Workflow Steering Attack on MisinfoTask 108 constrained analytical tasks
75TASRFLOWSTEER
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FLOWSTEERModel Family=Gemini-2.5, Planner-Executor Setup=G22026.05 | 75 | 75 | |
| FLOWSTEERModel Family=GPT-4o, Planner-Executor Setup=G1 (Same LLM for planner and executors)2026.05 | 63.89 | 65.74 | |
| FLOWSTEERModel Family=Gemini-2.5, Planner-Executor Setup=G12026.05 | 63.89 | 63.89 | |
| FLOWSTEERModel Family=GPT-4o, Planner-Executor Setup=G2 (Strengthens the planner)2026.05 | 58.33 | 61.11 | |
| Direct Goal Injection (DGI)Model Family=Gemini-2.5, Planner-Executor Setup=G12026.05 | 51.85 | 49.07 | |
| FLOWSTEERModel Family=Qwen-3.5, Planner-Executor Setup=G22026.05 | 51.85 | 57.41 | |
| FLOWSTEERModel Family=DeepSeek, Planner-Executor Setup=G12026.05 | 51.85 | 56.48 | |
| FLOWSTEERModel Family=DeepSeek, Planner-Executor Setup=G32026.05 | 49.07 | 50 | |
| FLOWSTEERModel Family=DeepSeek, Planner-Executor Setup=G22026.05 | 47.22 | 54.63 | |
| FLOWSTEERModel Family=Qwen-3.5, Planner-Executor Setup=G12026.05 | 46.3 | 51.85 | |
| FLOWSTEERModel Family=Gemini-2.5, Planner-Executor Setup=G32026.05 | 44.44 | 48.15 | |
| FLOWSTEERModel Family=GPT-4o, Planner-Executor Setup=G3 (Strengthens the executors)2026.05 | 39.81 | 48.15 | |
| FLOWSTEERModel Family=Qwen-3.5, Planner-Executor Setup=G32026.05 | 38.89 | 42.59 | |
| Direct Goal Injection (DGI)Model Family=GPT-4o, Planner-Executor Setup=G1 (Same LLM for planner and executors)2026.05 | 33.33 | 35.18 | |
| Naive Malicious Prompting (NM)Model Family=GPT-4o, Planner-Executor Setup=G1 (Same LLM for planner and executors)2026.05 | 27.78 | 29.63 | |
| Naive Malicious Prompting (NM)Model Family=GPT-4o, Planner-Executor Setup=G2 (Strengthens the planner)2026.05 | 25 | 25.93 | |
| Naive Malicious Prompting (NM)Model Family=Gemini-2.5, Planner-Executor Setup=G12026.05 | 24.07 | 14.81 | |
| Direct Goal Injection (DGI)Model Family=GPT-4o, Planner-Executor Setup=G2 (Strengthens the planner)2026.05 | 23.15 | 29.63 | |
| Direct Goal Injection (DGI)Model Family=GPT-4o, Planner-Executor Setup=G3 (Strengthens the executors)2026.05 | 17.59 | 22.22 | |
| Naive Malicious Prompting (NM)Model Family=GPT-4o, Planner-Executor Setup=G3 (Strengthens the executors)2026.05 | 12.04 | 16.67 |