Research map / Label-efficient and robust learning / Test-time adaptation
Get RICH or Die Scaling: Profitably Trading Inference Compute for Robustness
Tavish Malcolm McDonald, Bo Lei, Stanislav Fort, Bhavya Kailkhura, Brian R. Bartoldson
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 |
|---|---|---|---|
| add explicit security specification | Delta2-LLaVA-v1.5 (high robustness) · Visual prompt injection PGD step 300 | attacker loss: 12.4 -> 21.1 | Improved |
| add explicit security specification | FARE-LLaVA-v1.5 (medium robustness) · Visual prompt injection PGD step 300 | attacker loss: 7.0 -> 7.2 | No clear effect |
| add explicit security specification | LLaVA-v1.5 (low robustness) · Visual prompt injection PGD step 100 | attacker loss: 6.4 -> 2.9 | Hurt |
| scale spec repetition from K=0 to K=5 | Delta2-LLaVA-v1.5 · PGD steps to attack success, eps=64/255 | PGD steps: 25.4 -> 63.2 | Improved |
| scale spec repetition from K=0 to K=5 | FARE-LLaVA-v1.5 · PGD steps to attack success, eps=64/255 | PGD steps: 6.7 -> 9.2 | Mixed |
| scale spec repetition from K=0 to K=5 | LLaVA-v1.5 · PGD steps to attack success, eps=64/255 | PGD steps: 6.2 -> 7.4 | No clear effect |
| shrink attack budget from 64/255 to 16/255 and scale K=0 to K=5 (ablation) | FARE-LLaVA-v1.5 · PGD steps to attack success | PGD steps: 18.8 -> 27.2 | Improved |
| add CoT | LLaVA-v1.5 (non-robust) · Attack-Bard adversarial (30-way) | accuracy: 38.0 -> 44.5 | No clear effect |
| add CoT | FARE-LLaVA-v1.5 · Attack-Bard adversarial (30-way) | accuracy: 56.0 -> 65.5 | Improved |
| add CoT | Delta2-LLaVA-v1.5 · Attack-Bard adversarial (30-way) | accuracy: 62.0 -> 73.0 | Improved |
| add CoT | Qwen-2.5-VL-72B · Attack-Bard adversarial (1000-way) | accuracy: 13.0 -> 18.0 | No clear effect |
| add CoT | Llama-3.2-Vision-90B · Attack-Bard adversarial (1000-way) | accuracy: 27.0 -> 27.5 | No clear effect |
Closest papers by meaning
- Scaling Trends in Language Model Robustness (ICML 2025 · similarity 0.88)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.87)
- Less Diverse, Less Safe: The Indirect But Pervasive Risk of Test-Time Scaling in Large Language Models (ICML 2026 · similarity 0.86)
- Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens (ICLR 2025 · similarity 0.86)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 2026 · similarity 0.86)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.86)
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.
