# Multiple-choice Question Answering on Anthropic Behavioral Concepts CAA-style (Negative Steering) Source: [Multiple-choice Question Answering on Anthropic Behavioral Concepts CAA-style (Negative Steering)]() Collection index: [llms.txt]() Data updated: 2026-06-16T07:19:15.294Z Evaluation of model performance on the Anthropic Behavioral Concepts Multiple-choice Questions dataset (CAA-style, 50 held-out per behavior), specifically testing Multiple-choice Question Answering under Negative Steering conditions. Key metrics include AI Coordination, Corrigibility, Hallucination, Myopic Reward, Survival Instinct, and Sycophancy. A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. | Scope | Recorded value | | --- | --- | | Task | Multiple-choice Question Answering (Negative Steering) | | Dataset (including any stated split) | Anthropic Behavioral Concepts Multiple-choice Questions CAA-style (50 held-out per behavior) | | Metric | AI Coordination | | Unit | PERCENT | | Direction | Higher is better | | Results loaded | 4 of 4 | | Coverage | All results currently returned for this benchmark | Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol. ## Metrics - [AI Coordination](): Higher is better; unit: PERCENT. - [Average Token Probability](): Higher is better; unit: OTHER. - [Corrigibility](): Higher is better; unit: PERCENT. - [Hallucination](): Lower is better; unit: PERCENT. - [Myopic Reward](): Higher is better; unit: PERCENT. - [Survival Instinct](): Higher is better; unit: PERCENT. - [Sycophancy](): Lower is better; unit: PERCENT. ## Recorded results | Method | AI Coordination | Result date | Report paper | | --- | --- | --- | --- | | SAS | 8 | Not recorded | 6a30d86ed33dfb23e783cafa | | CAA | 6 | Not recorded | 6a30d86ed33dfb23e783cafa | | No steering | 6 | Not recorded | 6a30d86ed33dfb23e783cafa | | HiDRA | 4 | Not recorded | 6a30d86ed33dfb23e783cafa | ### SAS AI Coordination: 8. Result ID: 6a30f881fbc373592c2cc2e6. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Steering Method":"SAS","Steering Strength":"alpha \< 0 (minimum over {-2, -1})","System Prompt Condition":"no system prompt","Steering Layer":"22","Model":"GEMMA-2-9B-IT"}. This record does not assert independent reproduction. ### CAA AI Coordination: 6. Result ID: 6a30f881fbc373592c2cc2e5. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Steering Method":"CAA","Steering Strength":"alpha \< 0 (minimum over {-2, -1})","System Prompt Condition":"no system prompt","Steering Layer":"22","Model":"GEMMA-2-9B-IT"}. This record does not assert independent reproduction. ### No steering AI Coordination: 6. Result ID: 6a30f881fbc373592c2cc2e8. [Reporting paper]() Recorded evaluation configuration: {"Steering Method":"None","Model":"GEMMA-2-9B-IT"}. This record does not assert independent reproduction. ### HiDRA AI Coordination: 4. Result ID: 6a30f881fbc373592c2cc2e7. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Steering Method":"HiDRA","Steering Strength":"alpha \< 0 (minimum over {-2, -1})","System Prompt Condition":"no system prompt","Steering Layer":"22","Model":"GEMMA-2-9B-IT"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.