AROMA
A framework for controlling vision-language models that separates localization from control. Through techniques such as magnitude-normalized causal subspace steering and iterative minimal-effort steering, it enables selective intervention to improve model accuracy.
Timeline (last 90 days)
The paper "AROMA: Adaptive Representation-Oriented Multimodal Actuation for Vision-Language Model Control" proposes a framework that separates mechanism localization from behavioral control.
AROMA decomposes VLM control into five independent questions: where to intervene (where), which geometry and gain to use (which geometry and gain), how to reach the target (how), which target to pursue (which target), and when to act (when).
AROMA employs a magnitude-normalized causal subspace steering method and improves exact counting accuracy by 8.00 percentage points on procedural data it never touched.
Under a strict evaluation protocol in which both targets and gates are frozen in advance, targeted MEA raises the accuracy of the previously frozen policy from 34.444% to 38.730% (paired p=4.63e-7) on an entirely new natural Stage-2 cohort (N=630).
Across all 7,992 samples, editing only 63 inputs raises accuracy from 65.153% to 65.741%.
Through prospective routing, cross-attention-head and cross-model replication, and comparison with attribution-guided baselines, the paper systematically characterizes the independent properties of five stages: localization, actuation, reachability, target quality and selective deployment.
The paper notes that the breadth of cross-model replication, the generalization boundaries of routing strategies, and the performance of target quality evaluation on more complex tasks are open questions requiring further validation.
All Facts (7)
The paper "AROMA: Adaptive Representation-Oriented Multimodal Actuation for Vision-Language Model Control" proposes a framework that separates mechanism localization from behavioral control.
50%UnverifiedAROMA decomposes VLM control into five independent questions: where to intervene (where), which geometry and gain to use (which geometry and gain), how to reach the target (how), which target to pursue (which target), and when to act (when).
50%UnverifiedAROMA employs a magnitude-normalized causal subspace steering method and improves exact counting accuracy by 8.00 percentage points on procedural data it never touched.
50%UnverifiedUnder a strict evaluation protocol in which both targets and gates are frozen in advance, targeted MEA raises the accuracy of the previously frozen policy from 34.444% to 38.730% (paired p=4.63e-7) on an entirely new natural Stage-2 cohort (N=630).
50%UnverifiedAcross all 7,992 samples, editing only 63 inputs raises accuracy from 65.153% to 65.741%.
50%UnverifiedThrough prospective routing, cross-attention-head and cross-model replication, and comparison with attribution-guided baselines, the paper systematically characterizes the independent properties of five stages: localization, actuation, reachability, target quality and selective deployment.
50%UnverifiedThe paper notes that the breadth of cross-model replication, the generalization boundaries of routing strategies, and the performance of target quality evaluation on more complex tasks are open questions requiring further validation.
50%