Heuristic Set Evolution (EoH-S)

• Evolution of Heuristic Set (EoH-S) is an LLM-driven framework that automatically creates a diverse portfolio of heuristics tailored for varied optimization tasks.
• It employs complementary-aware search combined with greedy population management to ensure each heuristic uniquely contributes to improved instance coverage.
• Empirical results on tasks like TSP, OBP, and CVRP show EoH-S achieving up to a 60% performance boost over traditional single-heuristic methods.

Searching arXiv for papers on EoH-S and closely related LLM-based automated heuristic design. Evolution of Heuristic Set (EoH-S) is an LLM-based automated heuristic design framework that targets the automatic generation of a small complementary set of heuristics, rather than a single heuristic, for a given optimization task. In the formulation associated with Automated Heuristic Set Design (AHSD), a heuristic set $H = \{h_1,\dots,h_k\}\subseteq D$ is evaluated by the instance-wise best performance it can attain across a diverse instance set $I$, with the objective $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$, where $f^*_{H,i}=\min_{h\in H} f_i(h)$ and lower values are better (Liu et al., 5 Aug 2025). This shifts the unit of design from an individual heuristic to a portfolio whose members are explicitly selected for complementarity. Subsequent work interprets EoH-S more broadly as the evolution of an entire heuristic population or portfolio, sometimes augmented with reasoning traces, adversarial instances, or quality-diversity archives, but the core idea remains the same: the evolving object is a set of heuristics whose members cover different instance regimes and failure modes (Qiu et al., 12 Jun 2026, Liu et al., 30 Jun 2026, Gestrin et al., 28 May 2026).

1. Conceptual foundation and formal objective

The immediate predecessor of EoH-S is Evolution of Heuristics (EoH), which already maintained a population of heuristics represented as paired natural-language “thoughts” and executable code, but ultimately selected and deployed a single best heuristic per task (Liu et al., 2024). EoH-S generalizes this single-heuristic view by treating diversity across instances as a first-class design target. In the AHSD formulation, the task $\mathcal{T}$ is associated with an instance set $I=\{i_1,\dots,i_m\}$, a search space $D$ of heuristics, and a set size $1<k\ll m$. A heuristic set $H=\{h_1,\dots,h_k\}\subseteq D$ is scored by

$$f^*_{H,i} = \min_{h \in H} f_i(h), \qquad \mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i},$$

where $I$0 is the loss of heuristic $I$1 on instance $I$2 (Liu et al., 5 Aug 2025).

This formulation makes explicit that a heuristic set is valuable when each instance can be served well by at least one member of the set. In that sense, complementarity is not an auxiliary notion but the primary optimization target. The AHSD objective is proved NP-hard via reduction from the Discrete Clustering Problem. It is also monotone and supermodular: adding heuristics never worsens $I$3, and the marginal decrease in $I$4 from adding a new heuristic exhibits diminishing returns as the set grows (Liu et al., 5 Aug 2025). This property underwrites greedy subset construction within a finite candidate pool.

A related but broader interpretation appears in later work. RAISE explicitly distinguishes EoH, which optimizes a single heuristic, from EoH-S, which learns a portfolio of diverse heuristics to improve cross-distribution performance (Liu et al., 30 Jun 2026). MeEvo, in turn, characterizes “Evolution of Heuristic Set (EoH-S)” as the dynamics of the entire tuple $I$5, where $I$6 is the set of heuristics, $I$7 their reasoning traces, $I$8 their fitness values, and $I$9 their error logs (Qiu et al., 12 Jun 2026). This suggests a broader conceptual shift from “best heuristic search” to “knowledge-bearing population evolution.”

2. The EoH-S algorithmic framework

EoH-S as introduced for AHSD is an LLM-driven memetic evolutionary framework with two central mechanisms: complementary-aware memetic search and complementary population management (Liu et al., 5 Aug 2025). The representation remains aligned with EoH: each heuristic is a natural-language thought paired with a Python implementation conforming to a task-specific interface (Liu et al., 2024, Liu et al., 5 Aug 2025). The difference is that evaluation, selection, and reproduction are driven by set-level complementarity rather than average single-heuristic fitness alone.

The framework initializes a population $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$0 of heuristics using LLM prompts that specify the task description and function template. It then iterates until a fixed evaluation budget $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$1 is exhausted. In each round, it generates $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$2 offspring from the current population of size $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$3, evaluates them on all training instances, and applies Complementary Population Management (CPM) to select a new population of size $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$4 from the $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$5 combined parent-offspring pool (Liu et al., 5 Aug 2025).

Complementary-aware Search (CS) selects two parent heuristics $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$6 that are maximally complementary according to the Manhattan distance between their instance-wise performance vectors,

$\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$7

and prompts the LLM to create a new heuristic “that is different from the given ones” (Liu et al., 5 Aug 2025). Local Search (LS), by contrast, selects a single parent via weighted random sampling biased toward better average performance and asks the LLM to create an improved version of that heuristic (Liu et al., 5 Aug 2025). EoH-S uses CS and LS with equal probability, thereby balancing exploration of new specializations and exploitation of strong existing heuristics.

CPM then greedily constructs the next population by maximizing marginal improvement in set quality. Given a partial set $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$8, the delta CPI of a candidate $\mathcal{F}(H)=\frac{1}{m}\sum_{i\in I} f^*_{H,i}$9 is

$f^*_{H,i}=\min_{h\in H} f_i(h)$0

which measures how much the candidate improves current instance-wise best performance across the training distribution (Liu et al., 5 Aug 2025). The first selected heuristic is the best singleton by average performance; subsequent selections maximize $f^*_{H,i}=\min_{h\in H} f_i(h)$1. This is the operational core of EoH-S’s complementary population management.

3. Relation to earlier and later LLM-AHD systems

EoH-S sits within a lineage of LLM-based automated heuristic design frameworks but differs from them in what is optimized and preserved. Original EoH evolves a population of thoughts and codes using prompt strategies such as initialization, exploration, and modification, then keeps the single best heuristic (Liu et al., 2024). ReEvo adds reflective feedback but still optimizes toward a single incumbent, and HSEvo emphasizes exploration-exploitation balance through diversity metrics and harmony search while remaining primarily a single-heuristic framework (Dat et al., 2024). MEoH extends EoH to multi-objective search and returns a non-dominated set of heuristics, but its focus is objective trade-offs such as quality and runtime rather than explicit complementarity over instance coverage (Yao et al., 2024).

EoH-S’s distinctive contribution is to formulate automated heuristic set design directly as a set optimization problem and to tie both search and population update to the marginal contribution of each heuristic to the current set (Liu et al., 5 Aug 2025). This makes it closer in spirit to algorithm portfolios and hyper-heuristics than to classical evolutionary program search, even though it retains LLM-mediated code generation and mutation.

Later systems generalize or critique this design in different directions. RAISE argues that EoH-S improves cross-distribution robustness relative to single-heuristic AHD because it evolves a portfolio of heuristics, but still depends on a predefined set of diverse training instances and lacks explicit worst-case robustness optimization (Liu et al., 30 Jun 2026). The AST-based operator paper treats EoH-S as a framework whose heuristic set can be improved by replacing one-stage semantic variation with a two-stage operator: first generate structurally diverse, often invalid AST variants, then repair them with an LLM (Shengming et al., 3 Apr 2026). MeEvo goes further by coupling Natural Evolution over heuristic code with Metacognitive Evolution over reasoning traces and historical failures, effectively turning the evolving object into the full heuristic set plus its cognitive history (Qiu et al., 12 Jun 2026).

The symbolic planning work on domain-independent heuristic evolution does not use the EoH-S name as its primary method label, but it realizes a strong portfolio interpretation by maintaining a MAP-Elites archive of heuristics distributed over an informedness-speed trade-off and showing that the resulting heuristic suite spans the Pareto frontier (Gestrin et al., 28 May 2026). This suggests that EoH-S can also be understood as a quality-diversity process over heuristic repertoires.

4. Theoretical properties, complementarity, and selection principles

The AHSD objective underlying EoH-S has three consequences of algorithmic significance. First, because $f^*_{H,i}=\min_{h\in H} f_i(h)$2 aggregates instance-wise minima, a heuristic with mediocre average performance can still be essential if it dominates on a specific subset of difficult instances (Liu et al., 5 Aug 2025). Second, monotonicity ensures that enlarging the set cannot worsen the objective, while supermodularity implies diminishing returns from adding more heuristics (Liu et al., 5 Aug 2025). Third, greedy subset construction over a finite pool admits a constant-factor approximation guarantee relative to the optimal size-$f^*_{H,i}=\min_{h\in H} f_i(h)$3 subset of that pool (Liu et al., 5 Aug 2025).

These properties clarify why naive “top-$f^*_{H,i}=\min_{h\in H} f_i(h)$4 by average score” selection is insufficient. Such a rule favors redundant heuristics clustered around the same region of instance space. EoH-S instead uses a marginal-gain criterion that rewards instance-specific improvements not already captured by the current set (Liu et al., 5 Aug 2025). The resulting heuristic set is small but intentionally non-redundant.

This complementarity principle reappears, in modified form, across later work. In RAISE, the evolving heuristic set is continually re-scored against an expanding adversarial instance set $f^*_{H,i}=\min_{h\in H} f_i(h)$5, so complementarity emerges from robustness to a changing evaluation environment rather than from CPI-based subset selection (Liu et al., 30 Jun 2026). In HSEvo, diversity is measured in embedding space using the Shannon–Wiener Diversity Index (SWDI) and the Cumulative Diversity Index (CDI), emphasizing spread across heuristic families and global archive coverage (Dat et al., 2024). In planning, MAP-Elites preserves heuristics across niches defined by informedness and speed, making complementarity a function of behavioral descriptors rather than training-instance minima (Gestrin et al., 28 May 2026). These variants do not replace AHSD’s set objective, but they show that EoH-S admits multiple operationalizations of “complementarity.”

5. Empirical performance and benchmark evidence

The original EoH-S paper evaluates the method on three tasks: Online Bin Packing (OBP), Traveling Salesman Problem (TSP), and Capacitated Vehicle Routing Problem (CVRP), with diverse training and testing distributions (Liu et al., 5 Aug 2025). In all three domains, EoH-S consistently outperforms state-of-the-art single-heuristic baselines and also beats “Top10” sets assembled post hoc from the histories of methods such as EoH, FunSearch, and ReEvo (Liu et al., 5 Aug 2025). On classical benchmarks and out-of-distribution tests, reported improvements reach up to 60% (Liu et al., 5 Aug 2025).

The central empirical claim is not merely that the best EoH-S member is stronger, but that the final set exhibits substantially better Complementary Performance Index as its size increases. Figure-level analyses in the paper show that EoH-S’s CPI drops much faster with set size than competing methods, indicating that each additional heuristic adds meaningful coverage, whereas many heuristics in baseline-generated top-$f^*_{H,i}=\min_{h\in H} f_i(h)$6 sets are effectively redundant (Liu et al., 5 Aug 2025). Ablation studies show that removing Complementary Population Management largely collapses the advantage, bringing performance close to that of the strongest baseline (Liu et al., 5 Aug 2025). Removing either Complementary-aware Search or Local Search is also harmful, but less so than removing CPM (Liu et al., 5 Aug 2025).

Related evidence from subsequent work reinforces the same interpretation. RAISE reports that EoH-S is stronger than EoH and BestFit on Online Bin Packing under multiple distributions, with average waste 1.828% versus 1.948% for BestFit on $f^*_{H,i}=\min_{h\in H} f_i(h)$7, though it is still surpassed by robust adversarial instance search under distribution shift (Liu et al., 30 Jun 2026). The AST-based operator paper reports that replacing EoH-S’s semantic operators with a two-stage AST-destruction-and-repair mechanism improves both optimization quality and convergence speed on TSP and OBP, while reducing LLM token consumption on TSP from 1,493,820 to 821,698 in the EoH-S-i configuration (Shengming et al., 3 Apr 2026). These results suggest that EoH-S is a strong base framework but sensitive to the design of its search operators.

6. Extensions, reinterpretations, and limitations

A major extension appears in MeEvo, which recasts the heuristic set not just as code but as the tuple $f^*_{H,i}=\min_{h\in H} f_i(h)$8, where reasoning traces and error traces become heritable material (Qiu et al., 12 Jun 2026). In this view, EoH-S is implemented as dual-layer evolution: Natural Evolution performs LLM-mediated crossover and mutation on code while logging traces and errors, and Metacognitive Evolution reflects on the accumulated history to produce new heuristics that re-enter the parent pool (Qiu et al., 12 Jun 2026). This broadens the meaning of heuristic-set evolution from portfolio optimization to bilevel evolution over both executable programs and design knowledge.

RAISE offers a different extension. It identifies a limitation of EoH-S: although it evolves a set of heuristics and trains on a diverse instance set, it does not explicitly optimize for robustness under constrained distribution shift (Liu et al., 30 Jun 2026). RAISE therefore adds an LLM-free inner adversarial instance search that finds hard instances within an $f^*_{H,i}=\min_{h\in H} f_i(h)$9-ball around nominal training instances and continuously updates the evaluation pool (Liu et al., 30 Jun 2026). This turns the evolving heuristic set into a robust portfolio shaped by minimax pressure rather than only static diversity.

Another extension is structural. The AST-based paper argues that one-stage semantic operators constrain search to valid-code regions and limit the structural diversity of heuristic sets (Shengming et al., 3 Apr 2026). By allowing invalid intermediate AST variants and repairing them afterward, the method enlarges the reachable search space and produces more diverse heuristic sets for EoH-S (Shengming et al., 3 Apr 2026). The planning work suggests yet another avenue: rather than defining complementarity by instance coverage, one can define it over behavioral descriptors and maintain an archive that spans a Pareto frontier of heuristic properties (Gestrin et al., 28 May 2026).

Despite these expansions, several limitations remain recurrent. EoH-S and related systems are computationally expensive because each heuristic must be executed over many instances, and LLM generation remains a significant cost (Liu et al., 5 Aug 2025, Qiu et al., 12 Jun 2026). The final set in original EoH-S is evaluated with oracle selection—each instance is assumed to be served by its best member—so deployment may require either trying all heuristics or learning a separate selector (Liu et al., 5 Aug 2025). Robustness beyond the training distribution remains imperfect without explicit adversarial or multi-distribution mechanisms (Liu et al., 30 Jun 2026). Finally, set design is typically still offline; most frameworks do not evolve online switching policies or cooperative interaction among heuristics within a single solver run, though later papers identify this as a natural next step (Liu et al., 5 Aug 2025, Liu et al., 30 Jun 2026).

7. Broader significance

EoH-S marks a transition in LLM-based automated heuristic design from single-program search to structured portfolio construction. Its AHSD formulation makes explicit that in heterogeneous optimization domains, the appropriate design target is often a small complementary set rather than a universal heuristic (Liu et al., 5 Aug 2025). The monotone-supermodular objective, the CPI-based greedy population management, and the complementary-aware memetic search collectively provide one of the first principled formulations of heuristic-set evolution in the LLM era (Liu et al., 5 Aug 2025).

Subsequent work indicates that this idea is fertile rather than final. It can be combined with metacognitive reflection over reasoning traces (Qiu et al., 12 Jun 2026), constrained adversarial instance search for robustness (Liu et al., 30 Jun 2026), AST-based structural operators for expanded search spaces (Shengming et al., 3 Apr 2026), or quality-diversity archives for behavioral repertoires (Gestrin et al., 28 May 2026). Taken together, these developments suggest that EoH-S is best understood not only as a specific algorithm introduced in 2025, but also as a methodological shift: the evolving object in automated heuristic design is increasingly the heuristic set itself, together with the memory, instance distribution, and behavioral structure that make that set effective.