2026-08-11 · IROS 2026

Neural Introspection Gating for Adaptive KV-Cache Reuse in Vision-Language-Action Models

Zhijie Wu, Kento Kawaharazuka, Kei Okada

Published at IROS 2026. 0 citations, as of the last refresh.

Abstract

Vision-Language-Action(VLA) models map camera images and language instructions directly to motor commands through a single autoregressive transformer. In real-time control, they still spend substantial compute recomputing key-value(KV) representations for visual tokens that barely change across neighboring frames. Recent work such as VLA-Cache reduces that cost by reusing KV states for visually static patches, but its policy relies only on observation-space heuristics and does not account for the model's own uncertainty. We propose Gated VLA-Cache, a lightweight, training-free extension that augments visual-similarity caching with neural introspection. The method monitors the logit margin between the top two predicted action tokens, a zero-cost confidence signal available during decoding. When the margin drops below a threshold, the cache is invalidated and a full recompute is triggered. Evaluated on four LIBERO benchmark suites with both OpenVLA and OpenVLA-OFT, Gated VLA-Cache improves reliability when blind caching hurts. On LIBERO-Goal and LIBERO-Long, it recovers over 100% of the lost accuracy while retaining 80% of the compute savings.

arXiv comment: 6 pages, 5 figures, Accepted in IROS 2026. Project Page: https://zjw4321.github.io/neural-introspection-gating-page/

Ten-minute slide kit

Six slides is the whole talk: what was broken, what people tried, what these authors did, what the numbers say, where it falls over, and the sentence people should remember.

SLIDE 1

The problem

Vision-Language-Action (VLA) models – use a single autoregressive transformer that ingests a camera frame and a language instruction, then directly outputs discretized motor commands. Because they build on large visionlanguage models (VLMs), VLAs generalize across manipulation tasks without much task-specific engineering. However, running a 7-billion-parameter model at the 10–20 Hz rates needed for manipulation is computationally

SLIDE 2

What came before

VLA –, models unify perception, language understanding, and action prediction within a single transformer. RT-2 pioneered the approach by fine-tuning a VLM to output actions as text tokens. OpenVLA is a 7B-parameter model built on Prismatic VLMs that performs well on diverse manipulation

SLIDE 3

The method

Not recoverable from the parsed text. Read this section in the paper yourself.

SLIDE 4

What they measured

We evaluate on two VLA architectures: (a) Open-VLA , a 7B-parameter VLA based on the Prismatic VLM architecture with a Llama-2 backbone, and (b) OpenVLA-OFT , which uses optimized fine-tuning with parallel action-dimension decoding for improved downstream performance. We evaluate on four LIBERO task families: *Spatial* (10 spatial relationship reasoning tasks), *Object* (10 object type identification tasks), *Goal* (10 goalconditioned tasks), and *Long* (10 long-horizon multi-step tasks). Following the standard…

SLIDE 5

Where it breaks

The pattern is simple: Gated VLA-Cache helps most on tasks where VLA-Cache causes clear accuracy degradation. On LIBERO-Goal (−3.2%) and LIBERO-Long (−3.8%), where cache staleness compounds over long manipulation sequences, our gating recovers over 100% of the lost accuracy (106% and 121%, respectively). On suites where caching is already safe (Spatial, Object), the gating rarely triggers and adds minimal

SLIDE 6

One-line takeaway

Gated VLA-Cache is proposed, a lightweight, training-free extension that augments visual-similarity caching with neural introspection that improves reliability when blind caching hurts.

Assembled from the paper's own PDF, parsed with its layout intact so tables and equations survive, 31,369 characters of it, then split on the paper's own section headings. Extractive, not generated: every sentence here is lifted from the paper. Check it before you present it.