Reverse-engineering swarms: recover code from masked outputs for interpretable models
cephaloform · x · 2026-09-26
The author describes a twist on generation-based swarms: instead of generating and predicting, the swarm must produce a program that reproduces full outputs from partially masked ones. Once trained, the swarm acts as a masked model of any 8-bit data-generating process, yielding roughly interpretable symbolic models since it outputs code.
More from Research
- Inworld eval lead details how to evaluate free-form voice steering in TTS models — rdesh26 · 2026-09-26
- What Is Old Is New Again: Jev and Looped Transformers Revive Classic AI Ideas — bendee983 · 2026-09-26
- XTC Sampling Gets Formal Paper Treatment, Already Shipping in llama.cpp and ExLlamaV2 — ziv_ravid · 2026-09-26
- XTC Sampling Is Already Built Into llama.cpp and Other Open-Source Inference Engines — ziv_ravid · 2026-09-26
- XTC Sampling Paper Accepted to NeurIPS 2026, Targets Creative Blandness — ziv_ravid · 2026-09-26
- New paper: Provably Complete Generalized Planning with LLMs — JFPuget · 2026-09-26