AutoMR Framework: Adaptive Automation Methods

• AutoMR Framework is a versatile set of automated pipelines that adaptively optimize reasoning, motion recognition, and recommendation tasks using domain-specific instantiations.
• The meta reasoning variant employs dynamic DAG-based skeleton search with reinforcement learning to generate context-aware reasoning chains for LLMs.
• The motion recognition and memory retrieval instantiations integrate AutoML techniques and external memory systems to achieve state-of-the-art performance on varied datasets.

AutoMR Framework

The AutoMR name refers to several distinct frameworks in recent literature, each occupying a unique application domain. The most prominent instantiations are: (1) a meta reasoning skeleton search method guiding LLM reasoning via DAG-based adaptive skeletons (Zhang et al., 5 Oct 2025), (2) a universal time series motion recognition pipeline integrating end-to-end signal processing, learning, and AutoML (Zhang et al., 21 Feb 2025), and (3) an external memory retrieval system designed to enhance LLM-based generative recommendation by leveraging long-term user interests (Wang et al., 23 Dec 2024). Each instantiation shares the goal of automating complex reasoning, retrieval, or modeling procedures, but they are otherwise methodologically and contextually independent.

1. Meta Reasoning Skeleton Search for LLMs

AutoMR (Zhang et al., 5 Oct 2025) is a framework for automatic meta reasoning in LLMs, achieving query-adaptive skeleton selection and orchestration for complex reasoning tasks.

Core Representation: DAG-Based Skeletons

Meta reasoning skeletons are formalized as single-source, edge-heterogeneous directed acyclic graphs (DAGs) $\alpha = (V, E, \tau, C)$:

• $V$: Nodes, each corresponding to a reasoning step $(i, c_i)$,
• $E$: Edges denoting logical progression,
• $\tau: E\rightarrow S$: Edge mapping to reasoning strategy set $S$ (e.g. Next, Reflect, Explore, Decompose, Summarize, Recall, Answer),
• $C$: Node contents (textual substates),
• The root node $n_0$ corresponds to the input query $q$.

The search space $\mathcal{A}$ comprises all valid DAG skeletons with total generated token count (excluding $n_0$) limited by a budget $T$: $\mathcal{A} = \left\{ \alpha = (V, E, \tau, C) \;\middle|\; \text{%%%%12%%%% single-source DAG},\; \tau:E\to S,\; \sum_{n_i\in V\setminus\{n_0\}} |c_i| \le T \right\}.$ This unifies prior sequential, tree, parallel, and hybrid meta reasoning skeleton designs as special cases (formally shown in the paper’s Proposition 1).

Dynamic Skeleton Sampling

The skeleton is not fixed a priori, but is constructed at inference time, expanding contextually as reasoning unfolds (Algorithm 1). For each prospective node $n_i$, incoming edge strategies are sampled per predecessor using a parameterized MLP: $$s_{(j,i)} \sim p_\theta(s_{(j,i)}| c_j, s_{(>j,i)}, c_{:i-1}),$$ where $c_j$ is the predecessor's content, $s_{(>j,i)}$ are previously sampled strategies targeting $n_i$, and $c_{:i-1}$ encodes the execution context. Skeleton extension halts if all such edges are ‘zero’ (terminated). Otherwise, the LLM is prompted to generate new reasoning content for $n_i$ based on selected strategies.

Policy Optimization

AutoMR optimizes the skeleton sampling policy via policy gradient (REINFORCE): $$\theta \leftarrow \theta + \frac{\eta}{MN} \sum_{i=1}^{N}\sum_{j=1}^{M} [r(a_i, \text{LLM}(q_i, \alpha_{q_i}^j)) \nabla_\theta \log P_\theta(\alpha_{q_i}^j|q_i) ],$$ where $r$ measures downstream task correctness (e.g., math QA, multiple-choice), $P_\theta$ is the sampling log-probability, and $(q_i, a_i)$ are queries and answers from dataset $\mathbb{D}$.

2. Universal Motion Recognition Pipeline

AutoMR (Zhang et al., 21 Feb 2025) names an end-to-end, automated pipeline for multimodal time series motion recognition. It focuses on:

• Handling variability in sensor data formats and time series structures,
• Integrating preprocessing, augmentation, model instantiation (via a model factory pattern), robust training management, and automated hyperparameter optimization (SMAC-based).

The pipeline uses QuartzNet (1D time-channel separable convolutional blocks originally developed for speech processing) as the backbone. Adaptable kernel sizes, dilation rates, depthwise separable convolutions, and capacity scaling ensure broad applicability, including datasets with highly varied temporal and sensor characteristics.

The AutoML component defines a hyperparameter configuration space (e.g., learning rate $\eta$, weight decay, dropout, batch size $b$, model architectural parameters) and drives optimization via sequential model-based algorithm configuration (SMAC), with ablation studies demonstrating that automated tuning can supersede manual expert efforts.

Evaluations on ten public datasets (e.g., SHREC2021, MHEALTH, UCI-HAR, Berkeley-MHAD, multiple OPPORTUNITY subsets) evidence SOTA or near-SOTA accuracy on 8/10 benchmarks, with particular robustness to sensor and format variability.

3. Memory Retrieval for Generative Recommendation

AutoMR (Wang et al., 23 Dec 2024) denotes an external long-term memory retrieval system for LLM-based generative recommenders. The framework targets the context window limitation of LLMs, which typically only have access to recent user interactions:

• Memory stores encoded hidden representations $z_m$ of previous user-item interactions:

  $z_m = LLM^{(:L)}(x_{u,h_m}),$

  where $x_{u,h_m}$ represents user $u$’s $m$-th historic interaction and $L$ denotes a cutoff layer.

• Retriever, a trainable 2-layer MLP, computes relevance scores $s_m$ for each $z_m$:

  $$s_m = \operatorname{softmax} ( \mathrm{MLP}(z_t)^\top M_{u,t} )_m,$$

  with $z_t$ encoding the recent context.

• Labeling for Retriever Training is performed by measuring the reduction in LLM perplexity for next-item prediction when each candidate $z_m$ is included:

  $$s_m' = \operatorname{softmax} \left( [\delta_1, ..., \delta_m, ...] \right )_m,\quad \delta_m = \prod_i LLM^{(L:)} (y_i | z_t, z_m; y_{<i}) - \prod_i LLM^{(L:)} (y_i | z_t; y_{<i})$$

• The retriever is supervised to match $S'$ via KL divergence.

This system consistently outperforms baselines—including BIGRec, ReLLa, TRSR, SASRec—on Recall@1, Recall@5, and NDCG@5 for both Amazon Book and Movie datasets, demonstrating particular effectiveness at leveraging distant (long-term) user behaviors for improved recommendation performance.

4. Comparison Across AutoMR Instantiations

While denoted identically, each “AutoMR” addresses fundamentally different technical problems using domain-adapted solution architectures:

Variant	Domain	Distinctive Mechanisms
Meta Reasoning Skeletons	LLM reasoning/meta-cognition	Adaptive DAG skeleton search and RL policy optimization
Time Series Motion Recognition	Multimodal time series	Universal pipeline, QuartzNet, SMAC, modularization
Memory-Retrieval Recommendation	Generative recommender LLMs	External memory, PPL-based retriever supervision

All three contribute formalisms for automating complex meta-level or cross-input reasoning, either via structural pattern search (DAG skeletons), universal pre-/postprocessing pipelines plus AutoML (motion), or optimization-driven memory retrieval (recommendation).

5. Experimental Validation and Distinctive Empirical Findings

• Meta Reasoning Skeletons (Zhang et al., 5 Oct 2025): AutoMR achieves highest accuracy on GSM8K (81.9%, LLaMA), MATH-500 (50.2%), and competitive gains across AMC, MMLU-Pro benchmarks, outperforming rStar, Meta-Reasoner, MaAS, and CoT with minimal computational overhead.
• Motion Recognition Pipeline (Zhang et al., 21 Feb 2025): State-of-the-art accuracy on 8/10 datasets; e.g., SHREC2021 (91.48% vs 89.93% prior SOTA), OPPORTUNITY subsets with improvements >5%. Ablation confirms efficacy of automated hyperparameter search.
• Memory Retrieval Recommendation (Wang et al., 23 Dec 2024): Top performance across all metrics—Recall@1, Recall@5, NDCG@5—on both Amazon Book and Movie datasets, surpassing recent semantic and summary-based retrievers.

6. Implications, Limitations, and Prospects

The recurring AutoMR paradigm, despite different technical instantiations, illustrates an emphasis on modular, automated selection/search components for optimizing structural aspects of reasoning, perception, or retrieval. In meta reasoning, this supports more human-like, context-adjusted reasoning chains. In motion and recommendation domains, it manifests as automated data/model selection/augmentation that increases coverage, efficiency, and robustness to input variability or long-range dependencies.

Limitations are domain- and instantiation-specific. For meta reasoning skeleton search, computational cost is primarily in LLM inference but is explicitly bounded below LLM fine-tuning overhead; for universal motion pipelines, spatially complex datasets (e.g., DB4, LMDHG) reveal limitations of 1D CNNs for spatial dependencies; for memory retrieval systems, generalization outside recommendation, and retriever scalability are ongoing questions.

This suggests that the AutoMR “brand” is converging on a class of pipelines where adaptive, search-driven, or retrieval-driven structural choices improve automation, generality, and domain transfer for complex sequential reasoning or modeling tasks.