Cerevla Consequence Aware Residual Governance

---
name: cerevla-consequence-aware-residual-governance
description: Use when deploying frozen action-chunked VLA policies on robots. Lightweight residual refinement plus predictive consequence evaluation with selective suppression — cerebellum-inspired governor that corrects chunked execution errors without retraining.
trigger: consequence-aware residual, VLA action chunk correction, frozen policy residual refinement, cerebellum-inspired robotics, governor suppression, recurrent state-space consequence prediction, LIBERO SO-101 residual
category: ai_collection
---

# CereVLA: Consequence-Aware Residual Governance for Frozen VLA Execution

**Source**: arXiv:2609.27468v1 (2026-09-23) — Zeng, Liang, Hu, Zhou, Wang et al. (7 authors).

## Problem

Action-chunked VLA policies improve inference efficiency, but **limited feedback within committed action chunks** leads to accumulated execution errors. Residual adaptation can correct deviations without retraining the VLA — but existing corrections optimize for **reference-action consistency** without considering their **downstream consequences**.

## CereVLA Mechanism (cerebellum-inspired, 3 stages)

1. **Flow-based residual refinement**: lightweight corrective actions generated by a flow model on top of the frozen VLA's chunked output.
2. **Predictive consequence evaluation**: a **recurrent state-space model (SSM)** predicts short-horizon and interval-horizon consequences of the proposed residual; a history-aware classifier judges favorability.
3. **Selective suppression (the governor)**: residuals predicted to be **unfavorable are selectively suppressed** by a lightweight governor — only beneficial corrections pass through.

The cerebellum analogy: fast corrective loops that evaluate consequences before committing motor corrections.

## Verified Results

| Benchmark | Result |
|---|---|
| LIBERO-10, LIBERO-GOAL | outperforms SOTA residual adaptation methods |
| SO-101 real robot | task success **57.5% → 90.0%** vs frozen SmolVLA baseline |
| Mean control steps (successful trials) | **−19.6%** |

## Why It Matters

- **Residual corrections can hurt**: a residual matching the reference action better may still cause worse downstream states (e.g., overcorrection into irreversible regions). Consequence filtering is the missing piece.
- Pattern generalizes beyond robotics: **any frozen policy + cheap corrector** setup (LLM decoding with verifiers, RL deployment with safety filters) can benefit from *predict-then-suppress* governance instead of always-apply correction.

## Implementation Checklist

1. Freeze the VLA; run chunked action execution as-is.
2. Attach flow-based residual generator conditioned on (obs, chunk progress, reference actions).
3. Train (or fine-tune small) recurrent SSM world model on execution rollouts to predict consequences of residual vs no-residual at short and interval horizons.
4. History-aware classifier: input = predicted consequence features + execution history; output = favorable/unfavorable.
5. Governor gate: apply residual only if favorable; log suppression rate as a diagnostic (high suppression = residual generator miscalibrated).
6. Evaluate: success rate + control steps (efficiency) — a good governor reduces both failures and wasted steps.

## Related Skills

- `memoryvla-temporal-modeling-robotic-manipulation` — VLA temporal modeling
- `turbovla-real-time-vla` — VLA inference efficiency (the chunking CereVLA corrects)
- `carl-capacitive-...` family — other residual/correction patterns