AutoEP: Zero-Shot Online Hyperparameter Control

• AutoEP is a training-free framework for online hyperparameter control that adapts dynamically during optimization.
• It transforms live search dynamics into quantitative Exploratory Landscape Analysis metrics to guide decision making.
• A multi-LLM Chain of Reasoning, with Strategist, Analyst, and Actuator roles, diagnoses state and generates precise hyperparameter adjustments.

AutoEP is a training-free, zero-shot framework for online hyperparameter control of metaheuristic algorithms that dynamically adapts algorithm settings during a run by combining online Exploratory Landscape Analysis (ELA) for real-time search-state sensing with a multi-LLM Chain of Reasoning (CoR) for state diagnosis and action generation (Xu et al., 27 Sep 2025). Its central premise is to replace learning a control policy from scratch with reasoning over live optimization evidence, so that control decisions are made during optimization in a closed loop rather than fixed offline before the run (Xu et al., 27 Sep 2025). In the formulation reported in the literature, AutoEP targets metaheuristics whose performance depends strongly on the evolving balance between exploration and exploitation, and it is evaluated on combinatorial optimization tasks using GA, PSO, and ACO families (Xu et al., 27 Sep 2025).

1. Problem formulation and motivation

Metaheuristic algorithms such as GA, PSO, and ACO depend heavily on hyperparameters that govern the balance between exploration and exploitation, and static settings are often suboptimal because the best control strategy changes as the search evolves (Xu et al., 27 Sep 2025). AutoEP is proposed to address this online control problem under the claim that manual or rule-based control is brittle and hard to generalize, RL-based controllers require huge sample budgets and often overfit, and previous LLM-based methods mostly work offline by generating algorithms or heuristics before the run rather than steering the search online (Xu et al., 27 Sep 2025).

The framework is explicitly motivated by two concerns. The first is the cost of training: learning-based hyperparameter tuners, especially deep RL, need massive numbers of algorithm runs to train, whereas AutoEP avoids training entirely (Xu et al., 27 Sep 2025). The second is transfer: because the system reasons from current search statistics rather than a learned policy tied to a specific training distribution, it is expected to transfer better across different problem instances, different benchmark sizes, and even different metaheuristic families (Xu et al., 27 Sep 2025). This suggests a shift in perspective from policy learning to online diagnosis and adjustment, with the pre-trained reasoning ability of LLMs used instead of task-specific controller optimization.

The main contributions are stated as fourfold: a zero-shot paradigm for algorithm control, grounding reasoning with real-time search dynamics, a collaborative multi-LLM reasoning pipeline, and state-of-the-art performance on multiple combinatorial optimization problems (Xu et al., 27 Sep 2025). A plausible implication is that AutoEP should be understood less as a standalone optimizer than as a control layer that can be attached to an existing metaheuristic.

2. Closed-loop architecture and operating cycle

AutoEP is described as a closed-loop control system with three stages: state-sensing, reasoning, and action-and-feedback (Xu et al., 27 Sep 2025). At each decision point, the framework computes ELA-based statistics from the current optimization population and combines them with historical records stored in an Experience Pool (Xu et al., 27 Sep 2025). These features and records are passed to the Chain of Reasoning, which decides whether the algorithm should emphasize exploration or exploitation, which hyperparameters to change, and by how much (Xu et al., 27 Sep 2025). The resulting settings are applied to the metaheuristic, the run continues, and the observed performance is stored back into the Experience Pool, producing what the paper calls an online sense–reason–act loop (Xu et al., 27 Sep 2025).

The end-to-end cycle is given explicitly as: run the metaheuristic; compute online ELA metrics; package ELA plus Experience Pool history into a prompt; Strategist provides a static map of parameter semantics; Analyst diagnoses current state; Actuator proposes new hyperparameters; apply settings to the metaheuristic; observe outcome; store state-action-result in Experience Pool; repeat (Xu et al., 27 Sep 2025). The paper characterizes this as in-context learning during optimization rather than model training (Xu et al., 27 Sep 2025).

This architecture is significant because it separates sensing, diagnosis, and actuation into modular stages. The control decision is therefore not a single undifferentiated prompt output but the product of structured intermediate representations: quantitative ELA descriptors, within-run history, and a static semantic map of hyperparameters (Xu et al., 27 Sep 2025). This suggests that AutoEP’s performance depends not only on LLM capability but also on the quality of the state abstraction and the decomposition of reasoning.

3. Online Exploratory Landscape Analysis as state sensing

The online ELA module is the grounding mechanism that transforms the black-box search state into a structured signal that an LLM can use (Xu et al., 27 Sep 2025). The reported feature set spans four categories: fitness distribution, fitness landscape structure, solution diversity, and search progress (Xu et al., 27 Sep 2025). The role of ELA is to provide objective, quantitative evidence of the current search state so that the LLM does not operate purely from abstract language priors, thereby mitigating hallucination and grounding reasoning in empirical data (Xu et al., 27 Sep 2025).

The paper specifies several ELA statistics. Skewness is defined as

$$S=\frac{\frac{1}{n}\sum_{i=1}^{n}(y_i-\bar{y})^3}{\left(\sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i-\bar{y})^2}\right)^3}$$

with the interpretation that $S \approx 0$ indicates a balanced distribution, $S>0$ indicates a long tail of low-quality solutions and intensifies exploitation, and $S<0$ indicates that the population may be converging too strongly and exploration should increase (Xu et al., 27 Sep 2025).

Kurtosis is defined as

$$K=\frac{\frac{1}{n}\sum_{i=1}^{n}(y_i-\bar{y})^4}{\left(\sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i-\bar{y})^2}\right)^4}-3$$

with $K>0$ interpreted as concentrated fitness with heavy tails and low diversity, implying a need for exploration, and $K<0$ as a flatter, more dispersed distribution for which exploitation may help (Xu et al., 27 Sep 2025).

For landscape structure, the framework uses an $R^2$ landscape fit,

$$R^2 = 1 - \frac{\sum_{i=1}^n (y_i - f(\vec{x}_i))^2}{\sum_{i=1}^n (y_i - \bar{y})^2},$$

where high $R^2 \approx 1$ is interpreted as a structured or funnel-like landscape favoring exploitation, and low $S \approx 0$0 as a rugged or multimodal landscape favoring exploration (Xu et al., 27 Sep 2025).

For diversity, AutoEP uses the dispersion ratio

$S \approx 0$1

with

$S \approx 0$2

The reported interpretation is that $S \approx 0$3 means elite solutions are tightly clustered, suggesting a single funnel and the need for exploitation, whereas $S \approx 0$4 suggests multimodality and the need for exploration (Xu et al., 27 Sep 2025).

The module also introduces a rate-of-change indicator $S \approx 0$5 over a window of $S \approx 0$6 generations. Although the LaTeX is described as partially garbled, the meaning is specified: $S \approx 0$7 indicates sufficient progress and supports more exploitation, while $S \approx 0$8 indicates stagnation and triggers more exploration (Xu et al., 27 Sep 2025). Taken together, these signals provide the LLM with a compact but semantically rich description of fitness structure, diversity, regularity, and recent progress.

4. Multi-LLM Chain of Reasoning

The second principal component is the Chain of Reasoning, a multi-LLM pipeline that decomposes control into specialized sub-tasks rather than using a single monolithic prompt (Xu et al., 27 Sep 2025). The paper argues that this is advantageous because complex control decisions need decomposition (Xu et al., 27 Sep 2025).

The first agent is the Strategist LLM, a one-time setup agent whose inputs are the problem description and the chosen metaheuristic (Xu et al., 27 Sep 2025). Its output is a static control map describing the qualitative effect of each hyperparameter. The examples given are mutation rate mapping to “boost exploration” and crossover probability mapping to “increase mixing” or “tune exploitation/exploration balance” (Xu et al., 27 Sep 2025). This map is generated once and reused throughout the run (Xu et al., 27 Sep 2025).

The second agent is the Analyst LLM, the state-diagnosis component. It takes current ELA features and Experience Pool history as input and outputs a strategic diagnosis such as “Increase Exploration” or “Increase Exploitation” (Xu et al., 27 Sep 2025). The Analyst is described as using multiple signals together and looking for consensus or conflict among indicators (Xu et al., 27 Sep 2025).

The third agent is the Actuator LLM, which converts strategy into action. It takes as input the Analyst’s directive, the Strategist’s control map, and historical examples in the Experience Pool, and outputs concrete hyperparameter values (Xu et al., 27 Sep 2025). The process is decomposed into parameter selection, which determines which hyperparameters to change, and magnitude determination, which decides how strongly to change them using in-context examples from history (Xu et al., 27 Sep 2025).

The pipeline is summarized as

$S \approx 0$9

This suggests that AutoEP’s central claim is not merely that LLMs can tune hyperparameters, but that a structured collaboration among LLM roles can serve as a substitute for an explicitly trained controller.

5. Experimental scope, benchmarks, and reported results

AutoEP is evaluated on three metaheuristics—GA, PSO, and ACO—and on four combinatorial optimization problems: TSP, CVRP, FSSP, and UAV-enabled IoT data collection or trajectory optimization (Xu et al., 27 Sep 2025). The datasets are TSPLIB for TSP, VRPLIB for CVRP, Taillard benchmarks for FSSP, and a benchmark from the cited UAV trajectory optimization formulation for the UAV task (Xu et al., 27 Sep 2025). The framework is applied as a plugin to GA, GA-2opt, PSO-2opt, and ACO, and is also tested when stacked on top of already enhanced methods such as GA-2opt + EoH and GA-2opt + ReEvo, with similar PSO and ACO variants reported in the appendix (Xu et al., 27 Sep 2025).

The baseline groups comprise manual and learning-based tuners—PT, GLEET, and BEA—LLM-enhanced metaheuristic methods—EoH and ReEvo—and neural combinatorial optimization methods—DACT and LEHD (Xu et al., 27 Sep 2025). The reported evaluation metrics are Opt.gap (%) and Time, with lower values better in both cases, and Traj.Length for the UAV task, again with lower better (Xu et al., 27 Sep 2025). The paper states that AutoEP adds only modest inference overhead, about 30 ms per decision and about 0.3 seconds per run in the setup appendix (Xu et al., 27 Sep 2025).

On TSP, AutoEP is reported to consistently outperform all tuning baselines on GA and GA-2opt across multiple TSPLIB instances, including eil51, Rd100, Kroa150, rd300, rat575, and dsj1000, and to surpass DACT and LEHD (Xu et al., 27 Sep 2025). The strongest claim is that GA-2opt + AutoEP achieves the best results across all tested TSP sizes and improves already enhanced variants such as GA-2opt+EoH and GA-2opt+ReEvo (Xu et al., 27 Sep 2025).

On CVRP, the framework improves both PSO-2opt and GA-2opt and outperforms PT, GLEET, BEA, EoH, and ReEvo on the tested instance sizes (Xu et al., 27 Sep 2025). On FSSP, it improves GA-2opt across all benchmark sizes and beats the tuning baselines, though the appendix is noted to show that specialized scheduling methods like PFSPNet_NEH remain strong on their own task (Xu et al., 27 Sep 2025). On UAV trajectory optimization, it is reported to give the strongest improvement among the ACO tuning methods and to improve ACO by 17.16% at 300 sensor nodes (Xu et al., 27 Sep 2025).

A further claim is that AutoEP often lifts classical metaheuristics to a level competitive with or better than neural optimization methods, and specifically that it achieves state-of-the-art results across the tested TSP datasets (Xu et al., 27 Sep 2025). Since the reported study is centered on combinatorial optimization, a plausible implication is that the framework is presently most substantiated for population-based discrete or hybrid search settings in which a meaningful search state can be summarized online.

6. Model choice, ablations, and limitations

A notable aspect of the reported experiments is the use of open-source LLMs. AutoEP uses Qwen3-30B in the main experiments, while EoH and ReEvo use GPT-3.5-turbo in their standard configurations; additional comparisons include GPT-o1, Claude 3.7, Gemini 2.5 Pro, and DeepSeek-R1 in the CoR ablation (Xu et al., 27 Sep 2025). The paper highlights that Qwen3-30B can match the performance of GPT-4-class reasoning systems when used inside the AutoEP framework (Xu et al., 27 Sep 2025). This suggests that the framework attributes a substantial share of performance to the control architecture rather than to a single frontier proprietary model.

The ablation study on TSP reports that removing ELA worsens performance significantly, removing CoR reduces performance to about the baseline level, removing both yields performance worse than the untuned baseline, and full AutoEP performs best by far (Xu et al., 27 Sep 2025). The paper interprets this as evidence that both sensing and reasoning are necessary (Xu et al., 27 Sep 2025). It also reports that CoR with open-source models achieves comparable performance to single large models such as GPT-o1, Claude 3.7, Gemini 2.5 Pro, and DeepSeek-R1, with an order-of-magnitude lower runtime (Xu et al., 27 Sep 2025).

The robustness analysis states that EoH and ReEvo degrade substantially with smaller models, whereas AutoEP is much more robust because its performance depends on the structured framework rather than just raw model capacity (Xu et al., 27 Sep 2025). A sensitivity study on the UAV problem finds that adjusting every iteration gives the best convergence, but updating every 3–5 iterations still provides strong gains with lower overhead (Xu et al., 27 Sep 2025).

The limitations are presented more cautiously. AutoEP relies on quality ELA features and a well-defined search state, assumes that the metaheuristic exposes meaningful population statistics, depends on prompt quality and model availability even though it is training-free, incurs added runtime overhead from repeated LLM calls, and does not claim formal convergence guarantees (Xu et al., 27 Sep 2025). A practical limitation noted in the study is that performance can depend on the underlying base algorithm and the specific tuning interface available (Xu et al., 27 Sep 2025). These constraints delimit the scope of the framework: it is an online adaptive controller whose effectiveness depends on the informativeness of its sensed state and on the degree to which the controlled optimizer permits semantically meaningful intervention.

7. Position within automated algorithm control

AutoEP is situated within a broader line of work on automated control of optimization algorithms, but its distinguishing feature is the combination of online ELA grounding and multi-LLM reasoning in a closed loop (Xu et al., 27 Sep 2025). The framework is explicitly described as reframing hyperparameter tuning from a data-hungry learning problem into a grounded reasoning problem (Xu et al., 27 Sep 2025). By coupling real-time sensing, structured diagnosis, actuation, and within-run memory, it presents an alternative to RL-based controllers and to offline LLM-generated heuristic design (Xu et al., 27 Sep 2025).

Its reported significance lies in three related claims. First, it shows that training-free online control can outperform strong tuners and LLM-based baselines on TSP, CVRP, FSSP, and UAV optimization (Xu et al., 27 Sep 2025). Second, it suggests that open-source LLMs can be competitive with GPT-4-class systems for this control task when embedded in an appropriate reasoning framework (Xu et al., 27 Sep 2025). Third, it advances a more general methodological thesis: reasoning over live optimization signals may be more practical, more transferable, and less sample-intensive than learning a controller from scratch (Xu et al., 27 Sep 2025).

In that sense, AutoEP is best understood as a framework for online adaptive hyperparameter evolution rather than a single tuning heuristic. Its defining elements are the conversion of search dynamics into quantitative ELA descriptors, the decomposition of reasoning into Strategist, Analyst, and Actuator roles, and the use of the Experience Pool as an in-context memory for repeated decision making during optimization (Xu et al., 27 Sep 2025).