Unassigned · 2026-10-08
Unrolled Flow Models for Reasoning
Faissal Izermine, Hanru Bai, Oscar Davis, T. Konstantin Rusch
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study when increasing flow integration steps helps reasoning. They prove a two-layer Transformer-defined flow can solve directed graph reachability, with needed steps tied to distance. Standard flow training supervises each time point independently and yields little gains from more steps; they propose training through an unrolled latent rollout with only a terminal loss. On ProsQA this reaches 97% accuracy and improves with more steps; for longer Sudoku/Maze rollouts, sphere retraction stabilizes dynamics and improves scaling.
Why it matters
Reasoning with adaptive test-time compute requires that later flow steps build on earlier ones. The paper pinpoints why standard flow objectives may not yield this property and provides a training approach (terminal-loss rollout training), plus a stabilization method (sphere retraction) for long rollouts. The results show practical benefits of deeper step usage on reasoning benchmarks and give a theoretical basis via a constructive reachability flow.
Method
- Propose Unrolled Flow Model (UFM): evolve a continuous latent answer through multiple Euler steps and decode only at the endpoint, training with a terminal loss so gradients train successive updates jointly.
- Train through an unrolled latent rollout over a randomly sampled subinterval of [0,1] (no intermediate vocabulary losses) and use truncated backpropagation for long rollouts.
- For long reasoning (Sudoku/Maze), stabilize dynamics by applying position-wise sphere retraction to control latent-state growth.
Limitation
The paper reports that selecting among multiple rollouts is less reliable for longer answers: “reliable selection remains challenging for longer answers” and “Selection is less effective on Maze-Hard.”
Abstract (from arXiv)
Flow matching enables language generation in few steps, but whether additional integration steps improve reasoning remains unclear. We prove that a flow parameterized by a two-layer Transformer can solve graph reachability, with the required number of integration steps increasing with the target's distance from the root. Yet, standard flow language models can fail to benefit from additional steps on reasoning tasks. We attribute this limitation to objectives that supervise each time point independently, without explicitly training successive steps to build on one another. To address this, we instead train through the model's own latent rollout over a randomly sampled subinterval of [0, 1], decoding only at the endpoint. On ProsQA, this raises accuracy to 97% and enables performance to improve with additional integration steps. For the longer rollouts required by reasoning tasks such as Sudoku and Maze, retracting the latent state onto a sphere stabilizes the dynamics and yields substantial gains over baselines with more than three times as many parameters. Sampling multiple rollouts further improves performance when paired with a parameter-free selection score, although reliable selection remains challenging for longer answers. Together, these results establish a theoretical basis for reasoning with flows and show how rollout training, stable latent dynamics, and rollout selection help realize this capacity in practice.
Related papers
- Parallelizable Probabilistic Latent Reasoning for Diverse Reasoning Trajectories
- Looped Reasoning Finishes Earlier Than You Think
- $\Psi$-0.6: Adaptive Streams of Thought in a Unified Vision-Language World Model
- Reasoning as a Distribution: Generative Plan Latents for Vision-Language-Action Models
- Recompose and Refine Latent Reasoning Flows for Vision-Language-Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-08 UTC
Unrolled Flow Models for Reasoning
Preprint posted: 2026-10-07 · Original record ↗
- arxiv
- 2610.09759
Explicit source versions (1)
- v1 · 2026-10-07
- Preprint available
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Flow matching enables language generation in few steps, but whether additional integration steps improve reasoning remains unclear. We prove that a flow parameterized by a two-layer Transformer can solve graph reachability, with the required number of integration steps increasing with the target's distance from the root. Yet, standard flow language models can fail to benefit from additional steps on reasoning tasks. We attribute this limitation to objectives that supervise each time point independently, without explicitly training successive steps to build on one another. To address this, we instead train through the model's own latent rollout over a randomly sampled subinterval of [0, 1], decoding only at the endpoint. On ProsQA, this raises accuracy to 97% and enables performance to improve with additional integration steps. For the longer rollouts required by reasoning tasks such as Sudoku and Maze, retracting the latent state onto a sphere stabilizes the dynamics and yields substantial gains over baselines with more than three times as many parameters. Sampling multiple rollouts further improves performance when paired with a parameter-free selection score, although reliable selection remains challenging for longer answers. Together, these results establish a theoretical basis for reasoning with flows and show how rollout training, stable latent dynamics, and rollout selection help realize this capacity in practice.