Language Unalignability: Why Some Concepts Resist Cross-Cultural Benchmark Evaluation

Shu-Kai Hsieh; Da-Chen Lian

cs.CLOct 6, 2026

Language Unalignability: Why Some Concepts Resist Cross-Cultural Benchmark Evaluation

Authors: Shu-Kai Hsieh, Da-Chen Lian

Organizations: National Taiwan University

Abstract

Current evaluation of multilingual Large Language Models (LLMs) rests on an implicit Translation-Isomorphism Assumption (TIA): that semantic structures across languages are congruent and mutually mappable without loss of information. We argue that this assumption is not merely violated in practice, but ill-posed in principle for a typologically identifiable class of concepts, including pragmatic markers, honorifics, and diachronically stratified terms. We formalize this failure using a usage-cloud framework, representing concepts as point sets of contextualized embeddings. We define $α$ -unalignability as the impossibility of any mapping that simultaneously preserves lexical faithfulness (centroid correspondence) and structural faithfulness (local neighborhood topology). We provide three layers of evidence. Behaviorally, we show that FLORES-200 translation failures are predicted by language family and resource class but not by script, and that LOBSTER reasoning scores vary by family. Mechanistically, we report a Representation-Intervention Gap (RIG) in a nine-model case study on Yami: the models' activations encode a regularity along which Yami groups with other low-resource and Austronesian languages, yet interventions on language-specific neurons show no demonstrated advantage over random masks: the regularity is visible but not usable by this intervention. Finally, we operationalize these findings into a multidimensional diagnostic profile: Cycle-Consistency, Pragmatic-Load Disagreement, Manifold-Curvature Mismatch, and RIG. We argue that collapsing cultural competence into a single scalar incentivizes "probabilistic flattening," and that recognizing the unalignable class is a precondition for AI that respects, rather than erases, cultural divergence. This suggests that multilingual alignment is not a single well-defined objective, but a set of mutually incompatible projections.

Figures & tables

Profile	Diagnosis
Low CC, low PLD, low MCM, low RIG	Genuinely alignable; standard translation suffices.
High CC, low PLD, low RIG	Lexical drift but pragmatic competence intact ( scaling-tractable ).
Low CC, high PLD	Pragmatic load lost. The most common failure of current benchmarks.
High MCM, high RIG	Geometry differs and the model cannot recruit the difference. The hard case.
Any high, RIG $\approx 0$	Failure is behavioral, not architectural; better data plausibly closes it.

Table 1: Diagnostic patterns of the four-metric unalignability profile.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Concepts as usage clouds, alignment as cloud mapping (visualization of the framework introduced in § 3 ). Top : a source cloud

\Phi_{A}(\textit{x\={\i}n k\v{u} le})

exhibits internal pragmatic differentiation, with four sub-regions corresponding to superior-to-subordinate, peer-to-peer, service-encounter, and ironic uses; the target cloud

\Phi_{B}(\textit{good job})

has no corresponding sub-structure. The alignment map

\tau

is judged on two conditions: (M1) centroid faithfulness,

\|\tau(\mu_{A})-\mu_{B}\|<\varepsilon

, and (M2) structural faithfulness, preservation of

\mathrm{rank}_{k}(v)

under

\tau

. Even when

\tau

satisfies (M1) by aligning centroids, (M2) is necessarily violated when the target lacks the source’s pragmatic dimensionality. Bottom : the three structural routes to a large infimum in Definition 3. Route A (pragmatic-load asymmetry): many sub-registers in the source collapse to one in the target. Route B (diachronic non-stationarity): the cloud drifts across time windows

t_{1}\to t_{2}

, so no synchronic gloss can simultaneously match both. Route C (categorial non-correspondence): the target cloud is a projection of the source onto a lower-dimensional subspace.

Figure 2: Illustrative example with synthetic clouds: the Procrustes residual $r_{\mathrm{Proc}}$ (Metric 3, Manifold-Curvature Mismatch) for a structured source cloud and a flat target cloud. Left : the raw usage clouds $\Phi_{\mathrm{cmn}}(\textit{x\={\i}n k\v{u} le})$ (blue), partitioned into distinct pragmatic regions (superior-to-subordinate, peer-to-peer, service encounter), and $\Phi_{\mathrm{eng}}(\textit{good job})$ (red). Right : the clouds after optimal Procrustes alignment (translation, rotation, and scaling); gray lines connect corresponding points. $r_{\mathrm{Proc}}$ is the sum of squared distances between corresponding points after alignment. Even the best linear map cannot stretch the flat target cloud over the sub-clusters of the source cloud, so $r_{\mathrm{Proc}}$ remains high: the internal rank structure and local neighborhood densities are incompatible, an (M2) violation.

Family	Sizes	Architecture	HuggingFace identifier
Qwen3	4B / 14B / 32B	Decoder-only	Qwen/Qwen3-{4B,14B,32B}
Gemma 3 PT	4B / 12B / 27B	Decoder-only	google/gemma-3-{4b,12b,27b}-pt
NLLB-200	3.3B	Enc–dec	facebook/nllb-200-3.3B ∗
SeaLLMs-v3	7B	Decoder-only	SeaLLMs/SeaLLMs-v3-7B
SEA-LION-v4	27B	Decoder-only	aisingapore/Gemma-SEA-LION-v4-27B

Appendix

Table 2: Models in the case study. ∗ Extended with a tao_Latn language token before fine-tuning.

ID	Condition	LAPE target	LAPE contrast	Fine-tune pair
Exp 0	Neg. Ctrl	{rus, vie, kor}	{cmn, arb, fil, eng}	Yami–Chinese
Exp 1	Yami	{tao}	{cmn, kor, rus, vie, arb}	Yami–Chinese
Exp 2a	Filipino	{fil}	{cmn, kor, rus, arb, vie, eng}	Yami–Filipino
Exp 2b	Pan-AN	{fil, ceb, ind, mri, plt}	{cmn, kor, rus, arb, vie, eng}	Yami–Filipino
Exp 3	English	{eng}	{cmn, kor, rus, fil}	Yami–English

Appendix

Table 3: Five neuron-sourcing conditions; LAPE selects the listed target-set neurons (relative to the listed contrast set) via Language Activation Probability Entropy. Each condition is fine-tuned on the listed parallel pair under three modes: Standard LoRA (no mask), LAPE (gradient-masked LoRA on the LAPE-selected neurons), and Random Mask (matched-sparsity LoRA on a random neuron subset; absent for Exp 0). The common evaluation benchmark is Yami

\leftrightarrow

Chinese for all conditions.

Model	Condition	Train Loss	Val. Loss	Steps	Time (h)
Qwen3-4B	Standard LoRA	0.628	2.048	5369	1.3
	Standard LoRA (Yami–Chinese)	0.491	1.947	5369	1.3
	Exp 0 (Neg. Ctrl) [LAPE]	0.981	2.061	8437	2.0
	Exp 1 (Yami) [LAPE]	1.076	2.040	7670	1.8
	Exp 1 (Yami) [Random]	0.977	2.071	8437	2.0
	Exp 2a (Filipino) [LAPE]	1.347	2.072	7670	2.2

Appendix

Table 4: Training summary across models and experimental conditions. Train/validation loss are final values; time is wall-clock hours. The Standard LoRA row averages the Yami–Chinese, Yami–English, and Yami–Filipino baselines; Standard LoRA (Yami–Chinese) is the matched baseline for Exp 1. LAPE/Random entries are single runs.

Comparison	Metric	$\Delta$	95% stab. int.	$d_{z}$	Agree
Exp 0 LAPE vs. Standard LoRA (cmn)	BLEU	$-$ 4.48	[ $-$ 5.86, $-$ 3.06]	$-$ 1.96	9/9
	chrF++	$-$ 5.97	[ $-$ 7.76, $-$ 4.14]	$-$ 2.04	9/9
	GLEU	$-$ 3.76	[ $-$ 4.96, $-$ 2.51]	$-$ 1.87	9/9
Exp 1 LAPE vs. Standard LoRA (cmn)	BLEU	$-$ 3.75	[ $-$ 4.65, $-$ 2.90]	$-$ 2.64	9/9
	chrF++	$-$ 4.69	[ $-$ 5.49, $-$ 3.83]	$-$ 3.48	9/9
	GLEU	$-$ 2.84	[ $-$ 3.53, $-$ 2.17]	$-$ 2.57	9/9

Appendix

Table 5: Main model-level paired-analysis summary for the 16 pre-specified comparisons on the common Yami

\leftrightarrow

Chinese benchmark.

\Delta

: model-level mean difference

\overline{A-B}

(BLEU and GLEU on

[0,100]

scale; chrF++ native). 95% stability interval: percentile interval over resampled model-level differences. Agree: number of trained models whose difference shares the sign of the overall estimate.

d_{z}

: paired-samples Cohen’s

d_{z}

(mean paired difference divided by SD of paired differences).

Comparison	BLEU	chrF++	GLEU
Exp 0 LAPE vs. Standard LoRA (cmn)	9/9	9/9	9/9
Exp 1 LAPE vs. Standard LoRA (cmn)	9/9	9/9	9/9
Exp 2a LAPE vs. Standard LoRA (fil)	9/9	9/9	9/9
Exp 2b LAPE vs. Standard LoRA (fil)	9/9	9/9	9/9
Exp 3 LAPE vs. Standard LoRA (eng)	8/9	8/9	8/9
Exp 1 LAPE vs. Random Mask	6/9	6/9	7/9

Appendix

Table 6: Sign agreement across the three metrics: number of the nine models whose model-level difference shares the sign of the overall estimate.

Figure 3: Forest plot of model-level mean paired BLEU differences (system A

-

system B) with 95% percentile stability intervals for all 16 comparisons on the Yami

\leftrightarrow

Chinese common benchmark. Labels give the number of the nine models sharing the sign of the estimate. Comparisons are grouped by family (LAPE-vs-Standard, LAPE-vs-Random, cross-condition, negative-control validation). The LAPE-vs-Random intervals all include zero: no demonstrated LAPE advantage, which is not evidence of equivalence.

Model	Top-5 nearest neighbors (cosine)	Closest non-AN	All-AN tighter?
Qwen3-4B	Plateau Malagasy (0.890), Cebuano (0.890), Filipino (0.888), Maori (0.856), Indonesian (0.821)	Arabic (0.797)	yes
Qwen3-14B	Filipino (0.867), Plateau Malagasy (0.866), Cebuano (0.864), Maori (0.840), Indonesian (0.797)	Arabic (0.778)	yes
Qwen3-32B	Filipino (0.880), Cebuano (0.876), Plateau Malagasy (0.864), Maori (0.840), Indonesian (0.808)	Arabic (0.786)	yes
Gemma-3-4B	Cebuano (0.789), Filipino (0.785), Plateau Malagasy (0.764), Maori (0.761), Indonesian (0.726)	Mandarin (0.707)	yes
Gemma-3-12B	Cebuano (0.807), Filipino (0.801), Plateau Malagasy (0.774), Maori (0.773), Indonesian (0.722)	Mandarin (0.692)	yes
Gemma-3-27B	Cebuano (0.777), Filipino (0.774), Plateau Malagasy (0.764), Maori (0.748), Indonesian (0.697)	Mandarin (0.681)	yes

Appendix

Table 7: Per-model Yami nearest neighbors. Cosine similarity computed per-layer between activation-probability profiles, then averaged across layers. The last column flags whether every Austronesian language (Cebuano, Filipino, Indonesian, Maori, Plateau Malagasy) is closer to Yami than every non-Austronesian language tested. The encoder–decoder NLLB-200 is the partial exception: Filipino, Cebuano, and Plateau Malagasy remain Yami’s nearest neighbors, but English and Korean rank above Maori and Indonesian.

Figure 4: Pairwise activation cosine similarity heatmap (Qwen3-4B; per-layer cosine, mean across layers). Yami’s highest similarities are to the Austronesian languages (Plateau Malagasy, Cebuano, Filipino, Maori, Indonesian); Mandarin, English, and the contrast set (Russian/Vietnamese/Korean/Arabic) are distal.

Figure 5: Left : hierarchical clustering (average linkage) of the sixteen non-Yami evaluation languages and Yami ( tao ) by activation distance (

1-

cosine similarity of per-layer activation profiles, averaged across layers), averaged across all nine models. The low-resource languages, including most of the Austronesian set (green labels) and the low-resource non-Austronesian controls, group with Yami, while the higher-resource languages (among them Austronesian Indonesian, ind ) form a separate cluster. Right : mean Chinese

\to

Yami BLEU for each LAPE condition, averaged across models. The clustering is not matched by a corresponding translation gain from LAPE targeting: the Filipino and Pan-AN conditions fall well below the negative control, and direct Yami targeting (Exp 1) exceeds it only slightly; the gap holds whether the clustering reflects Austronesian genealogy, corpus-resource scale, or a mixture of the two [ Lian, 2026 ] .

	Neg. Ctrl	Yami	Filipino	Pan-AN	English
Neg. Ctrl	1.000	0.000	0.007	0.009	0.003
Yami	0.000	1.000	0.152	0.128	0.013
Filipino	0.007	0.152	1.000	0.336	0.020
Pan-AN	0.009	0.128	0.336	1.000	0.018
English	0.003	0.013	0.020	0.018	1.000

Appendix

Table 8: Pairwise Jaccard similarity between experimental masks, averaged across all 9 models. Values close to 0 indicate distinct neuron populations; values close to 1 indicate near-identical masks.

Figure 6: Per-(model, mask-pair) Jaccard heatmap. The (Yami, Neg. Ctrl) cell is uniformly near zero across all nine models.

Mask	Mean Jaccard	Mean shared neurons
Korean	0.1088	138.2
Vietnamese	0.0783	102.0
Russian	0.0233	26.9
Neg. Ctrl (union)	0.0100	51.0
English	0.0014	7.3
Filipino	0.0003	3.8

Appendix

Table 9: Overlap between selected masks and the model’s LAPE-identified Mandarin mask, averaged across the nine models. This tests whether the proxy-language masks may be harming the common Yami

\leftrightarrow

Chinese benchmark by masking Mandarin-associated neurons. The proxy-language masks overlap it far less than the negative-control union does. Korean, Vietnamese, and Russian are separate single-language masks shown for reference.

Model	Top-1 default prediction	Count	Unique preds
Qwen3-4B	“small fish”	125	2652
Qwen3-14B	“hair”	137	2824
Qwen3-32B	“small boat”	270	2811
Gemma-3-4B	“very difficult”	57	2751
Gemma-3-12B	“hunting”	34	2861
Gemma-3-27B	“catch by hand”	45	2896

Appendix

Table 10: The Yami

\to

Chinese collapse defaults to a small Yami-domain Chinese vocabulary (English glosses shown), and the default token differs by model. Filipino-LAPE condition (Exp 2a), all 3,566 test items per model. NLLB differs qualitatively: it emits whitespace rather than a default token.

Model	Long ( $>$ 50 wc)	Max wc	CJK leak	Top token in long preds	( $\times$ )
Qwen3-4B	184	256	23	ya	2810
Qwen3-14B	121	191	1	a	2003
Qwen3-32B	29	253	2	mo	403
Gemma-3-4B	70	254	73	a	2080
Gemma-3-12B	55	256	23	am,	913
Gemma-3-27B	7	170	20	a	197

Appendix

Table 11: Pan-AN (Exp 2b) Chinese

\to

Yami degeneration is bimodal: some models produce long Yami particle streams (no lexical content); others leak Chinese characters into the Yami output. SEA-LION-v4 leaks the most CJK despite SEA-language pretraining; Gemma-3-27B is the most stable.

Condition	Empty preds / total	Rate
Exp 1 (Yami), Random Mask	0 / 3566	0.0%
Standard LoRA, Yami–Chinese	0 / 3566	0.0%
Exp 0 (Neg. Ctrl), LAPE	0 / 3566	0.0%
Exp 1 (Yami), LAPE	0 / 3566	0.0%
Standard LoRA, Yami–English	4 / 3566	0.1%
Base model (no fine-tuning)	9 / 3566	0.3%

Appendix

Table 12: NLLB Yami

\to

Chinese empty-prediction rate by condition (3,566 items per condition). Empty outputs are concentrated in the cross-pair conditions under sparse masking (LAPE or Random); the matched-pair and Standard-LoRA conditions stay at or below 1%.

Explore similar work

Jul 1, 2026cs.CL

MSQA: A Natively Sourced Multilingual and Multicultural SimpleQA Benchmark

Multilingual fluency often invites a stronger assumption: a model that can speak a user's language must also understand the culture encoded by that language. We call this the Illusion of Cultural Alignment. To test this assumption directly, we introduce MSQA, a benchmark of 1,064 natively sourced questions across 11 language groups, five cultural dimensions, and three difficulty tiers. Unlike translated benchmarks, MSQA targets locally grounded knowledge and reduces shortcuts from English-centric cross-lingual transfer. Evaluating 18 LLMs, we find substantial cultural degradation and a pronounced Locality Effect: cultural competence tracks pre-training exposure more closely than general reasoning ability. We further show that common inference-time remedies do not dissolve the illusion. Models remain overconfident on unfamiliar cultural questions, repeated sampling yields unstable rather than reliable correctness, and retrieval augmentation helps unevenly on long-tail facts. These findings indicate that cultural alignment cannot be inferred from multilingual ability alone and requires deeper intervention than calibration, sampling, or retrieval at inference time

Xianru Chen, Yukai Huang, Mingxiang Chen +6

ByteDance Seed · Nanjing University · M-A-P

Multilingual Benchmark Multilingual Inference-Time

Ask PDF

Apr 21, 2026cs.CL

AlignCultura: Towards Culturally Aligned Large Language Models?

Cultural alignment in Large Language Models (LLMs) is essential for producing contextually aware, respectful, and trustworthy outputs. Without it, models risk generating stereotyped, insensitive, or misleading responses that fail to reflect cultural diversity w.r.t Helpful, Harmless, and Honest (HHH) paradigm. Existing benchmarks represent early steps toward cultural alignment; yet, no benchmarks currently enables systematic evaluation of cultural alignment in line with UNESCO's principles of cultural diversity w.r.t HHH paradigm. Therefore, to address this gap, we built Align-Cultura, two-stage pipeline for cultural alignment. Stage I constructs CULTURAX, the HHH-English dataset grounded in the UNESCO cultural taxonomy, through Query Construction, which reclassifies prompts, expands underrepresented domains (or labels), and prevents data leakage with SimHash. Then, Response Generation pairs prompts with culturally grounded responses via two-stage rejection sampling. The final dataset contains 1,500 samples spanning 30 subdomains of tangible and intangible cultural forms. Stage II benchmarks CULTURAX on general-purpose models, culturally fine-tuned models, and open-weight LLMs (Qwen3-8B and DeepSeek-R1-Distill-Qwen-7B). Empirically, culturally fine-tuned models improve joint HHH by 4%-6%, reduce cultural failures by 18%, achieve 10%-12% efficiency gains, and limit leakage to 0.3%.

Gautam Siddharth Kashyap, Mark Dras, Usman Naseem

School of Computing, Macquarie University, Australia

Cultural Awareness Large Language Model Alignment Natural Language Datasets

Ask PDF

Apr 24, 2026cs.CL

Multilingual Refusal Alignment for Safer Large Language Models

As Large Language Models (LLMs) are deployed globally, ensuring their safety and alignment across multiple languages becomes paramount. However, safety behaviors often vary unpredictably between languages, posing significant challenges for consistent and ethical AI. In this work, we systematically investigate the dynamics of multilingual alignment, exploring whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities. We introduce RefusEU, a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models. Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.

Aleksandra Krasnodębska, Wojciech Kusa, Aldo Lipani

NASK National Research Institute, Warsaw, Poland · University College London, London, UK

Multilingual Safety Large Language Model Alignment Large Language Model Safety Refusals Mmlu-Pro Artificial Intelligence Systems

Ask PDF