Dynamic Relative Alignment Quality (DRAQ)

• DRAQ is a family of dynamic, task-dependent alignment quality metrics that use relative representations to robustly capture semantic alignment beyond absolute encoding.
• It integrates online optimization with virtual queues and Lyapunov control to jointly manage semantic accuracy, power consumption, and latency in communication systems.
• Extensions of DRAQ are applied in interpretable classification, graph-text alignment, language model refinement, and tabular generation to dynamically weight quality and trust in predictions.

Searching arXiv for the cited papers and related uses of “Dynamic Relative Alignment Quality” to ground the article. arxiv_search(query="(Fiorellino et al., 2024)", max_results=5) arxiv_search({"query":"(Fiorellino et al., 2024)","max_results":5}) search_arxiv(query="(Fiorellino et al., 2024)", max_results=5) Dynamic Relative Alignment Quality (DRAQ) denotes a family of alignment-quality measures that are dynamic, task-dependent, and defined relative to a reference structure rather than on absolute representations alone. In goal-oriented semantic communications, DRAQ is the time-varying “goodness” of semantic alignment induced by an encoder–anchor-set pair $(E_t,A_t)$, typically instantiated by validation accuracy or by a geometric alignment score in relative-representation space (Fiorellino et al., 2024). Related formulations use the same label, or closely analogous quality-aware mechanisms, for dynamic alignment in interpretable classification, adaptive graph-text alignment, quality-aware direct language-model alignment, and self-improving tabular generation (Böhle et al., 2021, Liu et al., 22 Oct 2025, Yu et al., 2024, Long et al., 21 Apr 2026).

1. Canonical formulation in goal-oriented semantic communications

The most explicit formalization arises in “Dynamic Relative Representations for Goal-Oriented Semantic Communications” (Fiorellino et al., 2024). For each encoder–anchor-set pair $(E,A)$, a sample $x$ is mapped to a relative representation

$$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$

where $A=\{a_1,\dots,a_{|A|}\}\subset X$ is the anchor set and $\mathrm{sim}$ may be cosine similarity. A decoder $D_A$ then maps $r(x;E,A)$ to the downstream estimate $\hat y$ for a task such as classification or detection (Fiorellino et al., 2024).

In this setting, DRAQ is the time-varying, task-dependent quality of semantic alignment afforded by $(E_t,A_t)$ at slot $(E,A)$0. For classification, it is measured by the validation-set $(E,A)$1–$(E,A)$2 accuracy,

$(E,A)$3

The same framework also permits a purely geometric alignment score based on cosine similarity between the current relative representation and a fixed reference relative representation, and it often tracks the long-term average

$(E,A)$4

(Fiorellino et al., 2024).

The central motivation is semantic mismatch. When two devices employ heterogeneous absolute encoders, semantically identical inputs may be placed differently in $(E,A)$5, degrading decoding if one transmits $(E,A)$6 directly. Relative representations instead preserve the pattern of similarities to a shared anchor set. In the formulation described for (Fiorellino et al., 2024), this makes alignment more stable across heterogeneous encoders and enables zero-shot stitching even when the transmitter-side encoder was never seen during decoder training. Goal orientation is therefore explicit: the objective is not bit-error rate or PSNR, but end-task effectiveness.

2. Optimization, constraints, and online control

The dynamic role of DRAQ is clearest in the time-slotted stochastic optimization of (Fiorellino et al., 2024). At each slot $(E,A)$7, the controller selects uplink rate $(E,A)$8, CPU clock $(E,A)$9, encoder $x$0, and anchor set $x$1 so as to minimize long-term average power

$x$2

subject to three classes of constraints: a delay-violation probability constraint, a long-term average DRAQ constraint $x$3, and an average-latency constraint $x$4, together with per-slot bounds on $x$5, $x$6, $x$7, and $x$8 (Fiorellino et al., 2024).

The power and latency models decompose into computation and transmission terms: $x$9 with $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$0, $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$1, and $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$2 (Fiorellino et al., 2024). The resulting control problem couples semantic quality, computation, and wireless transmission through the same decision variables.

The solution introduces virtual queues $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$3, $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$4, and $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$5 for the three constraints and constructs the Lyapunov drift-plus-penalty $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$6. This yields a per-slot objective

$$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$7

minimized with respect to $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$8 (Fiorellino et al., 2024). For each candidate $$r(x;E,A)=\bigl[\mathrm{sim}(E(x),E(a_1)),\dots,\mathrm{sim}(E(x),E(a_{|A|}))\bigr]\in\mathbb{R}^{|A|},$$9, the subproblem in $A=\{a_1,\dots,a_{|A|}\}\subset X$0 is convex and is solved via KKT conditions to obtain closed-form $A=\{a_1,\dots,a_{|A|}\}\subset X$1 and $A=\{a_1,\dots,a_{|A|}\}\subset X$2; the controller then searches over the finite set $A=\{a_1,\dots,a_{|A|}\}\subset X$3 to pick the slotwise minimizer.

Within this formulation, DRAQ is not an after-the-fact evaluation only. It is a queue-coupled control variable. The framework’s practical interpretation is correspondingly direct: one can specify a minimum acceptable semantic quality, then let the online controller choose encoder complexity, anchor-set size, compute clock, and rate so as to satisfy the DRAQ target with minimum energy and acceptable latency. Monitoring the virtual-queue backlogs $A=\{a_1,\dots,a_{|A|}\}\subset X$4 further indicates whether delay, latency, or semantic quality is the currently stressed constraint (Fiorellino et al., 2024).

3. Empirical behavior and system-level trade-offs

The empirical behavior reported for (Fiorellino et al., 2024) makes the operational meaning of DRAQ concrete. Static semantic-alignment accuracy $A=\{a_1,\dots,a_{|A|}\}\subset X$5 improves with anchor-set size: as $A=\{a_1,\dots,a_{|A|}\}\subset X$6 grows from $A=\{a_1,\dots,a_{|A|}\}\subset X$7 to $A=\{a_1,\dots,a_{|A|}\}\subset X$8, accuracy improves by up to $A=\{a_1,\dots,a_{|A|}\}\subset X$9 percentage points for five different encoders, and zero-shot stitching remains viable even when the transmitter-side encoder was never seen during decoder training.

The long-term trade-off between semantic quality, power, and latency is equally explicit. The reported power-versus-latency curves were evaluated for DRAQ targets $\mathrm{sim}$0; stricter accuracy targets increase CPU and transmit power, while relaxed latency budgets provide large power savings. Sample-path traces for instantaneous latency $\mathrm{sim}$1 and instantaneous DRAQ under a setting $\mathrm{sim}$2 s and $\mathrm{sim}$3 show that running averages remain below or above their targets, respectively, in the long run (Fiorellino et al., 2024).

A central resource knob is the anchor-set cardinality $\mathrm{sim}$4. Small $\mathrm{sim}$5 reduces uplink load but produces coarser semantics and lower DRAQ; large $\mathrm{sim}$6 increases load but yields more robust alignment. In the reported comparisons, the relative-representation framework with dynamic control sustains a $\mathrm{sim}$7–$\mathrm{sim}$8 percentage-point higher accuracy than a baseline using absolute representations at the same energy–delay operating point (Fiorellino et al., 2024).

These observations clarify a common misconception. In this line of work, DRAQ is not a proxy for conventional reconstruction fidelity. It is the task-centric quantity by which semantic communication is judged and controlled.

4. DRAQ as object-faithful dynamic alignment in interpretable classifiers

A distinct formulation appears in “Convolutional Dynamic Alignment Networks for Interpretable Classifications” (Böhle et al., 2021). The basic unit is the Dynamic Alignment Unit (DAU),

$\mathrm{sim}$9

with $D_A$0 and $D_A$1. Because the weight vector depends on the input, the DAU acts as a dynamic linear map. The relation

$D_A$2

shows that maximizing the DAU output forces the dynamic weight vectors to align with task-relevant patterns in the input (Böhle et al., 2021).

By composing such layers, the final network logit becomes a single dynamic linear map,

$D_A$3

which admits a per-dimension contribution decomposition $D_A$4. For a single-class logit, this simplifies to $D_A$5 (Böhle et al., 2021). The model therefore yields inherent, input-dependent contribution maps rather than post-hoc attributions.

In this framework, DRAQ is defined as a pair of coordinated measures: $D_A$6 where $D_A$7 is the localization score derived from the positive contribution map and $D_A$8 is the class probability after removing the $D_A$9 least important pixels (Böhle et al., 2021). The first component measures how well positive attribution mass localizes on the ground-truth object; the second measures whether the attribution ranking is robust when unimportant pixels are removed and destructive when important pixels are removed.

Empirically, CoDA-Nets outperform common post-hoc attribution methods including Grad, Grad-CAM, Integrated Gradients, RISE, LIME, and Occlusion. Reported localization scores are approximately $r(x;E,A)$0–$r(x;E,A)$1, versus roughly $r(x;E,A)$2–$r(x;E,A)$3 for Grad-CAM or Integrated Gradients; under least-important-first removal, the target-class probability changes by less than $r(x;E,A)$4 percentage points up to $r(x;E,A)$5 pixel removal, whereas baselines often lose more than $r(x;E,A)$6 percentage points (Böhle et al., 2021). Here DRAQ measures object-faithful alignment rather than communication accuracy, but the relative and dynamic principles are the same: input-dependent alignment is quantified by a task-relevant, reference-dependent metric.

5. Quality-aware extensions in language, graph-text, and tabular alignment

Several recent systems use closely related relative-quality mechanisms, even when the exact scalar form differs.

Domain	Relative quality construct	Dynamic role
Direct LM alignment	$r(x;E,A)$7	Refines DPO/IPO loss
Graph-text alignment	$r(x;E,A)$8, $r(x;E,A)$9, $\hat y$0	Reweights losses and filtering
Tabular generation	$\hat y$1, $\hat y$2	Repartitions and self-improves

In “Direct Alignment of LLMs via Quality-Aware Self-Refinement,” the key quantity is the refinement function

$\hat y$3

which estimates the relative qualities of preferred and dispreferred responses under a prompt-augmented input (Yu et al., 2024). This quantity is inserted into DPO and IPO through a stop-gradient term so that more informative preference tuples are weighted more effectively. Reported results include Sr-DPO win/tie/loss of $\hat y$4 against DPO on MT-Bench and $\hat y$5 on Vicuna-Bench, together with Open-LLM-Leaderboard gains from $\hat y$6 to $\hat y$7 for DPO and from $\hat y$8 to $\hat y$9 for IPO (Yu et al., 2024).

In “Learning Noise-Resilient and Transferable Graph-Text Alignment via Dynamic Quality Assessment,” the batch-level relative margin

$(E_t,A_t)$0

is smoothed by exponential moving averages, and the control factor

$(E_t,A_t)$1

estimates batch-level alignment quality (Liu et al., 22 Oct 2025). The total loss interpolates among a soft many-to-many loss, a subgraph-to-text contrastive loss, and a one-to-one CLIP-style loss through $(E_t,A_t)$2, $(E_t,A_t)$3, and $(E_t,A_t)$4. The same control variable governs low-confidence filtering via a retained subset $(E_t,A_t)$5. Theoretical statements in the paper establish signal boundedness, dynamic stability, and convergence of the total loss to a stationary point under a two-time-scale regime. Empirically, the filter’s precision at removing truly noisy synthetic pairs rises to greater than $(E_t,A_t)$6, and the controller $(E_t,A_t)$7 settles into a narrow band under both clean and noisy conditions (Liu et al., 22 Oct 2025).

In “Self-Improving Tabular LLMs via Iterative Group Alignment,” quality enters through an automated per-row signal $(E_t,A_t)$8, such as the indistinguishability score

$(E_t,A_t)$9

or a distance-based alternative $(E,A)$00 (Long et al., 21 Apr 2026). Generated samples are sorted by $(E,A)$01 and partitioned into high- and low-quality groups, and the model maximizes the group-relative advantage

$(E,A)$02

through the GRAA loss

$(E,A)$03

Theoretical results include bounded gradient norm, variance reduction of order $(E,A)$04, reward-agnostic stability, and convergence to stationarity under standard SGD assumptions. Reported experiments show improvements in fidelity, utility, and privacy across five UCI datasets, including Adult AUC $(E,A)$05, Beijing RMSE $(E,A)$06, and near-random distinguishability-attack success on Adult with DA $(E,A)$07 (Long et al., 21 Apr 2026).

Taken together, these formulations suggest that DRAQ is best understood not as a single invariant equation, but as a design pattern: a relative quality signal that is updated online and coupled to optimization.

6. Interpretation, scope, and recurring misconceptions

Across the cited works, DRAQ is not a universal fixed scalar. In (Fiorellino et al., 2024), it is validation accuracy or a geometric relative-representation score; in (Böhle et al., 2021), it is a pair $(E,A)$08; in (Liu et al., 22 Oct 2025), the analogous adaptive quality mechanism is mediated through $(E,A)$09, $(E,A)$10, and $(E,A)$11; and in (Long et al., 21 Apr 2026), quality is instantiated through $(E,A)$12 and the group-relative advantage. This suggests that the shared core lies in the words “dynamic,” “relative,” and “alignment quality,” not in a single canonical formula.

A second misconception is to treat “relative” as a minor implementation detail. In every formulation above, relative structure is the main mechanism by which alignment becomes robust. The reference may be an anchor set, a frozen model, a preferred/dispreferred pair, a batchwise negative set, a subgraph neighborhood, or a high-/low-quality group. What is measured is not raw representation quality in isolation, but quality relative to an external or internal semantic standard.

A third misconception is to treat the dynamic component as mere online bookkeeping. In the semantic-communications setting, DRAQ is embedded in virtual queues and Lyapunov control; in graph-text alignment, the controller $(E,A)$13 induces a negative-feedback process; in tabular generation, the quality signal is recomputed on newly generated synthetic data at every round; and in quality-aware direct preference optimization, the model performs on-the-fly self-judgment through a prompt-augmented input (Fiorellino et al., 2024, Liu et al., 22 Oct 2025, Long et al., 21 Apr 2026, Yu et al., 2024). The dynamic mechanism changes which samples are trusted, how strongly they are weighted, and which operating point is selected.

A plausible implication is that DRAQ occupies an intermediate layer between representation learning and systems control. It is simultaneously an evaluation criterion, a training signal, and a constraint interface. In semantic communication, it directly mediates energy–latency–effectiveness trade-offs; in interpretable classification, it quantifies whether dynamic weights align with discriminative objects; in graph-text and language-model alignment, it calibrates how much trust to place in relative supervision; and in tabular generation, it drives self-improvement without additional exposure of real records after initial supervised fine-tuning. Within that broader view, DRAQ is less a single benchmark than a domain-specific alignment-quality interface for adaptive learning and control.