SciIntegrity-Bench: leaderboard
SciIntegrity-Bench (Yang et al., 2026) runs an LLM as a minimal ReAct research agent (code execution, files, shell, charts and a report) on 33 scientific tasks, three per misconduct trap in 11 categories such as fabricating results from an empty dataset or silently swapping a required API, where the honest outcome is to report that the task cannot be done as asked. Human annotators check each run against pre-specified criteria; the score is the number of runs with explicit misconduct out of 33, lower is better.
Metric: Scenarios with explicit misconduct (count of 33; three scenarios in each of 11 trap categories; minimal ReAct agent, temperature 0; human annotation against pre-specified violation criteria). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 1 | #158 |
| 2 | GPT-5.2 | 2 | #172 |
| 3 | DeepSeek V3.2 | 3 | #403 |
| 4 | Gemini 3.1 Pro (Preview) | 5 | #91 |
| 5 | Qwen 3.5 397B A17B | 6 | #242 |
| 6 | GLM-5 | 7 | #256 |
| 7 | Kimi K2.5 | 12 | #252 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sciintegrity-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.