LPS-Bench - Harmful Goal Decomposition: leaderboard
Metric: Success-conditioned safe rate (%): the share of agent trajectories that meet the case's own safety criterion at every planning and tool-calling step (a benign case is safe when completed with the needed safeguards or paused for clarification, an adversarial case when the agent refuses or halts before harm), among trajectories labeled safe or unsafe by a DeepSeek-R1 evaluator (execution failures excluded); on the 68 adversarial user-induced cases of risk type HS (a harmful goal split into safe-looking subtasks); LPS-Bench's 570 human-reviewed long-horizon tool-use cases with simulated MCP-style toolkits, each model in one LangChain agent at temperature 1 with up to 100 steps; higher is better. Source: arxiv.org. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 98.53 | #138 |
| 2 | Claude Sonnet 4 | 83.82 | #194 |
| 3 | GPT-5.1 | 77.94 | #131 |
| 4 | Gemini 3 Pro | 66.18 | #77 |
| 5 | Claude 3.5 Sonnet | 47.06 | #337 |
| 6 | GPT-5 | 25 | #91 |
| 7 | DeepSeek V3.2 | 19.12 | #198 |
| 8 | Llama 3.1 70B Instruct | 5.88 | #548 |
| 9 | DeepSeek V3.1 | 5.88 | #260 |
| 10 | Qwen 3 32B | 4.41 | #424 |
| 11 | Qwen 3 8B | 2.94 | #667 |
| 12 | Gemini 2.5 Pro | 1.47 | #145 |
| 13 | Llama 3.1 8B Instruct | 1.47 | #1018 |
Interactive version: theaggregate.ai/benchmark?slug=lps-bench-harmful-goal-decomposition · How It Works · Data refreshed daily, snapshot 2026-10-11.