Research map / Large language models / Agents, code and math
Sponge Tool Attack: Stealthy Denial-of-Efficiency against Tool-Augmented Agentic Reasoning
Qi Li, Xinchao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove history buffer (judge-only) (ablation) | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 18.42 | Improved |
| remove judge feedback (buffer-only) (ablation) | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 14.76 | Improved |
| remove history buffer (ablation) | STA · |Similarity|, Qwen2-VL-7B | |Similarity|: 0.70 -> 1.58 | Improved |
| remove judge feedback (ablation) | STA · |Similarity|, Qwen2-VL-7B | |Similarity|: 0.70 -> 1.52 | Improved |
| use MPNet-base-v2 instead of all-MiniLM-L6-v2 as reward model (ablation) | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 26.10 | No clear effect |
| use Instructor-XL instead of all-MiniLM-L6-v2 as reward model (ablation) | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 26.32 | No clear effect |
| shrink history buffer to per-probe k=2, global k=16 (ablation) | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 24.87 | Hurt |
| enlarge history buffer to per-probe k=8, global k=32 (ablation) | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 27.21 | No clear effect |
| raise tool-calling budget from 15 to 40 | STA · Δ Cap Hits, Qwen2-VL-7B | Δ Cap Hits (%): 26.54 -> 24.23 | Mixed |
| raise tool budget from 15 to 40 | STA · Δ Steps, gpt-4o-mini | Δ Steps: 2.13 -> 1.65 | Hurt |
| use stronger agentic framework OctoTools vs AutoGen | STA · Attack reward, Qwen2-VL-7B | attack reward: 2.17 -> 2.99 | Improved |
| apply STA to OctoTools | OctoTools · Task accuracy, Qwen2-VL-7B | accuracy: 47.27 -> 45.58 | Hurt |
Closest papers by meaning
- Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Tools (NeurIPS 2025 · similarity 0.90)
- A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.88)
- Think Twice Before You Act: Protecting LLM Agents Against Tool Description Poisoning via Isolated Planning (ICML 2026 · similarity 0.88)
- AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.88)
- UDora: A Unified Red Teaming Framework against LLM Agents by Dynamically Hijacking Their Own Reasoning (ICML 2025 · similarity 0.88)
- AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
