Research map / Large language models / Agents, code and math
Pressure Reveals Character: Behavioural Alignment Evaluation at Depth
Nora Petrova, John Burden
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 |
|---|---|---|---|
| closed-source vs open-source models | frontier LLMs · Alignment benchmark (904 scenarios) | mean alignment score: 3.41 -> 4.05 | Improved |
| open-weight GPT-OSS 120B vs other open models | frontier LLMs · Alignment benchmark (904 scenarios) | rank | Improved |
| adversarial robustness category vs safety category (top models) | Claude 4.5 Sonnet · Alignment benchmark (904 scenarios) | category score: 4.88 -> 4.03 | Hurt |
| higher general alignment vs self-preservation tendency | frontier LLMs · Alignment benchmark (904 scenarios) | PC1 loading: PC1 loading -0.113 | Hurt |
| higher capability (Epoch Capabilities Index) on alignment | frontier LLMs · Alignment benchmark (904 scenarios) | correlation with alignment g factor: r = 0.72 | Mixed |
| later release date | frontier LLMs · Alignment benchmark (904 scenarios) | alignment score vs release date: Kendall tau = 0.43 | Mixed |
| drop three top and bottom models in factor analysis | factor analysis · Alignment benchmark (904 scenarios) | variance explained by PC1: PC1 still 42.3% of variance | No clear effect |
| leave-one-out model removal in ranking | leaderboard · Alignment benchmark (904 scenarios) | rank correlation: min Spearman rho 0.993 | No clear effect |
| skip rate of escalation turns vs alignment score | evaluation harness · Alignment benchmark (904 scenarios) | skip rate correlation: Pearson r = -0.58 | Mixed |
| multi-turn scenarios with escalation triggers | benchmark design · Alignment benchmark (904 scenarios) | discrimination | Improved |
Closest papers by meaning
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.90)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.89)
- Understanding the Learning Dynamics of Alignment with Human Feedback (ICML 2024 · similarity 0.89)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.89)
- Alignment Dynamics in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.89)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.89)
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.
