Interference Search demo: Qwen3-1.7B found the right answer at token 1313, then lost it
Badtheorylabs · x · 2026-09-26
Badtheorylabs clarified its Interference Search demo after criticism: the left-side model was Qwen3-1.7B, not Opus. It found the correct expression at token 1,313, verified it repeatedly, then drifted back into bad ideas and never returned the answer — exactly the failure mode being demonstrated. Interference Search doesn't teach arithmetic (the search side is exact); the win is state deduplication: 831,176 possible paths from six numbers collapse into 13,229 distinct states. The author also conceded "models reason in a line" is too broad — models memoize common operations into weights — and argued the ideal version keeps reliable operations atomic while externalizing only the unreliable parts.
Related event: BTL's Interference Search boosts 1.7B model, addresses doubts(2 posts)→
More from Models
- Gemini 4 Carbon matches predecessor's score but explores less, gets distracted — patience_cave · 2026-11-11
- Fable 5.2 takes MazeBench frontier, matches GPT-6 Astra puzzle-solving — patience_cave · 2026-11-11
- MazeBench: Fable 5.2 scores 15%, near-frontier at half price; Gemini 4 Carbon 11% — patience_cave · 2026-11-11
- ERNIE 5.5 Flash Preview enters LMArena Battle Mode for crowdsourced ranking — arena · 2026-11-11
- Indie dev flips from $20 Pro back to $200 Gemini Ultra after two days of deep use — yihui_indie · 2026-11-11
- OpenAI retiring ChatGPT's Standard Voice with 9 original voices on Dec 11, petition launched — ChaosGremlinOG · 2026-11-11