---
title: 'Preview Accept Discard (PAD): A Cross-Domain Schema'
url: https://www.emergentmind.com/topics/preview-accept-discard-pad
type: topic
---

# Preview Accept Discard (PAD): A Cross-Domain Schema

Searching arXiv for recent papers on “Preview Accept Discard” and related PAD usages.
“Preview Accept Discard” (PAD) is not a single, universally standardized technical term across arXiv literature. Rather, it functions as a natural descriptive pattern that recurs in several research areas: speculative decoding for large language models, predictive human–computer interaction, creative-history interfaces, and decision workflows in control and statistical inference. In the strictest naming sense, some papers use PAD as an acronym for unrelated terms such as “Position-Aware Drafting,” “Pose-Aware Diffusion,” “Pivot-Aware Speculative Decoding,” or “Presentation Attack Detection,” and several explicitly state that PAD does **not** mean “Preview Accept Discard” in their context [2604.27747], [2605.00345], [2511.00351]. At the same time, multiple works either explicitly define PAD as “Preview Accept Discard” or admit a strong algorithmic or interactional mapping to a preview–then–commit–or–reject loop [2511.10532], [2604.27747], [2501.04163]. This suggests that PAD is best understood as a cross-domain operational schema rather than a single canonical algorithm.

## 1. PAD as a cross-domain decision pattern

Across the cited literature, PAD denotes a sequence in which a system first exposes a tentative future state or candidate action, then either commits to some portion of it or rejects it. The strongest explicit definition appears in the predictive low-motion interaction paper “Preview, Accept or Discard? A Predictive Low-Motion Interaction Paradigm,” where PAD is introduced as “a zero-click interaction paradigm that lets users preview predicted GUI targets, cycle through a small set of ranked alternatives, and accept or discard them via key-release timing” [2511.10532]. In that paper, the three terms are literal interface primitives.

In other works, PAD is not the paper’s official acronym but is presented as an apt descriptive interpretation. “Position-Aware Drafting for Inference Acceleration in LLM-Based Generative List-Wise Recommendation” states that “Preview Accept Discard” maps naturally onto speculative decoding: the draft model previews a candidate block, the target model accepts the longest verified prefix, and the remaining suffix is discarded [2604.27747]. “HistoryPalette: Supporting Exploration and Reuse of Past Alternatives in Image Generation and Editing” similarly describes a workflow where creators preview alternatives, accept some by reuse or restoration, and retain unchosen ones in history rather than deleting them outright [2501.04163]. In stochastic control with disturbance preview, the PAD lens is interpretive rather than terminological, but the paper explicitly links finite-horizon preview to principled retention or discarding of future information based on its diminishing closed-loop influence [2412.06662].

A plausible implication is that PAD is most coherent when treated as an abstract interaction or decision structure with three functional stages: tentative exposure of an option, selective commitment, and explicit non-commitment. The exact semantics of “discard,” however, vary substantially across domains. In speculative decoding it means rejecting drafted suffix tokens [2604.27747]; in HCI it means canceling a pending action [2511.10532]; in creative tooling it often means archive or deprioritize rather than hard deletion [2501.04163].

## 2. Speculative decoding and generative recommendation

The most detailed algorithmic mapping from PAD to machine inference appears in “Position-Aware Drafting for Inference Acceleration in LLM-Based Generative List-Wise Recommendation” [2604.27747]. That paper does **not** define PAD as “Preview Accept Discard”; its method is called **PAD-Rec**, short for **Position-Aware Drafting for generative Recommendation**. Nonetheless, it repeatedly describes the standard speculative decoding loop in PAD-like terms.

In this setting, the target problem is inference latency in LLM-based generative list-wise recommendation. Each recommended item is represented by a fixed-length tuple of semantic-ID tokens, typically \(K=4\), often separated by commas or spaces, so recommendation generation requires many autoregressive steps [2604.27747]. The paper formalizes the flattened generation objective as
\[
p_\theta\big(X_{t_0:T}\mid X_{1:t_0-1}\big) = \prod_{t=t_0}^{T} p_\theta\big(x_t \mid X_{1:t-1}\big),
\]
where each \(x_t\) is either a semantic-ID token or a separator [2604.27747].

The PAD reading of the algorithm is direct. The **preview** is the candidate block or tree proposed by a smaller draft model up to speculation depth \(B\). The **accept** step is target-side verification, which “commits the longest accepted prefix.” The **discard** step removes the remaining drafted suffix after the first mismatch [2604.27747]. The paper explicitly states that the target LLM “verifies all candidates in one batch and commits the longest accepted prefix, advancing the output by multiple tokens per verification round” [2604.27747].

PAD-Rec’s contribution is to improve the preview stage rather than altering accept/reject logic. It adds two signals to the draft model: **Item Position Embeddings (IPE)**, encoding within-item slot identity, and **Step Position Embeddings (SPE)**, encoding speculation depth [2604.27747]. The key claim is that recommendation outputs are structured rather than token-homogeneous: a token’s meaning depends on whether it occupies slot 1, 2, 3, 4, a separator position, or context, and uncertainty increases with draft depth [2604.27747]. The within-item label is
\[
\ell_t \in \{1,\dots,K,\text{sep},\text{ctx}\},
\]
and the step embedding is indexed by draft depth \(j\) [2604.27747].

The paper further introduces lightweight gating. Item-slot information is fused through a learnable scalar \(g_{\text{item}}\), while step information uses a context-dependent gate
\[
g_{\text{step}(t)}=\sigma\!\left(\mathbf{w}^{\top}\mathbf{z}_{t-1}\right),
\]
so step-position influence can vary with local draft context [2604.27747]. This yields a draft-side distribution
\[
P^{(s)}_{\theta,\,j}\left(x_{t}\mid x_{< t}\right)
=\mathrm{Head}\Big( \mathcal{M}_{\theta}\big( \mathrm{Fuse}(\mathbf{f}_{< t},\,\mathbf{v}_{< t},\,\mathbf{s}_{j}) \big) \Big),
\]
with a trainable draft backbone and frozen LM head copied from the target [2604.27747].

Experimentally, PAD-Rec is described as a **lossless inference acceleration technique** that changes only the draft side while preserving target verification and therefore the target distribution [2604.27747]. On Beauty, Instruments, Games, and Yelp, it reports up to **3.1× wall-clock speedup** and about **5% average wall-clock speedup gain over strong SD baselines** [2604.27747]. The acceptance-side metric is the average accepted length \(\tau\), which directly quantifies how much of the speculative preview survives target verification [2604.27747]. A practical implication is that, in this literature, PAD is best read not as a new decision rule but as a refinement of preview quality under a fixed longest-prefix accept/discard protocol.

## 3. Utility-aware relaxation in speculative decoding

A distinct reinterpretation appears in “Reject Only Critical Tokens: Pivot-Aware Speculative Decoding,” whose acronym PAD means **Pivot-Aware Speculative Decoding**, not “Preview Accept Discard” [2511.00351]. This paper is relevant because it relaxes the acceptance criterion itself. Standard speculative decoding preserves the exact target-model distribution, but the paper argues that exact distribution matching is often too stringent and yields unnecessarily low acceptance rates [2511.00351].

The central shift is from distribution preservation to expected-utility preservation. It defines a binary utility function
\[
u(y, x) =
\begin{cases}
1, & \text{if } \text{Eval}(y, x) \geq \theta_{\text{eval}}, \\
0, & \text{otherwise},
\end{cases}
\]
and expected utility
\[
U(p,x) = \mathbb{E}_{\mathbf{y} \sim p(\cdot \mid \mathbf{x})} \left[u(\mathbf{y}, \mathbf{x})\right].
\]
A decoding strategy is \(\epsilon\)-utility preserving if its expected utility remains within \(\epsilon\) of the target model’s expected utility over the task distribution [2511.00351].

The paper introduces **pivot tokens**, defined as tokens whose forced acceptance reduces the target model’s downstream expected utility by at least \(\epsilon\) [2511.00351]. Rather than rejecting every token that standard speculative decoding would reject, the proposed algorithm rejects only those predicted to be pivotal. Non-pivotal tokens may be accepted despite standard SD disagreement, thereby increasing the acceptance ratio \(\eta=\tau/\gamma\) [2511.00351]. The practical inference procedure is still PAD-like in the informal sense: the system previews draft tokens with target-side verification features, accepts tokens that are either standard-SD-acceptable or predicted non-pivotal, and discards or recomputes only critical ones [2511.00351].

A key theoretical statement is that if the pivot classifier has **100% recall on pivot tokens**, then PAD is utility preserving with \(\epsilon=0\) [2511.00351]. This is explicitly one-sided: false positives hurt speed only, false negatives can hurt utility. Empirically, the paper reports **AUC = 0.865** for the pivot classifier and up to **2.5× speedup with comparable utility** [2511.00351]. The results are task-sensitive. On GSM8K, PAD at \(\sigma=0.7\) reaches acceptance ratio \(77.2\%\) and **2.46×** speedup with small accuracy loss relative to target-only decoding, whereas on AIME24 aggressive thresholds incur larger utility degradation [2511.00351].

This introduces a substantive controversy around PAD-like workflows in inference. One line of work treats accept/discard as correctness-preserving and immutable, as in lossless speculative decoding [2604.27747]. Another treats those boundaries as task-dependent and optimizable under utility rather than exact distribution [2511.00351]. The disagreement is methodological rather than terminological.

## 4. Predictive low-motion interaction in human–computer interaction

The HCI paper “Preview, Accept or Discard? A Predictive Low-Motion Interaction Paradigm” is the clearest case where PAD is a formal interaction design rather than an interpretive abstraction [2511.10532]. It proposes a predictive, keyboard-based, zero-click method for GUI target selection intended to reduce fine-motor motion relative to mouse or trackpad input [2511.10532].

The system predicts likely GUI targets from the current interface state, described abstractly as
\[
T \leftarrow Predict(DOM),
\]
ranks the top-\(N\) candidates, and lets the user interact through a press-hold-preview-release grammar [2511.10532]. While the designated key chord is held, the top-ranked candidate is previewed via an on-screen curved chord overlay. Cycling through alternatives is defined by
\[
i \leftarrow (i \bmod N)+1,
\]
triggered by a spacebar press while PAD mode is active [2511.10532]. Acceptance is encoded by nearly simultaneous key release; discard is encoded by sequential release, with a default **170 ms differential release window** separating the two cases [2511.10532].

The full interaction loop is summarized in the paper’s pseudocode: on GUI update the system predicts and ranks targets, during PAD mode it previews the current candidate, and on keyboard update it either advances to the next target, accepts the current one via `ClickOn(T[i])`, or discards it via `HideChord()` [2511.10532]. All prototype processing is client-side, and the UI includes **200 ms** default ease-in-out animations for accept/discard transitions [2511.10532].

The paper evaluates PAD in two settings. In a browser-based email client, predictions were hard-coded and always correct to isolate interaction properties from model error [2511.10532]. PAD reduced physical clicks to zero, and median travel distance saved was about **3,000 pixels per user for every five accepted chord suggestions on average**, corresponding to about **600 fewer pixels per replaced click** [2511.10532]. In an adapted ISO 9241-9 task, the crucial variable was not merely top-3 coverage but rank distribution quality. Under **PAD@100%(95-4-1)**—95% correct top-1, 4% top-2, 1% top-3—throughput was **4.8 bps** versus **4.2 bps** for trackpad. Under **PAD@100%(uniform)**—33%-33%-33%—throughput dropped to **2.7 bps** [2511.10532].

The main conclusion is that PAD substantially reduces hand motion, but time competitiveness with trackpad appears only when prediction quality is comparable to “the best spell-checkers” [2511.10532]. This is important because it shows that PAD is not inherently a speed-up paradigm. It is a movement-reduction paradigm whose practicality depends on how often the previewed top-ranked option is already correct.

## 5. Creative tooling, archived alternatives, and non-destructive discard

“HistoryPalette: Supporting Exploration and Reuse of Past Alternatives in Image Generation and Editing” does not name its system PAD, but the paper explicitly describes Preview–Accept–Discard as the dominant interaction pattern embodied by the interface [2501.04163]. The context is generative image creation, where creators rapidly generate, compare, reject, and later reuse many alternatives. Here PAD is not an algorithmic verifier but a history-management and exploration framework.

HistoryPalette organizes alternatives through five surfaces: **Generation View**, **Position Palette**, **Concept Palette**, **Time Palette**, and **Version Timeline** [2501.04163]. A user selects a canvas region, enters a prompt, and the system runs a Stable Diffusion XL + ControlNet inpainting pipeline four times to produce **four generated alternatives per request** [2501.04163]. Alternatives may be previewed in place on hover and accepted by reuse, restoration, paste, or rasterization. The system stores both full-project versions and partial-project alternatives defined by region, image, and prompt [2501.04163].

The paper’s most PAD-relevant claim is that discarded alternatives are generally not deleted. “All discarded and selected alternatives from the generation view are added to the palettes of alternatives for later use,” and “all filtering and removal are non-destructive such that users can recover filtered alternatives in the interface” [2501.04163]. This shifts the meaning of discard from destruction to archive or deprioritization. Failed generations identified by GPT-4V are hidden by default but remain recoverable and can be surfaced with lower opacity [2501.04163].

Preview is the dominant operation. Users hovered between **179 and 781** times per session, far more often than they committed to alternatives [2501.04163]. For creative professionals, concept palettes reduced the search space by up to **11×** and position palettes by **7×** relative to browsing the full set of alternatives [2501.04163]. Acceptance is graded rather than atomic: users can preview an old alternative, move or resize it, and then either **paste** it directly or **rasterize** it through an image-to-image refiner. In study usage, **10 of 12 participants used paste more often than rasterize** [2501.04163].

This domain complicates simplistic PAD terminology. The “discard” stage is rarely final. A previously rejected alternative may later become contextually appropriate after other parts of the composition change. The paper therefore supports a stronger general proposition: in creative systems, PAD often means **Preview–Accept–Archive/Deprioritize** rather than Preview–Accept–Delete [2501.04163].

## 6. Related PAD-like mechanisms in control, statistics, and verification

Several additional papers support PAD as a broader decision pattern even where the acronym denotes something else.

In “Stochastic LQR Design With Disturbance Preview,” the controller has finite preview of future disturbances and the optimal policy includes an anticausal feedforward term
\[
u_t^*=-K_xx_t-K_ww_t-K_vv_{t+1},
\]
with
\[
v_{t+1}= \sum_{j=t+1}^{t+p} (\hat A^\top)^{\,j-t}PB_ww_j
\]
in the infinite-horizon LTI case [2412.06662]. The paper explicitly interprets this as a mathematically principled way to decide which future disturbance information should be “accepted” into the control law and which can be discarded because its effect decays through powers of the stabilized closed-loop matrix \(\hat A\) [2412.06662]. The exact residual cost gap to full noncausal preview is
\[
J_{\infty,p}^*-J_{\infty,nc}^*
= \operatorname{trace}\!\left[ Y(\hat A^\top)^{p+1}X\hat A^{p+1} \right],
\]
which yields a quantitative discard criterion based on tolerance [2412.06662].

In online multiple testing, “ADDIS: an adaptive discarding algorithm for online FDR control with conservative nulls” defines a three-zone rule over \(p\)-values using rejection thresholds \(\alpha_t\), candidate thresholds \(\lambda_t\), and discarding thresholds \(\tau_t\) [1905.11465]. It is not framed as Preview–Accept–Discard, but the paper’s own synthesis maps naturally onto that interpretation: compare the new \(p\)-value to \(\tau_t\) as a screening step, discard it if \(P_t>\tau_t\), and otherwise continue with candidate and rejection logic [1905.11465]. The core estimator is
\[
\widehat{FDP}_{\textnormal{ADDIS}(t)}
= \frac{\sum_{j \leq t}\alpha_j\frac{1\{\lambda_j < P_j \leq \tau_j\}}{\tau_j-\lambda_j}}{|R(t)| \vee 1},
\]
and the method is designed to improve power when null \(p\)-values are conservative by explicitly discarding very large \(p\)-values [1905.11465].

In fingerprint biometrics, “Improving fingerprint presentation attack detection by an approach integrated into the personal verification stage” proposes a **Closeness Binary Code (CC)** add-on that refines PAD decisions during verification by comparing PAD embeddings with the enrolled templates of the claimed user [2504.11066]. The code
\[
B(x) = \{b_f,b_p,b_o\}
\]
summarizes same-finger, same-person, and generic-user closeness in the PAD feature space [2504.11066]. The paper remains binary in its final PAD decision, but it suggests a multi-stage operational point where suspicious cases could be blocked or escalated before matcher invocation. This suggests that PAD-like workflows are also relevant in biometric verification, though the paper itself does not formalize Preview–Accept–Discard as a named protocol [2504.11066].

## 7. Ambiguity, misreadings, and competing meanings of PAD

A major source of confusion is that PAD is polysemous across fields. Several papers in the supplied corpus explicitly deny the Preview–Accept–Discard expansion. “Pose-Aware Diffusion for 3D Generation” states that PAD stands for **Pose-Aware Diffusion**, not Preview/Accept/Discard [2605.00345]. “Pivot-Aware Speculative Decoding” states that PAD means **Pivot-Aware Speculative Decoding**, though an informal preview/accept/discard reading is “not entirely misleading at the algorithmic level” [2511.00351]. In face biometrics, PAD overwhelmingly denotes **Presentation Attack Detection**, as in self-supervised face anti-spoofing [2208.13070], continual PAD with VLP models [2512.19022], and concept-guided PAD with CLIP prompts [2607.01303]. Earlier machine-learning literature also uses PAD for **Prior Activation Distribution** [1907.02711], and environmental economics uses PAD for **Protected Area Downgrading** [2308.16397].

This multiplicity creates a recurrent misconception: that PAD denotes a single method family. The literature does not support that interpretation. The only paper in the supplied set that formally introduces PAD as Preview–Accept–Discard is the predictive low-motion HCI work [2511.10532]. Elsewhere, PAD is either another acronym or a descriptive overlay applied to an algorithmic process.

A second misconception is that PAD always implies hard rejection of unchosen options. Creative-interface work contradicts this directly, since rejected alternatives remain accessible and often useful later [2501.04163]. A plausible implication is that “discard” should be interpreted relative to the domain’s persistence requirements: irreversible in some inference loops, provisional in exploratory workflows, and asymptotically negligible in preview control.

## 8. Significance and general interpretation

Taken together, the literature suggests that Preview–Accept–Discard is best understood as a reusable architectural motif for systems that must expose uncertainty or tentative action without surrendering control. In machine inference, the preview stage often reduces expensive target-model calls by proposing likely continuations; the accept step preserves exactness or utility; the discard step bounds error propagation [2604.27747], [2511.00351]. In HCI, preview externalizes a predicted action before commitment, thereby preserving agency in AI-assisted interaction [2511.10532]. In creative systems, preview supports contextual re-evaluation of previously rejected alternatives, while discard becomes reversible history management [2501.04163].

This suggests three stable principles behind PAD-like systems. First, **preview is a low-cost exposure of a tentative future state**. Second, **acceptance is rarely equivalent to mere prediction**; it is typically gated by verification, human confirmation, or a threshold rule. Third, **discard is domain-dependent**, ranging from immediate suffix rejection to archival retention or tolerance-based truncation of negligible future influence.

The most conservative reading of the research record is therefore that PAD is not a unified technical field but a powerful cross-domain schema. Where it is formalized, as in predictive low-motion interaction, it denotes a concrete interaction grammar [2511.10532]. Where it is interpretive, as in speculative decoding, control, and creative reuse, it provides a concise way to describe how systems stage tentative options before commitment or rejection [2604.27747], [2412.06662], [2501.04163]. That breadth is the principal reason the term persists despite acronymic overload in neighboring literatures.

Source: https://www.emergentmind.com/topics/preview-accept-discard-pad