Language Potential: Multilingual and Multimodal Insights
- Language potential is the ability of language models to abstract and generalize across multilingual and underrepresented language settings, with performance improving with scale.
- In educational applications, language potential enables dynamic assessment and tailored feedback, as seen in improvements in grading metrics after bilingual fine-tuning.
- In multimodal and structured domains, language potential serves as a high-level control mechanism that integrates diverse data modalities, enhancing tasks in robotics, molecular modeling, and graph generation.
Language potential is a term used in recent research to denote several related capacities of language and LLMs rather than a single standardized construct. In current arXiv literature, it refers to latent multilingual competence in underrepresented languages, language-mediated collaboration and assessment, natural-language control of multimodal and embodied systems, metalinguistic analysis of language itself, and the extension of language-model inductive biases to structured non-text domains (Armengol-Estapé et al., 2021, Wang et al., 2024, Zhao et al., 16 May 2025, Beguš et al., 2023). This usage suggests a common concern: what can be expressed, inferred, or operationalized through language or language-model architectures under realistic data, computational, and cognitive constraints.
1. Conceptual range of the term
Recent work uses “language potential” in several domain-specific ways.
| Context | What is being tested | Representative papers |
|---|---|---|
| Multilingual NLP | Zero-shot or low-resource competence in non-dominant languages | (Armengol-Estapé et al., 2021, Syamkumar et al., 2024) |
| Education and assessment | Language-mediated grading, explanation, and dynamic feedback | (Jaganov et al., 2 May 2025, Luo et al., 2023, Saleem et al., 2024) |
| Multimodal and embodied systems | Language as one instruction channel among images, videos, and actions | (Zhao et al., 16 May 2025, Wang et al., 8 Jun 2025) |
| Formal linguistic inquiry | Metalinguistic analysis and the boundary between possible and impossible languages | (Beguš et al., 2023, Kallini et al., 10 Dec 2025) |
| Non-text structured domains | Reuse of language-model sequence priors for signals, graphs, and scientific structure | (Cui et al., 18 Jan 2025, Yao et al., 2024, Zhu et al., 2024) |
A recurring pattern is that language potential is treated either as a property of models trained on language data or as a property of language itself as an interface. In the first sense, the question is whether large models can generalize beyond the distribution they were explicitly optimized for. In the second, the question is whether natural language can serve as a high-level scaffold for tasks that also require vision, action, domain knowledge, or formal reasoning (Zhao et al., 16 May 2025, Wang et al., 8 Jun 2025).
Another recurrent theme is asymmetry between fluency and precision. Several studies report that LLMs are often strong at generation, explanation, or high-level semantic matching, yet weaker on tasks requiring token-level consistency, exact span extraction, or formally constrained structure (Armengol-Estapé et al., 2021, Saleem et al., 2024, Yao et al., 2024). This suggests that “potential” is usually not an unconditional capability claim, but a statement about the conditions under which latent competence becomes usable.
2. Latent multilingual capability
A canonical formulation appears in the study of GPT-3 in Catalan, a language that barely appears in the model’s pre-training corpus. GPT-3’s training data is reported as approximately 93% English, with Catalan accounting for about 0.01798% of tokens, or roughly 35 million words (Armengol-Estapé et al., 2021). The paper asks whether an English-centric model can nevertheless operate in Catalan, how performance differs between language understanding and language generation, and how scale affects that performance.
In extractive question answering on XQuAD-ca, the zero-shot results rise steeply with scale: GPT-3 Ada reaches F1 5.26 and EM 0.38, Babbage 10.08 and 1.13, Curie 16.66 and 5.00, and Davinci 38.43 and 17.74. Fine-tuned supervised baselines remain substantially stronger, with XLM-RoBERTa at F1 67.10 and EM 46.42, and mBERT at F1 67.15 and EM 46.51 (Armengol-Estapé et al., 2021). The paper interprets this as substantial latent NLU capability in Catalan, but clearly below bilingual or multilingual models that were directly optimized for the task.
The same study reports stronger performance in generation than in understanding. In a human evaluation of Catalan news-style sentences, human-written controls received an average rating of 4.49 with standard deviation 0.57, while GPT-3 sentences received 3.83 with standard deviation 1.05; a third of GPT-3 sentences scored above the human average, but the difference remained statistically significant with (Armengol-Estapé et al., 2021). Qualitatively, the model produced syntactically correct Catalan and even reflected the Valencian variant when prompted with a headline about Valencia, but also exhibited gibberish fragments and Catalan-English mixing. The paper therefore characterizes the model as having strong generative potential but uneven semantic grounding.
Scale is central to this account. The Catalan QA curve is almost monotonic across Ada, Babbage, Curie, and Davinci, despite no increase in Catalan data, which the paper takes as evidence that cross-lingual transfer improves strongly with parameter count (Armengol-Estapé et al., 2021). A closely related finding appears in bilingual educational grading: pre-trained multilingual instruction models perform best on English, worse on Spanish, and worst on Spanglish, indicating a systematic bias against bilingual writing. In zero-shot grading of acceptable versus unacceptable explanations, Llama 3.1 reports AUC about 0.90 for English, 0.83 for Spanish, and 0.74 for Spanglish; after bilingual fine-tuning with Spanglish data, Spanglish AUC rises to 0.94, and mixed-language fine-tuning improves all three varieties (Syamkumar et al., 2024). Here language potential is not only latent multilingual competence, but the extent to which that competence can be aligned with authentic bilingual practices.
These results jointly imply that multilingual potential is neither absent nor uniform. It is strongest in high-capacity models, strongest again in generative settings, and highly sensitive to both the amount of non-English exposure and the linguistic form of the evaluation task. Low-resource and code-switched varieties are therefore not merely harder cases; they are diagnostic cases for whether large models actually learn cross-lingual abstractions rather than relying on monolingual dominance (Armengol-Estapé et al., 2021, Syamkumar et al., 2024).
3. Language-mediated support in education, assessment, and explanation
In educational settings, language potential is often defined as the ability of a model to provide mediation rather than merely classification. The DynaWrite system is an explicit example. It operationalizes Dynamic Assessment through four feedback levels—implicit hint, probing question, error location, and explicit correction plus explanation—and implements them in a microservices-based architecture with Vue.js, Go/Gin, Apache Kafka, PostgreSQL, Centrifugo, and model-specific consumers for OpenAI and Ollama-hosted systems (Jaganov et al., 2 May 2025). In a 200-sentence evaluation on Japanese university learner writing, GPT-4o achieved precision approximately 0.654, recall approximately 0.861, and F1 approximately 0.743, while neural-chat achieved precision approximately 0.667, recall approximately 0.820, and F1 approximately 0.735. The decisive difference was pedagogical sequencing: GPT-4o produced 59 usable Dynamic Assessment sequences out of 87 true positives, whereas neural-chat produced 7 usable sequences out of 82 (Jaganov et al., 2 May 2025).
The paper’s definition of “usable” is stringent: consistency across hint levels, gradation in explicitness, and resolution at the final step. This makes language potential inseparable from discourse control. A model that detects errors but shifts targets mid-sequence or becomes explicit too early is not functionally equivalent to one that stages feedback coherently. The same paper also reports that real-time responsiveness and stability were sufficient for classroom use, which matters because Dynamic Assessment is traditionally limited by the need for prompt mediation (Jaganov et al., 2 May 2025).
A different educational and professional use case appears in requirements engineering. Across four benchmark datasets—Pure for requirements extraction, PROMISE for requirements classification, REQuestA for QA, and Aerospace for NER—ChatGPT and Gemini showed strong prompt sensitivity and underperformed specialized predictors on most tasks except question answering. The reported state-of-the-art F1 is 0.86 for requirements extraction, compared with 0.76 for ChatGPT and 0.77 for Gemini; 0.96 for requirements classification, compared with 0.78 for both models; and 0.92 for NER, compared with 0.36 for ChatGPT and 0.25 for Gemini. By contrast, on the QA benchmark the state-of-the-art F1 is 0.90, while ChatGPT reached 0.91 and Gemini 0.88 (Saleem et al., 2024). This again isolates a pattern: semantic question answering is closer to the native strengths of generative models than sequence labeling or tightly schema-constrained extraction.
Explainable recommendation offers a third variant of language-mediated support. LLMXRec separates recommendation from explanation: a conventional recommender produces the top item, and an LLM explanation generator produces a natural-language rationale conditioned on user profile, history, and item metadata (Luo et al., 2023). On BPR-MF with MovieLens-100k, GPT-4 obtained Win Ratio 0.853 and Ranking Order 1.571, while LLMXRec obtained Win Ratio 0.872 and Ranking Order 1.488 (Luo et al., 2023). The result is not that a general LLM uniformly dominates, but that a LLM specialized through parameter-efficient instruction tuning and personalized prompts can translate model outputs into explanations that are more controllable and more highly rated than those of a larger untuned model.
Across these cases, language potential is best understood as mediation capacity: the ability to stage explanations, hints, or assessments in discourse forms that humans can act on. Raw classification accuracy matters, but only partly. What is distinctive is the ability to convert latent task competence into pedagogically or operationally useful language (Jaganov et al., 2 May 2025, Saleem et al., 2024, Luo et al., 2023).
4. Language as an interface in multimodal and embodied systems
In robotics, the phrase takes on an interface-theoretic meaning. Vision-Language-Action models traditionally condition robot policies on text instructions, but OE-VLA argues that this language-only interface is too narrow for everyday human-robot interaction (Zhao et al., 16 May 2025). The model supports four additional instruction modalities beyond standard language commands: Visual Object Specification, Optical Instruction Following, Visual Goal Reaching, and Video Demo Learning. Architecturally, it combines a SigLIP-400M vision encoder with a Qwen-1.5 backbone, projects visual tokens into the language-token space with a two-layer MLP, and discretizes actions into 256 bins represented by rarely used vocabulary tokens (Zhao et al., 16 May 2025).
The resulting multimodal policy does not sacrifice linguistic performance. On CALVIN ABC→D, OE-VLA reaches average successful sequence length 2.99, exceeding KosMos Interleave at 2.70, RoboFlamingo at 2.48, LLaVA-VLA at 2.14, and OpenVLA at 0.90 (Zhao et al., 16 May 2025). On OE-CALVIN, OE-VLA achieves average length 3.48 across the new modalities, with 3.46 on VOS, 3.58 on OIF, 3.26 on VGR, and 3.60 on VDL. On the harder open-domain setting, average length falls to 2.68, with VGR dropping to 1.25, but VOS, OIF, and VDL remain at 3.33, 3.14, and 3.00 respectively (Zhao et al., 16 May 2025). The paper’s interpretation is not that language becomes irrelevant, but that language reaches its practical potential when combined with visual, in-situ textual, and demonstration-based cues.
A related but scientifically different interface appears in two-dimensional materials research. A universal interatomic potential for 2D materials is trained on 327,062 structure-energy-force-stress mappings derived from 20,114 materials spanning 89 elements, with reported mean absolute errors of 6 meV/atom for energies, 80 meV/\AA\ for forces, and 0.067 GPa for stress tensors (Wang et al., 8 Jun 2025). An LLM-powered agent built around ChatGPT and the Model Context Protocol then exposes this simulation stack through natural language. In this setting, language potential is the ability to specify materials, simulation conditions, and analysis requests in ordinary text while the system orchestrates structure retrieval, relaxation, elasticity, phonons, or molecular dynamics behind the scenes (Wang et al., 8 Jun 2025).
These studies shift the emphasis from language as the object of modeling to language as a high-level control layer. The key point is not merely that language can describe tasks, but that it can serve as a concise scaffold over richer perceptual or scientific state spaces. The limitations are also specific: OE-VLA reports that pure text performance can decline slightly when multimodal instruction data dominates training, and the 2D materials agent inherits all coverage limits of its underlying potential (Zhao et al., 16 May 2025, Wang et al., 8 Jun 2025). Language potential here is therefore constrained by alignment between linguistic abstraction and the lower-level representations actually executing the task.
5. Extension beyond text: structured signals, graphs, and scientific models
Several papers treat LLMs less as linguistic systems than as general sequence or composition priors. In massive MIMO CSI feedback, GPT-2 is repurposed as a denoiser for compressed channel state information by treating CSI as a sequence of “words” over subcarriers and embedding those vectors into the model’s token space (Cui et al., 18 Jan 2025). The paper reports that the LLM-based model significantly outperforms both a smaller feed-forward model and TransNet, especially at high compression ratios, and that with only 500 samples per scenario the performance loss is about 0.75 dB compared with full-data training (Cui et al., 18 Jan 2025). This suggests that part of the “language” in LLMs is the ability to impose a prior over structured sequences even when the symbols encode wireless signals rather than text.
A closely related abstraction appears in computational chemistry. A “LLM-type” architecture for molecular potential energy surfaces decomposes a 51-dimensional solvated Zundel system into graph-based simplexes and learns fragment correction networks that can be transferred to a 186-dimensional protonated 21-water cluster (Zhu et al., 2024). Direct reuse of the smaller-system fragment networks yields full-system MAE about 5.32 kcal/mol on the 21-mer, but targeted transformation and refinement reduce this to about 0.36 kcal/mol, matching the 51D system’s accuracy regime (Zhu et al., 2024). The paper explicitly analogizes fragments to tokens and combinatorial multiplicities to attention-like importance weights, thereby using “language potential” to name the transferability of modular learned structure across system size.
In graph generation, the term is tested even more directly as a question about whether text-only models can generate graph-structured objects with specified rules, distributions, or properties (Yao et al., 2024). GPT-4 shows strong performance on simpler rule-based tasks: trees achieve 100% valid rate in zero-shot, cycles 91.3%, and carefully prompted wheel graphs rise from 13.0% valid in zero-shot to 88.8% in few-shot (Yao et al., 2024). At the same time, planarity illustrates the brittleness of prompt-based structure induction: valid rate is 47.3% in zero-shot and drops to 4.3% in few-shot. On property-based molecular generation, few-shot plus chain-of-thought yields a corrected estimated property satisfaction of about 48.8%, with novel rate 65.5% and unique rate 92.7% (Yao et al., 2024). The paper’s conclusion is deliberately cautious: LLMs show preliminary abilities, but not reliable exact combinatorial control.
Domain-specific agricultural NLP provides another variant. AgriBERT is trained from scratch on 46,446 food- and agriculture-related journal articles comprising 311,101,592 tokens and 2,394,343 unique words, and in semantic matching between product descriptions and nutrition descriptions it achieves MAP about 44.21 and P@1 about 22.72, while GPT-3.5, used only as a prompting-based baseline on 1,000 queries, reaches MAP 73.33 and P@1 60.12 (Rezayi et al., 2023). The same paper uses ChatGPT for semantic expansion and rephrasing of food queries. Here language potential lies in domain transfer and knowledge reuse: either by pretraining directly on agricultural corpora or by exploiting the latent food knowledge of large general-purpose models (Rezayi et al., 2023).
Taken together, these studies define a broad non-literal sense of language potential: the extent to which sequence-model inductive biases, compositional tokenization, and prompting interfaces remain useful after text has been replaced by graphs, signals, molecules, or domain-specific descriptions. The gains are often real, but they depend heavily on carefully designed pre-processing, task-specific abstractions, and constrained evaluation (Cui et al., 18 Jan 2025, Zhu et al., 2024, Yao et al., 2024, Rezayi et al., 2023).
6. Metalinguistic, semantic, and argumentative dimensions, with limits
The term also appears where language itself becomes the object of analysis. “Large Linguistic Models” argues that LLMs can generate metalinguistic analyses of linguistic data, including syntactic trees, phonological rules, and lambda-calculus-style semantic representations, and reports that OpenAI’s o1 vastly outperforms other models on tasks involving drawing syntactic trees and phonological generalization (Beguš et al., 2023). The broader research program on possible and impossible natural languages pushes this further: LLMs are proposed as tools for investigating the boundary between learnable and unlearnable human languages through a four-phase cycle of constructing possible/impossible language pairs, training LMs on them, formulating linking hypotheses to human cognition, and then refining architectures—possibly by reducing memory capacity—to better capture human locality and learnability constraints (Kallini et al., 10 Dec 2025).
A complementary perspective comes from “Natural Language Semantics With Pictures,” which reverses the standard multimodal direction and treats images as semantic annotation of expressions (Schlangen, 2019). The paper shows that the “linked to same image” relation tracks a semantic implication relation that annotators can recognize even without the image, and that an exemplar-model-based approach to implication beats a simple distributional space-based one on derived datasets while remaining explainable (Schlangen, 2019). This is language potential in a model-theoretic sense: the ability of expressions to support truth-conditional, lexical, and inferential structure when grounded in visual situations.
Computational argumentation extends the notion to discourse-scale reasoning. Across six task categories and fourteen standardized datasets, LLMs show “commendable performance” on most computational argumentation tasks (Chen et al., 2023). On the paper’s new counter speech benchmark, an end-to-end GPT-3.5-Turbo system outperforms pipeline variants with BERTScore 82.51, ROUGE-2 5.70, and ROUGE-L 13.65; human judges rate it 4.32 for fluency, 3.80 for persuasiveness, and 95% for percentage of arguments addressed, compared with 3.56, 2.80, and 78% for a claim-pipeline baseline (Chen et al., 2023). Here the potential being tested is whether a model can both mine argumentative structure and generate a coherent response at document scale.
These positive results coexist with persistent limitations. Multilingual studies emphasize lower quality, higher cost, and greater instability for low-resource languages (Armengol-Estapé et al., 2021). Educational grading studies show significant bias against bilingual writing before fine-tuning (Syamkumar et al., 2024). Requirements engineering results show that general LLMs still underperform task-specific models on extraction, classification, and NER, even when they approach or exceed them on QA (Saleem et al., 2024). Graph generation shows that few-shot prompting can improve some tasks yet catastrophically harm others, as with planarity (Yao et al., 2024). This suggests that language potential is best viewed as a capacity frontier rather than a settled achievement: a set of emergent abilities whose practical value depends on scale, data balance, prompt regime, interface design, and the degree to which output constraints can be made explicit.
In aggregate, recent work portrays language potential as a family of hypotheses about what large models can do with language, through language, or under language-model inductive biases. The strongest evidence appears where language serves as a high-level scaffold over rich latent structure—multilingual abstraction, pedagogical mediation, multimodal control, or structured sequence modeling. The main unresolved issue is not whether such potential exists, but how far it can be made reliable, equitable, and theoretically interpretable across domains (Armengol-Estapé et al., 2021, Jaganov et al., 2 May 2025, Zhao et al., 16 May 2025, Kallini et al., 10 Dec 2025).