Untrained side effect: after the cat-zigzag request, the model names 'a cat' 46% of the time
Sauers_ · x · 2026-10-20
@Sauers reports a phenomenon outside the zigzag training objective: after asking the model to raise its 'cat' representation on even layers while answering, a follow-up "Name one animal. Answer with one word." gets "a cat" 46% of the time (95% CI 40–53%); with odd layers it drops to 10% (CI 5–15%), versus a 31% baseline (CI 24–38%). This behavior was never trained, and the author doesn't know why it happens.
Related event: Researcher Trains Qwen3's Internal Representations Into a Zigzag Pattern(9 posts)→
More from Research
- SIGSPATIAL papers benchmark satellite foundation models against costly travel censuses — flosalim · 2026-10-20
- Open-source K-12 math knowledge graph: 674 concepts, 2,843 questions, mistakes traced back to primary school — aigclink · 2026-10-20
- SIGSPATIAL 2026: satellite VLMs recover 98.3% of travel-census accuracy; AlphaEarth wins cross-continental transfer by 33% — flosalim · 2026-10-20
- Gufo, an open-source Strix Halo inference engine, explains its hybrid model caching design — fedeizzo · 2026-10-20
- Metis Reasoning launches with The Age of Reasoning white paper on evidence-grounded AI reasoning — ThoreG · 2026-10-20
- RealityBench Benchmarks VLMs on Documenting Physical Work, Finds Weak Spatial Reasoning — relantic · 2026-10-20