Qwen2.5-1.5B Steering Sweep Shows No Degradation Threshold

Nearby_Indication474 · reddit · 2026-07-15

This Reddit post documents a steering sweep test on Qwen2.5-1.5B, comparing the results with previous tests on TinyLlama.

• Under identical protocols, FUD questions, fixed parameters, and greedy decoding, Qwen's diversity curve plateaus early and remains stable, never dropping below the vanilla baseline within the tested range.
• The author attributes this to kernel logs: TinyLlama's cos(theta) turns negative in certain layers, acting as a "brake," whereas Qwen's cos(theta) stays positive across all 20 layers, functioning like an "accelerator."
• Conclusion: For this specific set of questions and architecture, TinyLlama exhibits a "braking" degradation threshold, while Qwen shows no such threshold within the tested parameters.

Original post →

More from Models

• Grok 4.5 is now free inside Cursor, the popular AI coding IDE — mark_k · 2026-07-21
• GPT often converges on the same near-miss ideas in math problems — yacineMTB · 2026-07-21
• Eno Reyes says model distillation is basically unstoppable — LangChain · 2026-07-21
• Sakana says multiple diffusion models plus MCTS beat test-time scaling on coding and math — SakanaAILabs · 2026-07-21
• OpenAI hackathon project stalls as Codex struggles on voice, while Claude spots the issue — ColleenMBrady · 2026-07-21
• Kimi K3 lands exactly on China’s 2-year AI capability trend line — peterwildeford · 2026-07-21

Models channel →