AToken: Token-Level Decision Making in AI

• AToken is a multifaceted concept defining token-level decision-making, applied in adaptive visual tokenization, speech processing, automotive signal interpretation, and economic pricing.
• In image tokenization, adaptive mechanisms allocate a variable number of tokens per image using self-budgeting strategies and dynamic Pareto weighting to balance fidelity and efficiency.
• AToken also underpins methods for adjacent token merging in ASR and token valuation in LLM post-training, linking technical signal processing to economic resource accounting.

AToken is a field-dependent term rather than a single standardized construct. In recent arXiv literature, it denotes several distinct token-centric objects: an adaptive, content-conditioned output in self-budgeting image tokenization; a unified visual tokenizer for images, videos, and 3D assets; adjacent token merging in transducer-based ASR; contiguous CAN-bit segments for automotive signal discovery; the practical accounting unit of foundation-model economics; and a shorthand associated with response-token valuation in LLM post-training (Lu et al., 5 Jun 2026, Lu et al., 17 Sep 2025, Li et al., 2023, Verma et al., 2018, Zhu, 10 Jun 2026, Qing et al., 1 Jun 2026).

1. Scope of the term

Across the cited sources, the term is used for non-equivalent objects at different layers of abstraction. In some papers it denotes a learned representation or merge operation inside a model; in others it denotes a signal unit in vehicle telemetry or the economic accounting unit used to meter AI services.

Usage of “AToken”	Definition in source	Paper
Adaptive visual token	Content-conditioned output token count with decodable prefixes	(Lu et al., 5 Jun 2026)
Unified visual tokenizer	Shared tokenizer for images, videos, and 3D assets	(Lu et al., 17 Sep 2025)
Adjacent Token Merging	Parameter-free neighboring-token merge module for ASR	(Li et al., 2023)
Automotive token	Contiguous CAN bit segment $[j_s, j_e]$	(Verma et al., 2018)
Economic token unit	Discrete unit of information, computation, and pricing	(Zhu, 10 Jun 2026)
Response-token valuation shorthand	Token-level valuation framework in LLM post-training	(Qing et al., 1 Jun 2026)

This multiplicity matters because the same label can refer to token creation, token compression, token accounting, or token valuation. The shared motif is token-level control, but the mathematical objects, objectives, and evaluation protocols differ substantially.

2. Adaptive visual tokens in self-budgeting tokenization

In "AdaTok: Self-Budgeting Image Tokenization with Quality-Preserving Dynamic Tokens," AToken refers to adaptive tokens: the tokenizer learns, in one pass, how many discrete tokens a given image needs and guarantees that any chosen prefix is decodable at that budget (Lu et al., 5 Jun 2026). The paper motivates this by arguing that fixed-length tokenization is suboptimal because visual complexity is heterogeneous, and it defines self-budgeting as the tokenizer itself outputting the budget per image in one forward pass. The core claim is a representation–allocation co-design: prefixes must remain decodable across budgets, and the tokenizer must learn which prefix each image needs.

The representation side is Prioritized Representation Learning. Training samples a budget uniformly from $B=\{32,64,96,128,160,192,224,256\}$, masks the tail $z_{b+1:N}$, and optimizes a multi-budget objective,

$$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$

Nested tail masking induces a coarse-to-fine ordering, while Multi-Head LoRA decoder heads resolve budget-dependent semantic shift through

$W^{(b)}=W_0+\alpha_b A_b B_b,$

with LoRA rank $r=16$. The allocation side is Adaptive Token Allocation: a 3-layer GeLU MLP over globally pooled encoder features selects a budget using deterministic-group GRPO, and Dynamic Pareto Weighting updates the fidelity–efficiency balance without manual sweeps. On ImageNet-1K, AdaTok-Full reaches rFID 1.31 at 256 tokens, while AdaTok-Adaptive attains rFID 1.50 using only $\sim 118$ tokens on average; in autoregressive image generation, the shorter adaptive representation yields $\sim 2.1\times$ throughput over a fixed 256-token decode (Lu et al., 5 Jun 2026).

A related but differently named line is "AdaTok: Adaptive Token Compression with Object-Aware Representations for Efficient Multimodal LLMs." That paper explicitly states that it introduces AdaTok, not “AToken,” but it is relevant because it also makes token count input-dependent through object-level merging before the LLM (Zhang et al., 18 Nov 2025). There, SAM masks define object regions, the compression ratio is $r=M/T$, and object features are computed by mask-weighted average pooling. The reported result is that the method averagely utilizes only 10% tokens while achieving almost 96% of the vanilla model’s performance. This suggests that, in current vision research, “adaptive token” often means not merely variable length, but a content-conditioned budget chosen inside the representation pipeline.

3. AToken as a unified tokenizer for vision

In "AToken: A Unified Tokenizer for Vision," AToken is a single transformer-based tokenizer designed to achieve both high-fidelity reconstruction and semantic understanding across images, videos, and 3D assets (Lu et al., 17 Sep 2025). The paper states that it is the first unified visual tokenizer to do so across all three modalities. Its central representational object is a sparse shared 4D latent space,

$$\bm{z}=\{(\bm{z}_i,\bm{p}_i)\}_{i=1}^L,\qquad \bm{p}_i=[t,x,y,z]\in\{0,\ldots,N-1\}^4,$$

with images occupying the $B=\{32,64,96,128,160,192,224,256\}$0 slice, videos stacking along $B=\{32,64,96,128,160,192,224,256\}$1 with $B=\{32,64,96,128,160,192,224,256\}$2, and 3D assets occupying active voxels at $B=\{32,64,96,128,160,192,224,256\}$3. The architecture is a pure transformer encoder/decoder with 4D Rotary Position Embeddings applied in every attention layer, 27 blocks in both encoder and decoder, hidden size $B=\{32,64,96,128,160,192,224,256\}$4, and 16 attention heads.

The training design is explicitly adversarial-free. Reconstruction uses pixel, LPIPS, Gram matrix, and CLIP-alignment losses for images, while videos and 3D use L1 for efficiency and inherit high-frequency detail from image training. The global objective is reported as

$B=\{32,64,96,128,160,192,224,256\}$5

with fixed weights $B=\{32,64,96,128,160,192,224,256\}$6, $B=\{32,64,96,128,160,192,224,256\}$7, and $B=\{32,64,96,128,160,192,224,256\}$8. A four-stage curriculum expands from images to videos, then 3D, and finally discrete tokenization through FSQ. The discrete variant partitions a 48D latent into 8 groups of 6 dimensions, each quantized to 4 scalar levels, giving $B=\{32,64,96,128,160,192,224,256\}$9 codewords per group.

Empirically, the paper reports 0.21 rFID with 82.2% ImageNet zero-shot accuracy for images, 3.01 rFVD with 40.2% MSRVTT retrieval for videos, and 28.28 PSNR with 90.9% zero-shot classification accuracy for 3D assets in the continuous setting. The discrete variant reports 0.38 rFID and 27.00 PSNR for images; 33.12 PSNR and rFVD 22.16 for videos; and 28.17 PSNR, LPIPS 0.063, and 91.3% classification for 3D. In downstream use, AToken is integrated with Lightning-DiT for image generation, with TokenBridge for discrete autoregressive image generation, with MMDiT-based text-to-video generation, and with MLLMs through an attention pool plus a small projector. The conceptual significance is that AToken treats tokenization as a shared multimodal substrate rather than a modality-specific compression stage.

4. AToken as adjacent token merging in automatic speech recognition

In "Accelerating Transducers through Adjacent Token Merging," AToken is identified with A-ToMe, a simple, parameter-free module that accelerates transducer-based ASR by adaptively shortening the encoder’s time axis while preserving temporal order (Li et al., 2023). It is inserted between the MHSA and FFN blocks in selected Transformer encoder layers and merges only neighboring acoustic tokens whose self-attention keys are highly similar. Neighbor similarity is measured by cosine similarity,

$z_{b+1:N}$0

and selected pairs are merged by averaging,

$z_{b+1:N}$1

A fixed threshold strategy merges pairs with $z_{b+1:N}$2, while a fixed-ratio strategy merges the top $z_{b+1:N}$3 fraction of adjacent pairs. A per-layer cap of 50% merged tokens prevents collapse.

The operational motivation is the quadratic encoder bottleneck of Transformer Transducers. In the reported system, the encoder consists of a VGG-like front-end with $z_{b+1:N}$4 subsampling followed by 18 Transformer layers with attention dimension 512, 8 heads, and FFN hidden dimension 2048; the predictor is a 2-layer LSTM with hidden size 1024; and the joint network combines encoder and predictor states. Because MHSA scales as $z_{b+1:N}$5 and joint-network fusion roughly as $z_{b+1:N}$6, reducing the time dimension $z_{b+1:N}$7 directly reduces both encoder and decoding cost. The method is placed every three encoder layers, specifically at layers 2, 5, 8, 11, 14, and 17.

On LibriSpeech, the baseline Transformer Transducer reports test-clean 2.79% WER, test-other 6.01% WER, GPU latency 1.07 s, and CPU latency 3.66 s. With fixed threshold $z_{b+1:N}$8, A-ToMe merges $z_{b+1:N}$9 of tokens, yields GPU 0.63 s and CPU 2.70 s, and reports test-clean 2.89% and test-other 5.96%. With fixed ratio $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$0 per layer, it merges $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$1, yields GPU 0.62 s and CPU 2.43 s, and reports test-clean 2.88% and test-other 6.02%. The paper summarizes this as reducing 57% of tokens and improving GPU inference speed by 70% without notable loss of accuracy. In long-form ASR, where history utterances are concatenated before the encoder, the gains grow with sequence length, indicating that A-ToMe is specifically targeting the quadratic bottleneck.

5. Automotive CAN tokenization and translation

In "ACTT: Automotive CAN Tokenization and Translation," AToken denotes a contiguous bit segment $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$2 in a 64-bit CAN data field, treated as an atomic unit for decoding (Verma et al., 2018). ACTT is vehicle-agnostic and leverages diagnostic ground truth from OBD-II and UDS to discover such tokens and learn their translations into vehicle function values. The inputs are AID traces, consisting of time-ordered 64-bit payloads $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$3 with timestamps, and DID traces, consisting of integer-valued diagnostic sequences $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$4 at lower sampling rate.

The procedure first classifies each bit position as constant-0, constant-1, or used, then enumerates all contiguous spans that do not include constant bits. For each candidate token and each endianness, integer traces are constructed via

$$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$5

The token trace is interpolated to diagnostic timestamps, and a linear model $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$6 is fit by least squares. Tokens are scored by $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$7; accepted matches record $$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$8. Because multiple candidates may overlap, the final selection is obtained by weighted interval scheduling with objective

$$L_{\mathrm{rep}}=\sum_{b\in B} w_b \cdot L_{\mathrm{rec}}^{(b)}(x,\hat{x}_b) +\lambda_p \sum_{b\in B} w_b \cdot L_{\mathrm{perc}}^{(b)} +\lambda_g \sum_{b\in B} w_b \cdot L_{\mathrm{adv}}^{(b)} +L_{\mathrm{commit}}.$$9

subject to non-overlap.

The evaluation covers three vehicles from model years 2008, 2015, and 2016, across two makes and three models, with gasoline and hybrid powertrains. In the exemplar 20-minute capture on a 2008 gasoline vehicle, the system used 25 AIDs and 31 DIDs queried continuously at 20 Hz. With $W^{(b)}=W_0+\alpha_b A_b B_b,$0, 69.6% of bits were constant, 16.8% were matched, 13.6% were unmatched, and the aggregate message packing score was 14.5% over the 16.8% matched bits, yielding an overall match score of 86.0%. With $W^{(b)}=W_0+\alpha_b A_b B_b,$1, 22.0% of bits were matched, 8.4% unmatched, and the overall match score was 72.9%. The paper highlights near-perfect matches for RPM and accelerator pedal position with $W^{(b)}=W_0+\alpha_b A_b B_b,$2, identifies wheel-speed tokens with learned coefficients $W^{(b)}=W_0+\alpha_b A_b B_b,$3 and $W^{(b)}=W_0+\alpha_b A_b B_b,$4, and notes that counters, checksums, and multiplexed frames remain difficult. Here, AToken is neither a learned latent nor a compression primitive; it is an interpretable signal hypothesis grounded in automotive telemetry.

6. AToken as the economic accounting unit of foundation models

In "AI Tokenomics: The Economics of Tokens, Computation, and Pricing in Foundation Models," AToken is the discrete unit through which foundation models represent information, consume computation, use memory, expend energy, and are priced (Zhu, 10 Jun 2026). Raw inputs $W^{(b)}=W_0+\alpha_b A_b B_b,$5 are tokenized via $W^{(b)}=W_0+\alpha_b A_b B_b,$6, and the total token footprint is

$W^{(b)}=W_0+\alpha_b A_b B_b,$7

where $W^{(b)}=W_0+\alpha_b A_b B_b,$8 denotes input tokens, $W^{(b)}=W_0+\alpha_b A_b B_b,$9 context tokens, $r=16$0 retrieval tokens, $r=16$1 output tokens, and $r=16$2 hidden or internal reasoning tokens. The paper links this accounting unit to technical resource use through approximate mappings such as $r=16$3 for FLOPs, $r=16$4 for KV-cache memory, and $r=16$5 for energy.

On the pricing side, the paper distinguishes input and output token prices and writes

$r=16$6

with generalized cost accounting

$r=16$7

At the workflow level, token demand for task $r=16$8 is modeled as $r=16$9, workflow quality as $\sim 118$0, and system utility as $\sim 118$1. Marginal productivity is expressed through adjoint aggregation,

$\sim 118$2

Risk-aware allocation solves

$\sim 118$3

with first-order condition

$\sim 118$4

The contribution is conceptual rather than algorithmic in the tokenizer sense. AToken becomes the common unit that ties together technical cost, enterprise accounting, hidden reasoning measurement, congestion pricing, and token-based market design. This suggests that “token” in foundation-model systems is not only a representational primitive but also an economic resource with measurable marginal value, shadow price, and risk exposure.

7. AToken as response-token valuation in LLM post-training

In "AlphaToken: Decoupling Adaptation and Stability for Path-Aware Response Token Valuation in LLM Post-Training," the details explicitly use “AlphaToken (AToken)” for a token-level valuation framework rather than a tokenizer (Qing et al., 1 Jun 2026). The training instance is a prompt $\sim 118$5 and response $\sim 118$6, and the aim is to assign each response token a scalar valuation $\sim 118$7 that quantifies its marginal generalization contribution. The framework decouples valuation into adaptation, which promotes target-task learning, and stability, which preserves pre-trained capabilities. The composite validation objective is

$\sim 118$8

and the token value is proportional to gradient alignment with target and retention objectives.

The path-aware part of the framework combines direct-path and downstream causal-path signals. Direct-path adaptation is computed by Ghost Dot-Product using rank-1 token gradients, while causal-path adaptation routes future losses through autoregressive dependencies with a Value-Propagation approximation. Because retention data are typically unavailable, stability is approximated by a Fisher-drift proxy anchored at the pre-trained reference model,

$\sim 118$9

with virtual retention gradient

$\sim 2.1\times$0

The total token score is then

$\sim 2.1\times$1

Low-value response tokens are masked by a within-batch top-$\sim 2.1\times$2 policy during SFT or DPO.

The reported defaults are $\sim 2.1\times$3, $\sim 2.1\times$4, causal window $\sim 2.1\times$5, last-$\sim 2.1\times$6 scoring layers with $\sim 2.1\times$7, validation size $\sim 2.1\times$8, and DPO $\sim 2.1\times$9. The diagonal Monte-Carlo Fisher is built once from 1,000 prompts, and experiments use $r=M/T$0 NVIDIA A100 GPUs. Across Llama-3.2-3B, Gemma-3-4B, and Qwen-3.5-9B, the paper reports state-of-the-art Overall in SFT, improving over the strongest baseline by 1.54, 2.80, and 2.51 points, respectively, and the best Overall in DPO, improving over the strongest competing baseline by 1.73, 1.54, and 1.83 points. In this usage, AToken is neither a modality token nor a pricing unit; it is a valuation target for selectively routing training signal to response positions that jointly support adaptation and retention.

Across these literatures, the common denominator is token-level decision-making: deciding how many tokens to emit, how to merge them, how to decode them, how to map them to physical signals, how to price them, or how to value them during optimization. The term’s breadth is therefore substantive rather than merely terminological. It marks a broader shift in contemporary research from treating tokens as fixed byproducts of preprocessing toward treating them as controllable objects of modeling, systems design, and optimization.