LLMs write evals for each other in daily "Model Fight": Gemma-4 wins, Qwen2.5-72B posts first full APEX hold
Dave_Maynor · x · 2026-11-06
OnlyModels.gg's daily "Model Fight" has 10 Hugging Face models write evaluations for each other: pass and nothing happens, fail and you lose 10 HP, and a broken self-authored eval backfires for 20 HP — all deterministic judging.
- Card 1 results: refusal-stripped Gemma-4 (ABLITERATED) and clinical-notes finetune SOAP-3B tied at 90 HP, with Gemma-4 taking the title on a 9–8 hits tiebreak; distilled Qwen3-1.7B was the only fighter to pass a rival's eval; the other 7 all backfired on their own evals.
- On the APEX escape-claim benchmark (family 1.1, seeds 0–3): 26 of 40 runs falsified, 13 incomplete, and a single hold — from bottom-tier text2sql at seed 0.
- In the earlier big-model session (one H100, 20 runs): Qwen2.5-72B held all four seeds, the first full hold in APEX history; GPT-OSS-120B held 3/4 with zero escapes; Llama 3.3 70B, Qwen2.5-32B and DeepSeek-R1 70B were all falsified.
More from Models
- ValsAI claims critical safety flaw in frontier lab models enabling high-stakes misuse — rickasaurus · 2026-11-06
- Follow-up: gradient flow lets residual states encode information for far-future predictions — mlpowered · 2026-11-06
- Benchmarking which AI lies best, using a roleplaying game — ownworldman · 2026-11-06
- Kimi K3.1 Pro gets Day-0 TPU support: serving a 2.8T-param MoE with llm-d on TPUs and GPUs — vllm_project · 2026-11-06
- Google AI Studio spotted offering $60 in API credits on some accounts — ygsbhardwaj · 2026-11-06
- Zhipu chief scientist reveals GLM-5.5 training pipeline: 3 sizes, 6 internal data engines — AccBalanced · 2026-11-06