Virtual Best Encoding (VBE) in ASP

Updated 13 July 2026

Virtual Best Encoding (VBE) is an oracle-style benchmark that computes the minimum runtime across diverse ASP encoding variants for each instance.
It employs LLM-generated streamliners to produce semantically diverse constraints that improve runtime by leveraging the complementary strengths of each variant.
Empirical evaluations on ASP Competition benchmarks show VBE achieves up to 4–5× speedups and notable reductions in total runtime over the original encoding.

Virtual Best Encoding (VBE) is an oracle-style benchmarking measure for comparing multiple encoding variants of the same problem. In the Answer Set Programming (ASP) setting, VBE records, for each benchmark instance, the fastest runtime achieved by any considered encoding, including the original encoding and streamlined variants obtained by adding candidate constraints (Voboril et al., 21 Apr 2026). Its principal role is to quantify the attainable “portfolio effect”: the gain that would be available if one could select the best encoding variant on a per-instance basis in advance, even though such oracle selection is not directly deployable (Voboril et al., 21 Apr 2026).

1. Definition and conceptual role

In the ASP formulation, VBE is defined over a set of encodings $\mathcal{E}$ and an instance $i$ by

$\mathrm{VBE}(i) = \min_{E \in \mathcal{E}} T(E,i),$

where $T(E,i)$ denotes the runtime of encoding $E$ on instance $i$ (Voboril et al., 21 Apr 2026). The set $\mathcal{E}$ contains the original encoding and its streamlined variants.

This construction is explicitly a benchmarking device rather than a solver architecture. The paper characterizes VBE as a performance upper bound that would be achievable if, for each instance, one could magically select the fastest encoding variant in advance (Voboril et al., 21 Apr 2026). Accordingly, VBE is used to estimate the “portfolio effect” and to demonstrate the complementarity of different streamliners: even when no individual streamliner is uniformly superior, the virtual aggregate may still yield substantial improvements (Voboril et al., 21 Apr 2026).

A corresponding aggregate measure can be formed over an instance set $\mathcal{I}$ : $\mathrm{VBE\_sum} = \sum_{i \in \mathcal{I}} \min_{E \in \mathcal{E}} T(E,i).$ For penalized scoring, the paper also describes PAR2-style evaluation, in which timeouts are assigned $2 \times$ the time limit before taking the per-instance minimum (Voboril et al., 21 Apr 2026).

2. Generation of streamlined encodings in ASP

The VBE construction in “Streamliners for Answer Set Programming” is embedded in a fully automated pipeline for generating streamliner constraints with LLMs (Voboril et al., 21 Apr 2026). The inputs are the original ASP encoding in Clingo format and a small set of satisfiable training instances solvable in seconds (Voboril et al., 21 Apr 2026).

The prompt asks the LLM to generate five symmetry-breaking or implied Clingo constraints in JSON format, while insisting on syntactic correctness and creativity (Voboril et al., 21 Apr 2026). The process adapts the StreamLLM approach, originally designed to generate streamliners for Constraint Programming, to ASP (Voboril et al., 21 Apr 2026). Multiple LLMs are sampled randomly for diversity; the examples named in the description are Claude, GPT-5 Mini, Gemini Pro, Mistral, and Deepseek (Voboril et al., 21 Apr 2026).

The paper emphasizes that the generated constraints are not merely syntactic variants. Different LLMs produce semantically diverse constraints, including symmetry breakers, implied constraints, and occasionally nontrivial or domain-specific pruning rules (Voboril et al., 21 Apr 2026). This is presented as evidence that the method captures genuine problem structure rather than superficial reformulation (Voboril et al., 21 Apr 2026).

3. Filtering candidates and selecting the training portfolio

Candidate constraints are subjected to a multi-stage filtering procedure before they contribute to VBE computation. First, baseline runtimes are recorded by running the original encoding on the training instances (Voboril et al., 21 Apr 2026). LLM-generated candidates are then screened.

Three classes of candidates are discarded: those that cause syntax errors; those that make any satisfiable training instance unsatisfiable; and those that do not improve runtime on any training instance, equivalently those that degrade performance on all training instances (Voboril et al., 21 Apr 2026). Only constraints that improve runtime on at least one training instance are retained (Voboril et al., 21 Apr 2026).

This search is run iteratively for 30 minutes, accumulating retained streamliners (Voboril et al., 21 Apr 2026). If the retained streamliners are denoted $i$ 0, then every triple $i$ 1 is evaluated on the training set through

$i$ 2

where $i$ 3 is the original encoding (Voboril et al., 21 Apr 2026). The selected triple is

$i$ 4

that is, the combination of three streamliners minimizing the total runtime over the training instances (Voboril et al., 21 Apr 2026).

This training-time selection criterion is itself VBE-based: it does not ask for a universally dominant streamliner, but for a small set whose per-instance complementarity yields the strongest aggregate effect (Voboril et al., 21 Apr 2026).

4. Benchmark-time computation and scoring conventions

On the test set, the described procedure runs the original encoding together with the three selected streamlined encodings; the paper’s detailed description also allows optionally running the encoding with all three streamliners together (Voboril et al., 21 Apr 2026). For a test instance $i$ 5, the benchmark-time VBE is therefore computed as

$i$ 6

in the version that includes the joint encoding (Voboril et al., 21 Apr 2026).

The total VBE can then be reported either as a sum of runtimes or as a penalized average such as PAR2, with unsolved instances penalized by $i$ 7 the cutoff (Voboril et al., 21 Apr 2026). The paper also provides pseudocode reflecting this evaluation pattern: runtime is measured for each encoding on each instance up to a time limit, transformed to a PAR2 value if needed, and minimized across encodings on a per-instance basis (Voboril et al., 21 Apr 2026).

The methodological significance of this evaluation is that it separates two questions. One concerns whether individual streamliners are useful in isolation; the other concerns whether the family of generated encodings exhibits enough heterogeneity to support a strong virtual portfolio. VBE is designed for the second question (Voboril et al., 21 Apr 2026).

5. Empirical results on ASP Competition benchmarks

The reported experiments use three ASP Competition benchmarks: Partner Units Problem, Sokoban, and Towers of Hanoi (Voboril et al., 21 Apr 2026). On these benchmarks, the VBE achieves speedups of up to $i$ 8-- $i$ 9 over the original encoding (Voboril et al., 21 Apr 2026).

Benchmark	Original	Best VBE
Partner Units Problem	17/26; 13,938 s	22/26; 3,302 s
Sokoban	12/18; 8,837 s	14/18; 2,178 s
Towers of Hanoi	31/46; 56,710 s	46/46; 11,833 s

For Partner Units Problem, a second run is also reported in which VBE solves 20 instances with a total time of 7,412 s, compared with 13,938 s for the original encoding (Voboril et al., 21 Apr 2026). For Sokoban, a second run yields 3,758 s against 8,837 s for the original encoding (Voboril et al., 21 Apr 2026). For Towers of Hanoi, the second-run VBE total is 16,053 s against 56,710 s for the original encoding (Voboril et al., 21 Apr 2026).

Summarized as reductions in total runtime, the reported VBE improvement ranges from 47% to 79%, while often solving more instances than the original encoding (Voboril et al., 21 Apr 2026). The abstract compresses this result as speedups of up to 4--5x (Voboril et al., 21 Apr 2026).

An important empirical observation is that no single streamliner dominates. The fact that VBE outperforms every single streamliner, including combined variants, is presented as evidence that the observed gains arise from complementarity rather than from the existence of one universally superior added constraint (Voboril et al., 21 Apr 2026).

6. Interpretation, significance, and limitations

The empirical behavior of VBE in this setting is tied to the nature of the generated streamliners. The paper states that LLMs can produce syntactically valid and semantically diverse streamliners, and that these constraints can yield real performance improvements even when they are logically redundant or quadratic in grounding size (Voboril et al., 21 Apr 2026). It further notes that manual reformulations intended to improve scaling, such as replacing $\mathrm{VBE}(i) = \min_{E \in \mathcal{E}} T(E,i),$ 0 rules with aggregates, sometimes resulted in worse empirical performance (Voboril et al., 21 Apr 2026). This suggests that the benchmarked utility of a streamliner is not reducible to a simple syntactic or asymptotic proxy.

The method is also notable for requiring only small training sets and no domain- or encoding-specific templates, in contrast to prior learning methods that need a “language bias” (Voboril et al., 21 Apr 2026). In this sense, VBE functions not only as an evaluation metric but also as a diagnostic for whether a template-free streamliner generation process is discovering a useful and diverse encoding family.

The study’s stated efficacy limit is that it is restricted to decision problems; extending the approach to optimization is framed as future work (Voboril et al., 21 Apr 2026). Because VBE is oracle-style, it should not be conflated with an online selection policy or a directly executable portfolio solver. Its practical value in the paper is as an upper-bound analysis of attainable gains under perfect per-instance selection (Voboril et al., 21 Apr 2026).

The phrase “virtual best encoding” also appears in a broader empirical sense outside ASP. In multilingual text classification, the 2017 study by Zhang and LeCun describes a virtual best encoding as a hypothetical aggregate in which an oracle always selects the best encoding method and model for each dataset; the paper does not provide an explicit VBE error figure, but uses the concept to underscore that no single encoding or model is always best across Chinese, English, Japanese, and Korean datasets (Zhang et al., 2017). In that usage, the notion is again an oracle upper bound over a family of encodings rather than a deployable fixed method.

The acronym “VBE” is also used for unrelated concepts. In DAO governance research, “Voting-Bloc Entropy” denotes a decentralization metric based on clustering token holders by aligned utility functions across elections (Austgen et al., 2023, Fábrega et al., 26 Sep 2025). In quantum algorithms, “Variational Block-Encoding” denotes a variational method for compiling block-encoding unitaries with near-optimal resource requirements (Rullkötter et al., 23 Jul 2025). These usages are terminologically unrelated to Virtual Best Encoding, even though all three employ the same acronym.

Within the ASP context, however, Virtual Best Encoding has a specific and narrow technical meaning: a per-instance minimum over a set of original and streamlined encodings, used to expose the full portfolio value of LLM-generated streamliner constraints (Voboril et al., 21 Apr 2026).