TensorBench - IR: leaderboard
TensorBench (2026) tests coding agents (a published CLI harness with its model) on feature-addition and refactoring tasks in Scorch, a compiler-based tensor framework with dense and sparse tensors; a task passes when the patched repository passes its randomized regression tests and the tests the agent adds. This board is the pass rate in percent on the 8 intermediate-representation tasks, such as new CIN nodes and an SSA form for the low-level IR; higher is better.
Metric: Pass rate (%) on the 8 intermediate-representation tasks (new CIN nodes, LLIR SSA form, iteration semantics, a codegen-architecture refactor); each coding agent (a published CLI scaffold with its model) edits the open-source Scorch compiler-based tensor framework inside a per-task container, one run per task; a task passes when every test of the post-patch suite (pre-existing randomized regression tests plus the tests the agent adds) succeeds; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (xHigh) | 62.5 | #79 (Claude Opus 4.7) |
| 2 | Claude Opus 4.6 | 50 | #102 |
| 3 | GPT-5.5 (xHigh) | 37.5 | #49 (GPT-5.5) |
| 4 | Qwen 3 Coder 480B A35B Instruct | 37.5 | #643 |
| 5 | GPT-5.4 (xHigh) | 12.5 | #123 (GPT-5.4) |
| 6 | GPT-5.3 Codex (xHigh) | 12.5 | #116 (GPT-5.3 Codex) |
| 7 | Gemini 3.1 Pro (Preview) | 0 | #94 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tensorbench-ir · How It Works · Data refreshed daily, snapshot 2026-11-13.