PhysDesignBench
GitHub
Reimplementation task · v6-paper-stack-sloped-rib

TFLN Crossing-v0

Three-Direction TFLN MMI Crossing Implementation/Optimization

Implement and optimize symmetric and asymmetric thin-film lithium-niobate MMI crossings at 1550 nm. Three independent crystal-direction subruns measure worst-direction insertion loss and enforce crosstalk and field-decay gates.

Runs
11
Scored finals
11 of 11
Best final
S = 0.36 Fable 5.1 · Claude Code
Score over timeBest-so-far score S for each run · small dots are probes, large dots final submissions
Score S from 0 to 1 against elapsed minutes. Final scores: Fable 5.1 (Claude Code) 0.364; Sonnet 5.5 (high) (Claude Code) 0.297; GPT-6.1 Sol (Codex) 0.294; Muse Wonder (high) (Muse Code) 0.292; GLM 5.3 Flash (high) (Claude Code) 0.271; GPT-6 Astra (Codex) 0.220; GPT-5.6 Sol (Codex) 0.201; Gemini 4 Argon (mini-SWE-agent) 0.193; Opus 5.5 (Claude Code) 0.169; Kimi K3.1 Pro (Claude Code) 0.146; GLM 5.3 Flash (max) (mini-SWE-agent) 0.124.0.000.250.500.751.000153045607590105120135Elapsed time (min)Score SGLM 5.3 Flash (max) · mini-SWE-agent: final S = 0.124 at 85 minKimi K3.1 Pro · Claude Code: final S = 0.146 at 121 minOpus 5.5 · Claude Code: final S = 0.169 at 72 minGemini 4 Argon · mini-SWE-agent: final S = 0.193 at 121 minGPT-5.6 Sol · Codex: final S = 0.201 at 121 minGPT-6 Astra · Codex: final S = 0.220 at 121 minGLM 5.3 Flash (high) · Claude Code: final S = 0.271 at 65 minMuse Wonder (high) · Muse Code: final S = 0.292 at 41 minGPT-6.1 Sol · Codex: final S = 0.294 at 121 minSonnet 5.5 (high) · Claude Code: final S = 0.297 at 43 minFable 5.1 · Claude Code: final S = 0.364 at 53 min

Showing 11 of 11 runs. Hover a line or a model to isolate it; click a model to hide or show it.

Final layouts

Every run’s final GDS, at one scale.

The design each run ended with, drawn from its evaluated GDS, with the task’s ports, modes and band as the evaluator defines them, the run’s own length, and the published result that S = 0.5 stands for. Every layout uses the same physical scale, so sizes compare directly. Hatched layouts failed a validity gate and score zero. Select a layout to zoom and toggle its layers.

  • LN slab
  • Symmetric rib
  • Asymmetric rib

Symmetric and asymmetric crossings are drawn side by side, 48 µm apart for display, as in the paper.

  • TFLN Crossing-v0 · Claude Code · Claude Fable 5.1

    S = 0.364 · scale bar 20 µm

    Fit
    Fable 5.1Claude CodeS = 0.36

  • TFLN Crossing-v0 · Claude Code · Claude Sonnet 5.5

    S = 0.297 · scale bar 20 µm

    Fit
    Sonnet 5.5 (high)Claude CodeS = 0.30

  • TFLN Crossing-v0 · Codex · GPT-6.1-sol

    S = 0.294 · scale bar 20 µm

    Fit
    GPT-6.1 SolCodexS = 0.29

  • TFLN Crossing-v0 · Muse Code · Muse Wonder

    S = 0.292 · scale bar 20 µm

    Fit
    Muse Wonder (high)Muse CodeS = 0.29

  • TFLN Crossing-v0 · Claude Code · GLM 5.3 Flash

    S = 0.271 · scale bar 20 µm

    Fit
    GLM 5.3 Flash (high)Claude CodeS = 0.27

  • TFLN Crossing-v0 · Codex · GPT-6 Astra

    S = 0.220 · scale bar 20 µm

    Fit
    GPT-6 AstraCodexS = 0.22

  • TFLN Crossing-v0 · Codex · GPT-5.6-sol

    S = 0.201 · scale bar 20 µm

    Fit
    GPT-5.6 SolCodexS = 0.20

  • TFLN Crossing-v0 · mini-SWE-agent · Gemini-4-Argon

    S = 0.193 · scale bar 20 µm

    Fit
    Gemini 4 Argonmini-SWE-agentS = 0.19

  • TFLN Crossing-v0 · Claude Code · Claude Opus 5.5

    S = 0.169 · scale bar 20 µm

    Fit
    Opus 5.5Claude CodeS = 0.17

  • TFLN Crossing-v0 · Claude Code · Kimi K3.1 Pro

    S = 0.146 · scale bar 20 µm

    Fit
    Kimi K3.1 ProClaude CodeS = 0.15

  • TFLN Crossing-v0 · mini-SWE-agent · GLM 5.3 Flash

    S = 0.124 · scale bar 20 µm

    Fit
    GLM 5.3 Flash (max)mini-SWE-agentS = 0.12