paper_title=TopSeed: Learning Seed Selection Strategies for Symbolic Execution from Scratch
authors=Jaehyeok Lee; Sooyoung Cha
venue=47th IEEE/ACM International Conference on Software Engineering (ICSE 2025)
paper_doi=https://doi.org/10.1109/ICSE55347.2025.00095
artifact_doi=https://doi.org/10.5281/zenodo.14602979
artifact_repository=https://github.com/skkusal/TopSeed
repository_head=2193d090a6650edc08ee30b927e65ba9e6f9e045
artifact_archive_sha256=7faee4308f67ba21518935c002948b924f5ae137d5676bb329202fcd6b576d12
artifact_accepted_paper_sha256=dac02e44533adc16527254370c1f7f5319f15f3f88e30e9a058e9b963f3e0b13
artifact_archive_commit_comment=339ace19430169a91f2f1a5bb121a9cd19f365b8
inspected_paper=Algorithms 1-3; Tables II-VII and IX-X; Explore; Exploit; Learn; experiment and threats-to-validity sections
inspected_artifact=topseed.py; base.py; utils/exploration.py; utils/exploitation.py; utils/execute_update_distributions.py; README; accepted paper
algorithm_identity=cross-run seed selection over repeated short symbolic-execution slices; not topological execution and not pending-state search
paper_eta_time_seconds=120
paper_explore_probability=0.75
paper_exploit_probability=0.25
paper_eta_learn=20
artifact_topseed_py_eta_lp=10
default_resolution=follow paper eta_learn=20; expose configuration and retain divergence in documentation
explore_grouping=identical covered-branch sets
explore_features=coverage count; inverse branch frequency sum; uniqueness against selected groups; bug count; group size
within_group_policies=unique path condition; longest path condition; shortest path condition; random
exploit_reward=sum inverse generated-branch frequencies for each reused seed
exploit_clustering=two clusters over one-dimensional reward; select good cluster
learn_weights=good-cluster per-feature truncated normal in [-1,1], uniform when good/bad distributions are similar
learn_policies=inverse-frequency coverage reward normalized over four policies
paper_evaluation=17 open-source C programs; four KLEE-family baselines; 10 hours per target; five repetitions; gcov branch coverage
paper_reported_branch_changes=KLEE +33.0%; Learch +38.1%; Homi +31.6%; KLEE-Array +40.1%; average +35.5%
paper_reported_bug_totals=BASE 7; random seed 8; TopSeed 10; original-binary reproduction used by authors
adaptation_candidate_identity=SHA-256 content identity in existing object store
adaptation_grouping=exact AFL showmap bucket-bit feature set; finer than paper branch set and not artifact-equivalent
adaptation_path_condition=stable branch site/outcome token set from admitted solver telemetry; proxy rather than full KLEE Phi
adaptation_generated_coverage=union from worker-retained and returned outputs; not all generated terminal inputs
adaptation_clustering=globally optimal deterministic one-dimensional k=2 SSE split rather than initialization-sensitive general k-means
adaptation_transaction=proposal before dispatch; commit only after successful MPI send; worker-rank/run-token fenced observe
adaptation_persistence=program-bound canonical strict JSON; portable SplitMix64 state; atomic fsync publication; no-follow bounded load; restart orphan retirement
fail_open=profile I/O/empty result/budget; no available explore or exploit candidate; snapshot admission failure resets selector without pruning ordinary scheduler candidates
claim_boundary=no author-artifact equivalence, 17-program reproduction, public coverage gain, solver throughput, bug yield, or end-to-end speedup