---
title: Interference Search
type: vocabulary
url: "https://www.envisioning.com/vocab/interference-search"
summary: A search architecture where many explicit reasoning states expand at once, merge when they coincide, and get pruned by a learned judge, instead of one linear transcript.
year: 2026
generality: 0.30
---

# Interference Search

A search architecture where many explicit reasoning states expand at once, merge when they coincide, and get pruned by a learned judge, instead of one linear transcript.
Interference Search is a search architecture that replaces a language model's single linear reasoning transcript with many explicit states expanded at once. An independent researcher working under the name Bad Theory Labs, based in Lagos, released it in September 2026 as open-source code, a paper draft, and a claims-to-evidence log on GitHub. There has been no peer review yet, and the numbers below are self-reported by the author on small, self-run benchmarks. Every live branch expands together rather than one token at a time. Branches reaching the same state merge into one. A small trained judge cancels branches that look like dead ends, and the survivors advance to the next level as a group. The author compares the design to a quantum computer exploring many paths before wrong ones cancel out, though the mechanism is ordinary frontier search plus a learned judge.

On 30 hard four-number "Countdown" arithmetic puzzles, Qwen3-1.7B reasoning in text solved 3 of 30. Interference Search, using the same small trained judge and budget, solved all 30, averaging 3 sequential steps against roughly 24 for the linear approach at a matched larger budget. The author reports answering a puzzle in about 5 milliseconds on a laptop CPU against about 21 seconds for the model reasoning in text, while noting this compares two different systems: the search uses a 100,000-parameter judge and never runs the language model at all. On 30 MBPP coding problems the model first got wrong, Interference Search solved 9, against 7 to 8 for linear strategies such as best-of-N sampling, a margin the author calls within noise.

The repository also records negative results. Telling the model which attempts had failed made it repeat them more, and resampling after a repeat did not help. Letting parallel streams attend to each other without merging or a judge added nothing over independent streams. The stated next step is training the architecture into a model with reinforcement learning and testing it on agent benchmarks such as SWE-bench and Terminal-Bench.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/interference-search)