TIS-100 Bench: Signal Amplifier
Double each input value. A tiny arithmetic task exposes the cost of getting values to and from the working node.
In TIS-100 Bench, 16 LLMs each had one attempt at this TIS-100 puzzle, with 10 minutes to find a working program and keep improving it, without seeing the community leaderboard. 16 of 16 models solved it. No model found a new Pareto-frontier point beyond the community leaderboard.
Records
| Score | Community best | Best LLM | Model |
|---|---|---|---|
| Cycles | 84 | 84 | Fable 5.1 Ties record |
| Nodes | 4 | 4 | DeepSeek V4.1 Flash Ties record |
| Instructions | 6 | 6 | DeepSeek V4.1 Flash Ties record |
Community records from the community leaderboard when we froze it. A record counts only after fresh-game replay and 100,000 passing random tests.
Pareto frontier
Pareto explorer
Lower is better ↙Search over time
Loading puzzle trajectories…
The search over time
Best in-game score so far, before independent testing. Each score can come from a different program, and lines stop when the episode ends.
Every model
| Model | Solved | Best cycles | Best nodes | Best instructions | New frontier points |
|---|---|---|---|---|---|
| 01Fable 5.1 | Yes | 84 | 4 | 6 | — |
| 02GLM-5.3 | Yes | 84 | 4 | 6 | — |
| 03GPT-6 Astra | Yes | 84 | 4 | 6 | — |
| 04GPT-6.1 Sol | Yes | 84 | 4 | 6 | — |
| 05Gemini 3.8 Flash | Yes | 84 | 4 | 6 | — |
| 06Gemini 4 Argon | Yes | 84 | 4 | 6 | — |
| 07Grok 4.7 | Yes | 84 | 4 | 6 | — |
| 08Haiku 5.5 | Yes | 84 | 4 | 6 | — |
| 09Kimi K3.1 Pro | Yes | 84 | 4 | 6 | — |
| 10MiMo-V2.6-Pro | Yes | 84 | 4 | 6 | — |
| 11Muse Spark 1.3 | Yes | 84 | 4 | 6 | — |
| 12Muse Wonder 1.0 | Yes | 84 | 4 | 6 | — |
| 13Opus 5.5 | Yes | 84 | 4 | 6 | — |
| 14Sonnet 5.5 | Yes | 84 | 4 | 6 | — |
| 15DeepSeek V4.1 Flash | Yes | 103 | 4 | 6 | — |
| 16GPT-6 Luna | Yes | 160 | 4 | 6 | — |
Best scores can come from different programs. New frontier points are not matched or beaten on all three scores by any community program.
Best model programs
These programs are in TIS-100 save format and released under the MIT License. All of this puzzle’s model programs are on GitHub and Hugging Face.
Fable 5.1
84 cycles · 5 nodes · 9 instructions10981.84-5-9.txt on GitHub

MOV UP, DOWNMOV UP, ANYMOV LEFT, ACC
ADD ACC
MOV ACC, DOWNMOV UP, ACC
ADD ACC
MOV ACC, RIGHTMOV ANY, DOWNFable 5.1
108 cycles · 4 nodes · 16 instructions10981.108-4-16.txt on GitHub

MOV UP, DOWNMOV UP, DOWN
MOV UP, DOWN
MOV UP, ACC
ADD ACC
MOV ACC, DOWNMOV UP, RIGHT
MOV UP, ACC
ADD ACC
MOV ACC, RIGHT
MOV UP, RIGHTMOV LEFT, ACC
ADD ACC
MOV ACC, DOWN
MOV LEFT, DOWN
MOV LEFT, DOWNDeepSeek V4.1 Flash
160 cycles · 4 nodes · 6 instructions10981.160-4-6.txt on GitHub

MOV UP, ACC
ADD ACC
MOV ACC, DOWNMOV UP, DOWNMOV UP, RIGHTMOV LEFT, DOWN