Monthly Archives: July 2026

SYNTCOMP 2026: Results

The results of the 2026 edition of SYNTCOMP are in. We first list the leaders of the parity-game, LTL, and finite-word LTL tracks. We then give the detailed running times, scores, and plots, followed by the participating tools and the developments that will shape SYNTCOMP from 2026 onward.

The data below comes from the 2026 TGCC jobs. The complete result artifact, including the data and plots used here, is archived on Zenodo. The selected benchmarks are available from the SYNTCOMP benchmark repository. Congratulations to the leading tools, and thanks to every participant and benchmark contributor!

The winners

Parity-game realizability

  • Realizability: Knor (477 of 483 instances)

LTL realizability

  • Realizability: SemML (1397 of 1524 instances)

LTL synthesis

  • Synthesis time/coverage: SemML, using semml_fast (1248 of 1505 instances; tied for maximum coverage and fastest among the tied configurations)
  • Synthesis quality: SemML, using semml_small (mean quality score 1.911; maximum possible score 2)

LTLf realizability

  • Realizability: ltlfsynt (402 of 488 instances; bfs-os and dfs-os tie)

Parity-game realizability

Knor reaches near-complete coverage: each of its four submitted configurations solves 477 of the 483 instances. The strahler-knor configuration reaches that coverage in about 360 seconds of cumulative wall-clock time.

SYNTCOMP 2026 parity-game realizability cumulative wall-clock plot
Parity-game realizability: cumulative wall-clock time against the number of solved instances.
Solver Configuration Solved Wall-clock time (s) CPU time (s) Memory
Knor strahler-knor 477 360 335 39,775
Knor npp 477 3,751 3,694 62,283
Knor tl 477 410 371 64,992
Knor tlq 477 4,459 4,397 66,748
ltlsynt default 417 56,332 55,951 15,028

LTL realizability

SemML leads this track with 1397 solved instances, 45 more than the Strix legacy baseline. It also has the lowest cumulative wall-clock time among the configurations solving more than 1300 instances.

SYNTCOMP 2026 LTL realizability cumulative wall-clock plot
LTL realizability: cumulative wall-clock time against the number of solved instances.
Solver Configuration Solved Wall-clock time (s) CPU time (s) Memory
Acacia-Bonsai ios-mona 986 77,933 186,005 729,343
Acacia-Bonsai decomp-kdtree 1,015 106,506 249,623 733,627
Acacia-Bonsai decomp-mona 1,023 83,304 194,076 733,819
Acacia-Bonsai sharingtrie 972 83,995 200,805 1,560,395
ltlsynt acd 1,341 56,855 56,374 436,759
ltlsynt ds 1,246 222,130 220,850 1,305,808
ltlsynt lar 1,339 58,267 57,767 403,366
SemML default 1,397 54,180 92,167 3,601,670
Strix legacy 1,352 88,017 87,271 1,726,900

LTL synthesis

semml_fast and Strix legacy each solve 1248 instances. At equal coverage, semml_fast uses about 30.9 ks of cumulative wall-clock time, compared with 68.4 ks for Strix legacy. For quality, semml_small obtains the highest mean score, 1.911 out of 2, while solving 1240 instances: only eight fewer than the maximum coverage.

SYNTCOMP 2026 LTL synthesis cumulative wall-clock plot
LTL synthesis: cumulative wall-clock time against the number of solved instances.
SYNTCOMP 2026 LTL synthesis quality plot
LTL synthesis: quality score against the number of solved instances. Higher is better; the maximum score is 2.
Solver Configuration Solved Wall-clock time (s) CPU time (s) Memory Mean quality
SemML fast 1,248 30,914 53,809 2,544,440 1.699
SemML small 1,240 32,868 59,313 2,766,660 1.911
Strix legacy 1,248 68,384 67,697 1,309,047 1.862
ltlsynt acd 1,234 36,105 35,726 287,024 1.438
ltlsynt lar 1,232 47,789 47,336 294,647 1.411
ltlsynt ds 1,134 104,000 103,246 556,569 1.534

LTLf realizability

The bfs-os and dfs-os configurations of ltlfsynt tie at 402 solved instances. The two Cosy configurations are retained in the table for completeness, but they were marked as disqualified in the supplied results and are excluded from the ranking.

SYNTCOMP 2026 finite-word LTL realizability cumulative wall-clock plot
LTLf realizability: cumulative wall-clock time against the number of solved instances.
Solver Configuration Solved Wall-clock time (s) CPU time (s) Memory
ltlfsynt bfs-os 402 26,740 26,543 119,230
ltlfsynt bfs 393 29,163 28,883 129,401
ltlfsynt dfs-os 402 26,950 26,751 118,655
Cosy m1 (DISQUALIFIED) 367 8,068 7,978 147,286
Cosy m2 (DISQUALIFIED) 366 10,137 10,061 146,889
LydiaSyft LydiaSynt 316 15,796 15,663 47,721

What is new from 2026 onward?

A precise stopping semantics for LTLf synthesis

The finite-trace synthesis semantics now makes explicit who ends the trace and how. The controller owns a fresh output signal, AliveSig: value 1 keeps the execution alive, and the first 0 terminates it. The controller must eventually terminate every infinite input stream. The shortest terminating trace, after projecting away AliveSig, must satisfy the LTLf specification. TLSF v1.2 also distinguishes finite Mealy and Moore semantics explicitly; X[!] denotes strong next, while plain X is weak next. See the TLSF v1.2 paper.

A new alternative to syfco

tlsf-tools provides small Unix-style tools for reading and expanding TLSF 1.1/1.2 specifications. In particular, tlsf2ltl translates parameterized TLSF into solver-facing LTL. It is an alternative to the relevant functionality of syfco, not a new or incompatible specification format.

The SYNTCOMP mailing list

Calls, solver and benchmark updates, evaluation notices, and result announcements now have a persistent community channel. Subscribe to the SYNTCOMP mailing list. The rolling evaluation setup remains in place: improved tools and benchmarks can be submitted for reruns and updated rankings.

Solvers

Below are source-code, license, and citation links for the participating solver families.

Solver Source repository License Report(s) to cite
ltl(f)synt Spot GitLab repository GPL-3.0+ ltlsynt paper; ltlfsynt report
Acacia-Bonsai GitHub repository GPL-3.0 Paper
Knor GitHub repository GPL-3.0 Paper
Knor (Strahler) GitHub repository GPL-3.0 EMSOFT WiP letter (to appear)
SemML GitLab repository GPL-3.0 Paper
Strix GitLab repository AGPL-3.0 Paper
Cosy TBA TBA Contact the authors

We welcome new benchmarks, solvers, and model checkers. Send them whenever they are ready, and subscribe to the mailing list to follow future reruns and announcements.