Intent2Tx (Retrieval, Multi-Step): leaderboard
Metric: Final weighted score (from 0-1, times 100): per-instance 0.1 x JSON format validity + 0.3 x logic score (correct contract address, name and function) + 0.2 x parameter score + 0.4 x pass@1 (exact end-to-end match), averaged over instances, generating an Ethereum transaction (contract, address, function, parameters, ETH value) as JSON from a natural-language intent and an address book, on Intent2Tx instances derived from 300 days of real Ethereum mainnet traces, multi-step intents, with the three most similar instances retrieved as demonstrations; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro | 59 |
| 2 | DeepSeek V3.2 | 59 |
| 3 | Gemini 2.5 Flash | 58 |
| 4 | Claude Opus 4.5 | 58 |
| 5 | Claude Haiku 4.5 | 57 |
| 6 | GLM-4.6 | 55 |
| 7 | Kimi K2 0905 | 55 |
| 8 | GPT-5.2 Codex | 55 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 54 |
| 10 | codestral-2508 | 52 |
| 11 | Llama 3.3 70B Instruct | 51 |
| 12 | GPT-5.2 | 51 |
| 13 | GLM-4 32B (0414) | 50 |
| 14 | Qwen3 Coder Next | 49 |
| 15 | Llama 3.1 8B Instruct | 41 |
Interactive version: theaggregate.ai/benchmark?slug=intent2tx-retrieval-multi-step · How It Works · Data refreshed daily, snapshot 2026-10-07.