StackTok: Accelerating VLMs Inference with Budget-Adaptive Visual Token Selection

简体中文

L1 core-mechanism diagnostic. Runs a reference-gated selector on one crop without loading the paper's five VLMs.

Paper information

Field	Details
Paper	StackTok: Accelerating VLMs Inference with Budget-Adaptive Visual Token Selection
Organization	First-author affiliation: Sichuan University
First public release	2026-09-15 (arXiv v1)
Authors' code	wongzbb/StackToK is public.
Adapter / method	stacktok
Local reproduction code	src/auto_research/foundation_latest_20260916_followup.py

Original paper

Motivation and method

Use query relevance as the objective and coverage as a budget-dependent support constraint, progressively switching between the two selection criteria.

flowchart LR
  I[Public input and current state] --> M[stacktok mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper

Figure 2: Training pipeline and key optimization steps. From the original paper; copyright remains with the authors. Click the image for the source.

Core computation and code mapping

The local reference kernel preserves the decisive ranking, gates, preference direction, credit weights, or state transitions and outputs auditable statistics. Three-seed fixtures test invariants and edge cases rather than interpreting randomized mini-suite scores as paper-level capability.

Paper results

Accuracy, speed, training loss, or benchmark findings belong to the paper. Local results are not directly compared with them or extrapolated to online gains.

Local reproduction

See the three-seed results. The common receipt marks diagnostic_only=true, ineligible for formal capability ranking.

Reproduction boundaries

The local implementation executes a single-crop reference-gated selector. It does not load the five VLMs in the paper.