Suppose you must bring a multilingual parser, built mostly from English data, to thirty low-resource languages, and you can pay for a fixed number of hand-labeled sentences in total. Recent evidence says the payoff from those sentences is large and very uneven. Adding about ten target sentences raises dependency parsing scores by 13 to 17 points on average. The gain is largest for languages structurally far from English, for example those that put the verb last or pack grammar into long words (Lauscher et al., 2020). The open question is how to divide the budget. That is a resource allocation problem: many "users" (languages) compete for a scarce resource (labels and fine-tuning compute), and each has an unknown, changing return curve. Wireless networking has spent a decade building learned allocators for exactly this shape of problem.
We propose to treat languages as users and the annotation budget as spectrum. A deep reinforcement learning agent (a policy trained by trial and error to maximize a long-run reward) decides, round by round, which languages receive new labeled sentences and which labeled pools go into the next fine-tuning run. Its state includes each language's position in typological space, meaning a vector describing its word order, morphology and related structural traits. The bet is that structural distance predicts both how much a language gains from labels and how much its labels help its neighbors. The second effect makes allocation decisions interact, and that interaction is what a sequential policy can exploit and a static rule cannot.
Core Claim
Under a fixed labeling budget and per-language minimum-quality floors, a typology-conditioned sequential allocator will reach higher mean parsing accuracy and fewer floor violations than either uniform allocation or a one-shot allocation predicted from typological distance. The advantage should come specifically from cross-language spillover. Labels for Turkish raise Uzbek or Kazakh scores, so the value of a label depends on what has already been bought. A second, falsifiable prediction follows from Lauscher et al.'s finding that semantic tasks depend on pretraining volume rather than structural distance. On XNLI, typology features should add little to the allocator. If they help as much there as on parsing, the agent is exploiting something other than structure.
Where This Sits
Two lines of cross-lingual work motivate the problem without solving it. Lauscher et al. (2020) use regressions over typological database features to show that few-shot gains are cheap and depend on distance, and they note that longer sentences make better few-shot sets. They analyze returns per language in isolation and never allocate a shared budget. Nooralahzadeh et al. (2020) meta-learn a language-neutral starting point (X-MAML) over auxiliary languages and observe that gains grow when auxiliaries share typology with the target. They pick auxiliary sets by hand, and the file lists automated, distance-based selection as unsolved. Our allocator's second action, choosing which pools enter fine-tuning, is that missing selection step, made sequential and budgeted.
For the typology representation we reuse Yu et al. (2021). They learn 50-dimensional language vectors from a word-reordering denoising task; these vectors cluster by family, predict structural features in the World Atlas of Language Structures, and improve zero-shot parsing by 3.4 UAS on average. Their stated weakness on non-Indo-European and Semitic languages constrains us. We will therefore compare their vectors against hand-coded URIEL features rather than assume the learned ones are better.
The allocation machinery comes from wireless work. Behmandpoor et al. (2026) separate two jobs in their ALCOR system: deciding which users are active in order to meet utility demands, and allocating resources among active users with an unconstrained neural network. This separation adapts to changing demands faster than meta-learning baselines. We borrow the separation directly. Per-language accuracy floors are enforced by dual variables that raise the priority of languages under their floor, while the policy network only optimizes the unconstrained objective. ALCOR guarantees demands only in expectation, so we report realized floor violations, not averages.
Choosing which language to label is discrete, while choosing how many sentences is continuous. That matches the hybrid action spaces handled by parameterized Q-networks (Jiang et al.) and by paired continuous and discrete soft actor-critic agents (Wang et al., 2024). Kalntis et al. (2025) show that adversarial bandits can beat heavier learners when the environment shifts unpredictably. This is a real constraint on our design: it supplies the strongest simple competitor, discussed below.
Research Program
We define the objective as
$$\max_{n_1,\dots,n_L} \frac{1}{L}\sum_{\ell=1}^{L} s_\ell(n_1,\dots,n_L) \quad \text{s.t.} \quad \sum_\ell n_\ell \le B,\;\; s_\ell \ge \tau_\ell,$$
where $n_\ell$ is the number of labeled sentences bought for language $\ell$, $s_\ell$ is that language's test score after fine-tuning on all purchased data, $B$ is the total budget, and $\tau_\ell$ is a minimum acceptable score. The key structural fact is that $s_\ell$ depends on every $n_{\ell'}$, not just its own. If spillover is negligible, the problem separates into $L$ independent curves, and a greedy marginal-gain rule is near-optimal. Measuring how far the problem is from separable is therefore our first deliverable.
The agent works in rounds. Its state combines four things: current development-set scores per language, typology vectors, remaining budget, and current dual variables. Its action chooses a language and a batch size, plus a binary mask over already-labeled pools for the next fine-tuning run. Its reward is the round's change in the Lagrangian objective. Each real environment step requires fine-tuning a multilingual encoder, which is expensive. Sample inefficiency of deep RL is the main bottleneck flagged in the file (Wang et al., 2024). We address it the way allocators are trained on channel simulators.
First, we fine-tune mBERT and XLM-R on a grid of joint allocations over a training pool of languages and record the resulting scores. We fit a learned surrogate for $s_\ell$ to these runs, conditioned on typology and allocation vectors. The agent trains on the surrogate. It is then evaluated with real fine-tuning on a held-out set of typologically diverse languages that the surrogate never saw. The gap between surrogate and real outcomes is our counterpart of the simulation-to-reality gap, and we report it directly.
Baselines
The naive reference point is uniform allocation, $B/L$ sentences per language, chosen by Lauscher et al.'s longer-sentence heuristic. All methods use that selection rule, so the comparison isolates how much each language gets rather than which sentences are picked. Below it sits zero-shot transfer, which shows how much of the possible gain any allocator captures.
The strongest existing method is a distance-driven greedy allocator built from Lauscher et al.'s analysis. We fit their style of regression from URIEL distance and current budget to expected gain, then spend the budget one batch at a time on the language with the highest predicted marginal gain per sentence. No paper in the file reports results for any cross-language budget allocation method. The relevant reported figures are per-language: 13 to 17 UAS gained from ten parsing sentences (Lauscher et al., 2020), and 68.98% zero-shot XNLI accuracy for X-MAML against 65.33% for multilingual BERT (Nooralahzadeh et al., 2020). We treat these as sanity anchors that our fine-tuning pipeline must reproduce at matched settings before any comparison counts.
The third baseline is a contextual bandit, a learner that picks one arm per round using side information. Here it is a Gaussian-process upper-confidence-bound learner over the same typology features. It re-estimates returns online but plans only one step ahead.
All methods receive the same budget $B$, the same encoder checkpoints and fine-tuning recipe, the same development-set evaluation cost (charged against the budget), and the same hyperparameter search budget, all tuned on training-pool languages only. Beating the greedy allocator means a higher mean held-out UAS/LAS at equal $B$ across at least three budgets, together with fewer floor violations, with the gap exceeding seed variance across five fine-tuning seeds.
Two ablations isolate the key ingredients. Setting the discount to zero makes the agent myopic and essentially reduces it to the contextual bandit; this tests whether planning for spillover matters. Removing typology vectors from the state tests whether structure, rather than score history alone, drives allocation. A further variant swaps the learned typology vectors for URIEL features.
Why not just…?
Why not just allocate from a typology lookup table? One could look up each language's distance from English and give distant languages proportionally more sentences. This rule ignores spillover and cannot respond to observed scores. In wireless settings, fixed mapping-guided allocation lost up to 13% in quality of service to adaptive allocation (Wang et al., 2024). That figure comes from a different domain, so we include the lookup rule as an extra baseline and let the parsing results decide.
Why not just use a contextual bandit? This is the real threat. If spillover between languages is weak, the problem separates, and a bandit with a typology kernel should match deep RL at a fraction of the cost. Kalntis et al. show bandits winning in exactly such shifting regimes. Our separability measurement settles it in advance. Using the grid runs, we estimate how much one language's labels move another's score. If cross-language effects are small relative to direct effects, we will report that the bandit suffices and recommend it.
Why not just meta-learn a better starting model and skip allocation? X-MAML improves zero-shot transfer, but it still requires choosing auxiliary languages, and it does not decide how to spend target labels. Lauscher et al. find that a handful of target sentences outweighs zero-shot improvements on parsing. We will run X-MAML initialization under every allocator to check whether allocation gains survive a stronger starting point.
Decisive Checks
We will use Universal Dependencies treebanks for a training pool and a disjoint held-out pool. The held-out pool covers verb-final, agglutinative and non-Latin-script languages such as Turkish, Japanese, Basque, Arabic, Hindi and Korean. We evaluate UAS and LAS with mBERT and XLM-R, and run XNLI as the semantic control.
The proposal succeeds on three conditions. The planning agent must beat the greedy allocator on held-out parsing at matched budgets. The myopic ablation must lose a meaningful part of that gain. And typology features must matter on parsing but not on XNLI.
The proposal fails in any of three ways: the problem proves nearly separable, the bandit ties the agent, or the surrogate-to-real gap erases the ranking. The first and third outcomes would still be useful findings. Either would establish, for the first time, how strongly cross-lingual labels interact, which tells practitioners whether coordinated budgeting is worth doing at all.
References
- From Zero to Hero: On the Limitations of Zero-Shot Language Transfer with Multilingual Transformers
- Zero-Shot Cross-Lingual Transfer with Meta Learning
- Language Embeddings for Typology and Cross-lingual Transfer Learning
- A Deep Learning-Based Resource Allocator for Communication Networks With Dynamic User Utility Demands
- Time Division Multiple Access–Non-Orthogonal Multiple Access-Assisted Heterogeneous Semantic and Bit Communications
- Adaptive Resource Allocation for Semantic Communication Networks
- Adaptive Resource Allocation for Virtualized Base Stations in O-RAN with Online Learning