AWARE Training-free · plug-and-play · no additional observation cues

AWARE: Action-Entropy-Weighted and Attention-Guided Retrieval of Observation Evidence for Vision-Language-Action Models

Jiabing Yang1,2,5, Yixiang Chen1,2, Yuan Xu1,2, Peiyan Li1,2, Zichen Wen3, Bowen Fang1,2, Chenhang Cui4, Tao Yu1,2,
Qisen Ma1,2, Kai Wang1,2, Ziheng He1,2, Wen Ye1,2, Xiangnan Wu1,2, Yingda Li1,2, Yan Huang1,2,5,†, Liang Wang1,2

1School of Artificial Intelligence, University of Chinese Academy of Sciences 2New Laboratory of Pattern Recognition (NLPR), Institute of Automation, Chinese Academy of Sciences 3Shanghai Jiao Tong University 4National University of Singapore 5FiveAges †Corresponding author

01Overview

Re-read the observation where the action is still undecided

TL;DR

Abstract

02Diagnosis

Grounding fades with depth, and far actions settle last

(a) Observation share of the action representation
(b) Action Entropy across the chunk
Show
03Method

The model's own FFN as a memory of the observation

Figure 2.

One forward pass with AWARE

# once per policy call
O = obs_tokens(obs)                 # keys = values = O
g = instr_weight(instr, O)          # Eq. (g)
for l in range(L - 1):
    H = entropy(lm_head(h_act[l]))   # per element
    w = alpha * l / (L - 1) * H       # Eq. (w)
    q = ffn_in(h_act[l + 1])          # own query
    inj = s * O.T @ (g * (O @ q))       # Eq. (9)
    out[l + 1] = (1 - w) * out[l + 1] + w * inj
# non-action tokens keep their native FFN output
  • Training-free
  • Plug-and-play
  • No extra sensors or modules
  • One strength  α
Where the entropy is read, per backbone family

Where AWARE writes, element by element

Injection weight of every OpenVLA-OFT action element (columns) at every layer (rows) for the selected variant and strength. Brighter cells receive more retrieved observation evidence.

Mean weight over the stack
w = 0 high Entropy from Figure 1b · success from Figure 5 and Table IV
04Results

Seven backbones, largest gains on long horizons

Base → w/ AWARE, per suiteEach segment runs from the base policy to the same checkpoint with AWARE; hover for the numbers and the strength α.
Table I
Gain by suiteΔ = w/ AWARE − Base, success points; Long is where the gain is largest for every backbone.

Strength α used in Table I

Where the gain is small or reversed

    05Analysis

    The gain comes from routing real evidence

    Component analysisAverage success over the four LIBERO suites when one part of the rule is removed; hover a bar for each suite.
    Table IV

    Diagnostic variants on LIBERO-LongGain over the base policy when the entropy routing or the injected content is replaced.
    Table V

    Sensitivity to the strength α
    Backbone
    Inference overheadOpenVLA-OFT on LIBERO-Long, one NVIDIA A100 GPU; the policy predicts chunks of H = 8, so the per-step latency is the wait per environment step.
    Table VI

    06Rollouts

    Base vs. w/ AWARE, side by side

    07Citation

    BibTeX

    @article{yang2026aware,
  title   = {AWARE: Action-Entropy-Weighted and Attention-Guided Retrieval of Observation Evidence for Vision-Language-Action Models},
  author  = {Yang, Jiabing and Chen, Yixiang and Xu, Yuan and Li, Peiyan and Wen, Zichen and Fang, Bowen and Cui, Chenhang and Yu, Tao and Ma, Qisen and Wang, Kai and He, Ziheng and Ye, Wen and Wu, Xiangnan and Li, Yingda and Huang, Yan and Wang, Liang},
  journal = {arXiv preprint arXiv:2602.18020},
  year    = {2026}
}