Flow Models as Iterative Reasoners: Recurrence Alone Lets Models Get Confidently Wrong
burkov · x · 2026-09-05
A paper from Georgia Tech, IBM, and MIT tackles reasoning problems (Sudoku, pathfinding, logic puzzles) that require mutually consistent decisions. It turns a flow model into an iterative reasoner by feeding each predicted answer back for later passes to revise. Key finding: recurrence alone is unreliable—after several iterations the model can lock onto a wrong answer with growing confidence, because training exposes it to cleaner intermediate states than those it generates at inference time.
Related event: Flow Reasoning Models Turn Flows Into Efficient Recurrent Reasoners(6 posts)→
More from Research
- One L-system rule grows a tree, fern, lightning, river and lungs in a 45s code film — prasenx · 2026-10-21
- MOSS-Image's text encoder is multimodal and accepts image inputs — RandumbRedditor1000 · 2026-10-21
- Skip the open access fees: NIH-funded papers must go on PMC with no embargo — anshulkundaje · 2026-10-21
- Creed: LLM code auditing infers developer beliefs to find semantic security flaws — mboehme_ · 2026-10-21
- 5-person team uses Claude and 300+ agents to design 318,484 proteins in 64 hours — 141_1337 · 2026-10-21
- ProVADA: one AI framework deimmunizes CD19 CAR binder and rescues cytosolic protease — BrianHie · 2026-10-21