LingoEDU: Adaptive Multilingual Tutoring

• LingoEDU is a framework enabling adaptive language support and bilingual tutoring, integrating explicit context compression and LLM-driven dialogue for non-native learners.
• It incorporates modular components such as a structural context compressor, relevance-guided selector, and bilingual glossary to ensure traceability and accurate instructional support.
• It employs real-time grammar feedback, unknown word detection, and personalized pedagogy to enhance learning outcomes with measurable improvements across various educational metrics.

LingoEDU is a technical framework and system family enabling adaptive language support, faithful context management, and precision-controlled instructional dialogue for non-native English-speaking learners, particularly in STEM and language education contexts. It integrates explicit context compression, bilingual scaffolding, unknown word detection, and linguistically regulated dialogue generation to optimize educational interactions using LLMs.

1. Architectural Foundations and Core Modules

LingoEDU’s architecture is structured around composable modules designed for high interoperability with LLM APIs and educational interfaces. Central components include:

• Structural Context Compressor: Implements an “EDU-based” linear-to-tree decomposition of input documents. Raw text is segmented into Elementary Discourse Units (EDUs), each anchored by precise character offsets, with structure represented by nodes $n_j = (h_j, l_j, \sigma_j)$ in a relation tree $\mathcal{T} = (V, E)$ (Zhou et al., 16 Dec 2025). This design guarantees explicit traceability and eliminates hallucinated content.
• Relevance-Guided Selector: For query-based compression, a lightweight reranker (e.g., Qwen3-Reranker) assigns scores $s_j = \phi_\theta(q, h_j \oplus t_{\rm rep})$ to tree nodes. Token-budgeted greedy selection yields a context span $\mathcal{D}'$ preserving both global document backbone and fine semantic details (Zhou et al., 16 Dec 2025).
• Language-Adaptive Tutoring Interface: Interfaces auto-detect user language or support manual toggling, with queries and responses handled in the user’s preferred language and programming keywords retained in English. Guardrails enforce hint-centric scaffolding, Socratic questioning, and stepwise debugging rather than direct disclosure of solutions (Molina et al., 2024, Xu et al., 18 Sep 2025).
• Bilingual Glossary and Terminology Support: Inline, hover, or sidebar glossary maps technical terms (e.g., “inheritance”, “loop”, “if”) to common equivalents in major languages, addressing the prevalent issue of NNES students lacking terminology knowledge in their L1 (Molina et al., 2024).
• Grammar and Feedback Engines: Modules such as LangLingual aggregate learner turns, dispatch them to the LLM with fixed error-diagnosis prompts, and return up to three improvement areas filtered by confidence $c \geq 0.3$ (Gupta et al., 27 Oct 2025).
• Unknown Word Detection (EyeLingo Integration): A fused transformer classifier combining contextual (BERT/RoBERTa), gaze-stream, and word-level knowledge to estimate $p_j = P(y_j = 1 \mid s, g)$—the per-token probability that a word is unknown to an ESL learner, enabling on-the-fly definition popups and vocabulary scaffolding (Ding et al., 14 Feb 2025).

2. Explicit Context Compression and Document Structure

LingoEDU introduces an explicit structure-then-select model for context management:

• EDU Segmentation: Documents are decomposed into $N$ EDUs $e_i = (t_i, \mathsf{pos}_i, \mathsf{id}_i)$ by a learned function $$f: \{e_1, \dots, e_N\} \rightarrow \mathcal{T} = (V, E)$$. Each node $n_j$ contains a semantic summary $h_j$, depth $l_j$, and index range $\sigma_j$ precisely grounded in the source (Zhou et al., 16 Dec 2025).
• Downstream Integration: Compressed contexts yield measurable improvements across Multi-Document QA (+6.5% to +14.94%), Summarization (+12.68% to +19.75% ROUGE-1), and Deep Search (+32.53%; e.g., BrowseComp-ZH) when input to frontier LLMs (Zhou et al., 16 Dec 2025).
• Benchmarking (StructBench): Achieved Tree Edit Distance (TED) 5.67 and Document Level Accuracy (DLA) 46.77% on 248 documents, outperforming zero-shot GPT-4.1 and parser baselines (Zhou et al., 16 Dec 2025).

This approach contrasts with latent-vector and token-pruning compression by ensuring full human-readability, explicit traceability, and API compatibility, with none of the positional bias or content hallucination inherent in prior art.

3. Multilingual and Bilingual Tutoring Systems

LingoEDU features robust support for non-native English speakers in educational settings:

• Guardrail-Driven LLM Tutoring: Prompts ensure the tutor asks clarifying questions, emphasizes conceptual reasoning, and prevents academic integrity violations by not providing solutions (Molina et al., 2024).
• Language Fidelity: Tutor responses echo the input’s natural language, supporting code-mixing and multilingual framing, e.g., Java keywords embedded in Mandarin or Arabic explanations (Molina et al., 2024).
• Terminology Mapping: Glossaries address the observed lack of L1 computing vocabulary (reported by 66% of NNES students in (Molina et al., 2024)), with hover or sidebar features for instant translation.
• Help-Seeking Preferences: NNES users rank digital and online resources (CodeHelp, StackOverflow) above TAs, and demonstrate higher usage of multilingual and online help modalities.

Design guidelines include onboarding aids (multilingual demo videos) and monitoring underutilized cohorts for proactive engagement (Molina et al., 2024).

4. Linguistically Controlled Dialogue and Feedback Systems

LingoEDU’s language complexity control system is based on explicit, multi-feature instruction tuning and DPO optimization:

• Feature-Based Modulation: Three feature classes—readability (e.g., Flesch Reading Ease $F_{RE}$, Flesch-Kincaid Grade $F_{GK}$, Gunning Fog Index $G_F$, Coleman-Liau $C_L$), syntactic (tree depth $T_D$, leaf count $L_N$, etc.), lexical (simple/intermediate word ratios $S_W$, $I_W$)—quantify and regulate output difficulty (Xu et al., 18 Sep 2025).
• Dilaprix Metric for Difficulty: Normalized across percentiles, $$\mathrm{Dilaprix} = \frac{1}{|\mathcal{X}|} \sum_{x_i \in \mathcal{X}} \tau(x_i)$$ shows Pearson $r=0.950$ with human judgments, exceeding inter-expert reliability (Xu et al., 18 Sep 2025).
• Instruction Tuning: Training with explicit instruction blocks (target feature values), mapped from a user-set $t \in [0,1]$ (difficulty slider), conditions model output on desired complexity levels.
• Direct Preference Optimization (DPO): Robust preference alignment $\mathcal{L}_\text{DPO}$ ensures high response success rate (RSR up to 0.963), range and stability, contrasting with prompt-based methods (Xu et al., 18 Sep 2025).
• Implementation: Real-time UI presents a slider for $t$; the system inverts mapping for each feature, builds the prompt, and retrieves tuned LLM output, with fallback for low RSR (Xu et al., 18 Sep 2025).

A plausible implication is that instruction-tuned, feature-driven models can provide smooth and reliable adaptation of tutor dialogue complexity closely tracking learner proficiency trajectories.

5. Real-time Grammar, Error, and Vocabulary Assistance

LingoEDU incorporates several modules for feedback, error detection, and vocabulary scaffolding:

• LangLingual’s Feedback Workflow: Aggregates $N \geq 3$ learner turns, dispatches for LLM diagnosis, and captures up to three improvement areas above $c = 0.3$ confidence. Exercises are context-aware and adapt to session proficiency, with attempts logged and immediate hints provided (Gupta et al., 27 Oct 2025).
• Proficiency Tracker: Hybrid scoring: word-bank lookup (50k CSV, levels 1–14) and LLM-based prediction, combined as $$\mathrm{Level}_{\mathrm{combined}} = 0.4 \cdot \mathrm{Level}_{\mathrm{wb}} + 0.6 \cdot \mathrm{Level}_{\mathrm{LLM}}$$ (Gupta et al., 27 Oct 2025).
• Unknown Word Detection (EyeLingo): Two-stream transformer (text+gaze), high accuracy ($97.6\%$) and $F_1$ score (71.1% on eye-tracker), with near-real-time latency ($\leq 100$ ms). UI guidelines emphasize non-modal popup glossing, gaze-based debounce, and “I know this word” feedback for online personalization (Ding et al., 14 Feb 2025).
• Multimodal Support: Expansion to voice input via Whisper, and further extensions to multimodal document types (pending research into multimodal EDUs) (Gupta et al., 27 Oct 2025, Zhou et al., 16 Dec 2025).

Qualitative survey data report positive engagement, reduced grammatical errors, and higher motivation among learners using these support modules (Gupta et al., 27 Oct 2025).

6. Pedagogical Modeling, Act Taxonomies, and Adaptivity

LingoEDU’s CITS component leverages a pedagogically informed, linguistically annotated dialogue corpus—and a two-step tutoring model (act prediction $\rightarrow$ response generation):

• Dialogue-Act Lexicon (BIPED): 34 tutor acts (general, operational, assessment, teaching, engagement), 9 student acts. Annotated with Fleiss’ $\kappa$ = 0.70 agreement (Kwon et al., 2024).
• Model Pipeline: Transformer encoder predicts the next act (cross-entropy loss $\mathcal{L}_{\mathrm{act}}$), decoder generates a context-dependent utterance ($\mathcal{L}_{\mathrm{gen}}$).
• Multi-task Fine-Tuning: Four tasks (act prediction, act-conditioned generation, context-inference, minority-act augmentation) yield higher act accuracy (0.259), diversity, BLEU (15.87), and BERTScore (0.716), approaching human reference utterance length (Kwon et al., 2024).
• Personalization and Adaptivity: Learner profiling (pre-test, goals), act-prediction rates direct explanation for weaker learners, inferential or open-ended prompts for stronger learners. Explicit student profiling (e.g., CEFR) envisaged as an extension.

These structured pedagogical strategies support robust, context-sensitive interaction patterns, scaffolding, error feedback, and sustained engagement.

7. Empirical Results, Limitations, and Future Directions

Standardized evaluations report measurable gains in dialogue accuracy, controllability, structure fidelity, and user engagement:

Module/System	Key Performance Metric	Reported Value
Context Compressor	TED / DLA / Cost / Latency	5.67 / 46.77% / $0.0007$ / 1.2s (Zhou et al., 16 Dec 2025)
LLM-Difficulty Control	Pearson r (Dilaprix)	0.950 (Xu et al., 18 Sep 2025)
EyeLingo	Accuracy / F1 (Eye Tracker)	97.6% / 71.1% (Ding et al., 14 Feb 2025)
LangLingual	Learner Engagement	100% found exercises useful (Gupta et al., 27 Oct 2025)
BIPED CITS	Act Accuracy / BLEU / BERTScore	0.259 / 15.87 / 0.716 (Kwon et al., 2024)

This suggests that explicit, feature-annotated, and structurally grounded modules are essential for high reliability and personalization in educational LLM systems.

Limitations include EDU segmentation challenges for multimodal content, dependence on language-specific glossaries, and the need for further CEFR alignment. Human-in-the-loop annotation is required for deep hierarchies, but semi-supervised or active learning could reduce costs. Future research may address dynamic compression for real-time agent memory and multimodal discourse unit extraction.

Editor’s term: LingoEDU denotes the family of explicit, pedagogically annotated, multilingual scaffolding and structure-aware context compression techniques for robust, adaptive, and interpretable LLM-driven educational dialogue systems.