Attention Viterbi Decoding (AVD)

• Attention Viterbi Decoding (AVD) is a framework that supplements classical Viterbi recursion with nonlocal information, overcoming the limits of first-order Markov assumptions.
• It leverages methods such as K-best list decoding with periodic transformer language model rescoring and branch-and-bound search for exact non-Markov MAP inference.
• The approach also considers infinite-horizon stability via nodes and barriers, offering insights into online consistency and efficient chunkwise decoding.

Attention Viterbi Decoding (AVD) is not a standardized algorithmic label in the cited literature, but the term is plausibly used as an umbrella for Viterbi-style sequence decoding augmented by information that is not captured by the classical first-order Markov recursion. In the sources most directly relevant to this usage, two technically distinct directions appear. One direction injects semantic context from a transformer LLM into trellis decoding through periodic path reranking and pruning, as in the ByT5-guided decoder for convolutional codes in "LLM-Viterbi: Semantic-Aware Decoding for Convolutional Codes" (Li et al., 21 Apr 2026). The other direction addresses exact maximum a posteriori sequence decoding when the sequence of interest is not itself Markov because auxiliary variables have been marginalized out, requiring branch-and-bound rather than the standard Viterbi dynamic program, as in "Branch-and-bound method for calculating Viterbi path in triplet Markov models" (Soop et al., 25 Jul 2025). A foundational theoretical antecedent is the study of infinite-horizon Viterbi decoding for pairwise Markov models, which establishes conditions under which finite MAP decoders stabilize to an infinite Viterbi path (Lember et al., 2017). Taken together, these works locate AVD at the intersection of classical trellis decoding, generalized probabilistic sequence models, and context-aware or nonlocal scoring.

1. Terminological scope and conceptual definition

In the available literature, no cited paper introduces a method explicitly named “Attention Viterbi Decoding” in the narrow sense of embedding an attention operator directly into the Viterbi recursion. The semantic-aware decoder based on ByT5 is explicitly described as related in spirit to such a notion, but “not really” equivalent to a standard AVD because the LLM does not directly alter branch transition probabilities at every trellis step; instead it performs periodic path reranking and pruning over retained candidate paths (Li et al., 21 Apr 2026). Likewise, the triplet Markov model paper states that its central contribution is not an “attention” mechanism but a branch-and-bound framework for exact decoding when richer dependencies invalidate classical first-order Viterbi recursion (Soop et al., 25 Jul 2025).

This suggests that AVD is best treated as a broader methodological category rather than a single canonical algorithm. Under that broader reading, the defining feature is the supplementation of local dynamic-programming scores with nonlocal information: semantic priors from a transformer, marginalized latent-state effects, generalized pairwise couplings, or horizon-stability constraints. The common theme is that standard Viterbi assumes a local optimal-substructure property, whereas AVD-like settings arise precisely when the score of a partial path cannot be summarized adequately by the current discrete state alone.

A useful taxonomy emerges from the cited work. One strand keeps the classical trellis intact and adds external sequence priors, retaining tractability by using $K$-best or list-style decoding plus periodic rescoring. Another strand abandons the assumption that the decoded sequence is Markov and replaces direct recursion with admissible search. A third, more theoretical strand studies whether Viterbi decisions stabilize as the observation horizon grows, using nodes and barriers to characterize asymptotic path consistency.

2. Classical Viterbi structure and the source of AVD-like generalizations

The classical Viterbi algorithm applies when the score of a prefix can be summarized by the current state. In the hidden-state setting discussed in the triplet Markov model paper, the standard recursion can be written as

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$

with update

$$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$

or, in backward notation for a Markov chain $X$,

$$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$

The key computational fact is that a prefix needs only its terminal state to support optimal continuation (Soop et al., 25 Jul 2025).

The pairwise Markov model literature shows that this locality can survive in some generalized settings. For a homogeneous Markov chain $Z_k=(X_k,Y_k)$ on $\mathcal X\times\mathcal Y$, finite-horizon MAP decoding still admits a Viterbi recursion: $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$ followed by backtracking from $v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$ (Lember et al., 2017). Relative to an HMM, the local score now depends on $(x_t,x_{t+1},y',y)$, but the DP remains first-order in time because the hidden variable indexed by $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$0 is still the finite-state object on which optimal-substructure closes.

AVD-like complications begin when this closure fails. In the triplet Markov model setting, the quantity of interest is the marginal process $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$1 of a joint Markov chain $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$2, with

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$3

After conditioning on an observed $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$4 in a triplet Markov model $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$5, the decoding objective becomes

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$6

with

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$7

Although $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$8 is a non-homogeneous Markov chain, the marginal process $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$9 generally is not. The paper emphasizes that one does not typically have a factorization

$$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$0

nor the PMM/HMM-style conditional Markov property

$$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$1

This is exactly the structural gap in which AVD-like ideas become relevant: the decoder requires information beyond the current discrete state, whether supplied by an auxiliary latent state, a LLM, or an admissible relaxation (Soop et al., 25 Jul 2025).

3. Semantic-aware decoding with transformer LLMs

The clearest modern AVD-adjacent realization in the cited literature is the decoder in "LLM-Viterbi: Semantic-Aware Decoding for Convolutional Codes" (Li et al., 21 Apr 2026). Its baseline is standard maximum-likelihood Viterbi decoding for a convolutional code over an AWGN channel with BPSK signaling, with

$$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$2

where $$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$3 is i.i.d. Gaussian with variance $$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$4. The classical trellis recursion is expressed as

$$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$5

and the branch metric is the Euclidean distance

$$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$6

The proposed modification does not redefine the trellis. Instead, it preserves the convolutional-code state graph and replaces single-survivor decoding with $$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$7-best Viterbi, maintaining the $$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$8 best paths at each state and therefore at most $$\delta_{t+1}(j)=\max_i \delta_t(i)\, a_{ij}\, b_j(y_{t+1}),$$9 active paths across the trellis. The default settings are $X$0 and semantic evaluation interval $X$1 characters (Li et al., 21 Apr 2026).

The semantic model is a fine-tuned ByT5 LLM. The paper states that standard ByT5 follows an encoder-decoder Transformer architecture and that, in this work, ByT5 is utilized in decoder-only mode, where the encoder is bypassed and the decoder directly processes the character sequence. The LM supplies autoregressive probabilities

$X$2

Because bits are grouped into 8-bit ASCII characters, the decoder converts partial bit paths into text at character boundaries and evaluates linguistic plausibility on the resulting prefixes.

The paper formulates the intended integrated MAP criterion as

$X$3

and, under AWGN, uses

$X$4

Taking logarithms yields the path score

$X$5

The first term represents channel reliability, and the second linguistic plausibility. Importantly, the paper does not introduce a tunable fusion weight multiplying the LM term (Li et al., 21 Apr 2026).

Semantic information enters only periodically. At character position $X$6, the decoder groups all surviving paths by shared prefix $X$7, computes

$X$8

once per unique prefix group, selects the highest-scoring prefix, keeps all paths with that prefix, and prunes the rest. This “prefix-sharing” strategy reduces LM calls substantially while preserving multiple channel-consistent descendants of the selected text prefix (Li et al., 21 Apr 2026).

The empirical results show that this architecture can materially improve both coding performance and semantic fidelity. For convolutional codes with constraint length $X$9, the paper reports approximately $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$0 dB more coding gain in BLER and over $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$1 improvements in semantic similarity relative to conventional Viterbi decoding. At SNR $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$2 dB for $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$3, BLER is reported as $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$4 for LLM-Viterbi, $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$5 for standard Viterbi, and $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$6 for Viterbi plus one-shot LLM correction. At SNR $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$7 dB, semantic similarity scores are $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$8, $$\delta_t(x_{t-1})=\max_{x_t} p(x_t\mid x_{t-1})\,\delta_{t+1}(x_t).$$9, and $Z_k=(X_k,Y_k)$0, respectively; at SNR $Z_k=(X_k,Y_k)$1 dB, LLM-Viterbi reaches $Z_k=(X_k,Y_k)$2, and for SNR $Z_k=(X_k,Y_k)$3 dB all methods approach $Z_k=(X_k,Y_k)$4 (Li et al., 21 Apr 2026).

These results support a precise interpretation. The decoder is transformer-guided and context-aware, but the attention mechanism remains internal to ByT5 rather than being inserted into the trellis recursion itself. Accordingly, the method is more accurately characterized as $Z_k=(X_k,Y_k)$5-best or list Viterbi with periodic transformer LM posterior rescoring and prefix-group pruning than as an attention-weighted dynamic program in the strict sense.

4. Exact decoding beyond Markovian Viterbi recursion

A second major AVD-relevant direction concerns cases in which no exact Viterbi recursion exists over the sequence of interest. In triplet Markov models, decoding seeks the MAP sequence of $Z_k=(X_k,Y_k)$6 after marginalizing a nuisance latent process $Z_k=(X_k,Y_k)$7. The central point is that maximization over $Z_k=(X_k,Y_k)$8 and summation over $Z_k=(X_k,Y_k)$9 do not commute in the way required by the standard semiring-based Viterbi derivation. The cited paper writes, in essence,

$\mathcal X\times\mathcal Y$0

reflecting the generic inequality

$\mathcal X\times\mathcal Y$1

The result is that one cannot perform exact dynamic programming on $\mathcal X\times\mathcal Y$2 alone (Soop et al., 25 Jul 2025).

The proposed remedy is an exact branch-and-bound search over prefixes

$\mathcal X\times\mathcal Y$3

with branch objective

$\mathcal X\times\mathcal Y$4

Nodes are expanded by appending one symbol,

$\mathcal X\times\mathcal Y$5

and pruned whenever their upper bound is dominated by an already available lower bound. If $\mathcal X\times\mathcal Y$6, node $\mathcal X\times\mathcal Y$7 can be safely discarded. Exactness depends on admissible upper and lower bounds (Soop et al., 25 Jul 2025).

Several families of such bounds are developed. The simplest are

$\mathcal X\times\mathcal Y$8

A richer family uses power sums

$\mathcal X\times\mathcal Y$9

leading to

$$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$0

A Samuelson-type sharpening combines $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$1 and $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$2: $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$3 The paper also introduces swapped max-sum (SMS) upper bounds based on

$$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$4

implemented through blockwise maximization over $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$5-suffixes and dynamic programming on augmented states $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$6 (Soop et al., 25 Jul 2025).

Lower bounds come from tractable approximate continuations. The most important is the $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$7-Viterbi approximation, which replaces the non-Markov marginal by an $$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$8-th order Markov approximation

$$\delta_1(y)=p(x_1,y_1=y),\qquad \delta_{t+1}(y)=\max_{y'} \delta_t(y')\, q(x_{t+1},y\mid x_t,y'),$$9

then solves

$v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$0

using higher-order Viterbi recursion,

$v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$1

with argmax backpointers

$v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$2

A second heuristic, called UX-Viterbi, runs ordinary Viterbi on the augmented Markov chain $v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$3, producing a feasible $v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$4-sequence that yields a valid lower bound though generally not the true MAP path for $v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$5 (Soop et al., 25 Jul 2025).

The technical engine behind all of these bounds is the ability to compute marginal probabilities for fixed $v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$6 using the joint Markov property of $v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$7. The paper employs forward variables

$v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$8

and backward variables

$v_n=\arg\max_{y\in\mathcal Y}\delta_n(y)$9

so that

$(x_t,x_{t+1},y',y)$0

This is not standard Viterbi on $(x_t,x_{t+1},y',y)$1; it is dynamic programming on an augmented latent structure used to support systematic search (Soop et al., 25 Jul 2025).

In this sense, the branch-and-bound framework is AVD-relevant not because it uses attention, but because it formalizes a general decoding principle: when local optimal-substructure is destroyed by richer dependencies, exact MAP inference can still be approached through admissible search driven by augmented-state computations and controlled relaxations.

5. Infinite-horizon Viterbi paths, barriers, and stability

A third strand relevant to AVD concerns the asymptotic stability of MAP decoding. The pairwise Markov model paper studies whether finite-horizon Viterbi paths stabilize coordinatewise as the observation horizon tends to infinity. For an observation sequence $(x_t,x_{t+1},y',y)$2, an infinite Viterbi path $(x_t,x_{t+1},y',y)$3 is defined by the property that for every $(x_t,x_{t+1},y',y)$4 there exists $(x_t,x_{t+1},y',y)$5 such that

$(x_t,x_{t+1},y',y)$6

Thus the infinite path is not defined as an infinite-horizon global argmax over hidden sequences, but as the eventual stabilized prefix limit of finite-horizon MAP paths (Lember et al., 2017).

The paper shows that this stabilization problem is nontrivial. It gives an HMM example in which, for almost every observation sequence, there is no infinite Viterbi path because the finite MAP path flips infinitely often between two constant-state explanations. It also gives an example in which an infinite Viterbi path exists almost surely even though there are no nodes or barriers. These examples establish that infinite-horizon stabilization is neither automatic nor reducible to a single simple structural condition (Lember et al., 2017).

The core machinery uses nodes, strong nodes, and barriers. A time $(x_t,x_{t+1},y',y)$7 is an $(x_t,x_{t+1},y',y)$8-node of order $(x_t,x_{t+1},y',y)$9 if, after observing out to $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$00,

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$01

where

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$02

If the inequality is strict for all relevant competitors, the node is strong. A barrier is an observation block $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$03 that forces such a node regardless of surrounding observations; if the block occurs ending at time $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$04, then time $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$05 is an $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$06-node of order $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$07 (Lember et al., 2017).

Barriers matter because infinitely many separated nodes allow piecewise construction of the infinite path. Once node times $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$08 are available, one can optimize each segment under endpoint constraints, and only the last unfinished segment changes as the horizon extends. This yields finite-lookahead stabilization and supports a regenerative view of the decoded process (Lember et al., 2017).

The paper’s main abstract criterion constructs barrier sets via dominance conditions A1–A3. A more concrete existence route, Theorem 3.1, assumes a strictly positive measure $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$09, lower semicontinuous and bounded transition densities $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$10, structural conditions B1–B2, and Harris recurrence. Under these assumptions the authors derive a set $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$11 of strong barriers that occurs infinitely often almost surely, implying existence of an infinite Viterbi path and Viterbi process (Lember et al., 2017).

For AVD-like research, these results are conceptually important because they isolate a question often obscured in finite-length decoding benchmarks: whether adding more future context can keep changing earlier decisions indefinitely. The barrier formalism provides a rigorous language for online-stable decoding, regeneration points, and chunkwise computation in generalized sequence models.

6. Complexity, limitations, and methodological significance

The three cited strands also clarify the computational trade space of AVD-like methods.

For semantic-aware list Viterbi, the principal burden comes from retaining multiple paths and invoking a transformer periodically. The paper reports average decoding times per block of $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$12 ms, $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$13 ms, and $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$14 ms for standard Viterbi at $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$15, compared with $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$16 ms, $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$17 ms, and $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$18 ms for the proposed LLM-Viterbi. It therefore improves BLER and semantic similarity at substantial latency cost, especially as trellis size grows with $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$19 (Li et al., 21 Apr 2026).

For exact non-Markov MAP decoding, worst-case complexity remains exponential. The triplet Markov model paper states exhaustive search costs

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$20

or, with prefix reuse,

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$21

The simple bound has $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$22 time per node; one power-sum implementation has preparation

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$23

with per-node cost $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$24; $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$25-SMS preparation costs

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$26

and $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$27-Viterbi preparation

$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$28

with $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$29 time per node once precomputed. The paper is explicit that the maximal marginal problem is NP-hard, even for HMMs, by reduction from maximum clique (Soop et al., 25 Jul 2025).

Empirically, however, branch-and-bound can reduce search considerably. On $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$30 random TMMs with $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$31 and $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$32, average $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$33 visited-node counts drop from $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$34 for exhaustive search to $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$35 with simple bounds, $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$36 with $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$37-PS, and $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$38 with $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$39-SMS. The strongest practical results come from combining upper bounds with a $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$40-Viterbi lower bound: $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$41-PS $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$42 $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$43-Viterbi gives $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$44, and $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$45-SMS $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$46 $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$47-Viterbi gives $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$48. The paper concludes that $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$49-Viterbi lower bounds should essentially always be included, while no upper bound is uniformly dominant; it also stresses that $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$50-Viterbi is not monotone in $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$51, and larger $$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$52 can even produce a zero-probability path (Soop et al., 25 Jul 2025).

A comparative summary is useful.

Direction	Core mechanism	Relation to AVD
Semantic-aware list Viterbi (Li et al., 21 Apr 2026)	$$\delta_t(i)=\max_{x_{1:t-1}} p(x_{1:t-1},X_t=i,y_{1:t}),$$53-best trellis decoding with periodic ByT5 rescoring and prefix-group pruning	Transformer-guided and context-aware, but not branch-wise attention inside DP
Non-Markov exact MAP decoding (Soop et al., 25 Jul 2025)	Branch-and-bound with admissible bounds computed from augmented-state DP	Relevant as “Viterbi beyond local optimal-substructure,” not as an attention method
Infinite-horizon PMM theory (Lember et al., 2017)	Nodes, strong nodes, barriers, and recurrence arguments	Foundational for stability and online consistency of generalized Viterbi decoding

The most careful general conclusion is therefore terminological and methodological. “Attention Viterbi Decoding” is not represented here as a single established formalism. Rather, the cited work supports a broader conception in which Viterbi-style decoding is extended by nonlocal sequence information or by richer dependency structures that exceed the scope of ordinary first-order dynamic programming. In one realization, the nonlocal information is semantic and supplied by a transformer LM; in another, it is the residual dependence induced by marginalized auxiliary variables; in a third, it concerns asymptotic stabilization of the decoder itself. A plausible implication is that future work deserving the AVD label in a stricter sense would need to combine these threads: explicit context-sensitive scoring, trellis-compatible admissible summaries of nonlocal information, and guarantees about decoding stability or approximation quality.