Research map / Large language models / AI and society
Control Tax: The Price of Keeping AI in Check
Mikhail Terekhov, Zhen Ning David Liu, Caglar Gulcehre, Samuel Albanie
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 |
|---|---|---|---|
| use a stronger, more expensive monitor (o3-mini instead of Gemini 2.0 Flash) | trusted monitoring with sequential auditing · GPT-4o attacker, 5% audit budget | safety: 0.70 -> 0.85 at 45x cost | Improved |
| give attacker knowledge of monitor identity or examples of detected backdoors (ablation) | attacker prompting · APPS backdooring | monitor AUC | No clear effect |
| use stronger attacker models (given backdoors filtered by public tests) | attacker · APPS backdooring | monitor AUC | No clear effect |
| use 1-100 suspiciousness scale instead of 1-10 (ablation) | monitor prompt · APPS backdooring | AUC | No clear effect |
| add compound score combining suspiciousness and correctness queries (ablation) | monitor · APPS backdooring | AUC: never loses more than 1% of AUC of S; sometimes gains a few percentage points | Mixed |
| switch from sequential to batch auditing | trusted monitoring · APPS backdooring | safety | Mixed |
| conceal the deployment horizon N from the attacker | auditing protocol · theory (SNAC) | red team win probability | Improved |
| front-load audit budget (non-increasing allocation) | sequential auditing · theory (SNAC) | safety | Improved |
Closest papers by meaning
- Attack Selection In Agentic AI Control Evaluations Meaningfully Decreases Safety (NeurIPS 2026 · similarity 0.85)
- Position: Require Frontier AI Labs To Release Small "Analog" Models (NeurIPS 2025 · similarity 0.85)
- AI Control: Improving Safety Despite Intentional Subversion (ICML 2024 · similarity 0.85)
- Combining Cost Constrained Runtime Monitors for AI Safety (NeurIPS 2025 · similarity 0.84)
- Cost-of-Pass: An Economic Framework for Evaluating Language Models (ICLR 2026 · similarity 0.83)
- PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents (NeurIPS 2026 · similarity 0.83)
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.
