LARM: Multifaceted Constructs in AI Research

• LARM is a polysemous acronym in machine learning, covering distinct constructs such as modules, models, formal reward systems, and evaluation metrics across various domains.
• It encapsulates techniques including lightweight attentional modules for pose estimation, large auto-regressive policies in embodied control, language-aligned reward machines for reinforcement learning, and loop audio recurrent strategies in ASR.
• Disambiguation is critical in LARM applications, ensuring clear citation practices and method comparisons across human pose estimation, speech recognition, time-series anomaly detection, and sparse-view 3D reconstruction.

LARM is not a single canonical research term on arXiv; it is a polysemous acronym used for several technically distinct constructs across machine learning and adjacent domains. Current uses include a Lightweight Attentional Representation Module in human pose estimation, a Large Auto-Regressive Model for long-horizon embodied control, a language-aligned reward machine in compositional reinforcement learning, a Loop Audio Recurrent Model for automatic speech recognition, the aLignment AccuRacy Metric for time-series anomaly detection, and a Large Articulated-Object Reconstruction Model for sparse-view 3D reconstruction (Guo et al., 5 Jun 2025, Li et al., 2024, Castanyer et al., 16 Oct 2025, Kaloga et al., 3 Jun 2026, Wagner et al., 20 Oct 2025, Yuan et al., 14 Nov 2025). The term therefore functions less as a stable concept than as an acronym whose meaning is determined entirely by disciplinary context.

1. Acronymal scope and disambiguation

In recent arXiv usage, “LARM” spans module-level, model-level, formalism-level, and metric-level meanings. The following table summarizes the principal definitions that are explicitly introduced under that acronym.

Expansion	Domain	Core role
Lightweight Attentional Representation Module	Human pose estimation	Global-context modeling inside MobileViM
Large Auto-Regressive Model	Embodied intelligence	Autoregressive skill prediction in Minecraft
language-aligned reward machine	Reinforcement learning	RM state descriptions embedded in language space
Loop Audio Recurrent Model	Speech recognition	Test-time compute scaling via recurrent encoder depth
aLignment AccuRacy Metric	TSAD evaluation	Property-driven anomaly-detection metric
Large Articulated-Object Reconstruction Model	3D vision	Sparse-view articulated reconstruction

The coexistence of these meanings has two implications. First, citation-level disambiguation is essential: the acronym alone is insufficiently specific for scholarly reference. Second, some papers use “LARM” as the principal system name, whereas others use it only as an internal component or derived formal object. This distinction is material when comparing claims across fields, because a module-level LARM and a full-model LARM are not commensurate objects.

2. LARM in lightweight pose estimation

In "LGM-Pose" (Guo et al., 5 Jun 2025), LARM denotes the Lightweight Attentional Representation Module, the global-context component inside the proposed MobileViM Block. Its purpose is to provide long-range dependency modeling for lightweight top-down pose estimation without incurring the memory and latency cost of Transformer self-attention. The design targets a specific gap identified by the paper: multi-branch lightweight CNNs retain detail but remain structurally redundant and latency-heavy, while Transformer hybrids improve global reasoning but are too expensive for real-time edge deployment.

Architecturally, LARM operates on a projected feature map $\mathbf{X}_L \in \mathbb{R}^{H \times W \times d}$, partitions it into non-overlapping patches of size $h \times w$, and defines

$P = wh,\qquad N = \frac{HW}{P}.$

After patching and tensor reformatting, LARM performs two complementary MLP-based integrations. The first mixes information between patches at corresponding intra-patch positions by exposing the patch index $N$ as the MLP mixing dimension. The second mixes information within patches by permuting the tensor so that the intra-patch position dimension $P$ becomes the MLP mixing axis. The only explicit MLP equation given is

$$\mathbf{x}_G = \text{Linear}\left(\text{GELU}\left(\text{Linear}\left(\text{LayerNorm}(\mathbf{x}_U)\right)\right)\right),$$

and the module uses the Non-Parametric Transformation Operation (NPT-Op) to realize patch partitioning, flattening, permutation, and folding without learnable weights.

This design is lightweight for three stated reasons: global interaction is implemented with MLPs rather than self-attention; the tensor transformations are parameter-free; and only two MLP blocks are used. The ablation reported on MPII isolates the effect of replacing a MobileViT Block with a MobileViM Block built around LARM: parameters decrease from 2.5M to 1.1M, computation from 1.4 GFLOPs to 0.9 GFLOPs, PCKh increases from 87.9 to 88.4, CPU speed rises from 20.1 FPS to 24.4 FPS, and GPU speed from 69.4 FPS to 79.9 FPS (Guo et al., 5 Jun 2025). Within that paper, LARM is therefore best understood as a patchwise global-reasoning substitute for self-attention in lightweight pose backbones.

3. LARM in long-horizon embodied control

In "LARM: Large Auto-Regressive Model for Long-Horizon Embodied Intelligence" (Li et al., 2024), LARM denotes a multimodal autoregressive policy model for Minecraft. The paper’s central design choice is to predict the next executable action directly rather than natural-language text. This is implemented through a single skill token matched by cosine similarity against a library of 106 skills, each embedded by MineCLIP into vectors of length $L = 512$. The input combines serialized state text—task description, time, weather, position, inventory, last skill, success/failure, and game response—with 4 views captured at pitch $0^\circ$ and yaw $0^\circ, 90^\circ, 180^\circ, 270^\circ$.

The implementation described in the paper uses a backbone with about 7B parameters, including a frozen BERT-initialized tokenizer, a frozen CLIP image encoder, a trainable Q-Former, a trainable vision-language projector, 32 decoder layers initialized from LLaVA weights, trainable LoRA adapters, and a final projector producing the skill token. The image pathway uses $N=4$, $h \times w$0, $h \times w$1, and a reduced image-token count $h \times w$2. Action selection is described through cosine matching against the skill library; a faithful reconstruction of the decision rule from the text is

$h \times w$3

A notable clarification in the paper is that, despite the abstract’s wording, the actual training framework presented is not referee RL. The described method is a two-phase supervised pipeline: Minecraft Wiki pre-training on a 34G multimodal dataset yielding 63,666 sentences, followed by ARTNS/ARNTS skill-prediction training on 2,589 data pairs, augmented by target transmission, target rephrasing, and skill roll-back (Li et al., 2024). Empirically, LARM is reported as the first method in that comparison to craft enchanted diamond tools, with 2/3 success on the enchanted-tool benchmark, and it runs 6.8× faster than LLaVA-1.6-7B in the controlled inference comparison, reducing average inference time from 5.78 s to 0.85 s (Li et al., 2024). In this usage, LARM is a full embodied policy model whose significance lies in action-native output rather than free-form language generation.

4. LARM as a language-aligned reward formalism

In "ARM-FM: Automated Reward Machines via Foundation Models for Compositional Reinforcement Learning" (Castanyer et al., 16 Oct 2025), LARM denotes a language-aligned reward machine rather than a standalone policy architecture. The underlying reward machine is defined as

$h \times w$4

where $h \times w$5 is the RM state set, $h \times w$6 the initial state, $h \times w$7 the event alphabet, $h \times w$8 the deterministic transition function, $h \times w$9 the RM reward, $P = wh,\qquad N = \frac{HW}{P}.$0 the final states, and $P = wh,\qquad N = \frac{HW}{P}.$1 the labeling function. ARM-FM augments this structure by pairing each RM state $P = wh,\qquad N = \frac{HW}{P}.$2 with a natural-language instruction $P = wh,\qquad N = \frac{HW}{P}.$3 and an embedding

$P = wh,\qquad N = \frac{HW}{P}.$4

The paper’s claim is that this language alignment converts the RM from a pure reward-specification device into a compositional interface for cross-task skill reuse. The foundation model generates three artifacts from task description and environment context: the RM itself, executable Python labeling functions, and state-level natural-language descriptions. RL then operates in the product space $P = wh,\qquad N = \frac{HW}{P}.$5, with the policy conditioned on environment state and current RM-state embedding: $P = wh,\qquad N = \frac{HW}{P}.$6 The rollout uses event detection $P = wh,\qquad N = \frac{HW}{P}.$7, RM transition $P = wh,\qquad N = \frac{HW}{P}.$8, and total reward $P = wh,\qquad N = \frac{HW}{P}.$9 (Castanyer et al., 16 Oct 2025).

This LARM usage is conceptually distinct from the embodied-policy LARM above. Here the acronym names a structured objective representation whose states are semantically grounded in language, supporting multitask RL and zero-shot compositional generalization. The paper’s experiments in MiniGrid/BabyAI, Craftium, Meta-World, and XLand-MiniGrid position that alignment as the key mechanism differentiating ARM-FM from earlier automated RM-generation approaches (Castanyer et al., 16 Oct 2025).

5. LARM in audio and speech

In "Test-Time Compute Scaling for ASR with Depth-Conditioned Looped Transformers" (Kaloga et al., 3 Jun 2026), LARM denotes the Loop Audio Recurrent Model, a shared-parameter ASR architecture that turns recurrent encoder depth into a runtime-controllable test-time compute axis. The recurrent encoder can be written as

$N$0

with sparse CTC supervision applied only at checkpoint loops

$N$1

The architecture adds a supervision-clock embedding, FiLM depth conditioning using normalized depth $N$2, and delayed soft-posterior feedback: $N$3

The model is designed to make recurrent steps specialize despite shared weights. On LibriSpeech 100h, a 7.7M-parameter LARM with 4 shared blocks and $N$4 achieves 11.34 / 31.84 WER on test-clean / test-other, versus 26.78 / 52.82 for the parameter-matched 4-block standard encoder and 14.43 / 37.23 for a deeper 16-block standard encoder. On 960h, the same LARM configuration reaches 4.59 / 11.75, compared with 14.07 / 29.82 for the 4-block baseline and 4.79 / 13.26 for the 16-block baseline; with a 4-gram KenLM, both LARM and the 16-block baseline reach 3.51 on test-clean, while LARM improves test-other from 9.87 to 9.38 (Kaloga et al., 3 Jun 2026). In this context, LARM names a parameter-efficient ASR mechanism for test-time compute scaling rather than an audio reasoning system.

A related but different usage appears in "Audio-Cogito" (Li et al., 14 Apr 2026). That paper’s named contribution is Audio-Cogito, not LARM; however, its discussion suggests that “LARM” can function as a broad category label if it is understood as Large Audio Reasoning Model. The paper’s actual contribution is a data-centric recipe—Cogito-Pipe plus self-distillation—for improving explicit audio Chain-of-Thought reasoning, rather than a system formally introduced under the name LARM (Li et al., 14 Apr 2026).

6. LARM as an evaluation metric for time-series anomaly detection

In "Formally Exploring Time-Series Anomaly Detection Evaluation Metrics" (Wagner et al., 20 Oct 2025), LARM denotes the aLignment AccuRacy Metric, a scalar TSAD evaluation metric constructed from a formal property-based framework. The paper first defines nine desirable properties, including detection of anomalies, redundant-alarm penalization, minimizing false positives, minimizing false alarms, invariance under permutation of false positives, keeping user’s trust, maximizing true positives, alarm timing, and early bias. It then proves that no existing metric among the 37 surveyed satisfies all nine simple properties.

LARM is defined as a combination of three components: an anomaly-reward term over $N$5, a false-alarm-count penalty $N$6, and a false-positive-mass penalty $N$7. The theorem requires $N$8 to be monotonically increasing with $N$9, to satisfy early bias, and to have codomain in $P$0, while $P$1 must be monotonically increasing with codomain in $P$2. The default choices suggested are

$P$3

The metric is intentionally not bounded to $P$4. The paper notes that the constant predictor $P$5 receives score 0, that negative values are possible when false alarms dominate useful detections, and that the metric is designed as a ranking functional rather than a normalized accuracy. It further introduces ALARM, an advanced variant with differentiated penalties for true false alarms, early alarms, and late alarms, using default tolerance $P$6 (Wagner et al., 20 Oct 2025). In this usage, LARM is neither a network nor a module but a formally specified evaluation criterion.

7. LARM in articulated 3D reconstruction, and neighboring acronym confusions

In "LARM: A Large Articulated-Object Reconstruction Model" (Yuan et al., 14 Nov 2025), LARM denotes a feedforward framework for reconstructing articulated 3D objects from sparse posed RGB images. The input is

$P$7

namely two articulation states with $P$8 posed images per state. LARM extends LVSM by concatenating patchified RGB content, patchified Plücker ray embeddings, and articulation-state scalars into a decoder-only transformer with 12 layers and hidden size 768. The model predicts target RGB, depth, foreground mask, and movable-part mask, and then uses synthesized RGB-D and masks for explicit mesh extraction and joint estimation. The training objective is

$P$9

with $$\mathbf{x}_G = \text{Linear}\left(\text{GELU}\left(\text{Linear}\left(\text{LayerNorm}(\mathbf{x}_U)\right)\right)\right),$$0, $$\mathbf{x}_G = \text{Linear}\left(\text{GELU}\left(\text{Linear}\left(\text{LayerNorm}(\mathbf{x}_U)\right)\right)\right),$$1, $$\mathbf{x}_G = \text{Linear}\left(\text{GELU}\left(\text{Linear}\left(\text{LayerNorm}(\mathbf{x}_U)\right)\right)\right),$$2, and $$\mathbf{x}_G = \text{Linear}\left(\text{GELU}\left(\text{Linear}\left(\text{LayerNorm}(\mathbf{x}_U)\right)\right)\right),$$3.

The reported results place this LARM at the intersection of novel-view synthesis and explicit articulated reconstruction. On the articulated reconstruction benchmark, its average Chamfer Distance is 0.030 and average F-Score is 0.929, improving over Paris at 0.046 / 0.913 and substantially surpassing single-view baselines such as Singapo at 0.098 / 0.633. On joint estimation, it reports 81.2%, 81.9%, 87.0%, and 84.6% across the four success metrics listed in the paper, and the full reconstruction pipeline takes about 90 seconds per object (Yuan et al., 14 Nov 2025). In this context, LARM denotes a unified articulated reconstruction model rather than a module or evaluation measure.

The acronym is frequently confused with orthographically similar but distinct terms. LLAMA is the Large Latin American Millimeter Array, a 12 m single-dish radiotelescope with VLBI capability in Salta, Argentina (Romero, 2020). LhARA is the Laser-hybrid Accelerator for Radiobiological Applications, a proposed staged radiobiology facility rather than a machine-learning model (Aymar et al., 2020). LARAF is the Large-scale Array for Radio Astronomy on the Farside, a lunar farside radio-array concept (Chen et al., 2024). The ProeTEX emergency-response paper, despite the title “Autonomous Long Range Monitoring System,” does not define the acronym LARM; its actual system acronym is LRS for Long Range communication System (Lanati et al., 2011). These near-matches reinforce that “LARM” is not semantically stable across arXiv and should be resolved by full expansion and citation rather than acronym alone.