest of identical size and tuning budget that receives five non-typological descriptors instead of typological features. A gap in its favor therefore reflects the typological features, rather than model capacity or tuning. Table 5: Shared setup underlying all experiments in Table 6. All typology experiments use the Grambank+WALS feature source and the one-decimal ATLAS target atn=24, which is 552 directed pairs, unless a row of Table 6 states a deviation. ComponentExact configuration Empirical targetthe directed BTS transfer matrix digitized from Figure C.2 of the ATLAS paper, read at one decimal Typological featuresGrambank with 195 features and WALS with 192, keyed by Glottocode; the union carries 387 features Modeling set ATLAS-24: the 24 ATLAS languages with Grambank coverage of at least0.70; the floor is Grambank-only, so WALS adds feature columns but no languages; all 24Ă 23 directed pairs with sÌž= t, which is 552 Input representation[M s ,M t , â] , the one-hot code profiles of source and target plus a per-feature disagreement block, as defined above Main modelrandom forest with 200 trees,random_state= 42, andmax_features= 0.5; under WALS alonemax_featuresislog2; all remaining parameters at scikit- learn defaults Ridge modelsα=1.0throughout: the similarity proxy, the GLM on similarity, script, genealogi- cal, and geographic distances, the script-only model, and the feature-importance ranker; the bias ridge standardizes its inputs first Cross-validationleave-one-language-out with 24 folds; a pair is held out exactly when either endpoint is the held-out language; each fold refits the model; held-out predictions are pooled over folds, and a pair predicted in two folds resolves to the fold of its alphabetically later language Metricsheld-out Pearson Ï and true R 2 = 1â S res /S tot Seeds and records seed 42 throughout, with seeds 43 and 44 for seed stability; every result CSV carries a_meta.jsonholding the resolved configuration and the git commit; locked environment: Python 3.12, scikit-learn 1.9, NumPy 2.5, pandas 3.0, SciPy 1.18 7 Table 6: Exact configuration of the five experiments E1 to E5 and their sub- analyses. Each experiment addresses the research question named in its group header, with E1 to E4 probing RQ1 and E5 probing RQ2. Settings inherited from Table 5 are not repeated. Repository scripts preserve their own identifiers, so the experiment labels E1 to E5 are assigned here for readability. AnalysisExact configuration E1 (RQ1): held-out performance and capacity Tier comparisonfour tiers under identical folds: the similarity-proxy ridge, the GLM ridge, the 200-tree RF at scikit-learn defaults, and the tuned RF; metadata single- feature ridges for script, family, and Wikipedia and their combination, each with within-fold standardization on its own coverage-restricted uni- verse: script and Wikipedia cover 24 languages, the WALS family map covers 22 Distance baselinesfive non-typological distance baselines under the identical folds and met- rics, each a single-column ridge except where noted, with within-fold standardization on the full 24-language universe: the lang2vec cosine over the concatenated imputed syntax, phonology, and inventory vectors; the URIEL design with six columns, genetic from the language-family vector, geographic as the great-circle distance on Glottolog coordinates normal- ized by half the circumference, syntactic, phonological, and inventory as one minus the cosine of the respective vectors, and featural as their mean; and three Jaccard overlaps over per-language Wikipedia lead-paragraph corpora of 2,000 random articles per edition with a committed sha256 manifest, namely byte-level BPE vocabularies of 8,000 entries, pooled character 1-to-3-gram sets over the 5,000 most frequent word types, and the 10,000 most frequent word types; lang2vec 1.1.2 imputed feature sets cover all 24 languages, and the learned embeddings are unusable because seven of the 24 languages are absent and the embedding file does not load under NumPy 2 Leave-k-languages- out k â 1, 2, 3;k=1is the deterministic LOLO partition; fork â„ 2the tuned RF averages over 50 random disjoint partitions of the 24 languages and the remaining tiers over 25; a pair is held out when either endpoint is in the held-out group; the reportedÏis the partition mean and the interval is the partition spread; seed 42 Permutation nullfold-respecting: within each fold only the training labels are permuted and the held-out labels are untouched;B=200;p = (1 + #null â„ observed)/(B+1); seed 42 Cluster bootstrap 95% intervals by resampling the 24 source languages with replacement, B=2000, seed 42; paired bootstrap ofâÏagainst the script-only ridge and against the GLM on shared pairs; a language-level variant resamples the 24 languages and reweights each pair by the product of how often its two endpoints were drawn Feature sources Grambank with 195 features and Grambank+WALS with 387 run atn=24; WALS runs at n=22 because WALS lacks Glottocode mappings for two of the 24 languages; each source carries its own sweptmax_features, which is 0.5, 0.5, and log2 in the order above Shared-language sources all three sources refit on the 22 ATLAS-24 languages covered by the WALS mapping, 462 directed pairs; per-sourcemax_featuresmini- sweep over the seven grid values; identical LOLO folds; seed 42 WALS-block permuta- tion the WALS block of the wide matrix permuted across the 24 languages, preserving each featureâs marginal levels and missingness; ten permu- tations, each under the seven-valuemax_featuresmini-sweep on the production LOLO folds atn=24; forest seed 42, permutation seeds 1000 to 1009 8 Table 6: Exact configuration of Experiments E1 to E5, continued. AnalysisExact configuration Coverage-floor sweepthe per-language Grambank coverage floor lowered from0.70to0.60, 0.50,0.40, and0.00, admitting one language per step, namely Telugu, Thai, Indonesian, and Central Kurdish; every tier refit at each floor under the production configuration and the identical LOLO protocol; the tuned RF is additionally scored under the pure-unseen-language protocol at each floor; seed 42 Digitization-noise ro- bustness 100 refits of the tuned RF on targets perturbed by independent U (â0.05, 0.05)noise per cell, forest seed fixed at 42, scored on the clean target Pure-unseen-language protocol repeated grouped 5-fold over languages with five repeats; test pairs are those whose both languages are held out, so training pairs never touch a held-out language; seed 42, with three seeds in the refinement stage Matched-capacity con- trol a non-typological RF over the one-hot per-language metadata profile: script with 9 levels, WALS family, WALS genus, ak-means geographic region with 6 clusters, and a Wikipedia article-count quartile bin; missing values get an explicit unknown level, which favors metadata and is the conservative direction; the same pair design, 200-tree RF, and LOLO folds as the typology model; the budget-matched row averages LOLOÏ over 20 random 5-feature typology subsets drawn from the top 15 of the global importance ranking, seed 42 Matched-tuning con- trol the metadata RF swept over the same curated grid as the typology model; all 168 configurations screened under LOLO at 200 trees, seed 42; top 8 refined over seeds 42 to 46; the winner by trueR 2 ismax_features sqrtwithmax_depth10,min_samples_leaf1, andmax_samples 1.0; leave-2-languages-out over 50 random partitions; paired cluster boot- strap ofâÏagainst the tuned typology model with source-language blocks, B=2000, seed 42 Hyperparameter sweepcurated grid overmax_featuresâ0.2, 0.3, 0.5, 0.7, 1.0, sqrt, log2, max_depth â None, 6, 10, 16,min_samples_leaf â 1, 2, 5, and max_samplesâ0.6, 1.0; screening under LOLO at 200 trees and one seed; refinement of the top 8 configurations plus the scikit-learn-default and previously tuned references under both protocols at 200 and 500 trees, with five LOLO seeds and three unseen-language seeds; the metric of record is trueR 2 with Pearson as tie-break; the per-source winner maximizes the combined rank of LOLO and unseen-language R 2 E2 (RQ1): within-script typology isolation Script stratificationpooled LOLO predictions split into same-script, cross-script, and Latinâ Latin pairs; four models: the script-only ridge, the GLM with the script column, the GLM without the script column, and the tuned RF; run under Grambank+WALS and under Grambank alone; subgroup 95% intervals by cluster bootstrap, B=2000 E3 (RQ1): out-of-distribution taxonomy Grouped holdoutleave-one-script-out with groups by writing script and leave-one-family- out with groups by WALS major family; the fold for groupgholds out every pair touching a language ofg; per-groupÏis scored when the group has at least three held-out pairs; the macro value is the unweighted mean over groups; 95% intervals by one-sample bootstrap over held-out groups, B=2000 Composition-matched baseline pooled LOLO predictions restricted to the held-out pairs of each script or family group, macro-averaged and pair-weighted, compared against grouped holdout on the same groups 9 Table 6: Exact configuration of Experiments E1 to E5, continued. AnalysisExact configuration Symmetriccross- check per-factor Mantelrand MRM withB=999permutations of the response- matrix labels on the 276 upper-triangle cells; the response is symmetrized BTS, the mean of the two directions, treated as a distance; predictors are minâmax-normalized typological, genealogical, geographic, and script distances; partial Mantel is deliberately not used because of its Type-I error inflation E4 (RQ1): where the signal lives Feature rankingridge permutation importance withα=1and seed 42; the importance of a feature is the drop in held-outÏwhen its one-hot columns and disagreement bit are permuted within each foldâs training set; the global ranking pools this over folds, and the stability analysis keeps the 24 per-fold rankings Top-k sweepk â1, 5, 10, 25, 50, 100, 200plus the full 387; features ordered by the global ranking; the same tuned RF; the ranking is scored on the evaluation folds, so entries with k < 387 are optimistically biased Leak-free selectioneach fold takes its own top-kby mean importance over the other 23 folds, so the held-out language touches neither the fit nor the ranking; the same RF; seed stability over seeds 42, 43, and 44 Importance stabilitymean pairwise Spearman correlation over the 24 per-fold importance vec- tors and mean pairwise top-25 Jaccard overlap; 95% interval by bootstrap over fold pairs, B=2000 E5 (RQ2): the best pre-training source Operationalizationsthe candidate set and the target set are both the full ATLAS-24; empirical is the mean raw BTS per source over targets; the debiased residual is the mean residualT, with an on-scale variant Ë T + Ë B; centrality is the mean typological similarity to the other 23 languages on the Grambank+WALS features; the OOD-generalizer is the mean raw BTS to targets from a different WALS major family; the on-scale reweighted model is the 200- tree RF on raw BTS with sample weight1/|block|per unordered script- by-genus block; the top 10 sources are reported per operationalization, and the full 24 in Table 19 Bias layerridge withα=1on standardized script difference, genealogical distance, minâmax-normalized geographic distance, andlog 1pWikipedia article counts of source and target, fitted in-sample on all 552 pairs; the residual T = BTSâ Ë Bis mean-zero by construction; robustness is probed by refitting on all 31 non-empty descriptor subsets, Table 18 Leak-free bias layer Ë Brefit out-of-fold under the LOLO folds, each pairâs bias predicted by a ridge that never saw its held-out language, and the residual ranking recomputed as the raw mean residual and through the production rf T Residual permutation null the full operationalization, bias ridge, residual, rf T , and per-source rank- ing, rerun on permuted BTS labels,B=200, seed 42, recording the top- source margin over the median source and the gap over English per replicate Residual gap testthe difference of mean residuals between two sources is bootstrapped over targets, resampling targets and pairing both sources on the same resample, B=2000 Compute and provenance All experiments ran on a single CPU of a local workstation, and no GPU, cluster, or cloud resources were used. The models are scikit-learn random forests and ridge regressions on 552 directed pairs, so each configuration fits in seconds at this scale. Grambank v1.0 and WALS Online are distributed under C-BY 4.0, lang2vec under C-BY-SA 4.0, scikit-learn under the BSD 3-Clause license, and Wikipedia text under C-BY-SA 4.0. The ATLAS paper is 10 published at ICLR 2026, and we digitize the values of its Figure C.2 for analysis only. Our released code is MIT-licensed. Table 7: The 387 typological features used in this work, in natural ID order within each source: 195 Grambank parameters and 192 WALS features. IDFeatureIDFeature Grambank (195 features) GB020Are there definite or specific articles?GB021Do indefinite nominals commonly have indefinite articles? GB022Are there prenominal articles?GB023Are there postnominal articles? GB024 What is the order of numeral and noun in the NP?GB025What is the order of adnominal demonstrative and noun? GB026 Can adnominal property words occur discontinu- ously? GB027Are nominal conjunction and comitative expressed by different elements? GB028Is there a distinction between inclusive and exclu- sive? GB030 Is there a gender distinction in independent 3rd person pronouns? GB031Is there a dual or unit augmented form (in addition to plural or augmented) for all person categories in the pronoun system? GB035Are there three or more distance contrasts in demonstratives? GB036Do demonstratives show an elevation distinction?GB037 Do demonstratives show a visible-nonvisible dis- tinction? GB038Are there demonstrative classifiers?GB039Is there nonphonological allomorphy of noun num- ber markers? GB041Are there several nouns (more than three) which are suppletive for number? GB042Is there productive overt morphological singular marking on nouns? GB043Is there productive morphological dual marking on nouns? GB044 Is there productive morphological plural marking on nouns? GB046Is there an associative plural marker for nouns?GB047Is there a productive morphological pattern for deriving an action/state noun from a verb? GB048Is there a productive morphological pattern for deriving an agent noun from a verb? GB049 Is there a productive morphological pattern for deriving an object noun from a verb? GB051Is there a gender/noun class system where sex is a factor in class assignment? GB052Is there a gender/noun class system where shape is a factor in class assignment? GB053Is there a gender/noun class system where animacy is a factor in class assignment? GB054 Is there a gender/noun class system where plant status is a factor in class assignment? GB057Are there numeral classifiers?GB058Are there possessive classifiers? GB059Is the adnominal possessive construction different for alienable and inalienable nouns? GB065What is the pragmatically unmarked order of ad- nominal possessor noun and possessed noun? GB068 Do core adjectives (defined semantically as prop- erty concepts such as value, shape, age, dimension) act like verbs in predicative position? GB069Do core adjectives (defined semantically as prop- erty concepts; value, shape, age, dimension) used attributively require the same morphological treat- ment as verbs? GB070 Are there morphological cases for non-pronominal core arguments (i.e. S/A/P)? GB071Are there morphological cases for pronominal core arguments (i.e. S/A/P)? GB072Are there morphological cases for oblique non- pronominal NPs (i.e. not S/A/P)? GB073Are there morphological cases for independent oblique personal pronominal arguments (i.e. not S/A/P)? GB074Are there prepositions?GB075Are there postpositions? GB079Do verbs have prefixes/proclitics, other than those that only mark A, S or P (do include portmanteau: A & S + TAM)? GB080Do verbs have suffixes/enclitics, other than those that only mark A, S or P (do include portmanteau: A & S + TAM)? GB081Is there productive infixation in verbs?GB082Is there overt morphological marking of present tense on verbs? GB083Is there overt morphological marking on the verb dedicated to past tense? GB084Is there overt morphological marking on the verb dedicated to future tense? GB086Is a morphological distinction between perfective and imperfective aspect available on verbs? GB089Can the S argument be indexed by a suffix/enclitic on the verb in the simple main clause? GB090Can the S argument be indexed by a prefix/proclitic on the verb in the simple main clause? GB091Can the A argument be indexed by a suffix/enclitic on the verb in the simple main clause? GB092Can the A argument be indexed by a pre- fix/proclitic on the verb in the simple main clause? GB093Can the P argument be indexed by a suffix/enclitic on the verb in the simple main clause? GB094Can the P argument be indexed by a prefix/proclitic on the verb in the simple main clause? GB095Are variations in marking strategies of core partici- pants based on TAM distinctions? GB096Are variations in marking strategies of core partici- pants based on verb classes? GB098Are variations in marking strategies of core partici- pants based on person distinctions? 11 (continued) IDFeatureIDFeature GB099Can verb stems alter according to the person of a core participant? GB103 Is there a benefactive applicative marker on the verb (including indexing)? GB104Is there an instrumental applicative marker on the verb (including indexing)? GB105 Can the recipient in a ditransitive construction be marked like the monotransitive patient? GB107Can standard negation be marked by an affix, clitic or modification of the verb? GB108Is there directional or locative morphological mark- ing on verbs? GB109Is there verb suppletion for participant number?GB110Is there verb suppletion for tense or aspect? GB111Are there conjugation classes?GB113 Are there verbal affixes or clitics that turn intransi- tive verbs into transitive ones? GB114 Is there a phonologically bound reflexive marker on the verb? GB115Is there a phonologically bound reciprocal marker on the verb? GB116 Do verbs classify the shape, size or consistency of absolutive arguments by means of incorporated nouns, verbal affixes or suppletive verb stems? GB117Is there a copula for predicate nominals? GB118Are there serial verb constructions?GB119 Can mood be marked by an inflecting word ("aux- iliary verb")? GB120Can aspect be marked by an inflecting word ("aux- iliary verb")? GB121 Can tense be marked by an inflecting word ("auxil- iary verb")? GB122Is verb compounding a regular process?GB123Are there verb-adjunct (aka light-verb) construc- tions? GB124Is incorporation of nouns into verbs a productive intransitivizing process? GB126Is there an existential verb? GB127 Are different posture verbs used obligatorily de- pending on an inanimate locatumâs shape or posi- tion (e.g. âto lieâ vs. âto standâ)? GB129Is there a notably small number, i.e. about 100 or less, of verb roots in the language? GB130What is the pragmatically unmarked order of S and V in intransitive clauses? GB131Is a pragmatically unmarked constituent order verb- initial for transitive clauses? GB132Is a pragmatically unmarked constituent order verb- medial for transitive clauses? GB133 Is a pragmatically unmarked constituent order verb- final for transitive clauses? GB134Is the order of constituents the same in main and subordinate clauses? GB135Do clausal objects usually occur in the same posi- tion as nominal objects? GB136Is the order of core argument (i.e. S/A/P) con- stituents fixed? GB137Can standard negation be marked clause-finally? GB138Can standard negation be marked clause-initially?GB139Is there a difference between imperative (pro- hibitive) and declarative negation constructions? GB140Is verbal predication marked by the same negator as all of the following types of predication: loca- tional, existential and nominal? GB146 Is there a morpho-syntactic distinction between predicates expressing controlled versus uncon- trolled events or states? GB147Is there a morphological passive marked on the lexical verb? GB148Is there a morphological antipassive marked on the lexical verb? GB149 Is there a morphologically marked inverse on verbs? GB150Is there clause chaining? GB151Is there an overt verb marker dedicated to sig- nalling coreference or noncoreference between the subject of one clause and an argument of an adja- cent clause ("switch reference")? GB152Is there a morphologically marked distinction be- tween simultaneous and sequential clauses? GB155 Are causatives formed by affixes or clitics on verbs? GB156Is there a causative construction involving an ele- ment that is unmistakably grammaticalized from a verb for âto sayâ? GB158Are verbs reduplicated?GB159Are nouns reduplicated? GB160Are elements apart from verbs or nouns redupli- cated? GB165Is there productive morphological trial marking on nouns? GB166Is there productive morphological paucal marking on nouns? GB167Is there a logophoric pronoun? GB170Can an adnominal property word agree with the noun in gender/noun class? GB171Can an adnominal demonstrative agree with the noun in gender/noun class? GB172Can an article agree with the noun in gender/noun class? GB177Can the verb carry a marker of animacy of argu- ment, unrelated to any gender/noun class of the argument visible in the NP domain? GB184 Can an adnominal property word agree with the noun in number? GB185Can an adnominal demonstrative agree with the noun in number? GB186Can an article agree with the noun in number?GB187Is there any productive diminutive marking on the noun (exclude marking by system of nominal clas- sification only)? 12 (continued) IDFeatureIDFeature GB188Is there any productive augmentative marking on the noun (exclude marking by system of nominal classification only)? GB192Is there a gender system where a nounâs phonolog- ical properties are a factor in class assignment? GB193What is the order of adnominal property word and noun? GB196Is there a male/female distinction in 2nd person independent pronouns? GB197Is there a male/female distinction in 1st person independent pronouns? GB198 Can an adnominal numeral agree with the noun in gender/noun class? GB203What is the order of the adnominal collective uni- versal quantifier (âallâ) and the noun? GB204Do collective (âallâ) and distributive (âeveryâ) uni- versal quantifiers differ in their forms or their syn- tactic positions? GB250Can predicative possession be expressed with a transitive âhabeoâ verb? GB252 Can predicative possession be expressed with an S-like possessum and a locative-coded possessor? GB253Can predicative possession be expressed with an S-like possessum and a dative-coded possessor? GB254Can predicative possession be expressed with an S-like possessum and a possessor that is coded like an adnominal possessor? GB256Can predicative possession be expressed with an S-like possessor and a possessum that is coded like a comitative argument? GB257Can polar interrogation be marked by intonation only? GB260Can polar interrogation be indicated by a special word order? GB262Is there a clause-initial polar interrogative particle? GB263Is there a clause-final polar interrogative particle?GB264Is there a polar interrogative particle that most com- monly occurs neither clause-initially nor clause- finally? GB265 Is there a comparative construction that includes a form that elsewhere means âsurpass, exceedâ? GB266 Is there a comparative construction that employs a marker of the standard which elsewhere has a locational meaning? GB270Can comparatives be expressed using two con- joined clauses? GB273 Is there a comparative construction with a stan- dard marker that elsewhere has neither a locational meaning nor a âsurpass/exceedâ meaning? GB275Is there a bound comparative degree marker on the property word in a comparative construction? GB276Is there a non-bound comparative degree marker modifying the property word in a comparative con- struction? GB285Can polar interrogation be marked by a question particle and verbal morphology? GB286Can polar interrogation be indicated by overt ver- bal morphology only? GB291Can polar interrogation be marked by tone?GB296Is there a phonologically or morphosyntactically definable class of ideophones that includes ideo- phones depicting imagery beyond sound? GB297Can polar interrogation be indicated by a V-not-V construction? GB298Can standard negation be marked by an inflecting word ("auxiliary verb")? GB299 Can standard negation be marked by a non- inflecting word ("auxiliary particle")? GB300Does the verb for âgiveâ have suppletive verb forms? GB301Is there an inclusory construction?GB302Is there a phonologically free passive marker ("par- ticle" or "auxiliary")? GB303 Is there a phonologically free antipassive marker ("particle" or "auxiliary")? GB304Can the agent be expressed overtly in a passive clause? GB305Is there a phonologically independent reflexive pro- noun? GB306Is there a phonologically independent non-bipartite reciprocal pronoun? GB309Are there multiple past or multiple future tenses, distinguishing distance from Time of Reference? GB312Is there overt morphological marking on the verb dedicated to mood? GB313Are there special adnominal possessive pronouns that are not formed by an otherwise regular pro- cess? GB314Can augmentative meaning be expressed produc- tively by a shift of gender/noun class? GB315Can diminutive meaning be expressed productively by a shift of gender/noun class? GB316Is singular number regularly marked in the noun phrase by a dedicated phonologically free element? GB317Is dual number regularly marked in the noun phrase by a dedicated phonologically free element? GB318Is plural number regularly marked in the noun phrase by a dedicated phonologically free element? GB319Is trial number regularly marked in the noun phrase by a dedicated phonologically free element? GB320Is paucal number regularly marked in the noun phrase by a dedicated phonologically free element? GB321Is there a large class of nouns whose gender/noun class is not phonologically or semantically pre- dictable? GB322Is there grammatical marking of direct evidence (perceived with the senses)? GB323Is there grammatical marking of indirect evidence (hearsay, inference, etc.)? GB324Is there an interrogative verb for content interroga- tives (who?, what?, etc.)? GB325Is there a count/mass distinction in interrogative quantifiers? GB326Do (nominal) content interrogatives normally or frequently occur in situ? 13 (continued) IDFeatureIDFeature GB327Can the relative clause follow the noun?GB328Can the relative clause precede the noun? GB329Are there internally-headed relative clauses?GB330Are there correlative relative clauses? GB331Are there non-adjacent relative clauses?GB333Is there a decimal numeral system? GB334 Is there synchronic evidence for any element of a quinary numeral system? GB335Is there synchronic evidence for any element of a vigesimal numeral system? GB336Is there a body-part tallying system?GB400Are all person categories neutralized in some voice, tense, aspect, mood and/or negation? GB401Is there a class of patient-labile verbs?GB402Does the verb for âseeâ have suppletive verb forms? GB403Does the verb for âcomeâ have suppletive verb forms? GB408Is there any accusative alignment of flagging? GB409Is there any ergative alignment of flagging?GB410Is there any neutral alignment of flagging? GB415Is there a politeness distinction in 2nd person forms? GB421Is there a preposed complementizer in comple- ments of verbs of thinking and/or knowing? GB422Is there a postposed complementizer in comple- ments of verbs of thinking and/or knowing? GB430 Can adnominal possession be marked by a prefix on the possessor? GB431Can adnominal possession be marked by a prefix on the possessed noun? GB432Can adnominal possession be marked by a suffix on the possessor? GB433 Can adnominal possession be marked by a suffix on the possessed noun? GB519Can mood be marked by a non-inflecting word ("auxiliary particle")? GB520Can aspect be marked by a non-inflecting word ("auxiliary particle")? GB521Can tense be marked by a non-inflecting word ("auxiliary particle")? GB522Can the S or A argument be omitted from a pragmatically unmarked clause when the refer- ent is inferrable from context ("pro-drop" or "null anaphora")? WALS (192 features) 1AConsonant Inventories2AVowel Quality Inventories 3AConsonant-Vowel Ratio4AVoicing in Plosives and Fricatives 5AVoicing and Gaps in Plosive Systems6AUvular Consonants 7AGlottalized Consonants8ALateral Consonants 9AThe Velar Nasal10AVowel Nasalization 10BNasal Vowels in West Africa11AFront Rounded Vowels 12ASyllable Structure13ATone 14AFixed Stress Locations15AWeight-Sensitive Stress 16A Weight Factors in Weight-Sensitive Stress Systems17ARhythm Types 18AAbsence of Common Consonants19APresence of Uncommon Consonants 20AFusion of Selected Inflectional Formatives21AExponence of Selected Inflectional Formatives 21BExponence of Tense-Aspect-Mood Inflection22AInflectional Synthesis of the Verb 23ALocus of Marking in the Clause24ALocus of Marking in Possessive Noun Phrases 25ALocus of Marking: Whole-language Typology25BZero Marking of A and P Arguments 26A Prefixing vs. Suffixing in Inflectional Morphology27AReduplication 28ACase Syncretism29ASyncretism in Verbal Person/Number Marking 30ANumber of Genders31ASex-based and Non-sex-based Gender Systems 32ASystems of Gender Assignment33ACoding of Nominal Plurality 34AOccurrence of Nominal Plurality35APlurality in Independent Personal Pronouns 36AThe Associative Plural37ADefinite Articles 38AIndefinite Articles39A Inclusive/Exclusive Distinction in Independent Pronouns 39BInclusive/Exclusive Forms in Pama-Nyungan40AInclusive/Exclusive Distinction in Verbal Inflection 41ADistance Contrasts in Demonstratives42APronominal and Adnominal Demonstratives 43AThird Person Pronouns and Demonstratives44AGender Distinctions in Independent Personal Pro- nouns 45APoliteness Distinctions in Pronouns46AIndefinite Pronouns 47AIntensifiers and Reflexive Pronouns48APerson Marking on Adpositions 49ANumber of Cases50AAsymmetrical Case-Marking 51APosition of Case Affixes52AComitatives and Instrumentals 53AOrdinal Numerals54ADistributive Numerals 55ANumeral Classifiers56AConjunctions and Universal Quantifiers 57APosition of Pronominal Possessive Affixes58AObligatory Possessive Inflection 58BNumber of Possessive Nouns59APossessive Classification 60AGenitives, Adjectives and Relative Clauses61AAdjectives without Nouns 62AAction Nominal Constructions63ANoun Phrase Conjunction 64ANominal and Verbal Conjunction65APerfective/Imperfective Aspect 66AThe Past Tense67AThe Future Tense 14 (continued) IDFeatureIDFeature 68AThe Perfect69APosition of Tense-Aspect Affixes 70AThe Morphological Imperative71AThe Prohibitive 72AImperative-Hortative Systems73AThe Optative 74ASituational Possibility75AEpistemic Possibility 76A Overlap between Situational and Epistemic Modal Marking 77ASemantic Distinctions of Evidentiality 78ACoding of Evidentiality79ASuppletion According to Tense and Aspect 79BSuppletion in Imperatives and Hortatives80AVerbal Number and Suppletion 81AOrder of Subject, Object and Verb81BLanguages with two Dominant Orders of Subject, Object, and Verb 82AOrder of Subject and Verb83AOrder of Object and Verb 84AOrder of Object, Oblique, and Verb85AOrder of Adposition and Noun Phrase 86AOrder of Genitive and Noun87AOrder of Adjective and Noun 88AOrder of Demonstrative and Noun89AOrder of Numeral and Noun 90AOrder of Relative Clause and Noun90BPrenominal relative clauses 90CPostnominal relative clauses90DInternally-headed relative clauses 90ECorrelative relative clauses90FAdjoined relative clauses 90GDouble-headed relative clauses91AOrder of Degree Word and Adjective 92APosition of Polar Question Particles93A Position of Interrogative Phrases in Content Ques- tions 94AOrder of Adverbial Subordinator and Clause95ARelationship between the Order of Object and Verb and the Order of Adposition and Noun Phrase 96A Relationship between the Order of Object and Verb and the Order of Relative Clause and Noun 97ARelationship between the Order of Object and Verb and the Order of Adjective and Noun 98AAlignment of Case Marking of Full Noun Phrases99AAlignment of Case Marking of Pronouns 100AAlignment of Verbal Person Marking101AExpression of Pronominal Subjects 102AVerbal Person Marking103AThird Person Zero of Verbal Person Marking 104AOrder of Person Markers on the Verb105ADitransitive Constructions: The Verb âGiveâ 106AReciprocal Constructions107APassive Constructions 108AAntipassive Constructions108BProductivity of the Antipassive Construction 109AApplicative Constructions109BOther Roles of Applied Objects 110APeriphrastic Causative Constructions111ANonperiphrastic Causative Constructions 112ANegative Morphemes113ASymmetric and Asymmetric Standard Negation 114ASubtypes of Asymmetric Standard Negation115A Negative Indefinite Pronouns and Predicate Nega- tion 116APolar Questions117APredicative Possession 118APredicative Adjectives119ANominal and Locational Predication 120AZero Copula for Predicate Nominals121AComparative Constructions 122ARelativization on Subjects123ARelativization on Obliques 124AâWantâ Complement Subjects125APurpose Clauses 126AâWhenâ Clauses127AReason Clauses 128AUtterance Complement Clauses129AHand and Arm 130AFinger and Hand130B Cultural Categories of Languages with Identity of âFingerâ and âHandâ 131ANumeral Bases132ANumber of Non-Derived Basic Colour Categories 133ANumber of Basic Colour Categories134AGreen and Blue 135ARed and Yellow136AM-T Pronouns 136BM in First Person Singular137AN-M Pronouns 137BM in Second Person Singular138ATea 139AIrregular Negatives in Sign Languages140AQuestion Particles in Sign Languages 141AWriting Systems142APara-Linguistic Usages of Clicks 143AOrder of Negative Morpheme and Verb143BObligatory Double Negation 143COptional Double Negation143DOptional Triple Negation 143EPreverbal Negative Morphemes143FPostverbal Negative Morphemes 143G Minor morphological means of signaling negation144APosition of Negative Word With Respect to Subject, Object, and Verb 144BPosition of negative words relative to beginning and end of clause and with respect to adjacency to verb 144CLanguages with different word order in negative clauses 144DThe Position of Negative Morphemes in SVO Lan- guages 144E Multiple Negative Constructions in SVO Lan- guages 144FObligatory Double Negation in SVO languages144GOptional Double Negation in SVO languages 144HNegSVO Order144ISNegVO Order 144JSVNegO Order144KSVONeg Order 15 (continued) IDFeatureIDFeature 144LThe Position of Negative Morphemes in SOV Lan- guages 144M Multiple Negative Constructions in SOV Lan- guages 144NObligatory Double Negation in SOV languages144OOptional Double Negation in SOV languages 144PNegSOV Order144QSNegOV Order 144RSONegV Order144SSOVNeg Order 144T The Position of Negative Morphemes in Verb- Initial Languages 144UDouble negation in verb-initial languages 144VVerb-Initial with Preverbal Negative144W Verb-Initial with Negative that is Immediately Postverbal or between Subject and Object 144XVerb-Initial with Clause-Final Negative144Y The Position of Negative Morphemes in Object- Initial Languages B Ablations Table 8: Feature-source comparison under identical per-source tuning, from Experiment E1. The LOLO ordering is preserved on the pure-unseen-language metric, so the Grambank-only advantage is not overfitting. The LOLO columns use each sourceâs production configuration; the unseen-language column uses the per-source sweep winner. WALS runs atn=22; Grambank and Grambank+WALS run at n=24. Feature source#featLOLO ÏLOLO R 2 unseen-lang Ï Grambank (GB)1950.7800.6070.494 WALS1920.6480.4120.307 Grambank+WALS (operational)3870.7050.4920.405 Shared-language source comparisonTable 8 compares Grambank atn=24with WALS atn=22, so its rows rest on different language sets. Table 9 removes this caveat: all three sources are refit on the 22 languages that the WALS mapping covers, each after its ownmax_featuresmini-sweep, on identical LOLO folds. The ordering Grambank>Grambank+WALS>WALS is preserved, at Ï=0.716,0.673and0.648. The Grambank advantage therefore does not depend on the language set. The productionmax_features= 0.5, tuned atn=24, drops to0.576on the shared set. Removing two languages removes 90 directed pairs, from 552 down to 462, so a configuration tuned on the larger set no longer suits the smaller training problem. This is why each source carries its own sweep on its own language set. Table 9: Feature-source comparison on the 22 languages shared with WALS (Experiment E1), each source after its ownmax_featuresmini-sweep on identical LOLO folds. The order of Table 8 is preserved on identical languages. Feature sourcebest max_featuresLOLO Ï R 2 Grambank0.30.7160.511 Grambank+WALSsqrt0.6730.447 WALSlog20.6480.412 WALS-block permutation control Tables 8 and 9 show the union trailing Grambank under matched tuning. The permutation control asks whether that drop is generic column dilution at n=24, or a property of the real WALS data. The WALS block of the feature matrix is permuted across the 24 languages, so each language receives another languageâs real WALS profile. This destroys the feature-to-language association, while every featureâs marginal levels and missingness are preserved. Each permuted design is rerun under the same seven-valuemax_featuresmini-sweep on the production LOLO folds. Scrambling the WALS block recovers most of the unionâs drop: the per-design best configuration averagesÏ=0.751with standard deviation0.011over ten permutations, and all ten score above the real unionâs0.705, against Grambankâs0.780(Table 10). The unionâs drop therefore decomposes into roughly0.03of generic dilution, which any matched block of 192 16 extra columns causes, and roughly0.05that is specific to the real WALS structure. The forest fits WALS-to-BTS associations within the training folds that do not transfer to the held-out language. The WALS features are not noise in isolation, since WALS alone reachesÏ=0.648, but combined with Grambank at n=24 they subtract held-out signal rather than adding it. Table 10: WALS-block permutation control (Experiment E1) under identical LOLO folds and a per-designmax_featuresmini-sweep. The scrambled rows permute the WALS block across the 24 languages, preserving marginal levels and missingness; the mean and standard deviation are over ten permutations, every one of which scores above the real union. Feature design#featLOLO ÏR 2 Grambank1950.7800.607 Grambank+WALS, real block3870.7050.492 Grambank+WALS, block scrambled (10 perms)3870.751± 0.0110.561± 0.016 Coverage-floor sweepThe0.7Grambank coverage threshold defines ATLAS-24, so Table 11 refits every tier as the floor is lowered stepwise to zero. Each step admits exactly one further ATLAS language: Telugu at floor0.60, Thai at0.50, Indonesian at0.40, and Central Kurdish at0.00. Each admitted language is not one training example but2(nâ1)of them, since every ordered pair with an existing language becomes a directed training pair. The corpus therefore grows quadratically, from 552 pairs atn=24to 756 atn=28. Moderately lowering the floor raises the tuned RF fromR 2 =0.49 atn=24to0.68atn=26andn=27. The untuned RF rises with it, collapsing the tuning gap from 0.14to at most0.04. The added languages relieve the overfitting that separates the untuned from the tuned forest. Removing the floor entirely reverses the gain. The near-empty Central Kurdish profile, with Grambank coverage0.08, drops the untuned RF toR 2 =0.20and the tuned RF to0.48. The GLM stays at chance across the entire sweep, so the RF movement tracks the separability of one-hot feature profiles rather than a stronger smooth typological signal. The pure-unseen-language protocol, whose test pairs share no language with training, shows that the moderate floors also improve generalization. Unseen-languageÏrises from0.405at floor0.70to0.508at0.50and0.514 at0.40, nearly doublingR 2 from0.136to0.264. Adequately attested extra languages therefore improve prediction of languages the model has never seen, not just the LOLO score. At floor0.00the same protocol collapses toÏ=â0.012withR 2 =â0.20, below the mean predictor. The near-empty profile does not merely add noise, it breaks extrapolation entirely. A one-language change can also move the RF substantially. As throughout, atn=24every language appears in training through other pairs, so the tuned column reads as threshold sensitivity, while the unseen-language column isolates the floorâs role: excluding profiles too sparse to extrapolate from. Table 11: Coverage-floor sweep of the0.7Grambank threshold that defines ATLAS-24. Every tier is refit at each floor under the production configuration, and each lowering admits the listed language, reaching all 28 ATLAS languages with a Grambank profile at floor0.00. The remaining ten of the 38 ATLAS languages, among them Spanish and German, are absent from Grambank and cannot enter at any floor.R 2 is LOLO trueR 2 ; the final two columns are the tuned RFâs LOLOÏand pure-unseen-language Ï, whose ATLAS-24 values are the headline 0.705 and 0.405. Floor nlanguage addedglm R 2 full R 2 tuned R 2 tuned Ïunseen-lang Ï 0.7024â0.0140.3500.4920.7050.405 0.6025Telugu0.0030.5070.5290.7310.411 0.5026Thai â0.0090.658 0.6800.8240.508 0.4027Indonesian0.0010.6420.6760.823 0.514 0.0028Central Kurdish0.0330.2050.4760.700 â0.012 Matched-capacity control detailsThe metadata RF in Table 12 one-hot encodes, per language, the writing script with 9 levels, the WALS family, the WALS genus, ak-means geographic region from lat/lon with 6 clusters, and a Wikipedia article-count bin with 4 quantiles. Missing values receive an explicit unknown level, so the metadata model is informed of the same missingness the typology model is spared. This choice is biased in favor of metadata, which is the conservative direction for the 17 Table 12: Matched-capacity and matched-tuning controls from Experiment E1. A matched-capacity non-typological RF recoversÏâ0.58at the shared configuration and0.620under its own hyperparam- eter sweep, well above the single-feature ridges. Typologyâs full-capacity increment isâÏ=0.085 under matched tuning and disappears at a matched five-feature budget,0.597against0.620, which locates the advantage in feature breadth. Model (same RF, same folds)#featLOLO Ï R 2 metadata RF (script/family/genus/geo/wiki)50.5810.264 metadata RF, own-sweep tuned50.6200.375 typology RF (Grambank+WALS)3870.7050.492 typology RF, budget-matched to metadata50.597â comparison. The budget-matched typology row averages LOLOÏover 20 random 5-feature subsets drawn from the top informative features under a fixed seed. Code and per-experiment_meta.json records, holding the resolved config and git commit, accompany each result CSV. Matched-tuning control details The production metadata control inheritsmax_features= 0.5 from the typology winner, which is conservative for metadata. We therefore swept the metadata RF over the same curated grid as the typology model. We screened all 168 configurations under LOLO at 200 trees and seed 42, and refined the top eight over seeds 42 to 46. The winner by trueR 2 uses max_features sqrtwithmax_depth10, reaching LOLOÏ=0.620andR 2 =0.375at seed 42, and Ï=0.602±0.011over the five refinement seeds. Leave-2-languages-out gives0.621for the winner and0.611for the production configuration, against0.702for typology. The typology increment is therefore stable across seeds, configurations and holdout protocols, rather than an artifact of unequal tuning. A paired cluster bootstrap over source languages does not, however, separate the two models atn=24:âÏ=0.085at 95% CI[â0.11, 0.41], whose width reflects coarse language-level resampling at this scale. Digitization-noise robustness The target matrix is read from Figure C.2 at one decimal, so every BTS cell carries a rounding error of at most0.05. We perturbed each of the 552 targets with independentU (â0.05, 0.05)noise and reran the full tuned-RF LOLO on each noisy target, with the forest seed fixed so that only the labels vary. AcrossB=100replicates, LOLOÏaverages0.705with standard deviation0.007and worst case0.688.R 2 averages0.492with standard deviation0.011. The main reported correlation is therefore insensitive to digitization error at the reading precision. â0.20.00.20.40.6 LOLO PearsonÏ(permuted labels) 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 Null replicates observedÏ= 0.705 p= 0.005 (B= 200) â0.2â0.10.00.10.20.30.40.5 LOLO trueR 2 (permuted labels) 0 5 10 15 20 observedR 2 = 0.492 permutation nullobserved scorenull mean Figure 1: Label-permutation null versus observed LOLO scores for the tuned random forest on Grambank+WALS features (ATLAS-24,B=200replicates, seed 42). Left: PearsonÏ; right:R 2 . Bars show the fold-respecting permutation null, the dashed line its mean, and the solid line the observed score. The observed Ï exceeds all 200 null replicates (p=0.005). 18 Table 13: Symmetric distance regressions on the24Ă24upper triangle, the cross-check within Experiment E3. Script is a strong marginal correlate but provides no incremental directed power, since the script-only directed model is near chance. FactorMantel rMantel pMRM coefMRM p typological0.2660.0150.2490.281 script0.3340.0170.1490.035 genealogical0.2260.0510.0980.693 geographic0.1690.0780.0290.803 Composition-matched grouped holdout (E3)Macro-Ïunder LOSO and LOFO averages held-out sets that are dominated by cross-script pairs, where every tier is stronger, so the pooled LOLO value of0.705is not the right baseline for them. Table 14 scores pooled LOLO predictions on exactly the pairs each group contributes. Grouped holdout exceeds this matched baseline on the macro average for both axes, and matches it under pair weighting. Individual groups move in both directions. The Arabic script dips toÏ=0.459against0.683matched, and Latin to0.597against0.692, while the single-language script groups improve markedly: Greek at0.935against0.572, and Hebrew at0.914 against0.550. For a single-language group the grouped fold is that languageâs own LOLO fold, so its gain over the matched column reflects the mixed fold provenance of pooled predictions rather than a training difference. No group approaches the script-only model atÏ=0.138, so the dips are attenuations, not a collapse. Table 14: Composition-matched control for the grouped holdouts of Experiment E3. GroupedÏ retrains with every pair touching the group held out; matched LOLO scores the pooled LOLO predictions on exactly that groupâs held-out pairs. Macro rows are unweighted means over groups; weighted rows average by held-out pair count. AxisGroupgrouped Ïmatched LOLO Ï LOSOArabic0.4590.683 LOSOCyrillic0.9350.749 LOSODevanagari0.9120.622 LOSOGreek0.9350.572 LOSOHan0.7300.819 LOSOHangul0.7930.643 LOSOHebrew0.9140.550 LOSOJapanese0.7080.676 LOSOLatin0.5970.692 LOSOmacro0.7760.667 LOSOweighted0.6790.683 LOFOAfro-Asiatic0.9000.718 LOFOAltaic0.8380.676 LOFOAustro-Asiatic0.2430.267 LOFOAustronesian0.8920.801 LOFOIndo-European0.6930.723 LOFOJapanese0.7080.676 LOFOKorean0.7930.643 LOFONiger-Congo0.7600.715 LOFOSino-Tibetan0.7300.819 LOFOmacro0.7290.671 LOFOweighted0.7210.697 Leak-free top-kcontrol (E4) The sweepâs global feature ranking is scored on the same LOLO folds that are used for evaluation, so itsk<387entries are optimistically biased. In the control, foldi selects its own top-kby mean importance over the other folds only, so the held-out language touches neither the fit nor the ranking. Thek=200value falls from0.742to0.710. Across three seeds, the full model scoresÏ = 0.708± 0.004, while leak-freek=200scores0.720± 0.016, so the residual 19 gap is within seed noise. Per-fold selection also beats the global ranking at mid-rangek, for example 0.716 versus 0.627 at k=100, yet it still never significantly beats the full 387-feature model. Inferential statistics. Table 15 collects the permutation and bootstrap statistics behind the main results of E1, E3, E4, and E5. Table 15: Inferential statistics, with each row labeled by the experiment and research question it addresses. The permutation null usesB=200, and the cluster and paired bootstraps useB=2000. The final rowâs interval is the permuted-label nullâs central 95% range, not a bootstrap interval. Exp.RQQuantityEstimate95% CI E1RQ1LOLO Ï; permutation p=0.0050.705[0.51, 0.82] E1RQ1LOLO Ï; language-level bootstrap 95% CI0.705[0.50, 0.85] E1RQ1pure-unseen-language Ï (Grambank+WALS)0.405â E3RQ1LOSO / LOFO macro-Ï0.776 / 0.729[0.67, 0.88] / [0.59, 0.83] E4RQ1top-25 importance Jaccard (randomâ0.033)0.073[0.069, 0.078] E5RQ2residual gap, MarathiâEnglish (bootstrap over targets); p<0.001+0.236[0.19, 0.28] E5RQ2residual margin; null p=0.0050.246[0.065, 0.173] Residual-ranking validity controls Two controls ask whether the E5 residual flip is meaningful, or whether the operationalization produces decisive-looking rankings from nothing. The first removes the in-sample optimism of the bias layer, which explainsR 2 =0.12in-sample butâ0.006out-of- fold under the LOLO folds. Refitting Ë Bout-of-fold and recomputing the ranking, both as the raw mean residual and through the production rf T , leaves the flip intact: Marathi remains first at0.36, English remains at rank 9 or 10, and both rankings correlate with the production ranking at Spearman 0.99. The flip is therefore carried by the nuisance association itself, which replicates across folds at out-of-fold Pearson0.16, rather than by the layerâs in-sample optimism. The second control runs the entire operationalization on permuted BTS labels, refitting the bias ridge, the residual, the random forest, and the ranking per replicate. This quantifies the rankings the pipeline produces from meaningless labels. The null margin of the top source over the median source averages0.116 with interval[0.065, 0.173]overB=200replicates, and no replicate reached the observed0.246, so the rankingâs sharpness exceeds the null atp=0.005. The top-source gap over English specifically is weaker evidence: 11 of 200 null replicates exceed the observed rf T gap of0.237atp=0.055, because a random draw can seat English at the bottom of a null ranking. The residual ranking is thus sharper than what the operationalization produces on meaningless labels and stable under honest bias estimation. The identity of the top source is however weakly determined: the 200 permuted reruns place 24 different languages at rank one. E5 therefore reports the flip as a diagnostic rather than a recommendation. Additional distance baselinesWe compare the typology random forest against five further distance baselines drawn from the transfer and tokenization literature [7,11]. The lang2vec baseline scores each pair by the cosine of the concatenated imputed featural vectors of [7], covering syntax, phonology and inventory. The URIEL baseline feeds the six canonical distances of the same knowledge base, namely genetic, geographic, syntactic, phonological, inventory and their featural aggregate, as a six-column design to the same within-fold ridge as the metadata baselines. The geographic distance is the great-circle distance on Glottolog coordinates, because the lang2vec geographic feature is a vector of distances to fixed reference points rather than per-language coordinates. Three corpus-lexical baselines measure the Jaccard overlap between per-language artifacts of Wikipedia lead-paragraph corpora of 2,000 random articles per edition. These are byte-level BPE vocabularies of 8,000 entries trained per language following [13], pooled character one-to-three-gram sets over the 5,000 most frequent word types in the tradition of [2], and the 10,000 most frequent word types. The last two follow the tokenizer analysis of [11]. Table 16 scores all five under the protocol of the main text; every tier covers all 24 languages. All five sit at or below a trueR 2 of0.044. The charactern-gram and subword-tokenizer overlaps lead atÏof0.225and0.222, followed by the lang2vec cosine at 0.148, the word-vocabulary overlap at0.142, and the URIEL ridge at0.095, which is at chance. The typology random forest explains0.49of held-out variance, eleven times the best distance baseline. The script-adjacent lexical overlaps outscore the typological-distance baselines, which matches the 20 symmetric analysis, where script is the strongest marginal correlate of transfer. None of the five approaches the full typological feature set. Table 16: Additional distance baselines under the leave-one-language-out protocol of Table 1, on ATLAS-24. Every tier covers all 24 languages. The reference row repeats the tuned typology random forest from the main text. BaselineLOLO Ïtrue R 2 character n-gram overlap0.2250.044 subword tokenizer overlap0.2220.044 lang2vec featural cosine0.1480.012 word vocabulary overlap0.1420.010 URIEL six-distance ridge0.095 â0.019 tuned RF (typology, reference)0.7050.492 C Transfer matrices This section visualizes the transfer matrices behind the ATLAS-24 evaluation. Every panel is a 38Ă 38heatmap on the axis order of the digitized ATLAS Figure C.2. The cell in rowi, columnjis the bilingual transfer score (BTS) for transferring from the column language, the source, to the row language, the target. The diagonal, self-transfer, is undefined, and the 14 ATLAS languages without Grambank coverage of at least0.70are masked in grey. Axis labels are colored by language family, at the granularity of the original figureâs legend. All BTS panels, ground truth and four model tiers, share one diverging color scale spanning the ground-truth range [â1.60,0.50] BTS. Its neutral point is the midpoint of that range, not zero. Model panels show leave-one-language-out predictions. Each cell is produced by a training fold that excludes all pairs involving one of the cellâs two languages, so no cell is an in-sample fit. The difference panels share one symmetric scale of±1.35BTS. Red means the model underpredicts the ground truth, blue means it overpredicts. 21 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo ATLAS-38 ground truth (digitized Fig. C.2) â1.50 â1.25 â1.00 â0.75 â0.50 â0.25 0.00 0.25 0.50 BTS (source â target) language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 2: Digitized ATLAS-38 ground-truth transfer matrix (Figure C.2 of the ATLAS paper). This axis order, family coloring, and BTS color scale are shared by all model and difference panels in this section. 22 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo similarity proxy (Ridge·sim) â LOLO predictions â1.50 â1.25 â1.00 â0.75 â0.50 â0.25 0.00 0.25 0.50 BTS (source â target) language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 3: LOLO predictions of the similarity tier (ridge on typological similarity alone). Held-out Pearson Ï = 0.037, MAE = 0.277 BTS. 23 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo GLM (Ridge·sim+script+gene+geo) â LOLO predictions â1.50 â1.25 â1.00 â0.75 â0.50 â0.25 0.00 0.25 0.50 BTS (source â target) language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 4: LOLO predictions of the GLM tier (ridge on similarity, script, genealogical, and geographic distances). Held-out Pearson Ï = 0.162, MAE = 0.277 BTS. 24 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo full RandomForest â LOLO predictions â1.50 â1.25 â1.00 â0.75 â0.50 â0.25 0.00 0.25 0.50 BTS (source â target) language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 5: LOLO predictions of the full random forest (all typological features, default hyperparame- ters). Held-out Pearson Ï = 0.615, MAE = 0.203 BTS. 25 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo tuned RandomForest â LOLO predictions â1.50 â1.25 â1.00 â0.75 â0.50 â0.25 0.00 0.25 0.50 BTS (source â target) language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 6: LOLO predictions of the tuned random forest (the production model of the main ex- periments; max_features= 0.5). Held-out PearsonÏ = 0.705, MAE= 0.185BTS â this is the Ï = 0.705 of the main ATLAS-24 result. 26 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo ground truthâsimilarity proxy (Ridge·sim) â1.0 â0.5 0.0 0.5 1.0 â BTS = gt â prediction language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 7: Ground truth minus the LOLO predictions of the similarity tier, on the shared symmetric difference scale. MAE 0.277 BTS. 27 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo ground truthâGLM (Ridge·sim+script+gene+geo) â1.0 â0.5 0.0 0.5 1.0 â BTS = gt â prediction language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 8: Ground truth minus the LOLO predictions of the GLM tier. MAE 0.277 BTS. 28 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo ground truthâfull RandomForest â1.0 â0.5 0.0 0.5 1.0 â BTS = gt â prediction language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 9: Ground truth minus the LOLO predictions of the full random forest. MAE 0.203 BTS. 29 en ru es fr de pt zh vija sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it roar he hr uk te th yo en ru es fr de pt zh vi ja sw ko cy k tr af no sv el bn ckb fa gu mr ps ur fil id ms ca it ro ar he hr uk te th yo ground truthâtuned RandomForest â1.0 â0.5 0.0 0.5 1.0 â BTS = gt â prediction language family Austro-Asiatic Austronesian Bantu Celtic Dravidian Germanic Hellenic Indo-Iranian Japonic Koreanic Romance Semitic Sinitic Slavic Tai-Kadai Turkic Volta-Niger Figure 10: Ground truth minus the LOLO predictions of the tuned random forest. MAE 0.185 BTS. Note the shrinkage of the error relative to the weaker tiers: held-out error concentrates near the scaleâs neutral midpoint. 30 D Additional results Table 17: Terms of the bias layer Ë B, the ridge fit of BTS on five non-typological nuisance descriptors over the 552 ATLAS-24 directed pairs. Weights are on the standardized feature scale, and the mean absolute contribution is each featureâs contribution to Ë Baveraged over pairs. The layer explains R 2 =0.12of BTS in-sample at correlation0.35, but approximately zero under leave-one-language-out cross-validation. Its weights are therefore descriptive of ATLAS-24, not a stable law. Its mean equals the BTS mean ofâ0.31, so the residualT = BTSâ Ë Bis mean-zero. The source-resource and script terms dominate, while the target-resource term is negligible, so the bias is driven by source-side resource and script sharing. Nuisance descriptorstd. weightmean|contribution| (BTS) source resource (log)+0.0870.070 script difference â0.0770.071 genealogical distance â0.0340.027 geographic distance â0.0180.016 target resource (log)+0.0020.002 31 Table 18: Exhaustive feature-subset ablation of the ATLAS-24 bias layerB, grouped by subset sizek. Each row refitsBon that subset of the five nuisance descriptors and propagates to the debias residualT = BTSâ B. Abbreviations: scr=script difference, gene=genealogical distance, geo=geographic distance, res s /res t =log 1pWikipedia resource of source/target.R 2 (B,y)is the bias-explained variance of BTS, andâ„âTâ„is the relative shift ofTfrom the full-five residual. LOLO Ï is the leave-one-language-out Pearson of the debiased BTS predictor, an RF on T with the bias B added back. The production biased baseline isÏ = 0.705, and the no-bias-layer row refits within this pipeline, landing at0.701. RankÏis the Spearman correlation of the per-source pretraining ranking against the full-model ranking and against empirical BTS ground truth. The top source is English under every configuration;Ìž= marks a flip, and none is observed. Bias featuresR 2 (B,y) â„âTâ„ (%)LOLO Ïrank Ï (full)rank Ï (truth)top k = 1 (1 feature) geo0.01334.80.7190.9890.984English res s 0.06026.00.7180.9900.987English scr0.05128.00.7170.9790.984English gene0.02333.10.7120.9920.988English res t 0.00036.80.7050.9890.990English k = 2 (2 features) scr, res s 0.10612.70.7480.9890.987English scr, gene0.06624.80.7420.9780.982English scr, geo0.06325.30.7340.9780.974English gene, res s 0.07921.40.7290.9830.977English gene, res t 0.02333.10.7230.9930.992English scr, res t 0.05128.00.7220.9830.984English gene, geo0.02532.70.7200.9930.990English geo, res t 0.01334.80.7170.9940.988English geo, res s 0.06824.20.7130.9870.982English res s , res t 0.06125.90.7040.9950.990English k = 3 (3 features) scr, gene, res s 0.1184.50.7501.0000.995English scr, res s , res t 0.10612.60.7460.9950.993English scr, geo, res t 0.06325.30.7400.9870.979English scr, geo, res s 0.1148.10.7380.9950.993English scr, gene, res t 0.06624.80.7340.9890.989English scr, gene, geo0.06923.90.7330.9900.984English gene, geo, res s 0.08021.30.7290.9840.978English gene, res s , res t 0.08021.40.7270.9850.982English geo, res s , res t 0.06924.10.7190.9900.984English gene, geo, res t 0.02532.70.7140.9960.995English k = 4 (4 features) scr, gene, res s , res t 0.1184.40.7600.9990.994English scr, geo, res s , res t 0.1148.10.7570.9980.994English scr, gene, geo, res s 0.1200.50.7560.9970.995English scr, gene, geo, res t 0.06923.90.7330.9870.986English gene, geo, res s , res t 0.08021.20.7230.9910.985English k = 5 (5 features) scr, gene, geo, res s , res t 0.1200.00.7471.0000.995English none (no bias layer)0.00036.80.7010.9940.993English 32 Table 19: Full 24-language ranking of the ATLAS-24 candidate set under each operationalization of Experiment E5. The candidate set is used as both sources and targets. The columns are the empirical raw-BTS mean, the debiased residualT = BTSâ Ë B, the mean typological similarity to the targets, and the OOD-generalizer mean BTS to cross-family targets. Scores are on each columnâs own scale and are not comparable across columns; only the within-column order is meaningful. The on-scale reweighted random forest of Table 4 is omitted here because its full 24-language ranking tracks the empirical one at Spearman0.974, so its top sources are summarized in the main text only. The bold top row shows that the four methods disagree on the best source, which is the RQ2 result. RankempiricaldebiasedcentralityOOD-generalizer sourcescoresourcescoresourcescoresourcescore 1English0.022Marathi0.361Ukrainian0.717English-0.011 2Modern Hebrew-0.074Filipino0.355Portuguese0.700Modern Hebrew-0.065 3French-0.109Modern Hebrew0.328Serbian-Croatian-Bosnian0.679Filipino-0.114 4Filipino-0.113Swahili0.235Catalan0.678Vietnamese-0.114 5Standard Arabic-0.113Standard Arabic0.198Italian0.677Standard Arabic-0.120 6Portuguese-0.113Western Farsi0.187Southern Pashto0.673Western Farsi-0.122 7Marathi-0.113Modern Greek0.180Swedish0.662Turkish-0.124 8Western Farsi-0.113Turkish0.152English0.661Marathi-0.144 9Vietnamese-0.122Standard Malay0.152Modern Greek0.659French-0.144 10Catalan-0.122Vietnamese0.136Welsh0.659Portuguese-0.156 11Turkish-0.130English0.125French0.654Catalan-0.167 12Italian-0.170Catalan0.115Filipino0.654Standard Malay-0.171 13Standard Malay-0.174Serbian-Croatian-Bosnian0.115Standard Malay0.653Swahili-0.181 14Swahili-0.183Portuguese0.102Russian0.647Modern Greek-0.211 15Serbian-Croatian-Bosnian-0.200French0.057Modern Hebrew0.647Italian-0.233 16Swedish-0.204Ukrainian0.051Standard Arabic0.642Swedish-0.233 17Modern Greek-0.217Russian0.041Western Farsi0.623Serbian-Croatian-Bosnian-0.233 18Russian-0.217Italian0.022Marathi0.613Russian-0.267 19Ukrainian-0.230Swedish-0.036Turkish0.611Ukrainian-0.278 20Korean-0.752Korean-0.384Mandarin Chinese0.604Welsh-0.744 21Japanese-0.848Japanese-0.505Korean0.590Korean-0.776 22Welsh-0.883Welsh-0.589Japanese0.579Japanese-0.852 23Mandarin Chinese-1.087Southern Pashto-0.650Swahili0.575Mandarin Chinese-1.110 24Southern Pashto-1.217Mandarin Chinese-0.736Vietnamese0.573Southern Pashto-1.209 Table 20: Full 24-language ranking of the ATLAS-24 candidate set by positive-transfer coverage: the number of target languages with a positive transfer score. The columns are the empirical raw BTS out of 23 targets, the bias-removed residualT = BTSâ Ë Bout of 23 targets, and the OOD-generalizer raw BTS to cross-family targets, out of that sourceâs cross-family target count. Count ties are broken by the corresponding mean score in Table 19. Counts are not comparable across columns; only the within-column order is meaningful. Rankempiricalbias-removedOOD-generalizer sourcecountsourcecountsourcecount 1English8/23Marathi23/23Modern Hebrew3/20 2French5/23Filipino23/23English2/9 3Modern Hebrew3/23Modern Hebrew23/23French1/9 4Portuguese3/23Standard Arabic23/23Modern Greek1/9 5Catalan3/23Western Farsi23/23Filipino0/22 6Modern Greek2/23Modern Greek23/23Vietnamese0/21 7Standard Arabic1/23Standard Malay22/23Standard Arabic0/20 8Italian1/23Swahili21/23Western Farsi0/9 9Swedish1/23Turkish21/23Turkish0/21 10Russian1/23Vietnamese21/23Marathi0/9 11Ukrainian1/23English20/23Portuguese0/9 12Filipino0/23Catalan20/23Catalan0/9 13Marathi0/23Serbian-Croatian-Bosnian20/23Standard Malay0/21 14Western Farsi0/23Portuguese19/23Swahili0/21 15Vietnamese0/23Ukrainian17/23Italian0/9 16Turkish0/23Russian16/23Serbian-Croatian-Bosnian0/9 17Standard Malay0/23French15/23Swedish0/9 18Swahili0/23Italian14/23Russian0/9 19Serbian-Croatian-Bosnian0/23Swedish8/23Ukrainian0/9 20Korean0/23Korean0/23Welsh0/9 21Japanese0/23Japanese0/23Korean0/21 22Welsh0/23Welsh0/23Japanese0/21 23Mandarin Chinese0/23Southern Pashto0/23Mandarin Chinese0/21 24Southern Pashto0/23Mandarin Chinese0/23Southern Pashto0/22 33 Table 21: Full 24-language ranking of the ATLAS-24 candidate set by best-source coverage: the number of target languages for which the source attains the highest transfer score. The columns are the empirical raw BTS, the bias-removed residualT = BTSâ Ë B, and the raw BTS restricted to cross-family sources, the best outside-family option per target. A tie at a targetâs maximum counts for every co-best source, so column counts sum to more than 24: empirically, 11 of 24 targets have tied maxima, up to 9-way. Count ties between sources are broken by the corresponding mean score in Table 19. Only the within-column order is meaningful. Rankempiricalbias-removedOOD-generalizer sourcecountsourcecountsourcecount 1English14/24Marathi10/24Modern Hebrew14/24 2Modern Hebrew6/24Filipino8/24Swahili10/24 3Western Farsi5/24Modern Hebrew6/24English9/24 4French4/24Swahili0/24Vietnamese9/24 5Filipino4/24Standard Arabic0/24Standard Arabic9/24 6Marathi4/24Western Farsi0/24Turkish8/24 7Portuguese4/24Modern Greek0/24Western Farsi4/24 8Swahili4/24Turkish0/24Marathi3/24 9Vietnamese3/24Standard Malay0/24Standard Malay2/24 10Standard Arabic2/24Vietnamese0/24French1/24 11Catalan2/24English0/24Portuguese1/24 12Turkish2/24Catalan0/24Catalan1/24 13Italian1/24Serbian-Croatian-Bosnian0/24Modern Greek1/24 14Swedish1/24Portuguese0/24Swedish1/24 15Modern Greek1/24French0/24Filipino0/24 16Russian1/24Ukrainian0/24Italian0/24 17Ukrainian1/24Russian0/24Serbian-Croatian-Bosnian0/24 18Standard Malay0/24Italian0/24Russian0/24 19Serbian-Croatian-Bosnian0/24Swedish0/24Ukrainian0/24 20Korean0/24Korean0/24Welsh0/24 21Japanese0/24Japanese0/24Korean0/24 22Welsh0/24Welsh0/24Japanese0/24 23Mandarin Chinese0/24Southern Pashto0/24Mandarin Chinese0/24 24Southern Pashto0/24Mandarin Chinese0/24Southern Pashto0/24 34 Residual debiasing detail The bias layer explainsR 2 =0.12of BTS in-sample but roughly zero under LOLO, which is why its weights describe ATLAS-24 rather than a stable law. 35