# DORA: Dynamic Online Reinforcement Agent for Token Merging in Vision Transformers

arXiv: https://arxiv.org/abs/2605.11683 · HTML: https://mlsanity.com/paper/2605.11683/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Token pruning and merging for ViTs, one of 12 directions in Efficient LLM inference and serving.
- Builds on 19 of its 22 references that are in the map.
- First posted on arXiv 12 May 2026.

## Placed in

- Efficient LLM inference and serving / [Token pruning and merging for ViTs](https://mlsanity.com/inference/token-pruning-and-merging-for-vits/)
- Deep reinforcement learning methods / [RL with diffusion and flow models](https://mlsanity.com/deeprl/rl-with-diffusion-and-flow-models/)

## Builds on (most cited first, 19 of 19)

- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://mlsanity.com/paper/2103.14030/) (2021)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Training data-efficient image transformers & distillation through attention](https://mlsanity.com/paper/2012.12877/) (2020)
- [Benchmarking Neural Network Robustness to Common Corruptions and Perturbations](https://mlsanity.com/paper/1903.12261/) (2019)
- [The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization](https://mlsanity.com/paper/2006.16241/) (2020)
- [Is Space-Time Attention All You Need for Video Understanding?](https://mlsanity.com/paper/2102.05095/) (2021)
- [Natural Adversarial Examples](https://mlsanity.com/paper/1907.07174/) (2019)
- [Video Swin Transformer](https://mlsanity.com/paper/2106.13230/) (2021)
- [Token Merging: Your ViT But Faster](https://mlsanity.com/paper/2210.09461/) (2022)
- [DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification](https://mlsanity.com/paper/2106.02034/) (2021)
- [How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers](https://mlsanity.com/paper/2106.10270/) (2021)
- [Not All Patches are What You Need: Expediting Vision Transformers via Token Reorganizations](https://mlsanity.com/paper/2202.07800/) (2022)
- [AdaViT: Adaptive Tokens for Efficient Vision Transformer](https://mlsanity.com/paper/2112.07658/) (2021)
- [Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers](https://mlsanity.com/paper/2305.17328/) (2023)
- [Accelerating Transformers with Spectrum-Preserving Token Merging](https://mlsanity.com/paper/2405.16148/) (2024)
- [Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers](https://mlsanity.com/paper/2412.10569/) (2024)
- [Asymmetric Actor-Critic for Multi-turn LLM Agents](https://mlsanity.com/paper/2604.00304/) (2026)

## Abstract (arXiv)

Vision Transformers (ViTs) incur significant computational overhead due to the quadratic complexity of self-attention relative to the token sequence length. While existing token reduction methods mitigate this issue, they predominantly rely on fixed heuristic metrics, predefined ratios, or static offline masks, which lack the adaptability to capture input-dependent redundancy during inference. In this paper, we propose DORA (Dynamic Online Reinforcement Agent), the first reinforcement learning (RL)-driven online inference framework for dynamic token merging in ViTs. We formulate the merging process as a sequential Markov Decision Process (MDP), where a lightweight RL agent determines the merging strategy for each Transformer block based on the current feature state and layer-specific context. To balance computational efficiency and feature fidelity, the agent is optimized via a dense reward function incorporating a non-linear distillation-based penalty. We implement an asymmetric Actor-Critic architecture that utilizes a high-capacity Critic for stable offline training while retaining a minimal Actor head for low-computation online inference. Evaluations across multiple ViT scales (Tiny to Large) demonstrate that DORA improves the accuracy-efficiency Pareto front compared to current baselines. Under strict negligible accuracy-drop constraints (<= 0.05%), DORA achieves up to a 12.66% token merging rate, and delivers up to a 569.7% relative improvement over the most efficient baseline. On ImageNet-1K, under aligned accuracy constraints, DORA achieves up to a 76% relative improvement in computational savings compared to state-of-the-art methods. Furthermore, on out-of-distribution (OOD) benchmarks such as ImageNet-A and ImageNet-C, DORA attains a relative efficiency advantage of over 430%.