Archer: Adaptive Reuse of Cached Hidden States for Efficient Rollback in Diffusion Language Models

# Archer: Adaptive Reuse of Cached Hidden States for Efficient Rollback in Diffusion Language Models
Source: [https://arxiv.org/html/2608.08086](https://arxiv.org/html/2608.08086)
Xuning He1,2, Zinan Sheng3, Yongding Tao3, Huanyu Liu3, Ge Li3, Xue Jiang3, Yihong Dong1 1School of Computer Science, Shanghai Jiao Tong University 2College of Artificial Intelligence, Nankai University 3School of Computer Science, Peking University hxning@mail\.nankai\.edu\.cndongyh@sjtu\.edu\.cn

###### Abstract

Diffusion language models \(DLMs\) iteratively refine a sequence, allowing earlier predictions to be revised as context evolves\. This rollback capability distinguishes them from irreversible autoregressive generation, but makes inference costly\. Every denoising update alters the global context, forcing both prompt and response states to be recomputed even though only response tokens are revisable\. Key\-value \(KV\) caching could reduce this cost, yet conventional caching assumes immutable historical states and is therefore difficult to reconcile with rollback\. In this paper, we introduceAdaptiveReuse ofCachedHidden States forEfficientRollback \(Archer\), a training\-free KV caching method for rollback\-capable DLMs\. Archer asymmetrically keeps the mutable response synchronized with the current hypothesis while reusing prompt K/V within a bounded state neighborhood\. Although prompt representations also change under bidirectional attention, their token identities remain fixed; bounded reuse therefore amortizes repeated prompt computation without caching mutable response states\. It also delays feedback from tentative tokens, reducing premature reinforcement of transient high\-confidence errors and giving rollback more opportunity to correct them\. Our analysis characterizes prompt reuse as a reversibility\-aligned cache boundary, bounds its state\-dependent approximation error, and gives a decoder\-margin condition for preserving full\-refresh decisions\. Existing DLM acceleration often trades quality for speed\. Archer shifts this frontier, attaining the best mean performance of33\.63%33\.63\\%together with a2\.57×2\.57\\timesmean speedup on the main suite\. Across evaluated settings, it improves Pass@1 by up to3\.053\.05points and reaches up to2\.95×2\.95\\timesspeedup\. Controlled analyses connect the quality gain to delayed prompt feedback and validate state\-aware refresh\. Our code is available at[https://github\.com/Hxnng/Archer](https://github.com/Hxnng/Archer)\.

## 1Introduction

Diffusion language models \(DLMs\) generate through iterative, bidirectional denoising rather than an irreversible left\-to\-right factorization\(Austinet al\.,[2021a](https://arxiv.org/html/2608.08086#bib.bib1), Liet al\.,[2022](https://arxiv.org/html/2608.08086#bib.bib11), Sahooet al\.,[2024](https://arxiv.org/html/2608.08086#bib.bib2)\)\. This evolving state makes rollback possible: an earlier prediction can be reconsidered when later context reveals an inconsistency\. Recent rollback\-capable decoders show that this flexibility is particularly valuable for structured generation\(Wanget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib13), Honget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib18), Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\)\. Yet rollback requires many full\-sequence forward passes, and the cost grows rapidly with the prompt and output length\. Once rollback is used at scale, KV caching becomes necessary; the central question is how to introduce it without weakening the ability to revise the generation\.

The same rollback that makes DLMs attractive also breaks the premise behind conventional KV caching\. In autoregressive decoding, causal attention keeps previous contexts fixed, so their K/V states remain reusable\(Popeet al\.,[2023](https://arxiv.org/html/2608.08086#bib.bib7), Kwonet al\.,[2023](https://arxiv.org/html/2608.08086#bib.bib19)\)\. Under bidirectional attention, changing one generated token alters both the generation states and the prompt states that attend to it\. Full recomputation preserves the intended rollback process but repeatedly processes an unchanged prompt; caching the entire sequence saves this work but retains activations derived from a generation that may no longer exist\. Moreover, an eager prompt update immediately feeds every tentative token back into later predictions, which may reinforce a transient error before subsequent context can correct it\. The challenge is therefore to locate a cache boundary that saves computation while leaving every revisable generation state current\.

![Refer to caption](https://arxiv.org/html/2608.08086v1/figures/figure1.png)Figure 1:Archer reuses prompt K/V while recomputing the revisable response, reducing latency and improving Pass@1 on MBPP\.Existing methods improve either rollback quality or DLM efficiency, but do not resolve their interaction\. Token\-level rollback decoders improve quality but typically repeat full\-sequence computation\(Wanget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib13), Honget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib18), Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\)\. DLM accelerators reduce computation through blockwise generation, parallel token acceptance, delayed KV updates, or selective recomputation\(Arriolaet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib16), Wuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib8), Maet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib9), Liuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib10)\), but are designed for decoding without token\-level rollback\. Their generation caches must either be frequently invalidated or retain states derived from tokens that have already changed, so efficient reuse and unrestricted rollback remain at odds\. A method for scalable rollback must make this trade\-off explicit rather than assume that conventional cache reuse transfers unchanged\.

We introduceAdaptiveReuse ofCachedHidden States forEfficientRollback \(Archer\) to make this interaction explicit\. Archer caches only prompt K/V states and recomputes the complete generation at every step, preserving rollback for every generated token\. It refreshes the prompt cache when the current generation has moved sufficiently far from the state at which the cache was created\. This state\-aware policy amortizes prompt computation while preventing unbounded staleness\. Bounded reuse also forms a temporal anchor that delays feedback from tentative tokens and can reduce premature error reinforcement\. Our analysis identifies prompt states as a cache boundary aligned with rollback, derives the computational gain and a state\-dependent approximation bound, and gives a decoder\-margin condition under which Archer preserves the rollback decision of full recomputation\. Archer thus turns the tension between caching and rollback into a controlled lag\-and\-reset process that alternates local reuse with global synchronization\.

Across the main benchmarks, Archer breaks the usual quality–speed trade\-off, achieving the best average performance at33\.63%33\.63\\%, a2\.57×2\.57\\timesaverage speedup, and up to2\.95×2\.95\\timesspeedup on a single benchmark\. Across backbones, it improves Pass@1 by up to3\.053\.05points and accelerates every tested pair by1\.361\.36–1\.78×1\.78\\times\. Relative to Saber, it improves Pass@1 on the original MBPP and LiveCodeBench suites as well as on the MBPP\-ET and HumanEval\-ET versions, and reduces latency on all three benchmarks\. Controlled analyses support delayed prompt feedback and state\-aware refresh\. We1\)formulate the conflict between KV caching and rollback and identify prompt states as the appropriate boundary;2\)propose Archer with state\-aware refresh and characterize its efficiency, approximation error, and decision fidelity; and3\)show that rollback\-compatible caching moves the DLM quality–speed frontier rather than forcing a choice between the two\.

## 2Motivation

Rollback is a defining advantage of DLM decoding over an irreversible left\-to\-right process\(Wanget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib13), Honget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib18), Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\)\. It allows the current output to remain provisional until later context resolves earlier uncertainty, but every change is followed by another bidirectional forward pass over an increasingly long sequence\. As prompts and outputs scale, repeatedly recomputing the entire context is no longer practical, so KV caching becomes necessary for rollback to remain usable\. The design problem is to obtain this reuse without turning a revisable generation into fixed history\. The resulting systems question is not whether to cache, but where reuse can coexist with unrestricted revision inside the same decoder\.

![Refer to caption](https://arxiv.org/html/2608.08086v1/figures/figure2.png)Figure 2:Effect of prompt\-cache radiusKK\. Longer reuse reduces latency, while quality peaks at moderateKKand declines with excessive staleness\.The cache boundary must follow the semantics of rollback rather than the convenience of implementation\. Generation states represent precisely the part of the sequence that may change, and the affected positions are not known in advance\. At steptt, the rollback set is determined from the current context\-dependent logits, which can be summarized asℛt=ℬ​\(Fθ​\(p,gt\)\)\\mathcal\{R\}\_\{t\}=\\mathcal\{B\}\(F\_\{\\theta\}\(p,g\_\{t\}\)\)\. If the unreliable tokens were already known before evaluatingFθF\_\{\\theta\}, the decoder could remove them directly; rollback is needed because their reliability must first be reassessed as context evolves\. Generation caching therefore creates a circular dependency\. Determining which cached K/V and logits remain valid requires the same bidirectional generation computation that caching is intended to avoid, and one revised token can invalidate every state that depends on it\. Exact generation\-side reuse consequently approaches full recomputation, while approximate reuse introduces assumptions that may restrict rollback\. Preserving unrestricted rollback therefore keeps the mutable generation outside the persistent cache and recomputes it before every decoding decision\.

Prompt states provide the boundary that satisfies both requirements\. Their values depend on the current output, but their token identities remain fixed, so reuse introduces contextual delay without making any generated token permanent\. Archer therefore recomputes the complete generation at every step and reuses only prompt K/V\. The eager feedback loop is

gt⟶𝒞p​\(gt\)⟶zt\+1⟶gt\+1,g\_\{t\}\\longrightarrow\\mathcal\{C\}\_\{p\}\(g\_\{t\}\)\\longrightarrow z\_\{t\+1\}\\longrightarrow g\_\{t\+1\},\(1\)where𝒞p​\(gt\)\\mathcal\{C\}\_\{p\}\(g\_\{t\}\)is the prompt state induced by the current generation\. Replacing it temporarily with𝒞p​\(gr\)\\mathcal\{C\}\_\{p\}\(g\_\{r\}\)removes repeated prompt computation and delays the feedback of tentative tokens\. The former improves speed; the latter gives rollback more opportunity to correct transient high\-confidence errors before they reinforce themselves\. This asymmetry follows the semantics of revision rather than an arbitrary implementation choice\.

The cache lifetime controls whether this boundary yields a useful trade\-off\. Figure[2](https://arxiv.org/html/2608.08086#S2.F2)shows that increasing the reuse radiusKKconsistently reduces latency because prompt refreshes become less frequent\. Quality is non\-monotonic\. A very small radius behaves similarly to eager execution and provides little reuse or temporal separation\. A moderate radius delays premature feedback while keeping the prompt sufficiently representative of the current generation\. An excessively large radius withholds useful context for too long, allowing approximation error to dominate\. The rise\-and\-fall in quality shows that cache reuse must balance temporal anchoring against synchronization rather than maximize either one in isolation\.

Rollback and KV caching become compatible when reuse follows what the decoder is allowed to change\. Prompt\-only reuse keeps the generation current, delays feedback for a bounded interval, and restores the full context through synchronization\. Archer operationalizes this boundary as the state\-anchored lag\-and\-reset policy studied below\.

## 3Related Work

### 3\.1Rollback in Diffusion Language Models

DLMs replace left\-to\-right factorization with iterative denoising in continuous embeddings or discrete token spaces\(Liet al\.,[2022](https://arxiv.org/html/2608.08086#bib.bib11), Austinet al\.,[2021a](https://arxiv.org/html/2608.08086#bib.bib1)\)\. Advances in discrete objectives have substantially improved their modeling quality\(Louet al\.,[2024](https://arxiv.org/html/2608.08086#bib.bib15), Sahooet al\.,[2024](https://arxiv.org/html/2608.08086#bib.bib2), Liet al\.,[2024](https://arxiv.org/html/2608.08086#bib.bib12)\), and recent systems such as LLaDA, Dream, and DiffuCoder have scaled the paradigm to general language modeling and code generation\(Nieet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib3), Yeet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib4), Gonget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib5)\)\. Their changing intermediate states are not merely a sampling detail; they provide the rollback capability that motivates our systems design\.

Recent decoders increasingly exploit rollback through flexible sampling\. ReMDM derives a remasking transition, RemeDi learns to identify unreliable predictions, and WINO and Saber revisit token\-level decisions as context evolves\(Wanget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib13), Huanget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib20), Honget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib18), Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\)\. These token\-level methods generally process the revised sequence with another full forward pass\. Reversible Diffusion Decoding instead returns to earlier blocks using cached block states\(Wanget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib21)\), but does not address repeated prompt computation during fine\-grained revision inside a bidirectional generation region\. Archer treats that interaction as a systems constraint\. Because generated tokens may change again, it reuses only fixed prompt states, leaving the sampling policy and its correction mechanism unchanged\.

### 3\.2Efficient DLM Inference

Efficient DLM inference reduces either denoising steps or their per\-step cost\. Fast\-dLLM, EB\-Sampler, WINO, and Saber resolve multiple positions per pass according to confidence, entropy, or state changes\(Wuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib8), Ben\-Hamuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib17), Honget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib18), Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\)\. These methods shorten the trajectory, but each remaining pass still processes the full bidirectionally coupled sequence as prompts and outputs grow\. The bottleneck therefore shifts from how many iterations are executed to how much repeated context each surviving iteration processes\.

KV caching is exact for the immutable history of causal decoding\(Popeet al\.,[2023](https://arxiv.org/html/2608.08086#bib.bib7), Kwonet al\.,[2023](https://arxiv.org/html/2608.08086#bib.bib19)\), and fixed prompt modules can be shared across requests\(Gimet al\.,[2024](https://arxiv.org/html/2608.08086#bib.bib14)\)\. For DLMs, Block Diffusion creates cacheable semi\-autoregressive structure during training\(Arriolaet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib16)\), while training\-free methods use prefix or dual caches, delayed token\-level updates, and similarity\-guided partial recomputation\(Wuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib8), Maet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib9), Liuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib10)\)\. These approaches target conventional denoising and do not treat rollback as a cache\-invalidation event\. Directly introducing them into a rollback decoder can therefore retain generation\-side computation from tokens later re\-masked or replaced\. Archer derives the cache boundary from rollback semantics\. It amortizes prompt computation while recomputing every revisable generation state, so reuse neither commits a provisional token nor alters rollback semantics\.

## 4Archer

Archer realizes the preceding design as a training\-free cache controller\. It reuses prompt K/V, recomputes the complete generation region, and refreshes the cache according to response\-state drift, while leaving the DLM and rollback policy unchanged\. Figure[3](https://arxiv.org/html/2608.08086#S4.F3)makes the division explicit: prompt states are reused, while generation states remain revisable\.

### 4\.1Asymmetric State Reuse

Letp∈𝒱Pp\\in\\mathcal\{V\}^\{P\}be the prompt,gt∈\(𝒱∪\{𝙼𝙰𝚂𝙺\}\)Gg\_\{t\}\\in\(\\mathcal\{V\}\\cup\\\{\\mathtt\{MASK\}\\\}\)^\{G\}the current response, and𝒟\\mathcal\{D\}the rollback decoder\. Any auxiliary sampler state is suppressed for notation\. We write𝒞p​\(g\)\\mathcal\{C\}\_\{p\}\(g\)for the prompt K/V obtained under responseggandΦθ​\(g;𝒞\)\\Phi\_\{\\theta\}\(g;\\mathcal\{C\}\)for the generation forward using prompt cache𝒞\\mathcal\{C\}\. Eager decoding evaluates

zt⋆=Φθ​\(gt;𝒞p​\(gt\)\),gt\+1=𝒟​\(gt,zt⋆\)\.z\_\{t\}^\{\\star\}=\\Phi\_\{\\theta\}\(g\_\{t\};\\mathcal\{C\}\_\{p\}\(g\_\{t\}\)\),\\qquad g\_\{t\+1\}=\\mathcal\{D\}\(g\_\{t\},z\_\{t\}^\{\\star\}\)\.\(2\)
Suppose the current cache was created at responsegrg\_\{r\}\. Archer replaces the eager logits with

z^t=Φθ​\(gt;𝒞p​\(gr\)\)\.\\hat\{z\}\_\{t\}=\\Phi\_\{\\theta\}\(g\_\{t\};\\mathcal\{C\}\_\{p\}\(g\_\{r\}\)\)\.\(3\)The generation path is otherwise fresh\. At layerℓ\\ell, its queries attend to anchored prompt K/V and current generation K/V,

Attn⁡\(Qgℓ​\(gt\),\[Kpℓ​\(gr\)∥Kgℓ​\(gt\)\],\[Vpℓ​\(gr\)∥Vgℓ​\(gt\)\]\)\.\\operatorname\{Attn\}\\\!\\left\(Q\_\{g\}^\{\\ell\}\(g\_\{t\}\),\[K\_\{p\}^\{\\ell\}\(g\_\{r\}\)\\\|K\_\{g\}^\{\\ell\}\(g\_\{t\}\)\],\[V\_\{p\}^\{\\ell\}\(g\_\{r\}\)\\\|V\_\{g\}^\{\\ell\}\(g\_\{t\}\)\]\\right\)\.\(4\)Hence every generation embedding, hidden state, and logit reflectsgtg\_\{t\}; only response\-to\-prompt feedback remains anchored atgrg\_\{r\}\.

![Refer to caption](https://arxiv.org/html/2608.08086v1/figures/figure3.png)Figure 3:Archer’s state\-anchored prompt cache\. Response states are recomputed at every update, and prompt K/V is refreshed whendH​\(gt,gr\)≥Kd\_\{H\}\(g\_\{t\},g\_\{r\}\)\\geq K\.
### 4\.2State\-Anchored Refresh

Archer measures cache validity by the response change accumulated since the most recent refresh\. The resulting anchor\-relative distance is

Dt=dH​\(gt,gr\)=∑i=1G𝕀​\[gt,i≠gr,i\]D\_\{t\}=d\_\{H\}\(g\_\{t\},g\_\{r\}\)=\\sum\_\{i=1\}^\{G\}\\mathbb\{I\}\[g\_\{t,i\}\\neq g\_\{r,i\}\]\(5\)
Archer refreshes whenDt≥KD\_\{t\}\\geq K\. A refresh performs a full forward, replaces the prompt cache, and setsgr←gtg\_\{r\}\\leftarrow g\_\{t\}\. Otherwise Archer evaluates Eq\. \([3](https://arxiv.org/html/2608.08086#S4.E3)\)\. This makes synchronization depend on accumulated response drift rather than on the number of updates alone\. The logits passed to the original decoder are therefore

z~t=\{Φθ​\(gt;𝒞p​\(gt\)\),Dt≥K,Φθ​\(gt;𝒞p​\(gr\)\),Dt<K,gt\+1=𝒟​\(gt,z~t\)\.\\tilde\{z\}\_\{t\}=\\begin\{cases\}\\Phi\_\{\\theta\}\(g\_\{t\};\\mathcal\{C\}\_\{p\}\(g\_\{t\}\)\),&D\_\{t\}\\geq K,\\\\ \\Phi\_\{\\theta\}\(g\_\{t\};\\mathcal\{C\}\_\{p\}\(g\_\{r\}\)\),&D\_\{t\}<K,\\end\{cases\}\\qquad g\_\{t\+1\}=\\mathcal\{D\}\(g\_\{t\},\\tilde\{z\}\_\{t\}\)\.\(6\)The controller responds to net state drift rather than elapsed iterations\. Several negligible updates may continue to reuse an anchor, while one large rollback can trigger immediate synchronization\.

Algorithm[1](https://arxiv.org/html/2608.08086#alg1)gives the complete procedure\.FullForwardreturns exact logits and a new prompt cache;CachedForwardreconstructs the current generation using that cache\. Token acceptance, replacement, re\-masking, and termination remain entirely governed by𝒟\\mathcal\{D\}\.

Algorithm 1State\-anchored rollback decoding with Archer0:Prompt

pp, initial response

g0g\_\{0\}, rollback decoder

𝒟\\mathcal\{D\}, radius

KK
1:

g←g0g\\leftarrow g\_\{0\};

gr←g0g\_\{r\}\\leftarrow g\_\{0\}
2:

\(z,𝒞r\)←FullForward​\(p,g\)\(z,\\mathcal\{C\}\_\{r\}\)\\leftarrow\\textsc\{FullForward\}\(p,g\)
3:while

¬Terminated​\(g\)\\neg\\textsc\{Terminated\}\(g\)do

4:

g←𝒟​\(g,z\)g\\leftarrow\\mathcal\{D\}\(g,z\)
5:if

¬Terminated​\(g\)\\neg\\textsc\{Terminated\}\(g\)then

6:if

dH​\(g,gr\)≥Kd\_\{H\}\(g,g\_\{r\}\)\\geq Kthen

7:

\(z,𝒞r\)←FullForward​\(p,g\)\(z,\\mathcal\{C\}\_\{r\}\)\\leftarrow\\textsc\{FullForward\}\(p,g\)
8:

gr←gg\_\{r\}\\leftarrow g
9:else

10:

z←CachedForward​\(g;𝒞r\)z\\leftarrow\\textsc\{CachedForward\}\(g;\\mathcal\{C\}\_\{r\}\)
11:endif

12:endif

13:endwhile

14:return

gg

## 5Theoretical Analysis

Archer keeps the generation synchronized with the current hypothesis while allowing its interaction with the prompt to lag\. Bidirectional dependence makes this asymmetry necessary for rollback, and the resulting state radius links its computational saving to its effect on decoder behavior\.

### 5\.1Caching under Revision

LetΔ0\\Delta^\{0\}contain the generation positions changed by rollback, and letΔℓ\\Delta^\{\\ell\}contain all positions whose layer\-ℓ\\ellstates may depend onΔℓ−1\\Delta^\{\\ell\-1\}\. Dense bidirectional attention gives

Δ0≠∅⟹Δ1=\{1,…,P\+G\}\.\\Delta^\{0\}\\neq\\varnothing\\quad\\Longrightarrow\\quad\\Delta^\{1\}=\\\{1,\\ldots,P\+G\\\}\.\(7\)This is a statement about structural dependence rather than numerical magnitude, but it rules out a universal exact\-reuse guarantee for generation states after an arbitrary edit\.

###### Proposition 1\(Rollback\-compatible cache boundary\)\.

Under arbitrary response revision, exact generation\-state reuse requires recomputing its full bidirectional dependency closure\. Archer instead caches no generation state and therefore preserves every replacement and re\-masking action available to the original decoder\.

Archer avoids retaining invalidated generation computation because every revised token is re\-embedded and propagated through all layers before the next decision\. Prompt staleness may change the selected action, but it cannot make a response position immutable; rollback*capability*is preserved even when cached and eager trajectories differ\.

### 5\.2Efficiency and Fidelity

LetN=P\+GN=P\+G\. At layerℓ\\ell, letαℓ\\alpha\_\{\\ell\}collect projection and feed\-forward work per position andβℓ\\beta\_\{\\ell\}the cost of a query–key interaction\. The leading costs are

Cfull\\displaystyle C\_\{\\rm full\}=∑ℓ=1L\(αℓ​N\+βℓ​N2\),\\displaystyle=\\sum\_\{\\ell=1\}^\{L\}\\left\(\\alpha\_\{\\ell\}N\+\\beta\_\{\\ell\}N^\{2\}\\right\),\(8\)Ccache\\displaystyle C\_\{\\rm cache\}=∑ℓ=1L\(αℓ​G\+βℓ​G​N\)\+H,\\displaystyle=\\sum\_\{\\ell=1\}^\{L\}\\left\(\\alpha\_\{\\ell\}G\+\\beta\_\{\\ell\}GN\\right\)\+H,whereHHis cache overhead\. Archer retains the full receptive field while removing prompt queries, projections, and feed\-forward paths\. IgnoringHH, a cached step costs approximatelyG/\(P\+G\)G/\(P\+G\)of a full step\. With refresh fractionqKq\_\{K\},

Send​\(K\)≈\[qK\+\(1−qK\)​GP\+G\]−1\.S\_\{\\rm end\}\(K\)\\approx\\left\[q\_\{K\}\+\(1\-q\_\{K\}\)\\frac\{G\}\{P\+G\}\\right\]^\{\-1\}\.\(9\)The sequence partition determines the saving per cached step, while the refresh controller determines how often that saving is realized\.

Consider eager and cached logits at the same responsegtg\_\{t\}, where the prompt cache is their only difference\. LetE​\(g\)E\(g\)stack the response embeddings and assume locally that

‖𝒞p​\(g\)−𝒞p​\(g′\)‖≤LC​‖E​\(g\)−E​\(g′\)‖F,‖Φθ​\(g;𝒞\)−Φθ​\(g;𝒞′\)‖∞≤LΦ​‖𝒞−𝒞′‖\.\\begin\{gathered\}\\\|\\mathcal\{C\}\_\{p\}\(g\)\-\\mathcal\{C\}\_\{p\}\(g^\{\\prime\}\)\\\|\\leq L\_\{C\}\\\|E\(g\)\-E\(g^\{\\prime\}\)\\\|\_\{F\},\\\\ \\\|\\Phi\_\{\\theta\}\(g;\\mathcal\{C\}\)\-\\Phi\_\{\\theta\}\(g;\\mathcal\{C\}^\{\\prime\}\)\\\|\_\{\\infty\}\\leq L\_\{\\Phi\}\\\|\\mathcal\{C\}\-\\mathcal\{C\}^\{\\prime\}\\\|\.\\end\{gathered\}\(10\)If the embedding diameter is bounded byBB, responses at Hamming distanceDtD\_\{t\}satisfy‖E​\(gt\)−E​\(gr\)‖F≤B​Dt\\\|E\(g\_\{t\}\)\-E\(g\_\{r\}\)\\\|\_\{F\}\\leq B\\sqrt\{D\_\{t\}\}\. WithΓ=LΦ​LC​B\\Gamma=L\_\{\\Phi\}L\_\{C\}B, every cached step obeys

‖zt⋆−z^t‖∞≤Γ​Dt<Γ​K\.\\\|z\_\{t\}^\{\\star\}\-\\hat\{z\}\_\{t\}\\\|\_\{\\infty\}\\leq\\Gamma\\sqrt\{D\_\{t\}\}<\\Gamma\\sqrt\{K\}\.\(11\)ThusKKbounds the state perturbation that causes cache error rather than the elapsed time since refresh\. Consequently, cache age alone cannot determine validity\. Equally old caches may encode substantially different response changes since their anchors were created\.

The decoder need not preserve every logit; it only needs to preserve the next action\. Define local decision margin as

m𝒟\(g,z\)=infδ\{∥δ∥∞:𝒟\(g,z\+δ\)≠𝒟\(g,z\)\}\.m\_\{\\mathcal\{D\}\}\(g,z\)=\\inf\_\{\\delta\}\\left\\\{\\\|\\delta\\\|\_\{\\infty\}:\\mathcal\{D\}\(g,z\+\\delta\)\\neq\\mathcal\{D\}\(g,z\)\\right\\\}\.\(12\)Archer and eager decoding select same transition whenever

Γ​Dt<m𝒟​\(gt,zt⋆\),\\Gamma\\sqrt\{D\_\{t\}\}<m\_\{\\mathcal\{D\}\}\(g\_\{t\},z\_\{t\}^\{\\star\}\),\(13\)and the equality propagates through the trajectory when the condition holds at every cached step\. Otherwise the trajectories may differ, but Proposition[1](https://arxiv.org/html/2608.08086#Thmproposition1)still preserves revisability\. The relevant guarantee is decision fidelity under bounded state drift, not numerical identity\.

### 5\.3Prompt Reuse as Feedback Control

Prompt reuse also changes the local sensitivity of the decoder\. The eager and anchored logit maps are

zfresh​\(g\)\\displaystyle z^\{\\rm fresh\}\(g\)=Φθ​\(g;𝒞p​\(g\)\),\\displaystyle=\\Phi\_\{\\theta\}\(g;\\mathcal\{C\}\_\{p\}\(g\)\),\(14\)zranchor​\(g\)\\displaystyle z\_\{r\}^\{\\rm anchor\}\(g\)=Φθ​\(g;𝒞p​\(gr\)\)\.\\displaystyle=\\Phi\_\{\\theta\}\(g;\\mathcal\{C\}\_\{p\}\(g\_\{r\}\)\)\.Within a fixed anchor interval, the chain rule gives

Jfresh=∂gΦθ\+∂𝒞Φθ​∂g𝒞p,Janchor=∂gΦθ\.\\begin\{gathered\}J\_\{\\rm fresh\}=\\partial\_\{g\}\\Phi\_\{\\theta\}\+\\partial\_\{\\mathcal\{C\}\}\\Phi\_\{\\theta\}\\,\\partial\_\{g\}\\mathcal\{C\}\_\{p\},\\\\ J\_\{\\rm anchor\}=\\partial\_\{g\}\\Phi\_\{\\theta\}\.\\end\{gathered\}\(15\)Prompt reuse leaves direct generation interaction and rollback intact while temporarily suppressing the cross\-region feedback term∂𝒞Φθ​∂g𝒞p\\partial\_\{\\mathcal\{C\}\}\\Phi\_\{\\theta\}\\,\\partial\_\{g\}\\mathcal\{C\}\_\{p\}\. This constitutes temporal anchoring rather than freezing because current generation interactions remain fresh while only their return path through the prompt is delayed\.

If the cross\-region term amplifies a provisional error direction, delaying it reduces immediate self\-reinforcement and gives rollback additional evidence with which to revise the token\. The same delay harms the next decision when that term carries useful new context\. Refresh restores the full Jacobian and clears the accumulated discrepancy\. Stale prompts are not intrinsically more accurate, but bounded prompt staleness can provide short\-term regularization while periodic synchronization preserves long\-term consistency\.

The radiusKKcontrols the frequency of saved prompt computation, the size of the decision\-relevant approximation, and the duration for which cross\-region feedback is withheld\. Archer thereby balances efficiency, trajectory fidelity, and correction dynamics without caching the mutable object that rollback is designed to revise\.

## 6Experimental Results

Our experiments comprise a main comparison, a cross\-backbone evaluation, and two controlled analyses of delayed prompt feedback and cache refresh\. For a direct comparison with Saber, we follow its code\-generation setting and evaluate on MBPP\(Austinet al\.,[2021b](https://arxiv.org/html/2608.08086#bib.bib23)\), LiveCodeBench\(Jainet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib25)\), and HumanEval\(Chenet al\.,[2021a](https://arxiv.org/html/2608.08086#bib.bib22)\)\. Their executable tests provide a precise measure of whether revisions preserve functional correctness\. This benchmark choice ensures comparability rather than limiting the mechanism to code: Archer observes only generation\-state changes and uses neither code\-specific structure nor execution feedback\. The technical supplement provides the complete protocol, proofs, sensitivity results, and intervention analyses needed to reproduce and interpret this evaluation\.

Table 1:Comparison with DLM sampling and cache baselines\. Base/ET denote Pass@1 \(%\) on the original/Extended Test Cases versions of MBPP and HumanEval; Overall averages five quality scores and three benchmark speedups\. Speedup is relative to LLaDA\-Confidence; bold/underline mark first/second\.### 6\.1Main Results

We compare Archer with LLaDA\(Nieet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib3)\), Fast\-dLLM, dKV\-Cache, and dLLM\-Cache\(Wuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib8), Maet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib9), Liuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib10)\), and the rollback\-capable Saber\(Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\)\. All methods share the model revision, prompts, output length, deterministic decoding, and evaluator\. Archer usesK=11K=11on MBPP,K=10K=10on LiveCodeBench, andK=8K=8on HumanEval\. For MBPP and HumanEval, we report Pass@1 on the original benchmark \(Base\) and on its Extended Test Cases \(ET\) version, which retains the same tasks while adding edge\-case tests\(Donget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib24)\); LiveCodeBench uses its standard Pass@1\. Table[1](https://arxiv.org/html/2608.08086#S6.T1)reports these scores, latency, and speedup relative to LLaDA\-Confidence\.

Archer achieves the strongest overall quality–efficiency balance\. Its33\.63%33\.63\\%average performance is the best result, and its2\.57×2\.57\\timesaverage speedup is second only to aggressive parallel decoding, making it the only method in the top two for both metrics\.

The comparison with Saber isolates the benefit of prompt\-state reuse under the same rollback process\. Archer reduces mean latency from 8\.45 to 6\.20 seconds on MBPP, 17\.75 to 9\.81 seconds on LiveCodeBench, and 12\.13 to 8\.07 seconds on HumanEval, corresponding to1\.36×1\.36\\times,1\.81×1\.81\\times, and1\.50×1\.50\\timesspeedups over Saber\. Quality improves by\+0\.47\+0\.47points on MBPP Base and\+1\.17\+1\.17points on MBPP\-ET, and by\+2\.75\+2\.75points on LiveCodeBench\. On HumanEval, Archer improves HumanEval\-ET Pass@1 by\+0\.60\+0\.60points while Base Pass@1 decreases by1\.221\.22points\.

The remaining baselines expose the two ends of the trade\-off\. Fast\-dLLM with parallel decoding attains the lowest latency, but its average performance is2\.512\.51points below Archer\. dKV\-Cache\-Decode is the strongest competing method in average performance at33\.21%33\.21\\%, yet provides only1\.38×1\.38\\timesaverage speedup compared with Archer’s2\.57×2\.57\\times\. These comparisons place Archer on the strongest measured quality–efficiency frontier\.

### 6\.2Generalization across DLM Backbones

We apply Archer without training to LLaDA\-8B\-Instruct\(Nieet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib3)\), Dream\-v0\-Instruct\-7B\(Yeet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib4)\), and DiffuCoder\-7B\-cpGRPO\(Gonget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib5)\), and re\-run Saber on every setting, preserving each model’s native masking and logit conventions for a matched comparison\.

Table 2:Cross\-backbone comparison with Saber\. Base/ET denote Pass@1 \(%\) on the original/Extended Test Cases versions of MBPP and HumanEval; Time is seconds, Speedup is relative to Saber, and Overall averages six quality scores and three speedups\.Table[2](https://arxiv.org/html/2608.08086#S6.T2)shows that the acceleration transfers across all six model–benchmark pairs, with speedups ranging from1\.36×1\.36\\timesto1\.78×1\.78\\times\. Archer also improves Pass@1 in five settings, including a3\.053\.05\-point gain on DiffuCoder–HumanEval\. Averaged across backbones, Archer raises performance from45\.71%45\.71\\%to46\.37%46\.37\\%on MBPP and from41\.47%41\.47\\%to42\.38%42\.38\\%on HumanEval, while accelerating inference by1\.41×1\.41\\timesand1\.71×1\.71\\times, respectively\. Prompt reuse therefore improves the quality–speed frontier across independently trained DLMs rather than exploiting behavior specific to one backbone\.

### 6\.3Effect of Delayed Prompt Feedback

The end\-to\-end comparison does not isolate why controlled staleness can improve quality, because two decoding trajectories may diverge for many reasons after their first different update\. We therefore intervene at a shared state and vary only the timing of prompt feedback\. For every MBPP problem, we find the first token accepted with confidencep≥0\.9p\\geq 0\.9and clone the complete decoder state immediately after that acceptance\.*Fresh*then rebuilds prompt K/V before the next update, so the accepted token affects the prompt representation immediately\.*Cached*retains the preceding prompt snapshot, delaying that influence while leaving the response and rollback rule unchanged\. We follow the selected token for five updates and evaluate both final completions\. This paired construction turns feedback timing into the only controlled difference between the two branches\.

Table 3:Matched\-state feedback intervention on MBPP \(427 pairs\)\. Rev\.@5 is the five\-step revision rate; “Only pass” counts branch\-exclusive successes\.Table[3](https://arxiv.org/html/2608.08086#S6.T3)shows that Cached improves Pass@1 by1\.171\.17points on both MBPP Base and MBPP\-ET and uniquely solves 23 problems, compared with 18 for Fresh\. Importantly, both branches revise99\.53%99\.53\\%of the selected tokens, and do so after nearly the same number of updates\. The quality difference therefore cannot be explained by Cached disabling or postponing rollback itself\. It arises while the two branches retain the same correction mechanism but expose it to different prompt contexts, providing direct evidence that feedback timing can alter the functional outcome of rollback\-capable generation under otherwise identical correction rules\.

### 6\.4Why Refresh by State Distance?

Having shown that feedback timing matters, we next ask when a cached prompt state should be synchronized\. Archer measures how far the current response has moved from the cache anchor,Dt=dH​\(gt,gr\)D\_\{t\}=d\_\{H\}\(g\_\{t\},g\_\{r\}\)\. The simplest alternative is cache age,At=t−rA\_\{t\}=t\-r, which refreshes after a fixed number of updates\. Age treats all updates as equally damaging even though some change almost nothing and a single rollback may replace several response tokens\. State distance instead measures the change that can actually invalidate the anchored prompt context\. The relevant comparison is therefore not which signal best predicts small numerical logit drift, but which one better identifies reuse that changes the decoder’s next action\.

At sampled cached steps on MBPP, we execute an additional full forward from the identical response state\. This fresh computation is a shadow observation: it does not change any token, confidence, refresh decision, or random state on the main Archer trajectory\. We compare the cached and fresh Saber actions and their resulting next states, obtaining 5,122 paired probes while reproducing all 427 original completions, step counts, and main\-trajectory NFEs exactly\. We then measure how decision disagreement varies withDtD\_\{t\}and compare distance with age using partial Spearman correlations that control for the other signal\.

Table 4:Decision\-level cache validity on MBPP using 5,122 non\-intervening shadow forwards against full refresh\.Table[4](https://arxiv.org/html/2608.08086#S6.T4)shows a monotonic calibration pattern: action disagreement rises from19\.88%19\.88\\%to65\.09%65\.09\\%as the response moves away from its anchor\. After controlling for age, distance retains a partial correlation of0\.1450\.145with both action disagreement and next\-state distance; after controlling for distance, age falls to0\.0400\.040and0\.0450\.045\. Thus, two caches of the same age can have very different decision\-level validity\. Anchor distance is the more informative refresh signal for Archer because it tracks whether reuse changes the rollback transition rather than only the elapsed time since synchronization\.

## 7Conclusion

Rollback is a defining advantage of DLMs, but its practical value depends on avoiding repeated full\-sequence recomputation\. Archer makes rollback efficient by reusing fixed prompt K/V while recomputing the mutable response\. Across benchmarks and DLM backbones, Archer achieves the best overall performance with a2\.57×2\.57\\timesmean speedup, reaching up to2\.95×2\.95\\timesacceleration and\+3\.05\+3\.05Pass@1 points\. These results establish state\-aware prompt reuse as a practical basis for scalable revisable generation\. They also show that bounded cache staleness can moderate premature feedback rather than merely introduce approximation error\. Archer therefore reframes caching as a mechanism for improving both efficiency and correction dynamics in rollback\-capable DLMs as sequences and rollback horizons continue to grow\.

## 8Limitations

Archer deliberately adopts a conservative cache boundary: it reuses prompt states while recomputing every state derived from the mutable response\. This choice preserves unrestricted rollback, but it does not exhaust the possible computational savings\. Generation\-side caching remains a promising direction when additional structure is available, for example sparse attention, model\-specific validity certificates, or mechanisms that can identify an unchanged dependency region without first reproducing the full computation\. The central challenge is to obtain such reuse without treating a provisional token as immutable or silently changing the rollback transition\.

More broadly, prompt reuse is only one component of efficient revisable generation\. Archer uses a state\-distance controller and leaves the underlying decoder unchanged; future work could combine rollback\-compatible caching with adaptive token\- or layer\-level reuse, learned synchronization policies, parallel acceptance, and systems\-level attention optimizations\. Understanding which of these mechanisms can be composed while retaining reliable revision is an important open problem for scaling rollback\-capable DLMs to longer contexts and more demanding generation tasks\.

## References

- Block diffusion: interpolating between autoregressive and diffusion language models\.External Links:2503\.09573Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1)\.
- J\. Austin, D\. D\. Johnson, J\. Ho, D\. Tarlow, and R\. van den Berg \(2021a\)Structured denoising diffusion models in discrete state\-spaces\.InAdvances in Neural Information Processing Systems,Vol\.34\.Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p1.1),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1)\.
- J\. Austin, A\. Odena, M\. Nye, M\. Bosma, H\. Michalewski, D\. Dohan, E\. Jiang, C\. Cai, M\. Terry, Q\. Le, and C\. Sutton \(2021b\)Program synthesis with large language models\.External Links:2108\.07732Cited by:[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p1.1),[§6](https://arxiv.org/html/2608.08086#S6.p1.1)\.
- H\. Ben\-Hamu, I\. Gat, D\. Severo, N\. Nolte, and B\. Karrer \(2025\)Accelerated sampling from masked diffusion models via entropy bounded unmasking\.External Links:2505\.24857Cited by:[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p1.1)\.
- M\. Chen, J\. Tworek, H\. Jun, Q\. Yuan, H\. P\. d\. O\. Pinto, J\. Kaplan, H\. Edwards, Y\. Burda, N\. Joseph, G\. Brockman,et al\.\(2021a\)Evaluating large language models trained on code\.External Links:2107\.03374Cited by:[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p1.1),[§6](https://arxiv.org/html/2608.08086#S6.p1.1)\.
- Z\. Chen, W\. Chen, C\. Smiley, S\. Shah, I\. Borova, D\. Langdon, R\. Moussa, M\. Beane, T\. Huang, B\. Routledge, and W\. Y\. Wang \(2021b\)FinQA: a dataset of numerical reasoning over financial data\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,pp\. 3697–3711\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.300)Cited by:[Appendix D](https://arxiv.org/html/2608.08086#A4.p1.1),[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p2.1)\.
- Y\. Dong, J\. Ding, X\. Jiang, G\. Li, Z\. Li, and Z\. Jin \(2025\)CodeScore: evaluating code generation by learning code execution\.ACM Transactions on Software Engineering and Methodology34\(3\),pp\. 1–22\.External Links:[Document](https://dx.doi.org/10.1145/3695991)Cited by:[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p1.1),[§6\.1](https://arxiv.org/html/2608.08086#S6.SS1.p1.3)\.
- Y\. Dong, Z\. Ma, X\. Jiang, Z\. Fan, J\. Qian, Y\. Li, J\. Xiao, Z\. Jin, and G\. Li \(2026\)Saber: efficient sampling with adaptive acceleration and backtracking enhanced remasking for diffusion language model in code generation\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 3623–3642\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.165)Cited by:[§G\.2](https://arxiv.org/html/2608.08086#A7.SS2.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§2](https://arxiv.org/html/2608.08086#S2.p1.1),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p2.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p1.1),[§6\.1](https://arxiv.org/html/2608.08086#S6.SS1.p1.3)\.
- I\. Gim, G\. Chen, S\. Lee, N\. Sarda, A\. Khandelwal, and L\. Zhong \(2024\)Prompt cache: modular attention reuse for low\-latency inference\.InProceedings of Machine Learning and Systems,Vol\.6\.Cited by:[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1)\.
- S\. Gong, R\. Zhang, H\. Zheng, J\. Gu, N\. Jaitly, L\. Kong, and Y\. Zhang \(2025\)DiffuCoder: understanding and improving masked diffusion models for code generation\.External Links:2506\.20639Cited by:[§G\.4](https://arxiv.org/html/2608.08086#A7.SS4.p1.2),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1),[§6\.2](https://arxiv.org/html/2608.08086#S6.SS2.p1.1)\.
- D\. Hendrycks, C\. Burns, S\. Kadavath, A\. Arora, S\. Basart, E\. Tang, D\. Song, and J\. Steinhardt \(2021\)Measuring mathematical problem solving with the MATH dataset\.InProceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks,Vol\.1\.Cited by:[Appendix D](https://arxiv.org/html/2608.08086#A4.p1.1),[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p2.1)\.
- F\. Hong, G\. Yu, Y\. Ye, H\. Huang, H\. Zheng, Y\. Zhang, Y\. Wang, and J\. Yao \(2025\)Wide\-in, narrow\-out: revokable decoding for efficient and effective DLLMs\.External Links:2507\.18578Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§2](https://arxiv.org/html/2608.08086#S2.p1.1),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p2.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p1.1)\.
- Z\. Huang, Y\. Wang, Z\. Chen, and G\. Qi \(2025\)Don’t settle too early: self\-reflective remasking for diffusion language models\.External Links:2509\.23653Cited by:[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p2.1)\.
- N\. Jain, K\. Han, A\. Gu, W\. Li, F\. Yan, T\. Zhang, S\. Wang, A\. Solar\-Lezama, K\. Sen, and I\. Stoica \(2025\)LiveCodeBench: holistic and contamination free evaluation of large language models for code\.InInternational Conference on Learning Representations,Cited by:[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p1.1),[§6](https://arxiv.org/html/2608.08086#S6.p1.1)\.
- Q\. Jin, B\. Dhingra, Z\. Liu, W\. Cohen, and X\. Lu \(2019\)PubMedQA: a dataset for biomedical research question answering\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing,pp\. 2567–2577\.External Links:[Document](https://dx.doi.org/10.18653/v1/D19-1259)Cited by:[Appendix D](https://arxiv.org/html/2608.08086#A4.p1.1),[§G\.1](https://arxiv.org/html/2608.08086#A7.SS1.p2.1)\.
- W\. Kwon, Z\. Li, S\. Zhuang, Y\. Sheng, L\. Zheng, C\. H\. Yu, J\. E\. Gonzalez, H\. Zhang, and I\. Stoica \(2023\)Efficient memory management for large language model serving with PagedAttention\.InProceedings of the 29th Symposium on Operating Systems Principles,pp\. 611–626\.Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p2.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1)\.
- X\. L\. Li, J\. Thickstun, I\. Gulrajani, P\. Liang, and T\. B\. Hashimoto \(2022\)Diffusion\-LM improves controllable text generation\.External Links:2205\.14217Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p1.1),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1)\.
- Y\. Li, A\. Kirchmeyer, A\. Mehta, Y\. Qin, B\. Dadachev, K\. Papineni, S\. Kumar, and A\. Risteski \(2024\)Promises and pitfalls of generative masked language modeling: theoretical framework and practical guidelines\.InProceedings of the 41st International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.235,pp\. 27969–28017\.Cited by:[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1)\.
- Z\. Liu, Y\. Yang, Y\. Zhang, J\. Chen, C\. Zou, Q\. Wei, S\. Wang, Y\. Zhu, and L\. Zhang \(2025\)dLLM\-Cache: accelerating diffusion large language models with adaptive caching\.External Links:2506\.06295Cited by:[§G\.2](https://arxiv.org/html/2608.08086#A7.SS2.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1),[§6\.1](https://arxiv.org/html/2608.08086#S6.SS1.p1.3)\.
- A\. Lou, C\. Meng, and S\. Ermon \(2024\)Discrete diffusion modeling by estimating the ratios of the data distribution\.InProceedings of the 41st International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.235,pp\. 32819–32848\.Cited by:[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1)\.
- X\. Ma, R\. Yu, G\. Fang, and X\. Wang \(2025\)dKV\-Cache: the cache for diffusion language models\.External Links:2505\.15781Cited by:[§G\.2](https://arxiv.org/html/2608.08086#A7.SS2.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1),[§6\.1](https://arxiv.org/html/2608.08086#S6.SS1.p1.3)\.
- S\. Nie, F\. Zhu, Z\. You, X\. Zhang, J\. Ou, J\. Hu, J\. Zhou, Y\. Lin, J\. Wen, and C\. Li \(2025\)Large language diffusion models\.External Links:2502\.09992Cited by:[§G\.4](https://arxiv.org/html/2608.08086#A7.SS4.p1.2),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1),[§6\.1](https://arxiv.org/html/2608.08086#S6.SS1.p1.3),[§6\.2](https://arxiv.org/html/2608.08086#S6.SS2.p1.1)\.
- R\. Pope, S\. Douglas, A\. Chowdhery, J\. Devlin, J\. Bradbury, J\. Heek, K\. Xiao, S\. Agrawal, and J\. Dean \(2023\)Efficiently scaling transformer inference\.InProceedings of Machine Learning and Systems,Vol\.5\.Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p2.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1)\.
- S\. S\. Sahoo, M\. Arriola, Y\. Schiff, A\. Gokaslan, E\. Marroquin, J\. T\. Chiu, A\. Rush, and V\. Kuleshov \(2024\)Simple and effective masked diffusion language models\.External Links:2406\.07524Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p1.1),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1)\.
- G\. Wang, Y\. Schiff, S\. S\. Sahoo, and V\. Kuleshov \(2025\)Remasking discrete diffusion models with inference\-time scaling\.InAdvances in Neural Information Processing Systems,External Links:2503\.00307Cited by:[§1](https://arxiv.org/html/2608.08086#S1.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§2](https://arxiv.org/html/2608.08086#S2.p1.1),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p2.1)\.
- X\. Wang, M\. Zhang, S\. Cui, Z\. Chen, B\. Jiang, K\. Kuang, and M\. Lin \(2026\)Reversible diffusion decoding for diffusion language models\.External Links:2602\.00150Cited by:[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p2.1)\.
- C\. Wu, H\. Zhang, S\. Xue, Z\. Liu, S\. Diao, L\. Zhu, P\. Luo, S\. Han, and E\. Xie \(2025\)Fast\-dLLM: training\-free acceleration of diffusion LLM by enabling KV cache and parallel decoding\.External Links:2505\.22618Cited by:[§G\.2](https://arxiv.org/html/2608.08086#A7.SS2.p1.1),[§1](https://arxiv.org/html/2608.08086#S1.p3.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p1.1),[§3\.2](https://arxiv.org/html/2608.08086#S3.SS2.p2.1),[§6\.1](https://arxiv.org/html/2608.08086#S6.SS1.p1.3)\.
- J\. Ye, Z\. Xie, L\. Zheng, J\. Gao, Z\. Wu, X\. Jiang, Z\. Li, and L\. Kong \(2025\)Dream 7b: diffusion large language models\.External Links:2508\.15487Cited by:[§G\.4](https://arxiv.org/html/2608.08086#A7.SS4.p1.2),[§3\.1](https://arxiv.org/html/2608.08086#S3.SS1.p1.1),[§6\.2](https://arxiv.org/html/2608.08086#S6.SS2.p1.1)\.

## Appendix AFormal Analysis

This section supplies the assumptions and proofs underlying the claims in the main paper\. The analysis separates three notions that are easy to conflate\. Archer preserves the*ability*to revise any response position, but it does not claim bitwise equivalence with eager decoding\. Its prompt cache is an approximation whose local error is controlled by the distance from the cache anchor\. Whether that numerical error changes the trajectory depends on the margin of the rollback decision, not on logit error alone\.

### A\.1Notation and Decoder State

Letp∈𝒱Pp\\in\\mathcal\{V\}^\{P\}be a fixed prompt and letgt∈\(𝒱∪\{𝙼𝙰𝚂𝙺\}\)Gg\_\{t\}\\in\(\\mathcal\{V\}\\cup\\\{\\mathtt\{MASK\}\\\}\)^\{G\}be the response after updatett\. The complete sampler state is denoted byξt=\(gt,ut\)\\xi\_\{t\}=\(g\_\{t\},u\_\{t\}\), whereutu\_\{t\}collects confidence histories and any other state used by the rollback rule\. A full bidirectional forward can be written as

zt⋆=Φθ​\(gt;𝒞p​\(gt\)\),ξt\+1=𝒟​\(ξt,zt⋆\)\.z\_\{t\}^\{\\star\}=\\Phi\_\{\\theta\}\(g\_\{t\};\\mathcal\{C\}\_\{p\}\(g\_\{t\}\)\),\\qquad\\xi\_\{t\+1\}=\\mathcal\{D\}\(\\xi\_\{t\},z\_\{t\}^\{\\star\}\)\.\(16\)Here𝒞p​\(g\)\\mathcal\{C\}\_\{p\}\(g\)stacks the prompt K/V states from every transformer layer when the response isgg\. If the most recent cache was created atgrg\_\{r\}, Archer instead evaluates

z^t=Φθ​\(gt;𝒞p​\(gr\)\),ξ^t\+1=𝒟​\(ξt,z^t\)\.\\hat\{z\}\_\{t\}=\\Phi\_\{\\theta\}\(g\_\{t\};\\mathcal\{C\}\_\{p\}\(g\_\{r\}\)\),\\qquad\\hat\{\\xi\}\_\{t\+1\}=\\mathcal\{D\}\(\\xi\_\{t\},\\hat\{z\}\_\{t\}\)\.\(17\)Equations \([16](https://arxiv.org/html/2608.08086#A1.E16)\) and \([17](https://arxiv.org/html/2608.08086#A1.E17)\) compare the two forwards at the*same*sampler state\. This distinction is important: after their decisions differ, the two complete trajectories need not remain at the samegtg\_\{t\}\.

### A\.2Why Response\-State Reuse Conflicts with Arbitrary Rollback

The cache boundary follows from the dependency structure of bidirectional attention\. LetΔ0\\Delta^\{0\}be the response positions changed by a rollback, and define the structural dependency closure at layerℓ\\ellby

Δℓ=\{i:∃j∈Δℓ−1​with an attention edge​j→i\}\.\\Delta^\{\\ell\}=\\\{i:\\exists j\\in\\Delta^\{\\ell\-1\}\\text\{ with an attention edge \}j\\\!\\rightarrow\\\!i\\\}\.\(18\)
###### Proposition 2\(Dense rollback closure\)\.

For a transformer layer with dense bidirectional attention,Δ0≠∅\\Delta^\{0\}\\neq\\varnothingimpliesΔ1=\{1,…,P\+G\}\\Delta^\{1\}=\\\{1,\\ldots,P\+G\\\}\. Consequently, no nontrivial set of hidden states or K/V states from a previous response has a universal exact\-reuse guarantee after an arbitrary rollback\.

###### Proof\.

Every query positioniiattends to every key positionjj\. Choose anyj∈Δ0j\\in\\Delta^\{0\}\. The edgej→ij\\\!\\rightarrow\\\!iexists for everyii, so every position belongs toΔ1\\Delta^\{1\}\. Later layers inherit this full structural closure\. The claim concerns possible dependence rather than the magnitude of a particular numerical change\. Establishing that an affected state happens to remain identical would require evaluating the affected computation and therefore cannot provide a universal skip rule\. ∎

Proposition[2](https://arxiv.org/html/2608.08086#Thmproposition2)does not say that every possible response cache is useless\. Sparse attention, model\-specific certificates, or a custom approximation may permit additional reuse\. It says that a generic method cannot treat an accepted response token as immutable history while retaining arbitrary rollback under dense bidirectional attention\. Archer therefore stores no response hidden state across decoding updates, ensuring that every revised token is recomputed from the current state\.

###### Proposition 3\(Preservation of revisability\)\.

Suppose the original decoder𝒟\\mathcal\{D\}may replace or re\-mask any response position\. Archer leaves this action space unchanged: every response embedding, query, key, value, hidden state, and logit is recomputed before each update\. Prompt caching may change which action is selected, but cannot make a response position immutable\.

###### Proof\.

At a cached forward, response queries attend to cached prompt K/V and newly computed response K/V\. No response tensor from an earlierggis supplied to the model\. Archer then passes a full response\-logit tensor to the unmodified decoder\. Hence every action available to𝒟\\mathcal\{D\}under an eager forward remains representable\. Approximate prompt K/V can alter the logits and thus the chosen action, but they do not remove any response position from the decoder’s revision domain\. ∎

This is the precise sense in which Archer is rollback\-compatible\. It preserves revisability, not necessarily the eager trajectory\.

### A\.3End\-to\-End Computational Cost

LetN=P\+GN=P\+G, and letLLbe the number of transformer layers\. At layerℓ\\ell, writeαℓ\\alpha\_\{\\ell\}for the position\-wise projection and feed\-forward cost andβℓ\\beta\_\{\\ell\}for one query–key interaction\. A full forward has leading cost

Cfull=∑ℓ=1L\[αℓ​N\+βℓ​N2\]\.C\_\{\\rm full\}=\\sum\_\{\\ell=1\}^\{L\}\\left\[\\alpha\_\{\\ell\}N\+\\beta\_\{\\ell\}N^\{2\}\\right\]\.\(19\)An Archer cached forward evaluates onlyGGfresh query paths, while each query still attends to allNNpositions\. Including cache assembly and dispatch overheadHH, its cost is

Ccache=∑ℓ=1L\[αℓ​G\+βℓ​G​N\]\+H\.C\_\{\\rm cache\}=\\sum\_\{\\ell=1\}^\{L\}\\left\[\\alpha\_\{\\ell\}G\+\\beta\_\{\\ell\}GN\\right\]\+H\.\(20\)Thus Archer reduces prompt\-side projections, prompt queries, and prompt feed\-forward paths without shortening the receptive field of a response query\.

Let eager decoding useM0M\_\{0\}forward steps\. An Archer trajectory may have a different lengthMKM\_\{K\}because approximate logits can change a decoding decision\. IfRKR\_\{K\}of those steps are full refreshes, the cost\-model speedup is

S​\(K\)=M0​CfullRK​Cfull\+\(MK−RK\)​Ccache\.S\(K\)=\\frac\{M\_\{0\}C\_\{\\rm full\}\}\{R\_\{K\}C\_\{\\rm full\}\+\(M\_\{K\}\-R\_\{K\}\)C\_\{\\rm cache\}\}\.\(21\)For equal trajectory lengths, defineqK=RK/MKq\_\{K\}=R\_\{K\}/M\_\{K\}andη=Ccache/Cfull\\eta=C\_\{\\rm cache\}/C\_\{\\rm full\}\. Equation \([21](https://arxiv.org/html/2608.08086#A1.E21)\) reduces to

S​\(K\)=\[qK\+\(1−qK\)​η\]−1\.S\(K\)=\\left\[q\_\{K\}\+\(1\-q\_\{K\}\)\\eta\\right\]^\{\-1\}\.\(22\)IgnoringHHand layerwise constant differences givesη≈G/\(P\+G\)\\eta\\approx G/\(P\+G\)\. This yields the idealized expression in the main paper and the cached\-step ceiling1\+P/G1\+P/G\. Wall\-clock speed can depart from this ceiling because GPU kernels, memory movement, prompt\-length variation, refresh frequency, and trajectory length all remain visible in Eq\. \([21](https://arxiv.org/html/2608.08086#A1.E21)\)\.

### A\.4From Prompt Staleness to Logit Error

We next derive the local approximation bound\. LetHg,tℓH\_\{g,t\}^\{\\ell\}andH^g,tℓ\\hat\{H\}\_\{g,t\}^\{\\ell\}be eager and cached response hidden states after layerℓ\\ell, both evaluated atgtg\_\{t\}, and define

eℓ=‖Hg,tℓ−H^g,tℓ‖,e0=0\.e\_\{\\ell\}=\\\|H\_\{g,t\}^\{\\ell\}\-\\hat\{H\}\_\{g,t\}^\{\\ell\}\\\|,\\qquad e\_\{0\}=0\.\(23\)At layerℓ\\ell, prompt\-cache staleness is

stℓ=‖Kpℓ​\(gt\)−Kpℓ​\(gr\)‖\+‖Vpℓ​\(gt\)−Vpℓ​\(gr\)‖\.s\_\{t\}^\{\\ell\}=\\\|K\_\{p\}^\{\\ell\}\(g\_\{t\}\)\-K\_\{p\}^\{\\ell\}\(g\_\{r\}\)\\\|\+\\\|V\_\{p\}^\{\\ell\}\(g\_\{t\}\)\-V\_\{p\}^\{\\ell\}\(g\_\{r\}\)\\\|\.\(24\)Assume the response update at layerℓ\\ellis locally Lipschitz in its response input and prompt K/V\. For constantsaℓ,bℓ≥0a\_\{\\ell\},b\_\{\\ell\}\\geq 0,

‖Φℓ​\(H,C\)−Φℓ​\(H′,C′\)‖≤aℓ​‖H−H′‖\+bℓ​‖C−C′‖\.\\\|\\Phi\_\{\\ell\}\(H,C\)\-\\Phi\_\{\\ell\}\(H^\{\\prime\},C^\{\\prime\}\)\\\|\\leq a\_\{\\ell\}\\\|H\-H^\{\\prime\}\\\|\+b\_\{\\ell\}\\\|C\-C^\{\\prime\}\\\|\.\(25\)
###### Lemma 1\(Layerwise propagation\)\.

Under Eq\. \([25](https://arxiv.org/html/2608.08086#A1.E25)\),

eℓ≤aℓ​eℓ−1\+bℓ​stℓ\.e\_\{\\ell\}\\leq a\_\{\\ell\}e\_\{\\ell\-1\}\+b\_\{\\ell\}s\_\{t\}^\{\\ell\}\.\(26\)

###### Proof\.

Add and subtractΦℓ​\(H^g,tℓ−1,Ctℓ\)\\Phi\_\{\\ell\}\(\\hat\{H\}\_\{g,t\}^\{\\ell\-1\},C\_\{t\}^\{\\ell\}\)between the eager and cached updates\. The triangle inequality separates the error inherited from the previous response layer and the new error caused by prompt K/V\. Applying Eq\. \([25](https://arxiv.org/html/2608.08086#A1.E25)\) to the two terms yields Eq\. \([26](https://arxiv.org/html/2608.08086#A1.E26)\)\. ∎

Unrolling the recurrence and applying anLheadL\_\{\\rm head\}\-Lipschitz output head gives

‖zt⋆−z^t‖∞≤Lhead​∑ℓ=1Lbℓ​stℓ​∏j=ℓ\+1Laj\.\\\|z\_\{t\}^\{\\star\}\-\\hat\{z\}\_\{t\}\\\|\_\{\\infty\}\\leq L\_\{\\rm head\}\\sum\_\{\\ell=1\}^\{L\}b\_\{\\ell\}s\_\{t\}^\{\\ell\}\\prod\_\{j=\\ell\+1\}^\{L\}a\_\{j\}\.\(27\)This expression makes two points explicit\. Staleness can enter at every layer because each prompt representation depends on the response, and an early discrepancy can be amplified by later layers\.

To connect this bound to Archer’s controller, letE​\(g\)E\(g\)stack response token embeddings and assume their diameter is bounded byBB\. Then

‖E​\(gt\)−E​\(gr\)‖F≤B​dH​\(gt,gr\)\.\\\|E\(g\_\{t\}\)\-E\(g\_\{r\}\)\\\|\_\{F\}\\leq B\\sqrt\{d\_\{H\}\(g\_\{t\},g\_\{r\}\)\}\.\(28\)If the layer\-ℓ\\ellprompt\-cache map is locallyκℓ\\kappa\_\{\\ell\}\-Lipschitz inE​\(g\)E\(g\), thenstℓ≤κℓ​B​Dts\_\{t\}^\{\\ell\}\\leq\\kappa\_\{\\ell\}B\\sqrt\{D\_\{t\}\}\. Substitution into Eq\. \([27](https://arxiv.org/html/2608.08086#A1.E27)\) yields

‖zt⋆−z^t‖∞≤Γ​Dt,Γ=Lhead​B​∑ℓ=1Lbℓ​κℓ​∏j=ℓ\+1Laj\.\\\|z\_\{t\}^\{\\star\}\-\\hat\{z\}\_\{t\}\\\|\_\{\\infty\}\\leq\\Gamma\\sqrt\{D\_\{t\}\},\\qquad\\Gamma=L\_\{\\rm head\}B\\sum\_\{\\ell=1\}^\{L\}b\_\{\\ell\}\\kappa\_\{\\ell\}\\prod\_\{j=\\ell\+1\}^\{L\}a\_\{j\}\.\(29\)Archer reuses the cache only whileDt<KD\_\{t\}<K\. Because Hamming distance is integer\-valued, every cached step therefore satisfies

‖zt⋆−z^t‖∞≤Γ​K−1\.\\\|z\_\{t\}^\{\\star\}\-\\hat\{z\}\_\{t\}\\\|\_\{\\infty\}\\leq\\Gamma\\sqrt\{K\-1\}\.\(30\)The constants are local and generally unavailable for a large pretrained model, so Eq\. \([30](https://arxiv.org/html/2608.08086#A1.E30)\) is a structural guarantee rather than a numerical certificate\. It explains why the controller uses response drift: unlike elapsed time,DtD\_\{t\}appears directly in the perturbation bound\.

### A\.5Decision Fidelity and Trajectory Fidelity

Exact logits are stronger than the decoder requires\. For sampler stateξ\\xiand logitszz, define the local decision margin

m𝒟\(ξ,z\)=infδ\{∥δ∥∞:𝒟\(ξ,z\+δ\)≠𝒟\(ξ,z\)\}\.m\_\{\\mathcal\{D\}\}\(\\xi,z\)=\\inf\_\{\\delta\}\\left\\\{\\\|\\delta\\\|\_\{\\infty\}:\\mathcal\{D\}\(\\xi,z\+\\delta\)\\neq\\mathcal\{D\}\(\\xi,z\)\\right\\\}\.\(31\)
###### Proposition 4\(One\-step decision preservation\)\.

At a common stateξt\\xi\_\{t\}, Archer and eager decoding select the same next state whenever

Γ​Dt<m𝒟​\(ξt,zt⋆\)\.\\Gamma\\sqrt\{D\_\{t\}\}<m\_\{\\mathcal\{D\}\}\(\\xi\_\{t\},z\_\{t\}^\{\\star\}\)\.\(32\)

###### Proof\.

Equation \([29](https://arxiv.org/html/2608.08086#A1.E29)\) placesz^t\\hat\{z\}\_\{t\}inside the openℓ∞\\ell\_\{\\infty\}ball of radiusm𝒟​\(ξt,zt⋆\)m\_\{\\mathcal\{D\}\}\(\\xi\_\{t\},z\_\{t\}^\{\\star\}\)around the eager logits\. By definition of the margin,𝒟\\mathcal\{D\}is constant throughout this ball\. ∎

###### Corollary 1\(Trajectory preservation\)\.

If Eq\. \([32](https://arxiv.org/html/2608.08086#A1.E32)\) holds at every cached step of a deterministic run, Archer and eager decoding produce the same trajectory and completion\.

###### Proof\.

Both methods start from the same state\. Proposition[4](https://arxiv.org/html/2608.08086#Thmproposition4)preserves equality at cached steps, while a refresh evaluates the same full model at the same state\. Induction over decoder steps completes the proof\. ∎

The converse does not hold: a violation of the sufficient bound need not change the action\. This is why the shadow\-forward study measures action and next\-state disagreement rather than treating logit error alone as the operational failure criterion\.

### A\.6Temporal Anchoring as Delayed Feedback

The same approximation that creates logit error also changes the dynamics of revision\. To make this precise, consider a continuous relaxationy=E​\(g\)y=E\(g\)\. Within one anchor interval, the eager and anchored logit maps and their Jacobians satisfy

zfresh​\(y\)\\displaystyle z^\{\\rm fresh\}\(y\)=Φθ​\(y;𝒞p​\(y\)\),\\displaystyle=\\Phi\_\{\\theta\}\(y;\\mathcal\{C\}\_\{p\}\(y\)\),zranchor​\(y\)\\displaystyle z\_\{r\}^\{\\rm anchor\}\(y\)=Φθ​\(y;𝒞p​\(yr\)\),\\displaystyle=\\Phi\_\{\\theta\}\(y;\\mathcal\{C\}\_\{p\}\(y\_\{r\}\)\),\(33\)Jfresh\\displaystyle J\_\{\\rm fresh\}=∂yΦθ\+∂𝒞Φθ​∂y𝒞p,\\displaystyle=\\partial\_\{y\}\\Phi\_\{\\theta\}\+\\partial\_\{\\mathcal\{C\}\}\\Phi\_\{\\theta\}\\,\\partial\_\{y\}\\mathcal\{C\}\_\{p\},Janchor\\displaystyle J\_\{\\rm anchor\}=∂yΦθ\.\\displaystyle=\\partial\_\{y\}\\Phi\_\{\\theta\}\.\(34\)Archer does not suppress response–response interaction; that information is recomputed in∂yΦθ\\partial\_\{y\}\\Phi\_\{\\theta\}\. It temporarily suppresses only the indirect response\-to\-prompt\-to\-response path∂𝒞Φθ​∂y𝒞p\\partial\_\{\\mathcal\{C\}\}\\Phi\_\{\\theta\}\\,\\partial\_\{y\}\\mathcal\{C\}\_\{p\}\. A refresh restores this term by settingyr←yy\_\{r\}\\leftarrow y\.

Equation \([34](https://arxiv.org/html/2608.08086#A1.E34)\) supports a conditional, not universal, quality claim\. If the indirect term amplifies a provisional error direction, anchoring reduces its immediate gain and gives rollback another update in which to revise the token\. If the term instead carries useful new evidence, excessive anchoring delays that evidence and harms the next decision\. The non\-monotone quality curve in the main paper is consistent with these two regimes\. ModerateKKprovides short\-term inertia followed by reset, whereasK→∞K\\rightarrow\\inftyremoves the corrective synchronization that keeps the approximation aligned with the evolving response\.

### A\.7Scope of the Guarantees

The theory establishes four limited but useful facts\. Dense bidirectional attention denies a generic exactness guarantee for response\-state reuse after rollback; prompt\-only caching preserves the decoder’s revision domain; state distance bounds local prompt\-induced logit error under explicit smoothness assumptions; and a decision margin turns that error bound into a sufficient condition for trajectory fidelity\. It does not claim global Lipschitz constants for a pretrained DLM, statistical improvement from staleness, or exact equality with eager decoding\. The empirical analyses below test the decision\-level consequences that the theory deliberately leaves model dependent\.

## Appendix BComplete Results and Robustness

The main paper reports the operating points that best expose Archer’s quality–latency trade\-off\. This section supplies the complete refresh\-radius sweep behind those operating points\. No radius is selected from a hidden test\-only search\.

### B\.1Refresh\-Radius Sensitivity

The refresh radius controls how far the response may move from the state at which prompt K/V was constructed\. The limiting cases have direct interpretations\. AtK=1K=1, every response change invalidates the snapshot and Archer approaches eager Saber\. AtK=∞K=\\infty, the initial prompt snapshot is never synchronized again\.

Table 5:Refresh\-radius sweep on LLaDA\-8B\-Instruct\. Time is seconds per problem\.Table[5](https://arxiv.org/html/2608.08086#A2.T5)reports the full logarithmic sweep on all three code benchmarks\. Latency improves as refreshes become less frequent, whereas quality is non\-monotone\. Moderate radii preserve enough synchronization to avoid long\-lived context error while still delaying prompt\-mediated reinforcement\. The deterioration at large radii is therefore not an unexplained tuning artifact; it is the empirical counterpart of the approximation term in Eq\. \([30](https://arxiv.org/html/2608.08086#A1.E30)\)\. The sweep also makes clear that the same radius need not optimize every benchmark, which motivates reporting per\-task operating points rather than a universal “best”KK\.

## Appendix CMechanism and Controlled Analyses

The main paper establishes the two central mechanism results through matched feedback and shadow\-forward comparisons\. Here we retain the supporting diagnostics needed to interpret those interventions and verify that prompt reuse does not obtain speed by disabling rollback\.

### C\.1Rollback Remains Active

A response position is counted as revised if it is re\-masked or replaced after first receiving a non\-mask prediction\. Saber and Archer use the same rollback rule; the no\-rollback decoder provides a zero\-revision control\. We use the representative radiusK=8K=8throughout this diagnostic\. The MBPP measurement uses a fixed 150\-problem subset; HumanEval and LiveCodeBench use their complete evaluation sets\.

Table 6:Rollback activity withK=8K=8\.Table[6](https://arxiv.org/html/2608.08086#A3.T6)shows that Archer retains substantial revision activity and does not shorten the trajectory relative to Saber\. Its speedup therefore does not come from freezing response tokens or removing opportunities for correction\. This measurement supports Proposition[3](https://arxiv.org/html/2608.08086#Thmproposition3)at the realized trajectory level\.

### C\.2Matched\-State Feedback Intervention

End\-to\-end runs quickly diverge, so a conventional sampler comparison cannot isolate prompt\-feedback timing\. For each MBPP problem, we clone the decoder state after the first token accepted with confidence at least0\.90\.9\. Fresh immediately rebuilds prompt K/V, whereas Cached retains the pre\-acceptance snapshot\. The response state, rollback rule, and all other decoder variables remain matched\. We then follow the selected token for five updates and evaluate both final programs\.

Both branches revise99\.53%99\.53\\%of selected tokens at nearly the same time, yet Cached solves five more branch\-exclusive problems and improves Pass@1 by 1\.17 points on both MBPP Base and its Extended Test Cases version\. Delayed feedback thus changes functional trajectories without obtaining its gain by disabling rollback\. The intervention is local evidence rather than a universal monotonicity claim; the complete radius sweep shows that excessive delay can reverse the benefit\.

### C\.3Decision\-Level Cache Validity

We collect 5,122 non\-intervening shadow forwards from the MBPP run\. Each probe evaluates fresh and cached logits at the same decoder state and compares the resulting Saber action\. Shadow computations leave all 427 main responses, step counts, and NFEs unchanged\.

Action disagreement rises from 19\.88% to 65\.09% across distance bins, and next\-state distance rises from 0\.47 to 1\.90\. Cache age better describes low\-level logit drift, but anchor distance has the stronger partial association with the next decoding transition\. This distinction matches Archer’s objective: the controller need not minimize every floating\-point difference; it should identify when reuse is likely to alter rollback behavior\. The compact calibration and correlation results are reported in the main paper’s decision\-level cache\-validity analysis\.

## Appendix DGeneralization Beyond Code

Archer’s controller observes response\-state changes and uses neither code structure nor execution feedback\. We therefore extend the evaluation to mathematical, financial, and biomedical reasoning\. MATH\-500\(Hendryckset al\.,[2021](https://arxiv.org/html/2608.08086#bib.bib26)\)requires free\-form competition\-mathematics answers, FinQA\(Chenet al\.,[2021b](https://arxiv.org/html/2608.08086#bib.bib27)\)requires numerical reasoning over financial reports, and PubMedQA\(Jinet al\.,[2019](https://arxiv.org/html/2608.08086#bib.bib28)\)requires yes/no/maybe decisions from biomedical abstracts\. The three tasks differ substantially in prompt length and answer form, providing a direct test of whether prompt\-side reuse transfers beyond program synthesis\.

Table 7:Cross\-domain comparison with LLaDA\-8B\-Instruct\.We evaluate the fixed geometric sweepK∈\{4,8,12,16\}K\\in\\\{4,8,12,16\\\}and report a representative Pareto point for each dataset in Table[7](https://arxiv.org/html/2608.08086#A4.T7)\. The selected radii are1616,1616, and44for MATH\-500, FinQA, and PubMedQA, respectively\. Archer reduces latency in all three domains\. The Overall columns average the three accuracy scores and the three corresponding speedups\. On FinQA it improves numerical\-answer accuracy by0\.610\.61points while accelerating decoding by2\.50×2\.50\\times; on PubMedQA it improves label accuracy by0\.100\.10points at1\.48×1\.48\\timesspeedup\. On MATH\-500, a near\-quality\-preserving point obtains1\.27×1\.27\\timesspeedup with a0\.600\.60\-point accuracy change\. The same cache controller therefore produces useful quality–latency points across three non\-code domains without task\-specific logic\.

MATH\-500 is scored after canonical answer normalization, and PubMedQA uses exact normalized yes/no/maybe labels\. Because our FinQA decoder emits a free\-form number rather than an executable program, we compare the extracted answer with the goldexe\_ans; the reported value is numerical\-answer accuracy rather than official program accuracy\. Stating this distinction and the fixed candidate set makes the scope of the cross\-domain evidence explicit\.

## Appendix EScaling and Systems Analysis

Archer removes repeated prompt projections while recomputing the revisable response\. We test this systems claim through realized cache use and prompt\- and response\-length scaling\. At the reported MBPP and LiveCodeBench operating points, 81\.62% and 80\.57% of main\-trajectory model calls, respectively, use the cached prompt path\. Archer may execute slightly more NFEs than Saber, so the speedup comes from lower work per update rather than a shorter correction trajectory\.

### E\.1Prompt\-Length Scaling

We hold the response length atG=256G=256and increase the fixed prompt from 128 to 2,048 tokens\. At each length, full and cached updates start from the same state and are measured after warm\-up, with CUDA synchronization immediately before and after every timed call\.

Table 8:Prompt\-length scaling withG=256G=256andK=8K=8\. Time is in milliseconds\.Table[8](https://arxiv.org/html/2608.08086#A5.T8)shows that cached\-step speedup increases monotonically from1\.30×1\.30\\timesto4\.62×4\.62\\times\. Full\-forward cost grows rapidly with the prompt, whereas cached\-step time is dominated by the fixed response region\. The measured trend directly matches the complexity analysis: prompt reuse becomes more valuable as the immutable context occupies a larger fraction of the sequence\.

### E\.2Response\-Length Scaling

A complementary sweep fixesP=512P=512and varies the response budget\. Unlike the per\-update study, this comparison measures the complete realized trajectory, including cache construction and refreshes\. Archer remains faster at every tested response length, with gains of1\.461\.46–1\.67×1\.67\\times\. The benefit does not vanish when response computation dominates because the prompt is still revisited over many rollback updates\.

## Appendix FQualitative Analysis and Failure Cases

Aggregate paired outcomes show that feedback timing matters, but they do not show how two trajectories become functionally different\. We therefore manually inspected branch\-exclusive MBPP outcomes whose probe lies inside executable code rather than at the terminal token\. The accompanying case\-study figure uses two Cached\-only successes\.

![Refer to caption](https://arxiv.org/html/2608.08086v1/x1.png)Figure 4:Delayed prompt\-feedback case studies on MBPP\.In Task 452 \(loss\_amount\), Fresh settles on an unconditionaloriginal\_price \- sale\_pricereturn, which fails whenever the sale does not produce a loss\. Cached instead restores the necessary conditional and returns zero in the no\-loss case\. In Task 558 \(digit\_distance\_nums\), Fresh appliesabsdirectly to string characters, whereas Cached converts paired digits to integers before subtraction\. Both high\-confidence probed fragments are re\-masked within the next update in both branches; the final difference is therefore not caused by making a token immutable\.

The paired intervention also contains the complementary boundary case, Task 760, where Fresh reaches the correct set\-cardinality test and Cached adopts an incorrect duplicate detector\. Delayed feedback is thus an inductive bias, not a free accuracy guarantee\. Moderate anchoring can prevent transient evidence from being amplified too quickly, but it can also postpone useful evidence\.

## Appendix GDetailed Experimental Setup

This section specifies the evaluation protocol used throughout the paper\. The main experiments follow the code\-generation setting of Saber, while the supplementary cross\-domain study changes only the task prompt and evaluator\. Unless noted otherwise, we decode each problem once without demonstrations or test\-time sampling\.

### G\.1Datasets

Our primary comparison uses MBPP\(Austinet al\.,[2021b](https://arxiv.org/html/2608.08086#bib.bib23)\), HumanEval\(Chenet al\.,[2021a](https://arxiv.org/html/2608.08086#bib.bib22)\), their Extended Test Cases \(ET\) versions\(Donget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib24)\), and LiveCodeBench\(Jainet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib25)\)\. MBPP contains 427 sanitized Python synthesis problems and HumanEval contains 164 function\-completion tasks\. Their ET versions preserve the original problems but evaluate generated programs against additional edge\-case tests\. LiveCodeBench is contamination\-aware; we use all 400 tasks inrelease\_v1\. In each code benchmark, the model receives the task description and, where applicable, the function signature, public tests, or starter code\. The model must produce one executable Python completion that is evaluated without manual repair\.

To evaluate transfer beyond code, we additionally use the 500\-problem MATH\-500 test set derived from MATH\(Hendryckset al\.,[2021](https://arxiv.org/html/2608.08086#bib.bib26)\), the canonical FinQA test split\(Chenet al\.,[2021b](https://arxiv.org/html/2608.08086#bib.bib27)\), and the 1,000\-example expert\-labeled PubMedQA set\(Jinet al\.,[2019](https://arxiv.org/html/2608.08086#bib.bib28)\)\. MATH\-500 requests a boxed final answer; FinQA requests a numerical answer from a financial report and table; and PubMedQA requires one of*yes*,*no*, or*maybe*after reading the question and abstract\. These prompts preserve the same model template and response budget as the code experiments\.

### G\.2Baselines

We compare Archer with three classes of DLM decoding methods\. The first is standard confidence\-based LLaDA decoding, which executes the full denoising schedule without KV reuse\. The second includes efficient DLM methods that target standard, non\-rollback decoding: Fast\-dLLM in cache\-only and parallel forms\(Wuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib8)\), dKV\-Cache in decode and greedy modes\(Maet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib9)\), and dLLM\-Cache\(Liuet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib10)\)\. The third is Saber\(Donget al\.,[2026](https://arxiv.org/html/2608.08086#bib.bib6)\), which performs a full forward pass at every rollback update\. Archer uses exactly Saber’s acceptance, replacement, and re\-masking rules; the only difference is whether the prompt K/V state is reused or refreshed\. This pairing isolates the effect of the cache controller from a change in the rollback decoder\.

All baselines use the same task prompt, response budget, base\-model revision, and evaluator whenever their algorithms permit\. We follow the released configurations of Fast\-dLLM, dKV\-Cache, dLLM\-Cache, and Saber\. Fast\-dLLM is evaluated both with cache\-only transfer and factor\-one parallel transfer, and dKV\-Cache\-Greedy uses its released random ordering with seed 42\.

### G\.3Metrics

For code generation, Pass@1 is the fraction of tasks whose single extracted completion passes every test in the corresponding evaluator:

Pass​@​1=1N​∑i=1N𝕀​\[Passed​\(y^i\)\]\.\\mathrm\{Pass@1\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbb\{I\}\\\!\\left\[\\mathrm\{Passed\}\(\\widehat\{y\}\_\{i\}\)\\right\]\.\(35\)Here,NNis the number of tasks andy^i\\widehat\{y\}\_\{i\}is the decoded completion for taskii\. For MBPP and HumanEval, we compute this score separately on the original Base test suite and the Extended Test Cases \(ET\) suite; the latter evaluates the same completion against additional edge\-case tests\. LiveCodeBench Pass@1 is computed by its official evaluator\. MATH\-500, FinQA, and PubMedQA report normalized answer or label accuracy\.

We also report the mean number of response updates \(Steps\), synchronized generation time per example \(Time\), and Speedup\. Main\-table speedups are computed against confidence\-based LLaDA on the same benchmark; the cross\-backbone study instead uses Saber on the same model as its reference\. Timing includes cache construction, refreshes, cached forwards, and decoder control, but excludes model loading, tokenization, and program execution\.

### G\.4Implementation Details

The primary results use LLaDA\-8B\-Instruct\(Nieet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib3)\), pinned to revision6059b30\. The cross\-backbone study additionally evaluates Dream\-v0\-Instruct\-7B\(Yeet al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib4)\)and DiffuCoder\-7B\-cpGRPO\(Gonget al\.,[2025](https://arxiv.org/html/2608.08086#bib.bib5)\), preserving each model’s native mask identifier, logit alignment, and cache interface\. All experiments use temperature zero, batch size one, a generation length of 256, and a block length of 256\. Saber and Archer usen=2n=2andμ=2\\mu=2\.

The principal Archer operating points useK=11K=11on MBPP,K=10K=10on LiveCodeBench, andK=8K=8on HumanEval\. Cross\-backbone results useK=11/15K=11/15for LLaDA,17/917/9for Dream, and8/158/15for DiffuCoder on MBPP/HumanEval, respectively\. These operating points are accompanied by the logarithmic sensitivity sweep in Appendix[B\.1](https://arxiv.org/html/2608.08086#A2.SS1); they do not assert that a universal radius is optimal\. For the three cross\-domain benchmarks, we evaluate the fixed candidate setK∈\{4,8,12,16\}K\\in\\\{4,8,12,16\\\}and report the stated Pareto point\. All timings are measured with CUDA synchronization immediately before and after generation\.