Explicit Strategy-aware Learning (ESL)

• Explicit Strategy-aware Learning is an approach that makes task-specific strategies explicit in representations, feedback loops, or training objectives.
• It spans diverse applications including bundle construction, workplace language learning, embodied agents, and optimization modeling, demonstrating significant performance gains.
• Empirical studies report notable improvements and uncover domain-specific trade-offs such as sensitivity to graph depth, thresholding, and feedback modalities.

Searching arXiv for papers on "Explicit Strategy-aware Learning" and closely related uses of the term. Explicit Strategy-aware Learning (ESL) designates a family of approaches in which a task-relevant strategy is made explicit in the representation, feedback loop, memory, or training objective rather than being left implicit inside an end-to-end predictor. Recent arXiv usage spans bundle construction, workplace language learning, speaking-practice interfaces, embodied agents, optimization modeling, and strategic classification, but the term is not used uniformly across fields (Nguyen et al., 18 Jul 2025, Yang et al., 22 Sep 2025, Park et al., 16 Jan 2026, Wei et al., 25 Mar 2026, Zhao et al., 4 May 2026, Chen et al., 2019). The acronym is also polysemous: in other work, ESL denotes English as a Second Language or Epistemic Strategy Logic rather than Explicit Strategy-aware Learning (Liu et al., 2021, Belardinelli, 2014).

1. Terminological scope and core idea

Across papers, the common pattern is explicit treatment of the strategy-bearing structure of the problem. In RaMen, ESL denotes direct modeling of observable bundle strategy through multimodal item characteristics and item-item collaborative relations; in LingoQ, it denotes converting workers’ own LLM queries into deliberate review and retrieval practice; in AI Twin, it appears as strategy-aware feedback design; in ELITE, it denotes extraction and transfer of reusable procedural strategies; and in SAGE, it denotes explicit modeling-paradigm selection in optimization modeling (Nguyen et al., 18 Jul 2025, Yang et al., 22 Sep 2025, Park et al., 16 Jan 2026, Wei et al., 25 Mar 2026, Zhao et al., 4 May 2026).

Usage	Domain	Explicit object
RaMen (Nguyen et al., 18 Jul 2025)	Bundle construction	characteristics and direct item-item collaboration
LingoQ (Yang et al., 22 Sep 2025)	Workplace ESL learning	user queries, context, and review priorities
AI Twin (Park et al., 16 Jan 2026)	Speaking practice	feedback delivery strategy
ELITE (Wei et al., 25 Mar 2026)	Embodied agents	reusable strategies from trajectories
SAGE (Zhao et al., 4 May 2026)	Optimization modeling	modeling strategy
Other ESL uses (Liu et al., 2021, Belardinelli, 2014)	Educational NLP; formal logic	different meanings of the acronym

This usage pattern suggests that “explicit” does not refer to a single architecture. Rather, it refers to the decision to externalize a strategically important variable: bundle intent, learner review targets, procedural knowledge, modeling paradigm, or strategic response structure. A plausible implication is that ESL is best understood as a design stance rather than a single method family.

2. Explicit strategy representation in bundle construction

In bundle construction, RaMen defines Explicit Strategy-aware Learning as direct modeling of a bundle’s visible decision strategy from two complementary evidence sources: intrinsic item characteristics and extrinsic item-item collaborative relations (Nguyen et al., 18 Jul 2025). The ESL component contains a Characteristic Strategy Encoder and a Collaborative Strategy Encoder. The former projects image and text features into a shared latent space,

$$\boldsymbol{\mu}_i^v = \text{MLP}^v (\mathbf{m}_i^v), \quad \boldsymbol{\mu}_i^t = \text{MLP}^t (\mathbf{m}_i^t),$$

combines them with the item ID embedding to form a multimodal item representation,

$$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$

and refines this representation with task-specific attention to obtain characteristic vectors for items and bundles. The latter builds a homogeneous item-item graph from a co-purchase matrix $E = X^\top X$, thresholds edges with $\epsilon$, and applies attentive message passing that distinguishes target and source item roles. The two streams are fused as

$$\mathbf{g}_i = \gamma \mathbf{p}_i + (1-\gamma) \mathbf{c}_i,\quad \mathbf{g}_b = \gamma \mathbf{p}_b + (1-\gamma) \mathbf{c}_b.$$

RaMen contrasts ESL with Implicit Strategy-aware Learning (ISL), which models hidden shared intents through hyperedge dependencies and hypergraph message passing rather than directly observable evidence. The two views are coupled by the Multi-strategy Alignment & Discrimination module, which uses InfoNCE-style contrastive losses to align the same item or bundle across explicit and implicit spaces while preserving discrimination across different objects. Final bundle scoring combines both spaces,

$$\sigma_{b,i} = \mathbf{g}_b \cdot \mathbf{g}_i^\top + \boldsymbol{\phi}_b\cdot \boldsymbol{\phi}_i^\top,$$

and training uses

$$\mathcal{L} = \mathcal{L}_{NLL} + \lambda_1 \mathcal{L}_{MAD} + \lambda_2 \| \mathbf{\Theta} \|_2^2.$$

Empirically, the explicit component is central. The paper reports that RaMen improves over CLHE by up to $77.31\%$ in $R@20$ on Electronic, $66.61\%$ on Food, and $$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$0 on POG, and the ablation study shows that removing either the characteristic or collaborative encoder degrades performance substantially (Nguyen et al., 18 Jul 2025). The reported pattern is domain-dependent: on sparse POG, removing the characteristic branch causes larger drops, whereas on denser Electronic and Food, removing the collaborative branch hurts more. The paper also identifies two limitations specific to explicit strategy modeling in this setting: oversmoothing from excessive graph depth and sensitivity to the empirical edge threshold $$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$1.

3. Educational and language-learning instantiations

LingoQ gives one of the clearest educational realizations of Explicit Strategy-aware Learning because it turns routine workplace AI assistance into explicit study behavior (Yang et al., 22 Sep 2025). LingoQuery captures workers’ English-related queries, especially the three intents identified in the formative study—look up, translate, and proofread—and can also capture selected text, screenshots, and user-starred responses. A backend periodically filters non-English-related exchanges, generates two distinct questions per eligible conversation using few-shot prompts modeled after TOEFL, TOEIC, and GRE, evaluates them on the binary criteria of answerability and proficiency, and refines failed items up to three times. LingoQuiz then delivers 10 multiple-choice fill-in-the-blank questions per session, with 7 newly generated items and 3 review items selected by weighted probability based on repetition frequency, prior correctness, star markings, and recency. In a three-week deployment with 28 Korean native-speaking information workers, participants used LingoQuery for 13.2 days on average and LingoQuiz for 13.4 days, submitted 3,325 messages, and solved 7,155 quiz questions. QESE scores improved significantly overall, the mixed-effects analysis showed significant overall proficiency improvement, and the clearest test gains were for basic-level learners (CEFR A), who improved by about 4 points. Expert evaluation of the filtering pipeline reported Answerability: precision 0.91, recall 0.81, F1 0.86; Proficiency: precision 0.85, recall 0.92, F1 0.88.

AI Twin frames strategy awareness differently: the relevant strategic variable is the feedback regime used during speaking practice (Park et al., 16 Jan 2026). Rather than using explicit correction during conversation, it transcribes learner speech, rephrases it into more fluent English with gpt-4.1-mini, synthesizes the reformulation in the learner’s cloned voice using ElevenLabs, and lets the interlocutor respond to the rephrased version. The within-subject study compared three conditions—Explicit Feedback, AI Proxy, and AI Twin—with 20 adult South Korean ESL learners. The only statistically significant engagement effect was for emotional engagement,

$$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$2

with means 4.03 (1.12) for Explicit Feedback, 4.77 (0.67) for AI Proxy, and 4.83 (0.63) for AI Twin. Cognitive and behavioral engagement were not significant. Participants often found explicit correction discouraging, while rephrasing preserved conversational flow and reduced pressure. Yet the paper also reports an important tension: on “best for learning,” preferences split 9 for Explicit Feedback and 9 for AI Twin, whereas on likely actual use 10 chose AI Twin and 6 chose Explicit Feedback. This makes clear that strategy-aware design in pedagogy need not privilege overt correction; in this case, the strategy-aware choice was an implicit one.

A related educational direction appears in proficiency-aligned sentence simplification, where the system is trained to generate simplifications appropriate for a learner’s CEFR level while increasing vocabulary coverage for levels $$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$3 and $$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$4 (Li et al., 17 Feb 2025). The method uses reinforcement learning on a LLM without a parallel corpus, combining DNF-style lexical constraints, a dynamic token-level reward to prevent collapse onto a narrow vocabulary subset, a sentence-level reward model trained with pairwise ranking loss, and PPO with reference-model regularization. The paper reports that target-vocabulary frequency and diversity increase by more than 20\% relative to baselines while maintaining high simplification quality. Although the paper is primarily about CEFR-aware simplification, it explicitly presents the practical implication that explicit strategy-aware learning can be built into simplification systems so that input becomes a language-acquisition intervention rather than a purely readability-driven transformation.

4. Procedural strategy extraction and transfer in embodied agents

ELITE treats strategy as reusable procedural knowledge acquired from execution experience rather than static pretraining (Wei et al., 25 Mar 2026). Its core loop is: execute a task, reflect on the trajectory, update a strategy pool, retrieve relevant strategies for the next task, and condition planning on the retrieved knowledge. The self-reflective knowledge construction module analyzes a trajectory

$$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$5

together with instruction $$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$6 and outcome, producing success or failure reflections. A Context Consolidator then emits structured update operations

$$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$7

consisting of Add, Revise, and Remove, thereby maintaining an evolving strategy pool

$$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$8

Intent-aware retrieval is based on a coarse planning trace

$$\boldsymbol{\rho}_i = \Xi ( \mathbf{\Psi}_\rho ~(\boldsymbol{\mu}_i^v ~\| ~\boldsymbol{\mu}_i^t), ~\boldsymbol{\mu}_i^{id} ),$$9

and retrieves top-$E = X^\top X$0 strategies by cosine similarity between $E = X^\top X$1 and stored plan embeddings $E = X^\top X$2.

This explicit strategy memory is used to address specific embodied-agent failure modes identified in the paper: skipping critical steps, proposing invalid actions, and repeating mistakes. ELITE is evaluated on EB-ALFRED and EB-Habitat. In the online setting, it improves performance over base VLMs by 9\% and 5\% respectively without supervision; in the supervised setting on EB-ALFRED it achieves 70.8\% average success and outperforms GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro, the Qwen2.5-VL-72B-Instruct baseline, and prior training-based methods (Wei et al., 25 Mar 2026). The paper’s qualitative examples and ablations show that both retrieval relevance and strategy-pool refinement matter: removing intent-aware retrieval or context consolidation reduces performance. This suggests that, for embodied systems, explicit strategy awareness is not merely an explanatory layer but an operational memory and transfer mechanism.

5. Explicit modeling strategy in optimization program synthesis

SAGE makes Modeling Strategy explicit in both data construction and post-training for optimization modeling with reasoning LLMs (Zhao et al., 4 May 2026). The paper’s premise is that correct optimization modeling requires commitment to a high-level paradigm—such as flow-based, assignment-based, time-indexed, or big-$E = X^\top X$3—before variables and constraints are instantiated. To operationalize this, SAGE constructs a solver-verified multi-strategy dataset. For each problem $E = X^\top X$4, a teacher proposes candidate strategies

$E = X^\top X$5

and a reasoning teacher generates a strategy-conditioned reasoning trace and Gurobi program,

$E = X^\top X$6

Outputs are then solver-verified and semantically redundant strategies are removed with an LLM-as-Judge.

Post-training proceeds through SFT followed by Segment-Weighted GRPO. During RL, the model is required to emit a structured > block with <strategy>, <modeling>, and <check> segments, and token weights satisfy

$E = X^\top X$7

The total reward is

$E = X^\top X$8

Format compliance rewards correct segment tags and a final Python block; outcome reward depends on solver execution and correctness; efficiency reward is applied only when the solution matches ground truth and is computed as

$E = X^\top X$9

with $\epsilon$0 based on IterationCount for LPs and on a weighted combination of Gap and Nodes for MILPs.

Across eight benchmarks—NL4OPT, MAMO Easy, MAMO Complex, NLP4LP, ComplexOR, IndustryOR, OptiBench, and OptMATH—SAGE improves average pass@1 from 72.7 to 80.3 over the strongest open-source baseline in the table (Zhao et al., 4 May 2026). With multiple generations, it increases component-level diversity at pass@16 by 19–29\%, and at the largest scale it produces 14.2\% fewer constraints than the baseline. The reported structural-complexity comparison shows slightly more variables but fewer constraints and nonzeros, consistent with solver-efficient modeling. The ablation study further shows that removing RL, the structured template, segment weighting, or the efficiency reward degrades performance. In this setting, explicit strategy awareness functions as a planning layer that organizes formulation correctness, structural diversity, and solver efficiency.

6. Strategic interaction, formal semantics, and recurrent tensions

A related but distinct line of work studies strategy awareness in adversarial learning rather than representation learning. In strategic classification, agents manipulate their feature vectors within a bounded radius after observing the deployed classifier, so the learner must reason about best responses rather than passive data (Chen et al., 2019). The paper shows a sharp incompatibility between external regret and Stackelberg regret: there exist settings in which any action sequence with sublinear external regret incurs linear Stackelberg regret, and any action sequence with sublinear Stackelberg regret incurs linear external regret. Its strategy-aware algorithm, Grinder, uses geometric partitioning of the action space into upper, lower, and middle polytopes derived from observed manipulated reports, then applies an EXP3-style multiplicative-weights update over polytopes. The result is a nearly optimal Stackelberg-regret guarantee with dependence on the geometry of the induced partition. Here, strategy awareness means explicit accommodation of the agent’s response structure.

Epistemic Strategy Logic is also related in subject matter but is not a learning framework. It extends Strategy Logic with epistemic operators $\epsilon$1 so that formulas can express what agents know about their own and others’ strategies (Belardinelli, 2014). Its syntax includes temporal operators, knowledge, and strategy quantification,

$\epsilon$2

and the paper uses the de dicto/de re distinction to show that knowing a winning strategy exists is different from knowing which strategy to execute. The model-checking problem is PTIME-complete with respect to model size and NON-ELEMENTARYTIME with respect to formula size. This work is best read as a formal semantics of explicit strategic knowledge rather than as Explicit Strategy-aware Learning in the machine-learning sense.

Two recurrent misconceptions follow from this literature. First, ESL is not synonymous with explicit correction or overt intervention. AI Twin is explicitly strategy-aware precisely because it chooses implicit rephrasing during live conversation rather than direct correction (Park et al., 16 Jan 2026). Second, ESL is not a stable acronym. In “Solving ESL Sentence Completion Questions via Pre-trained Neural LLMs,” ESL denotes English as a Second Language, and the paper studies a BART-based solver that ranks completed candidate sentences via binary classification on a large K-12 exam dataset; it is educationally relevant, but it is not an explicit student-strategy model (Liu et al., 2021). Reported limitations also remain domain-specific: RaMen notes sensitivity to graph depth and fixed thresholding, LingoQ highlights privacy issues from screen capture and uncertain generalization beyond Korean–English, and AI Twin exposes a gap between emotional engagement and perceived learning value (Nguyen et al., 18 Jul 2025, Yang et al., 22 Sep 2025, Park et al., 16 Jan 2026). Taken together, these works indicate that Explicit Strategy-aware Learning is best understood as a cross-domain commitment to externalizing strategically consequential structure—observable, procedural, pedagogical, or adversarial—so that it can be represented, optimized, transferred, or reasoned about directly.