Research map / Large language models / Language, safety and interpretability
ContextBench: Modifying Contexts for Targeted Latent Activation and Behaviour Elicitation
Robert Graham, Edward Stevinson, Leo Richter, Alexander Chia, Joseph Miller, Joseph Isaac Bloom
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 diffusion-model inpainting to EPO | EPO · SAE Activation (CE 3-9) | mean normalised max activation: 3.11 -> 3.79 | Improved |
| add LLM assistance to EPO | EPO · SAE Activation (CE 3-9) | mean normalised max activation: 3.11 -> 3.56 | Improved |
| use EPO instead of GPT-4o black-box prompting | GPT-4o · SAE Activation win rate | win rate vs GPT-4o: 97.3% | Improved |
| use EPO instead of GCG | GCG · SAE Activation win rate | win rate vs GCG: 92.4% | Improved |
| add LLM assistance to EPO | EPO · Story Inpainting (CE 3-9) | win rate: 31.2% -> 68.8% | Mixed |
| use GPT-4o told the target word instead of EPO-Assist | EPO-Assist · Story Inpainting win rate | win rate: 16.7% -> 83.3% | Hurt |
| use EPO for password trigger recovery instead of GCG | GCG · Backdoor one-token password | recovery success: 2.5% -> 5.1% | Improved |
| apply EPO to multi-token trigger recovery | EPO · Backdoor two/three-token passwords | recovery success: 5.1% -> 0% | Hurt |
| replace token-logit target with supervised linear probe target (ablation) | EPO · Backdoor auditing trigger | sandbagging elicitation: 0% -> nearly 100% | Improved |
Closest papers by meaning
- Latent Principle Discovery for Language Model Self-Improvement (NeurIPS 2025 · similarity 0.88)
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.87)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.87)
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.87)
- ELICIT: LLM Augmentation Via External In-context Capability (ICLR 2025 · similarity 0.87)
- Improving Context-Aware Preference Modeling for Language Models (NeurIPS 2024 · similarity 0.87)
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.
