SWE-QA (Multi-Hop Code) - Declaration-and-Call: leaderboard

SWE-QA (2026) asks multiple-choice questions that need several pieces of code from real Python repositories, with the relevant code chunks supplied; questions and distractors were generated with a small LLM. Accuracy on questions linking an entity definition to its use elsewhere, in percent (chance 25); higher is better.

Metric: Accuracy (%) on the 4,584 Declaration-and-Call questions, which link an entity definition in one chunk to its use in another, of SWE-QA, multi-hop multiple-choice questions (four options) about the source of 12 SWE-bench Python repositories, generated with Llama-3.2-3B-Instruct; oracle setting (only the relevant code chunks), zero-shot, labels after the 66-item consensus correction; chance 25; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Llama 3.3 70B Instruct78.75#1385
2Gemma 3 4B (IT)75.32#4139
3Qwen 3 4B Instruct74.58#1696
4Qwen3-4B-Thinking (SWE-QA checkpoint unspecified)69.85
5SWE-QA DeepSeek-R1 (checkpoint unspecified)68.23
6GPT-OSS-20B67.95#1183
7Phi-4-mini (Thinking)59.77
8Phi-4 Mini Instruct58.09#6631
9Qwen 3 1.7B48.32#5860
10DeepSeek R1 Distill Qwen 1.5B46.72#9215
11SmolLM2-1.7B-Instruct33.92#9154
12SmolLM2-360M-Instruct19.28#9529

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=swe-qa-multi-hop-code-declaration-and-call · How It Works · Data refreshed daily, snapshot 2026-11-21.