Automated Heuristic Set Design (AHSD)

• Automated Heuristic Set Design (AHSD) is a formal framework that designs small, complementary sets of heuristics to effectively tackle heterogeneous instances and improve generalization.
• It leverages an objective function that is both monotone non-increasing and supermodular, is NP-hard via discrete k-center reduction, and attains a (1 - k/(e k - e))-approximation with greedy selection.
• AHSD underpins diverse algorithmic approaches—from EoH-S and MCTS-AHD to RAISE—that integrate evolutionary, tree-based, and adversarial strategies to enhance diversity, robustness, and transferability.

to AHSD. Since user only wants article, not extra. Need maybe integrate papers like HSEvo, RAISE robust set etc. Since topic AHSD, likely central papers are (Liu et al., 5 Aug 2025) (formal AHSD), maybe (Zhang et al., 2024) (set objective), (Terrazas et al., 2010) precursor, (Liu et al., 30 Jun 2026) robust set, plus related frameworks (EoH-S, MeLA, MeEvo, MCTS-AHD, PathWise, AHD Agent, RAISE, CoEvo-AHD). Need focus on AHSD, not general LLM-AHD. But article should cover related methodologies because AHSD field overlaps.

Need ensure claims from data verbatim. For example "objective function of AHSD is monotone and supermodular" exactly from (Liu et al., 5 Aug 2025). "NP-hard via reduction from discrete k-center." "greedy achieves (1 - k/(e k - e))-approximation" etc.

Need perhaps use formulas from paper. Keep precise. Use LaTeX.

Need section numbering in headings like "## 1. Conceptual origins". Intro no heading. Use technical prose.

Need maybe discuss misconception: AHSD is not merely running many independent AHD searches; explicit complementarity objective is central. This claim is supported by EoH-S details and figure 4? We can phrase factually: existing approaches only design a single heuristic... inducing poor generalization; AHSD addresses by generating small-sized complementary heuristic set (Liu et al., 5 Aug 2025). That's enough. Another misconception: small portfolio is not enough for robustness; RAISE shows EoH-S with 128 training instances outperformed by RAISE trained on 5 nominal instances. But careful EoH-S is robustness-aware AHD baseline. Actually details say "EoH-S uses 128 training instances yet is outperformed by RAISE trained on only 5 nominal instances." We can mention in robustness section. Need maybe say this concerns distribution shift rather than in-distribution complementarity.

Need decide whether to mention "Editor's term". Not necessary.

Need citations maybe at intro and specific sections. Could cite multiple papers per section. Remember not more than once per section per paper. That's okay.

Let's construct with tight compliance.

Potential intro paragraph: Automated Heuristic Set Design (AHSD) is a formulation of automated heuristic design in which the design object is not a single heuristic but a small complementary heuristic set. Given candidate space D, benchmark instances I, and set size k, AHSD seeks subset H of size k such that on each instance at least one heuristic in H performs well; in EoH-S this is formalized by minimizing complementary performance index $$\mathcal F(H)=\frac1m\sum_{i\in I}\min_{h\in H}f_i(h)$$, where $f_i(h)$ is instance-wise performance and lower is better (Liu et al., 5 Aug 2025). Relative to single-heuristic AHD, this formulation is motivated by instance diversity, distributional shift, and the empirical observation that one heuristic often generalizes poorly across varying sizes or distributions (Liu et al., 5 Aug 2025).

Section 1 origins: hyper-heuristics generate heuristics and self-assembly 2010. Then LLM-EPS benchmark and 2025 AHSD formulation. Need avoid claim "history" if not explicit? Could say "An earlier automated design methodology represented low-level TSP heuristics as Wang-tile-like components and treated assembled execution threads as heuristic pipelines" (Terrazas et al., 2010). Then "Recent LLM-based work recast heuristic search as evolutionary program search over executable programs or sets" (Zhang et al., 2024). Then "EoH-S introduced AHSD as new formulation" (Liu et al., 5 Aug 2025).

Section 2 formal problem. Use definitions from EoH-S, maybe mention alternative set objective from (Zhang et al., 2024) and robustness minimax from RAISE. Include formulas. Need maybe mention if k=1 reduces to single heuristic and if k=m choose best per instance. Then robust set formulation from RAIISE with epsilon-ball. Could say "For distribution shift, RAISE states an instance-level worst-case formulation..." formula. This broadens AHSD.

Section 3 mathematical properties. Monotone non-increasing, supermodular, NP-hard. Approx guarantee. Also perhaps mention relation to diversity and complementarity. Maybe mention Manhattan distance in EoH-S complementary-aware search and CPM pseudocode formula. Could be methodology though perhaps in next section. Let's keep section 3 properties and approximation.

Section 4 algorithmic architectures. Need discuss EoH-S and perhaps tree-based/agentic frameworks adapted to AHSD. Since article on AHSD, say current systems fall into population-based, tree-based, dual-population/co-evolutionary, and agentic/multi-agent. Need avoid inventing taxonomy maybe "Current AHSD-related systems include..." That's fine since papers represent these types. Need mention exact names and core mechanics: EoH-S complementary population management and memetic search (Liu et al., 5 Aug 2025); MCTS-AHD organizes every heuristic as node in search tree (Zheng et al., 15 Jan 2025); PoH frames heuristic design as MDP solved by MCTS (Wang et al., 17 Feb 2025); CogMCTS adds multi-round cognitive feedback and dual-track expansion (Wang et al., 9 Dec 2025); PathWise uses entailment graph as stateful memory (Gungordu et al., 28 Jan 2026); AHD Agent uses multi-turn agentic loop with tools (Lv et al., 9 May 2026); MeLA prompt evolution (Qiu et al., 28 Jul 2025); MeEvo couples natural and metacognitive evolution (Qiu et al., 12 Jun 2026); A2DEPT full program evolution (Chen et al., 27 Apr 2026). Need be concise maybe not too many details. But can mention "This suggests AHSD has become an umbrella over multiple search organizations." As inference.

Section 5 complementarity, diversity, robustness. Use HSEvo SWDI/CDI. Use AST-based operator expanding invalid region (Shengming et al., 3 Apr 2026), maybe also knowledge-first (Kiet et al., 7 May 2026) or LaGO modular forward-backward-update (Yu et al., 17 Feb 2026)—but maybe section 4 already crowded. Could mention complementarity/diversity/robustness specifically: HSEvo diversity metrics SWDI/CDI; AST-based operator; RAISE robust adversary instance search; knowledge-first search improves transfer and generalization (Kiet et al., 7 May 2026). Need careful not too many papers in one section? okay.

Section 6 domains and empirical record. Discuss domains where AHSD/AHD applied: TSP, bin packing, CVRP, FSSP, KP, etc. Need maybe cover networking case study from "Automated Heuristic Design for Network Operations" (Namvar et al., 27 May 2026). Since user topic AHSD, include application beyond COPs to network ops. Mention 5G LDPC decoding atomic task CNU and dense scalar objective. This shows extension beyond classical CO. Could also mention CoEvo-AHD on TTP/TPP, G-LNS on TSP/CVRP, robust OOD tasks, continuous transfer CAF, symbolic regression/protein engineering from knowledge-first paper. Need remain comprehensive but precise.

Could maybe include a small table of selected frameworks and design object. Table with 3 columns: framework, primary search object, notable mechanism. But detailed explanations in paragraphs before/after. Since user's audience okay. Need no more than 3 columns.

Need ensure no invented statistics. Use exact numbers only if from data. Could include up to 60% improvements from EoH-S, 10-heuristic set outperforms others even with 100 heuristics, etc. But maybe article doesn't need lots of stats. Maybe mention some key results in section 6. Also mention 5G case study produced on order 10^4-10^5 candidate programs and final best CNU statistically identical etc. But maybe that's AHD not AHSD. Still relevant as extension of frameworks.

Need manage citations once per section per paper. Could cite each paper once at most in section. Need not mechanical.

Let's draft carefully with factual statements tied to data. Need maybe not mention arxiv titles except maybe not necessary. But citations enough.

Potential sentences:

• "AHSD is NP-hard via reduction from discrete k-center." (Liu et al., 5 Aug 2025)
• "Within fixed candidate pool P, greedy addition by maximum marginal improvement achieves a $(1 - k/(e\,k-e))$-approximation." (Liu et al., 5 Aug 2025)
• "EoH-S maintains population n, uses CS selecting parent pair maximizing Manhattan distance over instance-wise performance, and LS selecting one parent by rank-based sampling." (Liu et al., 5 Aug 2025)
• "Complementary Population Management greedily selects next generation to minimize $\mathcal F$."
• "HSEvo measures diversity via SWDI and CDI..." (Dat et al., 2024)
• "RAISE casts robust AHSD as minimax over $\epsilon$-ball; on OBP exponential distribution RAISE(ε=0.005)=0.273% vs EoH=5.303% (~19.4x improvement)." (Liu et al., 30 Jun 2026)
• "CoEvo-AHD maintains two operator populations and synergy matrix S_{ij}." (Kuang et al., 30 May 2026)
• "G-LNS co-evolves destroy/repair operators..." (Zhao et al., 9 Feb 2026)
• "Network AHD uses four phases and asynchronous islands/samplers/evaluators/reset." (Namvar et al., 27 May 2026)

Need perhaps mention misconception/objectively:

• "AHSD should not be conflated with maintaining a population during search: in the formal AHSD objective, the returned set itself is the decision variable, and complementarity is evaluated instance-wise through the minimum over set members." That's directly inferable from formula; can say "A common misconception is to treat AHSD as merely keeping many candidates during optimization. In the formal AHSD problem, the set returned at the end is itself the optimization target." This is interpretation but backed. Use without citation maybe cite (Liu et al., 5 Aug 2025). It's a plausible explanatory statement maybe okay as interpretation not concrete; perhaps say "The formalism in EoH-S makes the distinction explicit..." safer.

Need maybe mention relation to AHD: "Some later frameworks are presented as AHD but explicitly summarize their components as applying equally to the goal of designing a set of heuristics rather than a single one." (Qiu et al., 12 Jun 2026) Could include in origins or methodologies.

Let's ensure all sections have factual-first then significance. Maybe section headings:

1. Historical emergence and relation to hyper-heuristics

2. Formal problem statements

3. Structural properties of the objective

4. Search architectures and algorithmic mechanisms

5. Diversity, robustness, and transfer

6. Application domains and empirical behavior

Need no heading at top.

Let's draft with moderate detail.

Automated Heuristic Set Design (AHSD) is a formulation of automated heuristic design in which the design target is a small complementary set of heuristics rather than a single heuristic. In the formulation introduced by EoH-S, one is given a design space $D$ of candidate heuristics, an instance set $I=\{i_1,\dots,i_m\}$, and a prescribed set size $k$, and seeks a subset $H\subseteq D$, $|H|=k$, such that on each instance $f_i(h)$0 at least one heuristic in $f_i(h)$1 performs well. The central objective is the complementary performance index (CPI), $f_i(h)$2, where $f_i(h)$3 is the performance score of heuristic $f_i(h)$4 on instance $f_i(h)$5 and lower is better (Liu et al., 5 Aug 2025). This formulation generalizes single-heuristic Automated Heuristic Design (AHD) to settings with heterogeneous instance distributions, cross-scale variation, and deployment scenarios in which complementarity matters as much as individual quality.

1. Historical emergence and conceptual scope

Earlier work on automated heuristic generation arose within the broader hyper-heuristics literature. A 2010 self-assembly methodology represented low-level TSP heuristics as Wang-tile-like components and treated assembled execution threads as heuristic pipelines, with the stated goal of presenting a methodology for the automated design of heuristics by means of self-assembly (Terrazas et al., 2010). In that setting, a library $f_i(h)$6 of low-level TSP heuristics was encoded into tiles, and pattern detection over execution threads was used to identify beneficial heuristic subsequences (Terrazas et al., 2010).

The LLM era shifted the dominant representation from symbolic composition of low-level moves to executable program search. A large-scale benchmark on LLM-based evolutionary program search described AHD, and explicitly AHSD when designing sets of heuristics, as a search over executable programs, with a common set objective written as

$f_i(h)$7

where $f_i(h)$8 is the cost of heuristic $f_i(h)$9 on instance $(1 - k/(e\,k-e))$0 (Zhang et al., 2024). That benchmark also argued empirically that standalone LLM generation is insufficient and that explicit evolutionary search is necessary (Zhang et al., 2024).

The term “Automated Heuristic Set Design” was then introduced as a new formulation for LLM-driven AHD by EoH-S, motivated by the observation that existing approaches only design a single heuristic to serve all problem instances, often inducing poor generalization across different distributions or settings (Liu et al., 5 Aug 2025). From that point onward, AHSD became both a formal optimization problem and a design principle: the output set is required to be complementary, not merely individually strong.

2. Formal problem statements

In the canonical AHSD formulation, $(1 - k/(e\,k-e))$1 denotes the performance of heuristic $(1 - k/(e\,k-e))$2 on instance $(1 - k/(e\,k-e))$3, and the set score is

$(1 - k/(e\,k-e))$4

The optimization problem is to minimize $(1 - k/(e\,k-e))$5 subject to $(1 - k/(e\,k-e))$6 (Liu et al., 5 Aug 2025). Two boundary cases clarify the formulation: if $(1 - k/(e\,k-e))$7, AHSD reduces to designing a single heuristic that minimizes average loss; if $(1 - k/(e\,k-e))$8, one can choose the best heuristic for each instance independently (Liu et al., 5 Aug 2025).

The formal distinction between AHD and AHSD is therefore not the use of a population during search, since many AHD systems maintain temporary populations. Rather, the returned set $(1 - k/(e\,k-e))$9 is itself the optimization variable, and complementarity is encoded directly through the per-instance minimum $\mathcal F$0 (Liu et al., 5 Aug 2025). This is the core mathematical feature that separates AHSD from population-based search procedures whose final output is a single incumbent heuristic.

Several later works broadened this basic formulation. MeLA defined AHSD more generally as finding a set $\mathcal F$1 maximizing an aggregate objective

$\mathcal F$2

where $\mathcal F$3, $\mathcal F$4, and $\mathcal F$5 denote solution quality, runtime, and success rate (Qiu et al., 28 Jul 2025). This formulation makes explicit that set design can be multi-criteria, although the EoH-S CPI objective remains the most detailed formalization of complementarity (Liu et al., 5 Aug 2025).

Robustness-oriented work extends AHSD to worst-case deployment neighborhoods. RAISE formulates robust AHSD as

$\mathcal F$6

where $\mathcal F$7 is an instance-level $\mathcal F$8-ball around the training set under a problem-specific distance (Liu et al., 30 Jun 2026). This replaces average-case complementarity with constrained worst-case complementarity and directly addresses distributional shift.

3. Structural properties and approximation theory

The EoH-S analysis establishes three central theoretical facts about the CPI objective. First, the AHSD problem is NP-hard via reduction from the discrete $\mathcal F$9-center problem (Liu et al., 5 Aug 2025). Second, the set function $\epsilon$0 is monotone non-increasing: for any $\epsilon$1, $\epsilon$2 (Liu et al., 5 Aug 2025). Third, $\epsilon$3 is supermodular: $\epsilon$4 for any $\epsilon$5 and any $\epsilon$6 (Liu et al., 5 Aug 2025).

These properties yield an approximation guarantee for greedy set construction. Writing the marginal improvement as

$\epsilon$7

the greedy algorithm that iteratively adds the heuristic with maximum marginal improvement achieves a $\epsilon$8-approximation to the optimum within any fixed candidate pool $\epsilon$9 (Liu et al., 5 Aug 2025). In practice, this means that one can decouple AHSD into two layers: an evolving candidate pool $D$0 produced by search, and a greedy extraction procedure that selects a complementary subset of size $D$1 from that pool (Liu et al., 5 Aug 2025).

This theory is important because it gives AHSD a structure that is stronger than a generic “diversity is good” intuition. Complementarity is not treated as an informal preference; it is captured by a set function with explicit monotonicity and supermodularity. A plausible implication is that many empirical design choices in later frameworks—pairwise distance metrics, diversity preservation, and complementary population management—can be read as computational surrogates for improving the marginal gains $D$2 under limited evaluation budgets.

4. Search architectures and algorithmic mechanisms

EoH-S is the first framework explicitly designed around the AHSD formulation. It evolves a population $D$3 of $D$4 heuristics and then extracts a final set of size $D$5 by greedy selection (Liu et al., 5 Aug 2025). Its memetic search has two operators. Complementary-aware Search selects two parent heuristics maximizing Manhattan distance in instance-wise performance,

$D$6

and asks the LLM to create a heuristic different from both (Liu et al., 5 Aug 2025). Local Search instead refines a single parent chosen by rank-based sampling (Liu et al., 5 Aug 2025). Its Complementary Population Management stage greedily retains $D$7 heuristics from a $D$8-sized pool by maximizing the marginal CPI reduction (Liu et al., 5 Aug 2025).

A number of related frameworks address AHSD either directly or by mechanisms that the authors state apply equally to heuristic sets. MCTS-AHD organizes every LLM-generated heuristic as a node in a tree rather than a flat population, with UCT-style selection, progressive widening, and LLM-driven crossover and mutation actions (Zheng et al., 15 Jan 2025). PoH formalizes heuristic design as an MDP solved by Monte Carlo Tree Search, where states are heuristic code strings, actions are textual improvement suggestions, and rewards are evaluation results (Wang et al., 17 Feb 2025). CogMCTS extends this line with multi-round cognitive feedback, positive and negative knowledge stores, dual-track node expansion, and elite heuristic management (Wang et al., 9 Dec 2025). PathWise treats heuristic generation as a sequential decision process over an entailment graph serving as stateful memory, with separate policy, world-model, and critic agents (Gungordu et al., 28 Jan 2026).

Other systems shift the search object or the evolutionary substrate. MeLA evolves prompts rather than code, using a problem analyzer, an error diagnosis system, and a metacognitive search engine that ranks prompts by the quality of heuristics they generate (Qiu et al., 28 Jul 2025). MeEvo cyclically couples Natural Evolution on code with Metacognitive Evolution on reasoning traces, explicitly recording fitness values, errors, and reasoning history into a shared archive (Qiu et al., 12 Jun 2026). A2DEPT moves from fixed heuristic templates to full-program evolution over executable solver programs represented in a program tree, with hybrid selection and a program-maintenance loop to ensure executability (Chen et al., 27 Apr 2026).

The resulting landscape is methodologically heterogeneous, but the common pattern is stable: heuristics are represented as compact executable artifacts, evaluated on instance sets, and selected not only for individual fitness but also for how they contribute to a broader search memory, archive, or final complementary set. This suggests that AHSD has become less a single algorithm than a problem class with several search organizations—population-based, tree-based, multi-agent, prompt-evolutionary, and full-program.

5. Complementarity, diversity, robustness, and transfer

Because AHSD optimizes sets, diversity is not incidental. HSEvo introduced two Shannon-entropy-inspired diversity metrics over an archive of vector-encoded heuristics: the Shannon–Wiener Diversity Index,

$D$9

and the Cumulative Diversity Index,

$I=\{i_1,\dots,i_m\}$0

where $I=\{i_1,\dots,i_m\}$1 are cluster proportions and $I=\{i_1,\dots,i_m\}$2 are normalized MST edge lengths in embedding space (Dat et al., 2024). In its experiments, EoH demonstrated higher diversity than FunSearch and ReEvo, while ReEvo yielded good objective scores but failed to optimize diversity effectively; HSEvo was designed to balance diversity and convergence by combining Flash Reflection with Harmony Search (Dat et al., 2024).

Search-space expansion is another route to complementarity. The two-stage AST-based operator of (Shengming et al., 3 Apr 2026) explicitly traverses invalid code regions $I=\{i_1,\dots,i_m\}$3 by performing subtree crossover and node-deletion mutation on Abstract Syntax Trees, then uses the LLM for repair (Shengming et al., 3 Apr 2026). Depending on the framework, either raw invalid variants or repaired heuristics are retained in the population (Shengming et al., 3 Apr 2026). The paper argues that this breaks “validity-induced boundaries” and enlarges the explored algorithm space (Shengming et al., 3 Apr 2026). A plausible implication is that AHSD benefits not only from selecting complementary heuristics but also from generating structurally diverse candidate families before selection.

Robustness under shift has become a distinct AHSD concern. RAISE embeds an LLM-free adversarial instance search within the outer heuristic-evolution loop, searching an $I=\{i_1,\dots,i_m\}$4-ball around the training set through a basis-distribution parameterization and radial boundary projection (Liu et al., 30 Jun 2026). It reports that existing LLM-based AHD methods degrade by up to $I=\{i_1,\dots,i_m\}$5 times under distribution shift, while RAISE maintains strong performance across tested distributions and scales; in Online Bin Packing on an Exponential test distribution, RAISE($I=\{i_1,\dots,i_m\}$6) obtains $I=\{i_1,\dots,i_m\}$7 waste versus $I=\{i_1,\dots,i_m\}$8 for EoH, approximately $I=\{i_1,\dots,i_m\}$9 better (Liu et al., 30 Jun 2026). The same work states that EoH-S uses 128 training instances yet is outperformed by RAISE trained on only 5 nominal instances (Liu et al., 30 Jun 2026). This does not negate the value of complementary heuristic sets, but it shows that in-distribution complementarity and out-of-distribution robustness are not equivalent objectives.

A related but conceptually different direction is knowledge-first search. “Back to the Beginning of Heuristic Design” distinguishes bottom-up, code-centric search from top-down, knowledge-first search, formalizes a distortion–compression trade-off, and reports that knowledge-first pipelines improve discovery efficiency, transfer, and generalization across combinatorial optimization and tasks beyond it (Kiet et al., 7 May 2026). In that framework, knowledge states $k$0 become the primary search object and code merely realizes them (Kiet et al., 7 May 2026). This suggests that future AHSD may return not only a set of executable heuristics but also an explicit set of reusable heuristic principles.

6. Domains, case studies, and empirical record

AHSD and closely related AHD frameworks have been evaluated on a wide range of combinatorial optimization problems. EoH-S reports results on Online Bin Packing, Traveling Salesman Problem, and Capacitated Vehicle Routing Problem, with population $k$1, budget $k$2 LLM evaluations, and up to $k$3 performance improvements over single-heuristic methods; its 10-heuristic set outperforms other methods even with 100 heuristics in the CPI analysis shown in Figure 1 of the paper (Liu et al., 5 Aug 2025). PoH reports state-of-the-art performance on TSP and Flow Shop Scheduling Problem, especially with large sizes (Wang et al., 17 Feb 2025). G-LNS extends the design target from single constructive operators to coupled destroy and repair operators for Large Neighborhood Search on TSP and CVRP, with co-evolved operator pools and synergy-aware evaluation (Zhao et al., 9 Feb 2026). CoEvo-AHD generalizes the set idea to dual-population co-evolution for bi-component coupled optimization, maintaining separate route and selection operator populations together with a synergy matrix $k$4 for Traveling Thief Problem and Traveling Purchaser Problem (Kuang et al., 30 May 2026).

The scope now extends beyond classical COPs. AHD Agent trains a 4B-parameter agentic policy that can decide whether to generate heuristics, call tools, or finalize a design, and reports strong results across eight domains including four held-out tasks and a continuous Cost-Aware Bayesian Optimization setting (Lv et al., 9 May 2026). LaGO decomposes heuristic search into forward evaluation, backward analytical feedback, and update, and reports up to $k$5 improvement in QYI on unseen test sets across PDPTW, Crew Pairing, DAOmap, and TSP (Yu et al., 17 Feb 2026). Knowledge-first search has also been evaluated on symbolic regression and protein engineering tasks (Kiet et al., 7 May 2026).

A particularly significant extension is to network operations. “Automated Heuristic Design for Network Operations” describes an end-to-end methodology that leverages LLMs inside an evolutionary loop to generate, evaluate, and refine tiny, human-interpretable code snippets for complex network functions (Namvar et al., 27 May 2026). The framework is organized into four phases: suitability analysis, atomic task selection, scoring function design, and evaluation methodology (Namvar et al., 27 May 2026). In a 5G LDPC decoding case study, the atomic task is the check-node-update function, and the objective is a dense hierarchical scalar

$k$6

where catastrophic evaluator errors, undecoded transport blocks, BER, and mean iteration count are weighted lexicographically (Namvar et al., 27 May 2026). The reported run generated approximately $k$7 programs over 48 hours on an asynchronous 16-GPU LLM plus 100-core evaluator cluster, and the final best heuristic was statistically identical to the industry-standard Boxplus–$k$8 baseline on the tested contexts (Namvar et al., 27 May 2026). Although this paper is framed as AHD rather than AHSD, its emphasis on libraries of context-tuned heuristics and reusable architecture makes it directly relevant to set design.

A recurring misconception is that AHSD is simply a larger-population version of AHD. The empirical record does not support that simplification. The defining improvement in EoH-S is not merely keeping more candidates, but optimizing complementarity explicitly through the CPI objective and corresponding population-management rule (Liu et al., 5 Aug 2025). At the same time, robustness work shows that even complementary sets can remain brittle under shift unless the training process actively searches for adversarial failure modes (Liu et al., 30 Jun 2026). AHSD is therefore best understood as a family of methods for constructing small, high-quality, mutually complementary heuristic portfolios under computational budgets, structural constraints, and increasingly, robustness requirements.