LLMs Play Zachtronics' TIS-100

133new frontier points

18 / 55puzzles alongside new frontier points

3outright sequence records

Each node is a small processor, and neighboring nodes have to delay for one another to toggle values. The equivalent scores a operating resolution on cycles, nodes, and instructions, and you commencement wondering whether one additional education could disappear. Lower is better, but improving one mark can create another worse, so all puzzle has a frontier of tradeoffs alternatively of one finest answer.

Undecimal Twin RounderSPEC44149538, among the eight tournament puzzles, rounds an input to the nearest multiple of N on one output and of 11−N on the other. When we froze the leaderboard, the finest community program took 2,694 cycles. Opus 5.5 passed its archetypal program following 6.5 minutes, at 5,669 cycles. It went under the document 16 minutes in and completed at 1,918, which is 29% faster. That speed disbursal one additional node and 28 additional instructions than the old record.

The center of the program is a two-speed search. Instead of subtracting N one stage at a time, nodes 5 and 6 subtract 4N until they overshoot, afterward add N rear until the value turns positive. Each drives a small coprocessor, node 2 or node 7, by sending it jump offsets through a port, so the coprocessor plant as a lookup array for 4N and N.

880episodes

95.5Moutput tokens

62,400game-tool calls

Multiplication: expend code to preserve cycles

Window sums: allocate the history

Indexed admission from stacks

We submitted selected programs to the leaderboard, and the community improved on them nearly immediately. On Undecimal Twin RounderSPEC44149538, @LongingForRest took Opus 5.5’s 1,918-cycle program and cut it to 1,744 by the next day. The document is now credited to both: LongingForRest/opus-5-5.

Neither got there alone. The models established new tradeoffs quickly, and players peruse those programs, worked out why they worked, and took them further.

Mathematics is seeing the identical shift. Terence Tao and colleagues used LLM-guided hunt throughout 67 problems and improved on the finest known results in several. In “After Math,” Silvia De Toffoli and Eamon Duede contend that a outcome lone becomes a resolution formerly mathematicians can comprehend it, and that AI fits improved as an aide than a competitor. TIS-100 is a small, measurable type of that future: a example finds a improved program, and group rotate it into item they understand.

Loading the icy benchmark snapshot…

Community leaderboardModelNew frontier point

Select a item alongside the mouse or keyboard to inspect its scores.

Suite. All 22 run puzzles, all 25 TIS-NET puzzles, and eight tournament puzzles from 2018: 19 uncomplicated (10 minutes), 17 average (30), and 19 difficult (45). Every example has one scored event per puzzle. Time limits differ by difficulty, so difficulty and hunt period can’t be separated.

Scoring. Each model–puzzle brace contributes one selected attempt, alongside no best-of-attempts scoring, so resolve rates are pass@1: a puzzle counts as solved whenever that one event saved a program that replays in a caller game. Unsolved puzzles remain in all average and mark zero.

New frontier points. The citation is the community leaderboard as of October 15, 2026. Ties don’t count. Models established 154 applicant points on 23 puzzles. Our figure pools and deduplicates them throughout models and keeps the 133 on 18 puzzles alongside fresh-game replay and 100,000 passing tests. Testing is powerful evidence of correctness, not proof.

Record proximity. The community document divided by the model’s finest score, averaged throughout the three scores and all puzzles, ×100. Each mark can arrive from a distinct program.

Limits. One event per brace compares the configurations as tested and doesn’t measure repeated-run variability. Spread between puzzles is assortment throughout tasks. Public puzzles may overlap alongside training data. Mechanic groups are our own labels.

Programs. Every example Pareto point’s program is community under the MIT License, in TIS-100 preserve format and named akin the community leaderboard’s records (<puzzle>.<cycles>-<nodes>-<instructions>.txt). They are mirrored as one document per program on GitHub and as a array on Hugging Face. Two points replay in the equivalent but neglect several random tests; they are included and never figure as new frontier points.

Resources. Costs are estimated from recorded October 2026 API catalog prices, including cache discounts. Tokens accumulate throughout responses, so repeated input counts again, and tokenizers differ. Game calls figure idiosyncratic dispatched tools, including errors, and exclude the automatic final submission.

@misc{wincon2026tis100bench, heading = {TIS-100 Bench: LLMs Push the Pareto Frontier in Programming Puzzle Games}, author = {{Wincon}}, twelvemonth = {2026}, duration = nov, url = {https://tis-100.wincon.dev/}, note = {Benchmark snapshot 2026-11-03-v5} }

Thanks to Zachtronics, the TIS-100 players, and the maintainers of the community leaderboard, TIS-100-CXX, and leaderboard bot. Their solutions and tools create this difference possible. Thanks additionally to @12345ieee, @LongingForRest, @Hersmunch, @easone, @18111398, and everyone on the Zachtronics Discord who discussed the submissions and kept improving them.