WeightBench: leaderboard
Models write programs that inspect or change neural-network weights. Scores cover several weight-space programming tasks under the published sample counts. The shorter GPT repair subset stays separate from the full task population.
Source: arxiv.org.
Benchmark results
Results are published separately for each benchmark or evaluation setting.
- WeightBench - Fix Non-Loop 9-Task Subset
- WeightBench - Macro SWB (38 tasks, no TR-02 loops)
- WeightBench - Macro SWB (42 tasks)
- WeightBench - TR-01 Read: Static Diagnostics (10)
- WeightBench - TR-02 Fix: Surgical Patching (13)
- WeightBench - TR-03 Meta: Infeasibility Verification (5)
- WeightBench - TR-04 Write: Zero-Shot Weight Synthesis (6)
- WeightBench - TR-05 Systems: Hardware-Aware Optimization (8)
Interactive version: theaggregate.ai/benchmark?slug=weightbench · How It Works · Data refreshed daily, snapshot 2026-11-17.