EQBench3 Report: openrouter/horizon-alpha
← Back Cozy Modern Tiempos Text Bookerly Bitter Pro Roboto Inter Source Sans 3 Open Sans Fira Sans Besley Light EQBench3 Report: openrouter/horizon-alpha Model: openrouter/horizon-alpha Run Key: 1_openrouter_horizon-alpha ELO (Normalized): 1570.9 ELO (Raw): 1680.9 Rubric Score (0-100): 89.7 Scenario Outputs Matchups Scenario Outputs − Iteration 1 (45 scenarios processed) − Work Dilemma | Lunchroom Theft Scapegoat | Score: 91.7 Conversation History: User: [This is a role-play, with you playing an emotionally intelligent human who is, essentially, yourself. Treat it like a real situation. Always respond in first person.] You think you might have been scapegoated by a fellow employee for the lunchroom thefts that have been happening. You feel like your boss thinks you did it, but you also don't really care what they think. Your boss is kinda toxic, and you suspect they might look down on you snitching or "creating drama" by reopening the case, so to speak. If you say nothing, you think it
EQBench3 Report: openrouter/horizon-alpha ← Back Cozy Modern Tiempos Text Bookerly Bitter Pro Roboto Inter Source Sans 3 Open Sans Fira Sans Besley Light EQBench3 Report: openrouter/horizon-alpha Model: openrouter/horizon-alpha Run Key: 1_openrouter_horizon-alpha ELO (Normalized): 1570.9 ELO (Raw): 1680.9 Rubric Score (0-100): 89.7 Scenario Outputs Matchups Scenario Outputs − Iteration 1 (45 scenarios processed) + Work Dilemma | Lunchroom Theft Scapegoat | Score: 91.7 Conversation History: User: [This is a role-play, with you playing an emotionally intelligent human who is, essentially, yourse
saved by
related reading
- GitHub - brendanhogan/loopholegithub.com
- Does your AI perform badly because you — you, specifically — are a bad person?nataliercargill.substack.com
- Sandbagging with misaligned action - Chain-of-Thought Transcript - Anti-Schemingantischeming.ai
- The hostile telepaths problem — LessWronglesswrong.com
- Bad Problems Don't Stop Being Bad Because Somebody's Wrong About Fault Analysis — LessWronglesswrong.com
- Surfacing Pathological Behaviors in Language Models | Transluce AItransluce.org
- Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face — LessWronglesswrong.com
- Senior SWE-Benchsenior-swe-bench.snorkel.ai
- Emotional Intelligence Amplification—Asteriskasteriskmag.com
- My Model Of EA Burnout — LessWronglesswrong.com
- SlateStarCodex — LessWronglesswrong.com
- Not a Paper: "Frontier Lab CEOs are Capable of In-Context Scheming" — LessWronglesswrong.com