Transfer-Localization in Multilingual Models
- Transfer-Localization Plane is a two-dimensional framework that jointly measures changes in cultural adaptation (ΔLocalize) and factual transfer (ΔTransfer) relative to an unaligned baseline.
- Empirical evaluations using benchmarks like gmmlu and blend reveal a trade-off where gains in cross-lingual factual transfer often come with cultural erasure.
- Interventions such as Surgical Steering and layer-wise analysis enable controlled adjustments in network representations to optimize both transfer and localization.
The transfer-localization plane is a two-dimensional analytical construct for studying transfer and localization jointly rather than as isolated metrics. In its most explicit formulation, introduced for multilingual LLMs, the plane places change in culturally adaptive performance on the horizontal axis and change in cross-lingual factual transfer on the vertical axis, both measured relative to an unaligned baseline model (Han et al., 29 Oct 2025). Related literature uses the same phrase more loosely for phase-diagram-like views of transport, localization, and transfer efficiency in disordered quantum systems (Anderson et al., 22 Sep 2025, Somoza et al., 2017). This suggests that the term functions as a general comparative device for exposing trade-offs that single-axis evaluation suppresses.
1. Formal definition in multilingual language-model evaluation
In the multilingual LLM setting, let be the unaligned baseline model and the same model after a cross-lingual alignment intervention. Two held-out multiple-choice benchmarks are used: universal or transfer tasks, indexed by , instantiated by Global MMLU (); and culturally-adaptive or localization tasks, indexed by , instantiated by a de-contextualized version of the Blend benchmark. The plane is defined through the coordinate pair
$\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$
$\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$
By construction, $\DeltaTransfer>0$ means that an alignment method improved cross-lingual transfer on universal facts, whereas $\DeltaLocalize<0$ means that the method caused “cultural erasure,” defined as worse performance on culturally situated questions. The Transfer-Localization Plane is then the set
$\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$
Each evaluated model is represented by a point 0 on this plane (Han et al., 29 Oct 2025).
The formalism is designed to make cross-lingual alignment legible as a two-objective problem. A method that improves factual transfer but degrades culturally situated response quality moves upward and leftward; a method that improves both objectives moves upward and rightward. This definition is notable because it treats cultural localization as an explicit evaluation target rather than as an unmeasured side effect of representational alignment.
2. Axes, benchmarks, and geometric interpretation
The practical realization of the two axes relies on benchmark design and normalized scoring. The 1 benchmark contains approximately 2K multiple-choice questions in six non-English languages: es, id, ko, el, zh, and ar. The localization benchmark, 3-decon, contains 4K culturally-adaptive questions per language and is created by stripping explicit location cues such as “in Greece” and requiring the model to infer them from the language. For each question, the log-likelihood of each multiple-choice option is computed, the 5 is selected, and accuracy is defined as the fraction correct. The reported quantity for each benchmark is normalized as
6
so that the unaligned baseline lies at the common origin 7 (Han et al., 29 Oct 2025).
The geometry of the plane is integral to its interpretation. The horizontal axis places 8 from more erasure on the left to better localization on the right, while the vertical axis places 9 from worse transfer below to better transfer above. The upper-left quadrant is identified as the “undesirable quadrant,” because it corresponds to positive transfer obtained at the cost of cultural erasure. The upper-right quadrant is the “ideal quadrant,” because it corresponds to simultaneous improvement on both axes. A scatter plot can contain one point per model-language pair or an average across languages, which makes the plane a visual proxy for a Pareto frontier of alignment methods (Han et al., 29 Oct 2025).
A common misconception is that better cross-lingual alignment should register as unambiguously better multilingual behavior. The plane is constructed precisely to test that assumption. By separating universal factual transfer from culturally adaptive performance, it reveals that representational convergence and culturally appropriate divergence need not move together.
3. Empirical trade-off frontier
When four representative cross-lingual alignment methods are evaluated—Multilingual Instruction Tuning (0), Middle-Layer Alignment (1), Cross-Lingual Optimization (2), and English Steering (3)—all four produce positive 4 on 5 across all six languages, and all four produce negative 6 on 7, indicating cultural erasure (Han et al., 29 Oct 2025).
| Method | 8 on gmmlu | 9 on blend |
|---|---|---|
| mist | 0 | 1 |
| midalign | 2 | 3 |
| clo | 4 | 5 |
| en | 6 | 7 |
The dominant empirical pattern is a trade-off frontier: stronger gains in factual transfer coincide with larger drops in localization. 8 provides the largest transfer gain, at approximately 9, but also incurs a localization drop of approximately 0. 1 produces the most severe erasure, approximately 2, while still increasing transfer by approximately 3. Even English Steering, which yields the smallest transfer gain at approximately 4, still hurts localization by approximately 5. None of the post-training alignment techniques enters the upper-right quadrant (Han et al., 29 Oct 2025).
This finding has two implications. First, factual transfer alone is an incomplete summary statistic for multilingual quality. Second, the plane operationalizes “cultural erasure” as a measurable regression rather than a qualitative concern. The framework therefore shifts evaluation from single-objective optimization to explicitly multi-objective diagnosis.
4. Layer-wise structure of transfer and localization
The plane is complemented by an internal-representation analysis that examines where in the network universal and culturally specific information are most steerable. The reported analyses include PCA on hidden activations at layers 6, 7, and 8 for both 9 and $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$0 inputs, together with angular alignment analyses of an English-steering vector $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$1 and a localization vector $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$2. The key observations are that representations for universal factual questions converge early and peak around middle layers, approximately $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$3; representations for cultural questions remain language-specific through the middle layers and merge only in deeper layers, for $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$4; and $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$5 and $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$6 are nearly parallel in shallow layers, $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$7, but become almost orthogonal around $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$8 (Han et al., 29 Oct 2025).
Two concrete vector constructions are specified. For a small set $\DeltaTransfer(M)=Acc(M;\mathrm{gmmlu})-Acc(M_0;\mathrm{gmmlu}),$9 of English–nonEnglish parallel prompts $\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$0, the English steering vector at layer $\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$1 is
$\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$2
and the steered activation is
$\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$3
Analogously, from pairs of de-contextualized and original culturally specific prompts $\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$4, the localization vector is
$\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$5
with the same additive steering rule $\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$6. The paper also specifies a Mid-Layer Representation Alignment loss at layer $\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$7,
$\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$8
The principal interpretive result is that universal factual transfer and culturally situated knowledge are optimally steerable at different depths. Layer-wise steering experiments confirm that applying $\DeltaLocalize(M)=Acc(M;\mathrm{blend})-Acc(M_0;\mathrm{blend}).$9 at $\DeltaTransfer>0$0 maximizes $\DeltaTransfer>0$1 with minimal side effect on $\DeltaTransfer>0$2, whereas applying $\DeltaTransfer>0$3 at $\DeltaTransfer>0$4 maximizes $\DeltaTransfer>0$5 without destroying $\DeltaTransfer>0$6 (Han et al., 29 Oct 2025).
5. Surgical Steering and controlled motion on the plane
The layer-wise analysis motivates a two-stage inference-time intervention called “Surgical Steering.” For each layer $\DeltaTransfer>0$7,
$\DeltaTransfer>0$8
with $\DeltaTransfer>0$9 and $\DeltaLocalize<0$0. The stated purpose of this targeted addition is to push a model simultaneously upward and rightward on the Transfer-Localization Plane by improving transfer at a middle layer and localization at a deeper layer (Han et al., 29 Oct 2025).
The reported effect on the $\DeltaLocalize<0$1-aligned model is an increase of approximately $\DeltaLocalize<0$2 in $\DeltaLocalize<0$3 and approximately $\DeltaLocalize<0$4 in $\DeltaLocalize<0$5 relative to $\DeltaLocalize<0$6 alone, moving the model from the undesirable quadrant into the top-right. Even heavily aligned models such as $\DeltaLocalize<0$7 and $\DeltaLocalize<0$8 remain partially steerable: applying Surgical Steering to them yields non-zero gains on both axes, although smaller than on unaligned or lightly aligned models (Han et al., 29 Oct 2025).
Within the logic of the plane, Surgical Steering is significant because it converts the plane from a passive evaluation device into an intervention space. The coordinates are no longer merely diagnostic; they become steerable targets. This does not remove the underlying trade-off, but it shows that the trade-off is not rigid across all layers of the network.
6. Related uses in transport and localization physics
A distinct but conceptually related use of the transfer-localization idea appears in transfer-tensor analyses of disordered quantum systems. In the Anderson and Aubry–André–Harper models, transfer tensors $\DeltaLocalize<0$9 are defined recursively from disorder-averaged dynamical maps and yield a discrete-time propagation with memory,
$\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$0
The central result is that ensemble averaging over static disorder produces non-Markovian memory effects even when each disorder realization is Markovian. Memory is required to cancel fictitious cross-terms that would correspond to re-drawing disorder at every time step, namely temporally uncorrelated dynamic disorder. In this setting, a “transfer–localization plane” is constructed with disorder strength on the horizontal axis and cumulative outgoing pseudoflux $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$1, or equivalently the effective spectral radius of the cumulative transfer-tensor map $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$2, on the vertical axis. A boundary curve $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$3 sharply separates localized and extended regions (Anderson et al., 22 Sep 2025).
The outgoing pseudoflux is defined from population-sector tensor elements by
$\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$4
In the Anderson model, as static disorder $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$5 increases, $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$6 first increases and then decreases when localization becomes strong. In the Aubry–André–Harper model, $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$7 for all $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$8, but $\{(x,y)\in\mathbb{R}^2 : x=\DeltaLocalize,\; y=\DeltaTransfer\}.$9 shows a sharp transition at 00: for 01, 02; for 03, one finds after a finite cutoff 04 that 05, so no outgoing flux appears beyond 06. The paper therefore concludes that eternal memory is necessary for localization but not sufficient (Anderson et al., 22 Sep 2025).
An earlier nanoring study uses the phrase more heuristically. There, one may imagine a two-dimensional parameter plane spanned by a localization measure such as ASS–IPR or steady-state 07, and a transfer-efficiency measure such as 08 or 09. The system then occupies qualitative quadrants such as high transfer/low localization or low transfer/high localization as disorder 10 and exciton–phonon coupling 11 vary. The paper identifies an optimal band near 12 and 13, whereas strong disorder or strong coupling leads to poor transfer and stronger localization (Somoza et al., 2017).
These physical constructions are not identical to the multilingual LLM plane. The former are disorder- or coupling-dependent transport diagrams, while the latter is a benchmark-normalized intervention space. The shared structure is the simultaneous treatment of a propagation-oriented quantity and a localization-oriented quantity within one coordinate system.
7. Terminological heterogeneity and adjacent usages
The phrase “transfer-localization” also appears in applied localization research with different semantics. In supervised sound-source localization, the direct-path relative transfer function is defined as
14
estimated from noisy and reverberant microphone signals via auto- and cross-power spectral densities under a convolutive transfer function approximation, and concatenated across frequencies into a feature vector for localization. The summary describes this pipeline as forming the desired “transfer–localization” plane, but the actual construction is a high-dimensional feature-to-direction mapping rather than the two-dimensional 15 plane used in multilingual evaluation (Li et al., 2015).
A different use appears in WiFi indoor localization through deep transfer learning. There, the model ingests CSI tensors of shape 16, achieves mean localization errors of 17 m in an office with no obstacles, 18 m in an office with obstacles, and 19 m in a sports hall, and evaluates transfer by freezing layers 20 through 21 and retraining only the fully connected block. That protocol saves approximately 22 of the model parameters from retraining, reduces training time by more than half, and requires only 23 of the new-environment training data to recover baseline accuracy (Li et al., 2021). Here “transfer” denotes cross-environment reuse of learned features, while “localization” denotes coordinate regression.
Taken together, these usages show that “Transfer-Localization Plane” is not yet a domain-invariant technical term. Its most rigorous and explicit mathematical formulation in the provided literature is the multilingual LLM framework that plots 24 against 25 to quantify the balance between factual transfer and cultural erasure (Han et al., 29 Oct 2025). In neighboring literatures, the phrase or its components name related but non-identical constructs: transport-versus-localization phase diagrams in disordered systems, acoustic transfer-function features for source localization, and transfer-learning protocols for localization models. The term therefore refers less to a single universal formalism than to a recurring strategy: representing transfer and localization jointly so that hidden trade-offs become observable.