Driving · 2026-10-21
Reasoning as a Distribution: Generative Plan Latents for Vision-Language-Action Models
Matteo Cercola, Katie Z Luo, Milan Ganai, John Alora, Simone Formentin, Marco Pavone
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 introduce Generative Plan Latent (GPL), a low-dimensional conditional distribution over task-level plans for Vision-Language-Action models. They argue that representing the reasoning state as addressable plans (sample, steer, and reweight by outcomes) preserves valid alternatives that vector-only representations collapse. Without retraining the VLA, steering increases commanded-mode execution from 0.48 to 0.94 in SimplerEnv; action-level guidance shows no reliable gain.
Why it matters
GPL targets a reasoning “decision bottleneck” where upstream models often commit to one plan per observation, hiding alternative valid decisions. The paper’s notion of addressability (coverage/steering/outcome adaptation) and the “redundancy principle” suggest when explicit plans should help and when they are redundant with what the observation already determines.
Method
- Insert a generative variable upstream of an action expert: a rectified-flow head models a conditional plan distribution pθ(z|c) using frozen VLA conditioning c=fϕ(o,l), then projects sampled plan tokens into the executor prefix.
- Enable steering by biasing the flow using a noise-conditioned classifier qψ(y|zt,t,c) with guidance strength w; changing commands does not require retraining the VLA or plan generator.
- Enable outcome adaptation by reweighting the plan prior toward higher-scoring candidate plans using deployment feedback via either a learned critic s(c,z)=gω(z) or a reward table over discrete modes; the plan generator and VLA remain fixed.
Limitation
The authors state that plan benefits are conditional: an explicit plan can influence behavior only when it carries decision information not already determined by the observation (redundancy principle).
Abstract (from arXiv)
Vision-Language-Action (VLA) policies increasingly reason before acting, but existing reasoning models compress this intermediate decision into a single representation. When several plans are valid, this commitment hides alternatives that could otherwise be selected or revised. We introduce the Generative Plan Latent (GPL), which represents the reasoning state as a low-dimensional conditional distribution over plans, such as which object, grasp, or route to use. This makes the reasoning state addressable: without retraining the VLA, plans can be sampled (coverage), selected on request (steering), and reweighted by deployment outcomes (outcome adaptation). We test all three on SimplerEnv manipulation with $\pi_0$, and extend the analysis to RoboCasa, LIBERO, and a maze domain. Steering the plan raises the fraction of executions that follow the request from 0.48 to 0.94, while the same guidance on the native action distribution shows no reliable gain. Finally, we show that an explicit plan is most useful when it carries decision information not already determined by the observation.
Related papers
- Recompose and Refine Latent Reasoning Flows for Vision-Language-Action Models
- Plan First, Act Later: Streaming Reasoning for Executable Memory in Robot Control
- WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
- LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence
- PLaW-VLA: Predictive Latent World Modeling for Vision-Language-Action Policies
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-20 · Indexed by Pub-AI: 2026-10-21 UTC
Reasoning as a Distribution: Generative Plan Latents for Vision-Language-Action Models
Preprint posted: 2026-10-20 · Original record ↗
- arxiv
- 2610.28166
Explicit source versions (1)
- v1 · 2026-10-20
- Preprint available
Source date: 2026-10-20 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · OAI arXiv created; not header datestamp (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Vision-Language-Action (VLA) policies increasingly reason before acting, but existing reasoning models compress this intermediate decision into a single representation. When several plans are valid, this commitment hides alternatives that could otherwise be selected or revised. We introduce the Generative Plan Latent (GPL), which represents the reasoning state as a low-dimensional conditional distribution over plans, such as which object, grasp, or route to use. This makes the reasoning state addressable: without retraining the VLA, plans can be sampled (coverage), selected on request (steering), and reweighted by deployment outcomes (outcome adaptation). We test all three on SimplerEnv manipulation with $π_0$, and extend the analysis to RoboCasa, LIBERO, and a maze domain. Steering the plan raises the fraction of executions that follow the request from 0.48 to 0.94, while the same guidance on the native action distribution shows no reliable gain. Finally, we show that an explicit plan is most useful when it carries decision information not already determined by the observation.