EvoSpeak: LLM-Enhanced GP Heuristics

• EvoSpeak is a framework that integrates genetic programming with large language models to evolve efficient, interpretable, and adaptive heuristics for dynamic scheduling.
• It leverages LLM-driven knowledge extraction to generate a high-quality, warm-start population that improves initialization quality and reduces performance variance.
• Its transfer mechanism and preference-aware optimization enable evolved heuristics to be effectively applied across similar job shop scheduling tasks.

EvoSpeak is a framework that combines genetic programming (GP) with LLMs to make automatically evolved heuristics more effective, faster to obtain, easier to interpret, and more transferable across related optimization tasks. It is developed in the setting of dynamic flexible job shop scheduling (DFJSS), where dispatching decisions must be made quickly under uncertainty, and it addresses a recurring tension in GP: symbolic heuristics are interpretable in principle, yet the most effective rules in dynamic, large-scale scenarios often become highly complex, slow to rediscover from scratch, and difficult for practitioners to inspect or trust (Xu et al., 3 Oct 2025).

1. Problem setting and motivation

EvoSpeak is motivated by four linked limitations of conventional GP-based heuristic evolution in DFJSS. First, high-performing GP heuristics tend to suffer from bloat: although they are tree-structured symbolic expressions, their depth and nesting can make them practically opaque. The paper gives a representative example,

$$h(s) = \frac{p_1(s) + p_2(s) \cdot \exp(\mathrm{PT} / \mathrm{TIS})}{\max\{1, p_3(s)\} + \log(1 + p_4(s))},$$

where $p_i(s)$ are symbolic subexpressions, $\mathrm{PT}$ is processing time, and $\mathrm{TIS}$ is time in system. Such heuristics may perform well while still obscuring the strategic principle they encode.

Second, standard GP typically initializes populations randomly. In simulation-based scheduling, evaluating one heuristic is already expensive, so uninformed initialization wastes computation. Third, heuristics evolved for one objective, utilization level, or environment are usually not reused for related tasks, even though those heuristics may encode transferable structural regularities. Fourth, real scheduling is preference-sensitive: objectives such as flowtime, tardiness, and weighted tardiness often compete, and different users may prefer different trade-offs. EvoSpeak is proposed as a response to these issues by making GP-generated heuristics reusable, explainable, and preference-aware rather than merely performant (Xu et al., 3 Oct 2025).

The framework is also positioned against heavier LLM-in-the-loop evolutionary systems that query an LLM repeatedly during mutation or crossover. EvoSpeak instead uses the LLM around the GP process, primarily before evolution and after evolution. This preserves a conventional, scalable GP search loop while still exploiting LLMs for knowledge extraction, initialization, transfer, and explanation.

2. Framework architecture and knowledge flow

EvoSpeak is organized as a four-stage pipeline. It begins with a reference set of high-quality heuristics,

$\mathcal{H} = \{h_1, h_2, \dots, h_C\},$

where each heuristic maps a scheduling state $s \in \mathcal{S}$ to a priority score. The LLM is then used to extract knowledge from those heuristics and synthesize an informed initial population aligned with task requirements and, in multi-objective settings, user preferences (Xu et al., 3 Oct 2025).

The paper formalizes this by introducing an LLM-driven mapping

$$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$

where $\Lambda$ encodes user preferences and $\mathcal{H}'$ is a synthesized set of heuristics. GP then optimizes according to a weighted fitness

$F(h) = \sum_{k=1}^{K} w_k f_k(h),$

and the final best heuristic $p_i(s)$0 is returned to the LLM for explanation:

$p_i(s)$1

The knowledge-extraction stage is described conceptually as learning symbolic motifs, recurrent structural patterns, operator preferences, and implicit decision principles from reference heuristics. The paper denotes this as

$p_i(s)$2

with $p_i(s)$3 representing latent heuristic knowledge. The crucial point is that $p_i(s)$4 is not implemented as a trained embedding or explicit vector representation; the process is prompt-based rather than gradient-based. Prompts include problem context, reference heuristics, terminal semantics, task-specific preferences, and explicit output constraints. The LLM used in the experiments is ChatGPT 4.0.

The warm-start population is then generated as

$p_i(s)$5

From that point onward, GP proceeds conventionally with tournament selection, crossover, mutation, and reproduction over $p_i(s)$6 generations. The paper’s algorithmic summary is correspondingly lightweight: initialize $p_i(s)$7 with the LLM, evolve with standard GP, track the best individual, and finally pass that heuristic back to the LLM for natural-language explanation.

Transfer is treated as a further use of the same mechanism. Source-task heuristics are summarized into source knowledge

$p_i(s)$8

which is then adapted to a target distribution and converted into a target initial population

$p_i(s)$9

The paper decomposes $\mathrm{PT}$0 conceptually into principle generalization, instance adaptation, and cross-domain mapping, but does not give a more explicit implementation than prompt-based adaptation.

3. Heuristic representation and GP mechanics

EvoSpeak uses the standard GP representation for DFJSS dispatching heuristics: tree-structured symbolic expressions of the form

$\mathrm{PT}$1

where the output score determines job or operation priority. The terminal set includes machine, operation, job, and transportation features: $\mathrm{PT}$2 (number of operations in the queue), $\mathrm{PT}$3 (work in the queue), $\mathrm{PT}$4 (machine waiting time), $\mathrm{PT}$5 (processing time), $\mathrm{PT}$6 (median processing time of the next operation), $\mathrm{PT}$7 (operation waiting time), $\mathrm{PT}$8 (work remaining), $\mathrm{PT}$9 (number of operations remaining), $\mathrm{TIS}$0 (relative due date), $\mathrm{TIS}$1 (job slack time), $\mathrm{TIS}$2 (job weight), $\mathrm{TIS}$3 (time in system), and $\mathrm{TIS}$4 (transportation time) (Xu et al., 3 Oct 2025).

The function set is

$\mathrm{TIS}$5

Division is protected: if the denominator is zero, the operator returns 1. This ensures numerical closure and stability during evolution.

The GP configuration is fully specified. EvoSpeak uses population size $\mathrm{TIS}$6, generations $\mathrm{TIS}$7, initial depth $\mathrm{TIS}$8, maximum tree depth $\mathrm{TIS}$9, terminal/non-terminal selection rate $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$0, crossover/mutation/reproduction $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$1, tournament size $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$2, and ramped half-and-half initialization. EvoSpeak changes only initialization, transfer, and explanation relative to the baseline; the underlying GP search remains standard.

This design choice is central to the framework’s identity. EvoSpeak does not use the LLM as a fitness evaluator, online mutation operator, or online crossover operator. It also does not fine-tune the LLM on a curated heuristic corpus. “Learning from heuristics” therefore means LLM-mediated extraction and reuse of symbolic knowledge from example heuristics, not training a new model.

4. DFJSS formulation and optimization objectives

The target problem is dynamic flexible job shop scheduling. Jobs are

$\mathcal{H} = \{h_1, h_2, \dots, h_C\},$3

with release time $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$4, weight $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$5, and due date $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$6. Each job has operations

$\mathcal{H} = \{h_1, h_2, \dots, h_C\},$7

and each operation $\mathcal{H} = \{h_1, h_2, \dots, h_C\},$8 can be processed on an eligible machine set

$\mathcal{H} = \{h_1, h_2, \dots, h_C\},$9

with machine-dependent processing time $s \in \mathcal{S}$0. Transportation delays may occur between operations (Xu et al., 3 Oct 2025).

The formulation includes operation precedence,

$s \in \mathcal{S}$1

non-preemption,

$s \in \mathcal{S}$2

machine-capacity constraints,

$s \in \mathcal{S}$3

and machine assignment

$s \in \mathcal{S}$4

The paper studies both single-objective and multi-objective settings. Tardiness is defined as

$s \in \mathcal{S}$5

with

$s \in \mathcal{S}$6

Flowtime is

$s \in \mathcal{S}$7

with

$s \in \mathcal{S}$8

The weighted objectives are

$s \in \mathcal{S}$9

Preference-aware optimization is implemented through weighted sums:

$$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$0

subject to

$$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$1

The experiments use preference vectors such as $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$2, $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$3, and $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$4.

The simulation environment uses 5,000 jobs per instance, including a 1,000-job warm-up period, 10 heterogeneous machines, machine processing rates randomly in $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$5, transportation speed fixed at 5, machine-to-entry/exit distances sampled uniformly from 35 to 500, Poisson job arrivals, 2 to 10 operations per job, workloads sampled from $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$6, job weights distributed as 20% weight 1, 60% weight 2, and 20% weight 4, and due dates generated as arrival time plus $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$7 times total processing time.

The study covers six single-objective scenarios and four multi-objective scenarios under utilization levels $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$8 and $$\mathcal{M}_{\mathrm{LLM}} : (\mathcal{H}, \Lambda) \rightarrow \mathcal{H}',$$9, with 50 training instances and 30 test instances per scenario.

5. Empirical results and evolutionary behavior

The core empirical claim is that LLM-guided warm-starting improves initial population quality and frequently improves final heuristic quality relative to standard GP. Fitness density plots show that EvoSpeak’s initial populations are usually shifted toward lower fitness values than standard GP, and in some settings also have a wider spread, which the authors interpret as simultaneously improving initial quality and useful diversity (Xu et al., 3 Oct 2025).

In single-objective experiments over 30 runs, EvoSpeak significantly outperforms GP in three of six scenarios while remaining comparable in the others. The reported results are:

Scenario	GP	EvoSpeak	Outcome
$\Lambda$0	858.72(51.49)	868.74(48.62)	no significant difference
$\Lambda$1	1050.41(69.10)	1040.05(57.40)	no significant difference
$\Lambda$2	571.00(12.09)	564.81(3.76)	significant improvement
$\Lambda$3	624.70(12.26)	621.87(3.95)	significant improvement
$\Lambda$4	450.03(13.44)	445.76(10.01)	no significant difference
$\Lambda$5	569.96(31.01)	559.13(13.86)	significant improvement

A salient secondary pattern is that EvoSpeak’s standard deviations are consistently lower, suggesting more stable runs. The paper interprets the stronger gains on mean-based objectives as evidence that transfer works best when source and target tasks are similar.

The multi-objective experiments are preference-conditioned. EvoSpeak28, trained under preference $\Lambda$6, generally performs best when tested under $\Lambda$7, with average Friedman rank 1.75. EvoSpeak55, trained under $\Lambda$8, performs best when tested under $\Lambda$9, with average rank 1.25. EvoSpeak82, trained under $\mathcal{H}'$0, generally also outperforms GP, although the paper notes an anomaly in which EvoSpeak55 sometimes performs even better than EvoSpeak82 under $\mathcal{H}'$1, suggesting that some generated initial individuals generalize unusually well across preference settings.

Convergence analyses show that EvoSpeak often starts from a better initial point and maintains strong progress. The paper further reports higher phenotypic diversity throughout evolution, defining phenotypic diversity as the percentage of individuals with unique fitness values. Standard GP populations lose diversity quickly, whereas EvoSpeak retains more behavioral variety across generations. The authors take this as evidence that knowledge-guided initialization does not collapse the search into a narrow basin but changes the exploration–exploitation balance.

A terminal-usage analysis under $\mathcal{H}'$2 further supports the claim that EvoSpeak is not generating arbitrary trees. EvoSpeak-generated initial heuristics emphasize $\mathcal{H}'$3 and $\mathcal{H}'$4 and avoid terminals such as $\mathcal{H}'$5, $\mathcal{H}'$6, and $\mathcal{H}'$7, whereas standard GP uses terminals much more uniformly because its initialization is random.

6. Interpretability, transfer, limitations, and scope

Interpretability is one of EvoSpeak’s stated contributions, but it is implemented qualitatively rather than through a dedicated human-evaluation benchmark. After evolution, the LLM converts the final symbolic heuristic into a human-readable report intended to explain the rule in scheduling terms rather than merely restating its algebra. The paper gives an illustrative explanation fragment: “PT plays a more important role in scenarios where optimizing flowtime is the primary objective...” This indicates that the LLM is being used as a symbolic-to-natural-language translator for domain users, not merely as a pretty-printer (Xu et al., 3 Oct 2025).

Transferability is empirically supported, but unevenly. The paper explicitly notes that EvoSpeak performs better on mean-based objectives than on max-based objectives, plausibly because the source heuristics used for initialization are more similar to the mean-based target tasks. This suggests that transfer depends materially on task similarity. A plausible implication is that EvoSpeak’s knowledge reuse is strongest when the source heuristic library already spans the relevant objective family and operating regime.

Several limitations are explicit. EvoSpeak depends on LLM output quality and may inherit LLM biases or inaccuracies. The operators $\mathcal{H}'$8 and $\mathcal{H}'$9 remain conceptual rather than fully procedural. There is no quantitative evaluation of explanation fidelity, readability, or usefulness, and no comparison with other LLM-plus-evolution frameworks such as ReEvo or EoH. Preference-aware optimization is handled with weighted sums rather than a Pareto-front method, so some multi-objective trade-offs remain outside the paper’s operational scope.

A common misconception is suggested by the title: EvoSpeak is not a speech, audio, or AAC system. By contrast, "EgoSpeak" studies when an egocentric conversational agent should initiate speech (Kim et al., 17 Feb 2025); "SpeakEasy" is a spoken conversational coaching chatbot for college students (Jeon et al., 2023); "Speak Ease" is an AAC prototype combining multimodal input, LLM-based interpretation, and personalized TTS (Xu et al., 21 Mar 2025); "EvolveCaptions" adapts ASR collaboratively for DHH speakers in real time (Wu et al., 2 Oct 2025); "VoiceX" uses a $F(h) = \sum_{k=1}^{K} w_k f_k(h),$0-ES to search a VITS speaker space for custom voice design (Mertes et al., 2024); and "VoiceCoach" is an evidence-based public-speaking modulation trainer built from 2,623 TED Talks (Wang et al., 2020). EvoSpeak belongs instead to interpretable optimization and heuristic evolution.

Within that domain, its significance lies in reframing GP as part of a larger human-centered pipeline in which the LLM functions as knowledge extractor, warm-start generator, transfer mechanism, and translator. The empirical evidence is strongest for improved initialization quality, lower variance, and preference-aware gains under DFJSS. The interpretability contribution is more qualitative, but it directly addresses a longstanding practical problem: symbolic heuristics may be structurally interpretable yet operationally unreadable. EvoSpeak’s central claim is that LLMs can help close that gap without replacing conventional GP search (Xu et al., 3 Oct 2025).