MAPLE: A Polysemous Research Framework

• MAPLE is a multifaceted acronym representing task-specific frameworks in diverse fields, such as latency estimation, prompt learning, and symbolic computing.
• It is employed in hardware-aware neural architecture search with quantitative descriptors to enable accurate, few-shot latency predictions.
• MAPLE also underpins modular approaches in multimodal prompt learning and reasoning systems, emphasizing compact, data-efficient, and explainable designs.

Searching arXiv for recent and relevant MAPLE papers to ground the article. arxiv_search.query({"search_query":"ti:MAPLE OR ti:MaPLe OR ti:Maple OR ti:Mapple", "start": 0, "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"}) Reviewing search results for the MAPLE variants relevant to the provided materials. MAPLE is an acronym used by multiple research systems across machine learning, computer architecture, scientific computing, symbolic computation, robotics, autonomous driving, distributed systems, and computer vision. In the arXiv literature, the name has been applied to methods as different as hardware-aware latency estimation for neural architecture search, multimodal prompt learning for CLIP adaptation, few-shot claim verification, table reasoning, dimensionality reduction, dexterous manipulation, hierarchical remote-sensing classification, whole-slide image classification, essay scoring, app-usage prediction, and end-to-end autonomous driving, as well as to Maple-based symbolic-computation workflows and processing elements for sparse tensor accelerators (Abbasi et al., 2021, Khattak et al., 2022, Zeng et al., 2024, Bai et al., 6 Jun 2025, Huang et al., 28 Jan 2026). The commonality is nominal rather than technical: each MAPLE instantiation defines a task-specific framework whose meaning is determined by its expansion and application domain.

1. Name, scope, and major usages

The acronym MAPLE has been expanded in several distinct ways. In hardware-aware machine learning, it denotes Microprocessor A Priori for Latency Estimation, a latency predictor for neural architecture search that incorporates quantitative hardware descriptors during training (Abbasi et al., 2021). In vision-language adaptation, MaPLe denotes Multi-modal Prompt Learning, a CLIP adaptation method that learns prompts in both the vision and language branches (Khattak et al., 2022). In explainable machine learning, MAPLE denotes Model Agnostic Supervised Local Explanations, a locally linear explanation system using random forests as both supervised neighborhoods and a feature-selection mechanism (Plumb et al., 2018).

Other uses are comparably specialized. In claim verification, MAPLE stands for Micro Analysis of Pairwise Language Evolution, a few-shot method that extracts semantic features from seq2seq training dynamics on unlabeled claim-evidence pairs (Zeng et al., 2024). In table reasoning, it denotes Multi-Agent Adaptive Planning with Long-term mEmory, a four-agent framework with Solver, Checker, Reflector, and Archiver components (Bai et al., 6 Jun 2025). In dimensionality reduction, MAPLE denotes manifold-aware projection with learned edges, a self-supervised graph-refinement extension of UMAP (Huang et al., 28 Jan 2026). In robotics, it refers to Encoding Manipulation Priors Learned from Egocentric Videos, a representation-learning framework for dexterous manipulation (Gavryushin et al., 8 Apr 2025). In autonomous driving, it denotes Latent Multi-Agent Play for End-to-End Autonomous Driving, a latent closed-loop rollout framework for VLA planners (Yasarla et al., 13 May 2026).

The acronym also appears outside machine learning proper. In sparse-accelerator design, Maple is a processing element for row-wise product tensor accelerators (Reshadi et al., 2023). In mobile systems, it denotes Mobile App Prediction Leveraging LLM Embeddings (Khaokaew et al., 2023). In remote sensing, Multi-Path Adaptive Propagation with Level-Aware Embeddings is a hierarchical multi-label classifier (Koloski et al., 31 Mar 2026). In pathology, Multi-scale Attribute-enhanced Prompt Learning is a few-shot whole-slide image classification framework (Zhou et al., 30 Sep 2025). In AES, MetA-learning framework for PrototypicaL cross-prompt Essay scoring is a prototypical meta-learning method (Albatarni et al., 19 Apr 2026). A closely related but distinct spelling, Mapple, denotes a DSL for mapping distributed heterogeneous applications (Wei et al., 23 Jul 2025).

This multiplicity makes MAPLE a strongly context-dependent research term. A plausible implication is that disambiguation by expansion or by arXiv identifier is usually necessary in technical writing.

2. Hardware-aware latency estimation

The MAPLE of Kim et al. formulates latency prediction as a hardware-aware regression problem,

$f(a, S; \theta) \mapsto \hat{y},$

where $a$ is the architecture encoding, $S$ is a quantitative hardware descriptor, and $\hat{y}$ is the predicted latency (Abbasi et al., 2021). The method was introduced to reduce the measurement burden in neural architecture search, where direct latency evaluation on every candidate architecture is expensive and hardware dependence limits transfer across devices.

Its distinguishing feature is the explicit incorporation of prior hardware characteristics at training time rather than post-hoc transfer learning or domain adaptation. The paper defines an original training set

$$\mathbb{X} = \Big\{ \big(a,S,y\big)\mid a\in \mathbb{A}, S\in \mathbb{S}, y\in \mathbb{Y}\Big\},$$

an adaptation set

$\hat{\mathbb{X} = \Big\{ \big(a,S,y\big)\mid a\in \hat{\mathbb{A}, S\in \hat{\mathbb{S}, y\in \hat{\mathbb{Y}\Big\},$

and a final training set

$\mathbb{T} = \mathbb{X} \cup \hat{\mathbb{X}.$

Because the adaptation set is much smaller, MAPLE minimizes a weighted mean absolute error objective in which target-device samples are weighted by $\frac{1}{\sqrt{|\hat{\mathbb{X}|}$ and initial-pool samples by $\frac{1}{\sqrt{|\mathbb{X}|}$ (Abbasi et al., 2021).

The hardware descriptor is built from ten CPU performance counters measured for each primitive operation in the NAS search space: CPU cycles, instructions, cache references, cache misses, L1 data cache loads, L1 data cache load misses, LLC load misses, LLC loads, LLC store misses, and LLC stores. In NAS-Bench-201, with 15 operations and 10 counters per operation, the descriptor has dimension 150 (Abbasi et al., 2021). The paper further argues that tightly coupled CPU–GPU I/O permits CPU counters to act as a proxy for GPU behavior, allowing one descriptor design to cover both CPUs and GPUs. The regression model is a dual-stream neural network in which the one-hot architecture encoding is mapped into a 32-dimensional embedding and then concatenated with the 150-dimensional hardware descriptor (Abbasi et al., 2021).

The reported empirical result is few-shot generalization to unseen hardware. On eight devices total—three CPUs and five GPUs—in leave-one-out evaluation, MAPLE with 3 adaptation samples attains mean $\pm 10\%$ accuracy of 0.87, compared with HELP’s 0.81 using 10 samples; with 10 samples, MAPLE reaches 0.93 mean accuracy, with a reported rise from about 0.85 to 0.94 in average $a$0 error-bound accuracy (Abbasi et al., 2021). The same work also shows that better latency prediction improves selection of Pareto-optimal architectures in NAS. This suggests that MAPLE is not only a regression model but a hardware-aware surrogate for downstream architecture optimization.

3. Prompt learning and representation alignment

In multimodal prompt learning, MaPLe addresses a different problem: adapting CLIP-like vision-LLMs to downstream tasks without full fine-tuning (Khattak et al., 2022). The paper’s core claim is that prompt learning in only one branch of CLIP is suboptimal because classification depends on alignment between image and text embeddings. MaPLe therefore learns prompts in both the language branch and the vision branch, coupling them through a linear projection $a$1 such that vision prompts $a$2 (Khattak et al., 2022).

For the first $a$3 layers, the language and vision branches receive coupled prompts, and the paper emphasizes “discouraging independent uni-modal solutions” by preventing separate, uncoordinated optimization of branch-specific prompts (Khattak et al., 2022). It also uses deep prompting across multiple early transformer stages rather than only at the input, with the explicit rationale that early layers progressively model stage-wise feature relationships. In the reported setup, the backbone is ViT-B/16 CLIP, the few-shot regime uses 16 shots per class, prompt depth is $a$4, prompt length is 2 for both language and vision, and training runs for 5 epochs with SGD at learning rate 0.0035 on a single NVIDIA A100 GPU (Khattak et al., 2022).

The paper evaluates MaPLe on base-to-novel generalization, cross-dataset transfer, and domain generalization. Averaged over 11 datasets, it improves over Co-CoOp from base 80.47 / novel 71.69 / HM 75.83 to base 82.28 / novel 75.14 / HM 78.55, corresponding to gains of +1.81 base accuracy, +3.45 novel accuracy, and +2.72 harmonic mean (Khattak et al., 2022). Cross-dataset average accuracy is reported as 66.30 for MaPLe, compared with 63.88 for CoOp and 65.74 for Co-CoOp (Khattak et al., 2022). The method adds only about 2.85% additional parameters over CLIP, with roughly comparable FLOPs and minimal inference overhead.

A related but distinct pathology MAPLE extends prompt learning to few-shot whole-slide image classification. There, the framework combines LLM-generated entity-level prompts, LLM-generated slide-level prompts, entity-guided cross-attention, cross-scale entity graph learning, and joint entity-level plus slide-level prediction (Zhou et al., 30 Sep 2025). The paper reports best results across three TCGA cohorts and argues that existing prompt-learning WSI methods fail to capture subtype-specific phenotypic variations of histological entities. This suggests that prompt learning under the MAPLE name has evolved from generic VLM adaptation toward domain-specific, hierarchical semantic alignment.

4. MAPLE as an interpretation and reasoning framework family

The interpretability MAPLE introduced by Plumb, Molitor, and Talwalkar is a local explanation framework combining a random-forest-based supervised neighborhood with local linear modeling and global feature selection (Plumb et al., 2018). For a query point $a$5, the weight assigned to training point $a$6 is

$a$7

where $a$8 indicates leaf co-membership in tree $a$9 (Plumb et al., 2018). Feature selection is performed using DStump scores

$S$0

and the final local predictor is a weighted least-squares model

$S$1

The same local linear model serves as both predictor and explanation, yielding what the authors call faithful self-explanations (Plumb et al., 2018).

The paper reports that MAPLE is at least as accurate as the tree-ensemble baselines on all but Music, and that its causal local explanation metric outperforms LIME on most UCI regression datasets, for example on Autompgs with explanation RMSE 0.042 versus 0.178 for LIME, and on Crimes with 0.047 versus 0.276 (Plumb et al., 2018). It also provides example-based explanations through training-point weights and diagnostics for global patterns through the local training distribution. A plausible implication is that this MAPLE occupies a hybrid position between local surrogate methods and neighborhood-based explanation systems.

The table-reasoning MAPLE likewise decomposes a difficult task into multiple specialized functions, but at agent level rather than model-component level. Its architecture consists of Solver, Checker, Reflector, and Archiver, with working memory $S$2 and long-term memory $S$3 (Bai et al., 6 Jun 2025). The Checker scores answer type, format, and evidence grounding; the Reflector produces a diagnostic summary $S$4 and improvement plan $S$5; and the Archiver stores structured memory notes for future reuse. On WikiTQ and TabFact, MAPLE achieves 74.01 and 73.39 on WikiTQ with LLaMA 3 and Qwen 2.5 respectively, and 90.66 and 86.02 on TabFact, with reported gains up to +8.26 and +7.41 over the best baseline (Bai et al., 6 Jun 2025). The ablation shows a progression from 45.58 baseline to 74.01 with all four agents on WikiTQ (Bai et al., 6 Jun 2025).

These two systems share no direct mechanics, but both use MAPLE to denote architectures that combine local structure with broader diagnostic or memory mechanisms. This suggests a recurring naming pattern in which MAPLE is associated with modular frameworks intended to improve faithfulness, verification, or transfer.

5. Learning with structured priors

Several later MAPLE systems use structured priors rather than generic end-to-end training. In claim verification, MAPLE trains a small T5-small seq2seq model in both claim-to-evidence and evidence-to-claim directions on unlabeled pairs, tracks the resulting “micro language evolution path,” transforms each triple $S$6 into semantic similarity features using all-mpnet-base-v2, and trains a logistic classifier on the few labeled examples (Zeng et al., 2024). The SemSim feature vector is

$S$7

Across FEVER, Climate FEVER, SciFact_oracle, and SciFact_retrieved, MAPLE is reported as strongest in the 1-shot to 5-shot regime, with about 0.6964 F1 at 5-shot on FEVER and about 0.4846 F1 at 5-shot on SciFact_retrieved (Zeng et al., 2024).

In dexterous manipulation, MAPLE learns manipulation priors from Ego4D by extracting contact frames, MANO-format hand pose $S$8, and projected fingertip contact points, then training a ViT-B/16 encoder with a 2-layer Transformer decoder to predict future contact points and tokenized 3D hand pose at contact (Gavryushin et al., 8 Apr 2025). The method uses either MAPLE-LN, which trains only LayerNorm parameters, or MAPLE-AP, which trains the full encoder with an additional temporal contrastive loss (Gavryushin et al., 8 Apr 2025). The paper reports that MAPLE achieves the best mean performance across seven dexterous tasks and outperforms DINO, HRP, VC-1, R3M, and VIP on average, with additional real-world gains on a Mimic hand mounted on a Franka Emika Panda arm (Gavryushin et al., 8 Apr 2025).

In remote sensing, the hierarchical multi-label MAPLE injects semantic priors from graph-aware textual descriptions into label-node embeddings,

$S$9

then combines a multi-token ViT, GraphSAGE-style propagation, adaptive visual-semantic fusion, and a level-aware CE/BCE loss (Koloski et al., 31 Mar 2026). On CORINE-aligned AID, DFC-15, and MLRSNet, it improves leaf-level AU$\hat{y}$0 from 84.21 to 87.25 on AID, 98.65 to 98.71 on DFC-15, and 96.17 to 96.71 on MLRSNet, while few-shot gains reach +41.9% on AID 16-shot and +38.1% on DFC-15 8-shot (Koloski et al., 31 Mar 2026). The paper reports only +0.7% GFLOPs and +2.6% parameters (Koloski et al., 31 Mar 2026).

These systems differ in domain and supervision, but all place structured inductive bias ahead of raw scaling. This suggests that MAPLE is often used for methods whose novelty lies in encoding domain priors—hardware counters, textual hierarchies, contact points, or semantic evolution—into compact trainable frameworks.

6. Systems, optimization, and scientific-computing usages

Beyond predictive modeling, MAPLE appears in systems papers where the emphasis is on execution, portability, or symbolic computation. In sparse accelerator design, Maple is a processing element for row-wise product sparse tensor accelerators that uses multiple MAC units, A Row Buffer (ARB), B Rows Buffer (BRB), and Partial Sum Buffer (PSB) to exploit local clusters of non-zero values under CSR format (Reshadi et al., 2023). Inserted into Matraptor and Extensor, Maple yields 50% and 60% energy benefit, 15% and 22% speedup, and 5.9x and 15.5x smaller area consumption compared with the baseline designs (Reshadi et al., 2023).

In mobile app prediction, MAPLE converts heterogeneous context into templated natural-language sentences, processes them with pretrained seq2seq models such as T5 and BART, and uses installed-app similarity to mitigate cold start (Khaokaew et al., 2023). On the China-telecom dataset, MAPLE reports A@1 = 0.5191, A@3 = 0.7385, A@5 = 0.8115, MRR@3 = 0.6169, and MRR@5 = 0.6338; on LSApp, it reports A@1 = 0.7157, A@3 = 0.8649, A@5 = 0.9150, MRR@3 = 0.7821, and MRR@5 = 0.7936 (Khaokaew et al., 2023). In cold-start evaluation, performance remains strong, and the paper reports that removing installed-app information causes a drop (Khaokaew et al., 2023).

In cross-prompt essay scoring, MAPLE uses prototypical networks over tasks defined by prompt, trait, and score level, with essay, prompt, rubric, and optional engineered features fused by an element-wise gate (Albatarni et al., 19 Apr 2026). It achieves new state of the art on ELLIPSE and LAILA, outperforming the best baseline by 8.5 QWK points and 3 QWK points on average respectively (Albatarni et al., 19 Apr 2026). In end-to-end autonomous driving, MAPLE performs multi-step latent rollout of ego and nearby reactive agents, then trains with supervised fine-tuning and GRPO-based RL using global, agent-specific, and diversity rewards (Yasarla et al., 13 May 2026). On Bench2Drive, MAPLE reports DS 85.2 and SR 67.1, while MAPLE† reports DS 88.3 and SR 70.5 (Yasarla et al., 13 May 2026).

Maple is also the name of the symbolic-computation platform in papers centered on general relativity and CAD. The GR paper describes a Maple 2017 plus GRTensorIII workflow for relativistic spherical models, including construction of the stress-energy tensor, Lorentz transformation, transformation to Bondi radiative coordinates, electromagnetic extension, conservation equations, and Einstein equations (Medina, 2022). The CAD papers report incremental CAD in Maple with verified Lazard projection and valuation (Cowen-Rivers et al., 2018), and ProjectionCAD version 3 with layered and variety sub-CAD functionality (England et al., 2015). These usages refer to the Maple CAS itself rather than an acronymic method, but they are part of the broader scholarly footprint of the term.

7. Ambiguity, citation practice, and conceptual patterning

Because MAPLE is highly polysemous in arXiv literature, citation practice requires unusually explicit disambiguation. “MAPLE” alone may refer to a latency predictor (Abbasi et al., 2021), a CLIP adaptation method (Khattak et al., 2022), a self-supervised dimensionality-reduction method (Huang et al., 28 Jan 2026), a table-QA framework (Bai et al., 6 Jun 2025), or a multi-agent autonomous-driving system (Yasarla et al., 13 May 2026). It may also denote the Maple symbolic environment in scientific-computing papers (Medina, 2022, Cowen-Rivers et al., 2018, England et al., 2015). A common misconception is therefore that MAPLE identifies a unified research program; the arXiv record instead shows a naming collision across largely unrelated subfields.

At a higher level, however, some recurrent design themes are visible. Many MAPLE systems are modular rather than monolithic: dual-stream predictors in latency estimation (Abbasi et al., 2021), coupled prompts in vision-language learning (Khattak et al., 2022), four-agent feedback loops in table reasoning (Bai et al., 6 Jun 2025), graph-refinement pipelines in dimensionality reduction (Huang et al., 28 Jan 2026), and multi-stage rollout plus RL in autonomous driving (Yasarla et al., 13 May 2026). Many also emphasize compactness and data efficiency: few-shot hardware adaptation with 3 samples (Abbasi et al., 2021), few-shot prompt tuning with modest parameter overhead (Khattak et al., 2022), few-shot claim verification via T5-small and LoRA (Zeng et al., 2024), and pathology or remote-sensing models with limited training examples (Zhou et al., 30 Sep 2025, Koloski et al., 31 Mar 2026). This suggests that the recurrence of the name is not purely accidental in spirit, even though the methods are independent.

In encyclopedic usage, MAPLE is therefore best treated as a disambiguated umbrella entry: a recurrent acronym in contemporary technical literature whose referent is determined entirely by local disciplinary context and formal expansion.