BTL claims 'Interference Search' architecture lifts 1.7B model from 3/30 to 23/30 on hard problems
CatAstro_Piyush · x · 2026-09-26
Badtheorylabs (BTL) claims a breakthrough: LLMs reason too linearly — one path at a time, scrambling and restarting when wrong. Their new architecture, Interference Search, makes reasoning work "like a quantum computer": many paths at once, converging paths merge, dead ends cancel out.
- On the same hard problems, a 1.7B model reasoning normally solved 3/30; Interference Search solved 23/30
- They observed the baseline model finding the right answer at token 1,313, re-checking it 9 more times, then running out of budget without answering; their method got there in 3 steps
- The authors argue parallel subagents are a poor approach; technical details promised soon
Caveat: no paper or third-party verification yet.
More from Models
- Frontier models have 'memoized' 2x2-digit multiplication, no intermediate steps needed — ctjlewis · 2026-09-26
- DL Weekly #474: GPT-6 Sol and Luna, AI Evals guide, and an RRSI paper — dl_weekly · 2026-09-26
- DeepSeek reportedly runs smaller-model inference on NVIDIA gaming GPUs, argues Teortaxes — teortaxesTex · 2026-09-26
- Price of intelligence collapsing 13x per year, says blogger citing Epoch AI data — LeviTurk · 2026-09-26
- OpenAI Showcases Proaction Using GPT-6 Astra to Build Fleet-Management Agents Faster — OpenAIDevs · 2026-09-26
- Opus 5.5 Tops RareBench at 52% as Grok 4.7 and GPT-6 Sol Show Regressions — danielmckinn0n · 2026-09-26