acceptodds
Under review as a conference paper at ICLR 2027

PEAL: Probabilistic Edit Alignment Loss for Discrete Sequence Prediction

Abstract

Discrete sequence prediction requires models to identify the correct elements and capture their sequential dependencies. However, position-wise cross-entropy (CE) can encourage local corrections that disrupt a correct subsequence when it is shifted relative to the target—a limitation we call CE's myopia. We introduce Probabilistic Edit Alignment Loss (PEAL), which formulates alignment between predicted and target sequences as cost optimization over monotone edit paths. Substitution costs decrease with the probability assigned to the aligned target class, while insertion and deletion incur fixed penalties. PEAL aggregates the costs of all valid paths using a differentiable soft minimum. These edit paths allow correctly ordered subsequences to remain aligned despite positional shifts. Our gradient analysis shows that the resulting gradient is a Gibbs-weighted combination of path-cost gradients, with lower-cost paths receiving larger weights. We further establish conditional empirical -consistency: under specified conditions, any model minimizing empirical PEAL risk within a fixed model class is also optimal for structured prediction within that class, as measured by empirical edit-distance risk. Experiments on automatic speech recognition, temporal action segmentation, and long-term action anticipation demonstrate improvements across multiple backbone architectures. Further analyses associate these gains with better relative-order preservation and longer correctly aligned subsequences.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Sequence alignment, Discrete Sequence Prediction, Structured prediction, H-consistency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gZgxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a new loss function, PEAL, which computes a soft minimum over the costs of all possible edit paths aligning a predicted sequence with a target sequence. Insertions and deletions have fixed costs, while substitutions have a prediction-dependent cost of (1−p)α(1-p)^\alpha. The authors show that, under restrictive conditions, minimizing PEAL serves as a surrogate for minimizing empirical edit distance within a fixed hypothesis class. The loss can be used either as an auxiliary regularization term in a combined training objective or as a fine-tuning objective following initial training with another loss, such as CTC.

Critical strengths

One strength of the paper is that the proposed loss is supported by a theoretical analysis linking it to edit-distance minimization, although the result relies on very restrictive assumptions.

Critical weaknesses

Clarity of the theoretical presentation. The paper would benefit from explicitly defining terms such as empirical risk, excess risk, and H-consistency in the main text rather than in the appendices, or at least referring readers to the relevant appendix, which is currently not done. Since the paper is primarily application-oriented, these concepts should be clearly introduced to make the theoretical discussion accessible to a broader audience.

Motivation and positioning relative to existing work. The positioning of the paper with respect to existing alignment-based objectives is not entirely clear to me. The limitation of position-wise cross-entropy described by the authors is valid, but CE is already known to be poorly suited to tasks where the alignment between predictions and targets is unknown. This is precisely why objectives such as CTC and other alignment-aware losses have been introduced.

I therefore struggle to identify the specific limitation of these existing methods that PEAL aims to address. In the related work, the authors state that the implications of alignment-based objectives for "structured prediction capability require further analysis." However, it is unclear what is meant by structured prediction capability beyond the alignment and sequence-level prediction properties already addressed by these methods. Existing alignment-based losses have demonstrated strong empirical effectiveness, and the paper does not clearly explain what remains unresolved or how PEAL fundamentally differs in this respect.

The theoretical connection between PEAL and edit-distance minimization is interesting, but given the restrictive assumptions under which it is established, I am not convinced that it sufficiently addresses this motivation. A clearer comparison with existing alignment-based objectives would help establish the contribution.

Questions for authors

Justification for using soft-min smoothing: In Section 2.2, the authors justify the use of a soft minimum by stating that "Direct hard minimization is nondifferentiable when the minimum-cost path changes." Could the authors clarify this justification? As I understand it, the envelope theorem would allow one to first identify the minimum-cost alignment path and then compute a gradient step with respect to its cost, at least when the optimal path is unique. Why is soft-min smoothing preferred over this approach?

Choice of ASR dataset: Could the authors explain why they chose to evaluate ASR on a 24-minute subset of the Gronings dataset rather than on a more widely used benchmark such as LibriSpeech?

Handling of ASR alignments: How are blank tokens, repeated labels, and differences between acoustic-frame and target-sequence lengths handled? Is PEAL computed over all possible edit paths between the full frame-level predictions and the target sequence, or is some preprocessing applied?

15 comments

Part 1

W1

Clarity of the theoretical presentation. The paper would benefit from explicitly defining terms such as empirical risk, excess risk, and H-consistency in the main text rather than in the appendices, or at least referring readers to the relevant appendix, which is currently not done. Since the paper is primarily application-oriented, these concepts should be clearly introduced to make the theoretical discussion accessible to a broader audience.

Response

Thank you for suggesting clearer theoretical definitions.

We will state the risk definitions explicitly before Theorem 1 in Section 2.4 (Conditional Optimality Guarantees for PEAL) and connect them to the assumptions used in the theorem.

Empirical risk. Fix a sample set S=(xk,yk)k=1N\mathcal S=\\{(x_k,y_k)\\}_{k=1}^N and a nonempty class HH of probabilistic predictors. In Theorem 1, prediction and target lengths are both nn, and insertion and deletion costs are one. Let y^h(x)\widehat y_h(x) be the position-wise argmax sequence, using a fixed tie-breaking rule. The normalized target loss and theoretical surrogate are

ℓE(h;x,y)=dED(y^h(x),y)n,ℓS,τ(h;x,y)=Lτ(h(x),y)+τlog⁡Kn,nn,\ell_{\mathrm E}(h;x,y)=\frac{d_{\mathrm{ED}}(\widehat y_h(x),y)}{n}, \qquad \ell_{\mathrm S,\tau}(h;x,y) =\frac{L_\tau(h(x),y)+\tau\log K_{n,n}}{n},

where Kn,nK_{n,n} is the number of valid edit paths. For either loss ℓ\ell, empirical risk is its sample average, R^ℓ(h)=N−1∑k=1Nℓ(h;xk,yk)\widehat R_\ell(h)=N^{-1}\sum_{k=1}^N\ell(h;x_k,y_k). The path-count correction is constant across HH at fixed lengths. Consequently, the theoretical surrogate and practical normalized loss Lτ/nL_\tau/n have identical empirical excess risks (Appendix H.2).

Excess risk. This is the difference in risk from the best value attainable within the same class,

E^ℓ,H(h)=R^ℓ(h)−inf⁡h′∈HR^ℓ(h′).\widehat{\mathcal E}_{\ell,H}(h) =\widehat R_\ell(h)-\inf_{h'\in H}\widehat R_\ell(h').

Using the same HH for both losses makes the comparison about their preferred models within that class [9].

Conditional empirical HH-consistency. For the fixed sample set and class, the guarantee means that any sequence of models (ht)⊂H(h_t)\subset H with E^S,τ,H(ht)→0\widehat{\mathcal E}_{\mathrm S,\tau,H}(h_t)\to0 also satisfies E^E,H(ht)→0\widehat{\mathcal E}_{\mathrm E,H}(h_t)\to0. In our result, this follows from

E^E,H(h)≤E^S,τ,H(h)1−βξ,τ/g,βξ,τ<g,\widehat{\mathcal E}_{\mathrm E,H}(h) \leq\frac{\widehat{\mathcal E}_{\mathrm S,\tau,H}(h)}{1-\beta_{\xi,\tau}/g}, \qquad \beta_{\xi,\tau} < g,

where gg is the smallest positive empirical edit-distance excess risk, when one exists, and βξ,τ\beta_{\xi,\tau} is the uniform approximation width. Its confidence condition requires max⁡cph,i,c(xk)≥1−ξ\max_c p_{h,i,c}(x_k)\geq1-\xi for every h∈Hh\in H, sample kk, and position ii, with 0<ξ<1/20 < \xi < 1/2. For NN equally weighted samples of length nn, unit edit costs give g≥1/(Nn)g\geq1/(Nn), yielding the sufficient condition in Theorem 1.

Appendix H.1 defines the sequences, decoding rule, losses, and risks; H.2 relates the practical and theoretical objectives; K derives the approximation bounds; and L proves the excess-risk transfer. The fixed unequal-length extension is described in the response to W2 below.


Part 2

W2

Motivation and positioning relative to existing work. The positioning of the paper with respect to existing alignment-based objectives is not entirely clear to me. The limitation of position-wise cross-entropy described by the authors is valid, but CE is already known to be poorly suited to tasks where the alignment between predictions and targets is unknown. This is precisely why objectives such as CTC and other alignment-aware losses have been introduced.

I therefore struggle to identify the specific limitation of these existing methods that PEAL aims to address. In the related work, the authors state that the implications of alignment-based objectives for "structured prediction capability require further analysis." However, it is unclear what is meant by structured prediction capability beyond the alignment and sequence-level prediction properties already addressed by these methods. Existing alignment-based losses have demonstrated strong empirical effectiveness, and the paper does not clearly explain what remains unresolved or how PEAL fundamentally differs in this respect.

The theoretical connection between PEAL and edit-distance minimization is interesting, but given the restrictive assumptions under which it is established, I am not convinced that it sufficiently addresses this motivation. A clearer comparison with existing alignment-based objectives would help establish the contribution.

Response

Thank you for these questions about PEAL's motivation and contribution.

PEAL aims to make further optimization build on correct sequence structure already present in a prediction. A locally shifted subsequence can contain useful learned content even when the complete prediction has missing or extra events. The central question is which parts of that prediction the training signal reinforces and which errors it penalizes. PEAL determines this signal through edit correspondences, allowing correct shifted matches to contribute positively while missing and surplus events remain charged. We address four aspects below. (1) We explain how existing alignment and sequence-training objectives assign supervision, and what PEAL changes in the resulting updates. (2) We measure the amount and continuity of correct structure retained in the predictions. (3) We test the contribution of edit correspondences through ablations, training trajectories, and comparisons with related objectives. (4) We connect the theoretical bounds to initialization, temperature selection, and the relationship between surrogate and edit-distance optima.

1. How the supervision changes subsequent optimization.

CTC [1] resolves unknown frame-to-label timing by summing the probabilities of paths that collapse exactly to the reference. It can reinforce correct shifted events through compatible frame-level paths. When retaining those events would require an insertion or deletion outside its collapse rules, however, that partial correspondence supplies no path to the CTC likelihood. Optimization then follows the available exact-transcription paths, which can assign different labels to the positions already supporting correct content.

Consider the target A B C D and four prediction positions favoring B C D X. The first three predictions already form a correct ordered subsequence. On this four-position grid, CTC has only one path collapsing to the four distinct target labels, namely A B C D, so each position is pushed toward its diagonal target. PEAL admits an edit path retaining the three matches B C D and charging for missing A and extra X. This path lets optimization strengthen the existing three-label subsequence without requiring those positions to first change their labels to obtain training credit.

This difference can be checked directly. Over {A,B,C,D,X,blank}\{A,B,C,D,X,\mathrm{blank}\}, let each position assign probability 0.95 to its favored label and 0.01 to every other label. With α=1\alpha=1 and unit gap costs, the PEAL path retaining B C D costs 2+3(0.05)=2.152+3(0.05)=2.15, whereas the diagonal path costs 4(0.99)=3.964(0.99)=3.96. At τ=0.1\tau=0.1, enumerating all 321 edit paths gives a negative logit gradient of approximately −0.0475-0.0475 for each of these three correct labels. CTC gives a positive gradient of +0.95+0.95 for each on the same grid, using unnormalized losses. Gradient descent thus reinforces B C D under PEAL and suppresses them under CTC in this example. In actual ASR training, CTC operates on the acoustic frames, where additional frames can support other exact-transcription alignments; PEAL adds edit supervision on the compact event sequence constructed in Q3.

Part 3

Soft-DTW [2] preserves ordered matches through repeated pairings that represent temporal stretching. For prediction A A B C and target A B C C, both predicted A events can match one target A, and the final predicted C can match both target C events (Appendix G.2, Figure 9). As these pairwise content costs approach zero, the warped path can have near-zero cost despite the extra A and missing C. Thus, further improving the matched content can leave an event-count error unrepresented in that path's cost. PEAL uses one-to-one matches and charges the remaining deletion and insertion. A prediction with correct content but incorrect event multiplicities consequently retains an edit penalty. Its gap costs also affect which competing matches receive gradient weight, keeping the supervision sensitive to event-count discrepancies as well as label agreement.

AXE [3] preserves shifted matches while supervising skipped predictions through −log⁡pi,blank-\log p_{i,\mathrm{blank}} and skipped targets through −δlog⁡pi,yj-\delta\log p_{i,y_j} at an existing prediction. In the B C D X example, a path retaining B C D must still assign missing A to an existing prediction. If A and B share the position favoring B, that path contains both −δlog⁡p1,A-\delta\log p_{1,A} and −log⁡p1,B-\log p_{1,B}. Increasing the probability of missing A then competes with strengthening the already correct B. PEAL charges for missing A without assigning it an additional likelihood target at the B position. On that edit path, the position receives matching supervision for B. This changes what the optimizer is asked to preserve during correction, and the retention measurements below assess the resulting sequence structure.

The resulting local signal is explicit. Let Ai,jA_{i,j} be the total Gibbs probability of paths pairing prediction ii with target jj. Sections 2.2–2.3 give

\qquad \frac{\partial L_\tau}{\partial p_{i,c}} =-\alpha(1-p_{i,c})^{\alpha-1}\sum_{j:y_j=c}A_{i,j}.$$ A correct shifted label receives support from paths matching it to its reference occurrence. When several low-cost paths share that match, their weights add; when they disagree, the competing correspondences contribute according to their complete path costs. Gap costs alter these weights and have zero direct probability derivatives. Consequently, PEAL can reinforce correspondences shared by plausible alignments while adapting supervision where they differ. This also distinguishes PEAL from methods that introduce edit-distance information at other stages of training. MRT [4] and normalized N-best MWER/EMBR [5] use a candidate's centered sequence error to weight its log-likelihood gradient. Under autoregressive factorization, the same scalar coefficient multiplies all token log-likelihood terms in that candidate. This weights the candidate's correct and incorrect parts together. For a fixed candidate set with equal errors, the centered risk gradient is zero, even though their probability assignments can still be improved. PEAL's probability-sensitive costs and pairwise marginals retain feedback about which reference labels each prediction supports. O-1 [6] selects an oracle and a 1-best candidate for error-weighted likelihood updates; its available supervision depends on those selected hypotheses. PEAL obtains its correspondence weights directly from the supplied probability matrix and reference. OCD [7] and prefix-based TLE [8] provide finer next-action feedback. OCD teaches optimal continuations after a generated prefix, and TLE learns scores from completion-loss increments. The supervision for a position is constructed from the prefix preceding it and the possible continuations. PEAL uses the complete current probability matrix, including downstream predictions, to decide which reference positions should supervise that position. It can therefore reinforce an earlier label because of its role in an ordered subsequence already supported by later predictions. This makes the correspondence information available to parallel prediction heads without introducing an autoregressive completion decoder. We will clarify these differences in supervision and their optimization consequences in the Introduction and Related Work.
Reviewer_K9qXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PEAL, an alternative training loss to standard teacher-forced cross entropy (CE) for neural sequence modelling. To compute PEAL:

  1. Make a prediction using the current model weights
  2. Create the edit distance alignment lattice from the reference to the prediction
    • Each insertion / deletion gets a cost of 1
    • Each match gets a cost of 0
    • Each substitution gets a cost of (1−pi,yj)α(1-p_{i,y_j})^\alpha, where pi,yjp_{i,y_j} is the model probability of the jj-th reference token as the ii-th predicted token.
  3. The costs of all alignment paths is combined with a softmin to produce the loss value

Training can be either with an interpolation of CE and PEAL (PEAL-Joint), or in two stages (PEAL-FT: CE first, then PEAL).

Experimental results on Ego4D (long-term action anticipation), Gronings (ASR), and Breakfast (temporal action segmentation), show PEAL outperforms CE and other baselines when evaluated with edit distance related metrics.

Critical strengths
  • The paper is well-motivated from CE's weakness: CE has been long known to be a surrogate for maximizing sequence accuracy (Goodman 1996), instead of minimizing edit distance or other per-position metrics. This paper provides another perspective of CE's weakness under H-consistency (Appendix K).
  • The proposed method is simple to implement compared to RL / EMBR based methods, and may be promising in effectiveness.
Critical weaknesses
  • There is a rich literature on sequence modelling with minimizing edit distance as an explicit goal. The connection of this work with them should be more sufficiently discussed. For example
  • The experiments compared against label smoothing (LS), focal loss (FL), and aligned cross entropy (AXE). Only AXE is arguably another method that attempts to more closely optimize for the evaluation metric. LS is merely a regularization method. As a result, it is not clear from results in Table 1 & 2, whether PEAL achieved better results because of direct optimization of edit distance, or because of indirect, more effective regularization (especially for Gronings). Comparisons with more closely related previous works would be necessary to strengthen the claim of PEAL's effectiveness.
Questions for authors

(Already covered in the previous section)

6 comments

PART 1

W1

There is a rich literature on sequence modelling with minimizing edit distance as an explicit goal. The connection of this work with them should be more sufficiently discussed. For example

  • Optimal Completion Distillation
  • EMBR
  • O1
  • Minimum Risk Training
  • Task Loss Estimation
  • Many direct RL applications: GOLD, actor-critic, Reward Augmented Maximum Likelihood

Response

Thank you for highlighting these related methods.

The listed methods incorporate task information through candidate-level weighting, prefix-conditioned supervision, or reward-weighted likelihood training. PEAL derives prediction-reference correspondence weights from a continuous edit-path objective.

For these comparisons, the relevant PEAL expressions from Sections 2.2 and 2.3 are

C_p(\pi) &=N_D(\pi)c_{\mathrm{del}}+N_I(\pi)c_{\mathrm{ins}} +\sum_{(i,j)\in S(\pi)}(1-p_{i,y_j})^\alpha,\\ L_{\mathrm{PEAL}} &=-\tau\log\sum_{\pi\in\Pi_{n,m}}\exp[-C_p(\pi)/\tau]. \end{aligned} \tag{1}$$ Here $p_{i,c}$ is the probability of class $c$ at prediction position $i$, $y_j$ is target label $j$, and $\Pi_{n,m}$ contains monotone paths on the prediction-target grid. $N_D$ and $N_I$ count deletions and insertions, and $S$ is the set of diagonal pairs. For $\alpha,\tau>0$, the position-level gradient is $$\begin{aligned} \frac{\partial L_{\mathrm{PEAL}}}{\partial p_{i,c}} &=-\alpha(1-p_{i,c})^{\alpha-1} \sum_{j:y_j=c}A_{i,j},\\ A_{i,j}&=\sum_{\pi:(i,j)\in S(\pi)} \frac{e^{-C_p(\pi)/\tau}}{\sum_{\pi'}e^{-C_p(\pi')/\tau}}. \end{aligned} \tag{2}$$ Here $A_{i,j}$ is the Gibbs marginal of paths aligning prediction position $i$ with reference position $j$. A prediction position can therefore receive supervision from different reference positions according to their global alignment weights. The loss and marginals are computed exactly on the supplied grid by dynamic programming. ### Candidate-Level Metric Weighting: MRT [4], MWER/EMBR [5], and O-1 [6] * **Core Difference (Candidate-Level Weights vs. Pairwise Alignment Weights):** * **Challenge:** These methods use task metrics to weight candidate sequence probabilities or likelihoods. MRT minimizes $\widetilde R_{\mathrm{MRT}} = \sum_{u\in\mathcal S} Q_u \Delta(u, y)$ over sampled candidates, where $Q_u = P_\theta(u\mid x)^\eta / \sum_{v\in\mathcal S} P_\theta(v\mid x)^\eta$. Its gradient multiplies each candidate's sequence log-likelihood gradient by a centered risk coefficient. Normalized N-best MWER uses the same weighted-expectation form with word-error counts. O-1 selects the lowest-WER oracle hypothesis and the highest-probability hypothesis from a beam, increasing the former's likelihood and penalizing the latter's with WER-dependent weights. In each case, task errors supply candidate-level coefficients for sequence log-likelihood gradients. * **Our Improvement:** PEAL defines a separate correspondence weight $A_{i,j}$ for each prediction-reference position pair. Equation (2) combines these weights with local probability-cost sensitivity, making the supervision at a prediction position depend on which reference positions it aligns with. ### Prefix-Conditioned Supervision: OCD [7], TLE [8], and Actor-Critic [10] * **Core Difference (Next-Action Supervision at Prefixes vs. Supervision over Complete Edit Paths):** * **Challenge:** These methods construct next-action supervision conditioned on a supplied or generated prefix. OCD computes optimal edit-distance completions by dynamic programming and distills a next-token teacher, uniform over optimal tokens in its original zero-temperature experiments. TLE defines the optimistic loss $\Delta_o(s)=\min_v\Delta(sv,y)$ and regresses next-action scores toward the increments $\Delta_o(sa)-\Delta_o(s)$; its greedy squared-error variant supervises all next actions at each generated prefix. Actor-Critic learns reference-conditioned estimates of future rewards under the actor's policy and uses them to weight next-token probability gradients. Their local training signals thus depend on the continuation of a discrete prefix. * **Our Improvement:** PEAL determines supervision by aggregating complete monotone edit paths on the supplied prediction-target grid. Its training signal follows directly from the differentiable path objective, so constructing it requires neither an autoregressive rollout nor a learned completion or value score. This also makes the objective directly applicable to parallel predictors.

PART 2

Reward-Weighted Likelihood Training: GOLD [9] and RAML [11]

  • Core Difference (Reweighting Demonstrations or Sampling Targets vs. Aligning to a Fixed Reference):
    • Challenge: These methods incorporate rewards by changing the weights or targets used in likelihood training. GOLD trains offline on reference demonstrations, weighting each demonstrated token's log-likelihood by an importance weight and a return. RAML samples augmented target strings from the fixed reward distribution qT(u∣y)∝exp⁡[r(u,y)/T]q_T(u\mid y)\propto\exp[r(u,y)/T] and minimizes −Eu∼qTlog⁡Pθ(u∣x)-\mathbb E_{u\sim q_T}\log P_\theta(u\mid x). GOLD therefore changes the influence of demonstrated tokens, while RAML changes the target strings that supervise the model.
    • Our Improvement: PEAL keeps the reference string fixed and places its Gibbs distribution over alternative edit alignments. Insertion and deletion steps absorb positional shifts, and the resulting marginals determine which reference positions supervise each prediction.

Empirical Comparison on Gronings ASR

We compare PEAL with CTC adaptations of OCD, O-1, and TLE using the XLS-R encoder and greedy CTC evaluation (main text Table 6; Appendix F). The training configurations are:

  • AR-OCD and AR-TLE: Both use a 256-dimensional GRU decoder sharing the XLS-R encoder. AR-OCD uses acoustic cross-attention and sampled prefixes to construct its optimal-completion teacher; AR-TLE uses greedy prefixes and MSE supervision against oracle edit-distance completion loss increments. Training uses LCTC+λLOCDL_{\mathrm{CTC}}+\lambda L_{\mathrm{OCD}} or LCTC+λLTLEL_{\mathrm{CTC}}+\lambda L_{\mathrm{TLE}} for 13,000 steps, with λ\lambda warmed up from 00 to 0.150.15 over 260 steps. Evaluation uses the CTC branch.
  • O-1: Width-8 CTC prefix beam search generates training candidates, whose differentiable posterior probabilities are computed by CTC dynamic programming. Starting from CTC checkpoint-12000, training optimizes 0.1LCTC+LO10.1 L_{\mathrm{CTC}} + L_{\mathrm{O1}} with length normalization for 2,000 steps at learning rate 10−510^{-5}.
  • PEAL: Grid positions are non-blank events formed by averaging frame posteriors within consecutive argmax segments. Gradients propagate through event averaging and the edit grid with segment boundaries fixed (Appendix C.4).
MethodTest WER ↓Test CER ↓
CTC Baseline32.53%8.22%
CTC + AR-OCD31.42%7.84%
CTC + O-1 (Fine-tuned)31.14%7.91%
CTC + AR-TLE33.38%8.42%
PEAL-Joint (CTC + PEAL)29.13%7.15%

Relative to AR-OCD, O-1, and AR-TLE in this CTC setting, PEAL-Joint reduces WER by 2.29, 2.01, and 4.25 percentage points, respectively, and CER by 0.69, 0.76, and 1.27 percentage points.


PART 3

W2

The experiments compared against label smoothing (LS), focal loss (FL), and aligned cross entropy (AXE). Only AXE is arguably another method that attempts to more closely optimize for the evaluation metric. LS is merely a regularization method. As a result, it is not clear from results in Table 1 & 2, whether PEAL achieved better results because of direct optimization of edit distance, or because of indirect, more effective regularization (especially for Gronings). Comparisons with more closely related previous works would be necessary to strengthen the claim of PEAL's effectiveness.

Response

Thank you for raising this concern about implicit regularization.

To distinguish direct edit-distance optimization from implicit regularization, we examine PEAL's alignment mechanism from three complementary angles: (1) fixed-position controls retain the probability-sensitive substitution cost while prohibiting cross-position alignment, isolating the contribution of dynamic edit correspondences beyond local probability reweighting; (2) training and validation edit-distance trajectories test whether PEAL directly improves sequence-level fitting on the training set rather than merely constraining generalization; and (3) joint combinations with LS [1] or FL [2] test whether conventional position-wise regularizers complement or compete with PEAL. Full PEAL outperforms the fixed-position controls in both LTA and ASR.

Operational Ablation Experiments: Restricting Cross-position Alignment

To isolate the contribution of explicit cross-position alignment from local probability-sensitive reweighting (a potential regularization-like effect), we construct a fixed-position ablation control. In full PEAL, global monotone edit paths dynamically establish correspondences between predicted positions ii and reference positions jj via insertion, deletion, and substitution operations. The control preserves the exact same local probability cost (1−pi,yj)α(1-p_{i,y_j})^\alpha—retaining probability-sensitive gradient reweighting—while disabling insertions and deletions to strictly prohibit cross-position alignment. For instance, when a missing initial symbol shifts a correct subsequence, the fixed-position control still forces supervision at misaligned indices, whereas full PEAL absorbs the missing element via an insertion and aligns the shifted subsequence to its matching reference labels. Comparing full PEAL against this control therefore directly measures the performance gain attributable to cross-position edit alignment beyond local probability reweighting.

LTA Operational Ablation: Table 4 in Section 3.3 compares different operational combinations on SlowFast and Qwen-3.5-4B (under joint training). When prediction and target sequences are equal in length, the number of insertions and deletions in any valid path must be equal; therefore, SS, I+SI+S, and D+SD+S all reduce to rigid diagonal matching/substitution, unable to absorb local shifts. The results are summarized below (normalized edit distance, lower is better):

Auxiliary ObjectiveSlowFast VerbSlowFast NounQwen VerbQwen NounQwen Action
Insert + Delete only (I+DI+D)0.75990.79780.67930.70460.8721
Diagonal Match/Sub only (SS / I+SI+S / D+SD+S)0.72020.78230.67940.68900.8653
Full PEAL (I+D+SI+D+S)0.68750.76280.66050.68780.8552

Full PEAL has lower edit distance than the diagonal variants across all reported heads and backbones. On SlowFast, verb edit distance decreases from 0.7202 to 0.6875 and noun edit distance decreases from 0.7823 to 0.7628. The I+DI+D-only condition does not use prediction probabilities in its auxiliary cost and performs worse than full PEAL. Together, these comparisons show the contribution of combining probability-dependent substitution costs with insertion and deletion operations.

Gronings ASR Operational Ablation: For low-resource ASR, we compared "base loss + fixed-position match/sub" against "base loss + full PEAL" (Appendix F.1, Table 15). The fixed-position control pairs the common prefix of prediction and target sequences position-by-position, penalizing the remaining length difference with unit insertion/deletion costs; only prefix matches provide gradients, while tail penalties provide no gradients. The results (in percentage) are:

Auxiliary ObjectiveTest WER ↓Test CER ↓
Fixed-position Match/Sub only30.86%7.74%
Full PEAL29.13%7.15%

Full PEAL reduces WER by 1.73 percentage points and CER by 0.59 percentage points relative to the fixed-position control. This result is consistent with the LTA operation ablation: in both settings, the full edit objective performs better than a control that restricts matching to fixed positions.

Reviewer_G5GsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the mismatch between position-wise cross-entropy and sequence-level evaluation objectives in discrete sequence prediction. Even when a prediction contains a subsequence with correct content and relative ordering, positional shifts may cause cross-entropy to encourage undesirable local corrections. The authors propose Probabilistic Edit Alignment Loss (PEAL), which defines costs over monotonic edit paths comprising insertions, deletions, and substitutions, and aggregates all paths using a soft minimum. Substitution costs depend on the model’s predicted probability for the aligned target class, while insertion and deletion costs are fixed. The authors derive PEAL’s gradient as a Gibbs-weighted combination of path-cost gradients. They also establish conditional empirical H\mathcal H-consistency with respect to edit distance under assumptions including equal-length sequences, uniformly high confidence across the hypothesis class, and sufficiently small approximation error. PEAL is evaluated through joint training or subsequent fine-tuning on long-term action anticipation, temporal action segmentation, and low-resource automatic speech recognition. The results generally support its practical utility, while further analyses associate the improvements with better relative-order preservation and longer correctly aligned subsequences.

Critical strengths
  1. The method directly addresses the mismatch between position-wise supervision and sequence-level evaluation, with supporting analysis and ablations. PEAL uses insertions and deletions to accommodate positional shifts while preserving one-to-one, order-preserving matches, making it suitable for discrete sequence tasks where repetitions, omissions, and substitutions need to be distinguished. Section 2.3 and Appendices G–H clearly explain how different paths contribute to the gradient. The ablation results in Table 4 further show that the complete edit-path formulation outperforms diagonal-path variants that cannot accommodate positional shifts. Together, these findings support the role of sequence alignment and make the method reasonably interpretable.

  2. The cross-task experiments demonstrate practical benefits, with substantial improvements in some settings. In Table 1, PEAL-Joint reduces SlowFast’s normalized verb edit distance from 0.7546 with CE to 0.6875, also outperforming AXE-Joint at 0.7140. For speech recognition, WER decreases from 0.3253 with CTC to 0.2913. In Table 2, PEAL outperforms the listed baselines on the temporal action segmentation edit score, F1 scores at all reported thresholds, and frame-wise accuracy. Together with the evaluations across backbones and training strategies and the order-preservation analyses, these results provide valuable initial evidence for PEAL as a reusable training objective.

Critical weaknesses
  1. The central consistency guarantee has a narrow scope and remains substantially disconnected from the main experimental configurations. Theorem 1 requires equal prediction and target lengths, unit insertion and deletion costs, and uniformly high confidence for every hypothesis at every training input and prediction position. It additionally requires the approximation error to be smaller than 1/(Nn)1/(Nn). Appendix J.3 explicitly notes that, for the default LTA setting of n=20n=20 and τ=0.1\tau=0.1, the temperature-related error term alone is approximately 0.16597, which already prevents this sufficient condition from being satisfied. The theorem also does not directly cover joint training or the sampling and Viterbi decoding procedures used in the experiments. Consequently, the result does not currently provide a consistency guarantee for the main experimental configurations or establish generalization to the test distribution. Although the authors disclose these limitations, they reduce the practical explanatory value of the theory as a central contribution. The concern is the scope of the guarantee; failure to satisfy a sufficient condition does not itself imply that the method is inconsistent.

  2. The experiments do not yet establish the stability of smaller gains or their dependence on hyperparameter tuning. The paper does not report results across multiple random seeds or confidence intervals. TAS is evaluated only on Breakfast split 1, and ASR uses approximately 24 minutes of training data. Some gains are small or inconsistent: for example, Qwen’s noun edit distance changes from 0.6861 with CE to 0.6878 with PEAL-Joint. Moreover, the default PEAL configuration in Table 6 achieves an average F1 of 67.330 on TAS, below Focal Loss’s 68.306 in Table 2, whereas the selected configuration reaches 69.439. The appendix describes an independent validation-based selection procedure, but does not sufficiently clarify whether the competing losses received comparable tuning budgets. These limitations make it difficult to assess whether some advantages are reproducible and how much of the improvement can be attributed to the objective itself.

Questions for authors
  1. Connection between theory and practical training: The appendix acknowledges that the default configuration does not satisfy the sufficient condition in Theorem 1. Is there a tighter approximation bound or risk-gap bound that could provide a valid guarantee for any actual experimental configuration? If not, what testable guidance does the theoretical result offer for practical parameter selection or training behavior? This would affect my assessment of the practical significance of the theoretical contribution.

  2. Stability of improvements and fairness of comparisons: Could you provide results across multiple random seeds for TAS and for the Qwen results where the differences are small, and clarify the hyperparameter search ranges, selection criteria, and training budgets for PEAL, AXE, Focal Loss, and the other objectives? In particular, does the difference between the default and selected configurations in Table 6 persist across repeated runs?

  3. Implementation details across tasks: What are the joint-loss weights λ\lambda, the length-normalization schemes for the individual losses, and the initialization and training budgets for PEAL-FT? For ASR, is PEAL applied directly to the raw acoustic-frame probability distributions or to a transformed sequence? How are blank symbols, consecutive repeated characters, and padding handled? These details affect the interpretation of loss scaling and alignment semantics, as well as reproducibility.

  4. LTA evaluation protocol: Appendix C.1 evaluates the minimum edit distance among five sampled candidates. Does PEAL retain its advantage under position-wise argmax decoding or single-candidate evaluation? This would help distinguish improvements in the alignment quality of individual predictions from changes in the sampling distribution and candidate diversity.

10 comments

Part 1

W1&Q1

Comment summary

The conditional consistency guarantee relies on assumptions about sequence length, edit costs, and prediction confidence. How do these assumptions relate to practical PEAL training? Can tighter approximation bounds or an analysis of the empirical risk gap certify a concrete experimental configuration? What guidance does the theory provide for choosing the training strategy, transition timing, temperature, and cost exponent, and how can this guidance be tested experimentally? The relationship between the guarantee and joint training, decoding, and generalization to unseen data also needs clarification.

Response:

Thank you for these questions about the theoretical guarantees and practical training.

We address these concerns through four connected analyses centered on our generalized approximation error bound βn,m\beta_{n,m}. (1) We extend the conditional empirical HH-consistency bound to unequal sequence lengths (n,m)(n,m), obtaining an approximation bound with a substitution term governed by prediction confidence ξ\xi and cost exponent α\alpha, and a smoothing term governed by temperature τ\tau. (2) The confidence-dependent term shows how reducing ξ\xi shrinks the approximation gap, motivating our two-stage CE-to-PEAL training (PEAL-FT), which we investigate through experiments comparing different transition timings. (3) The τ\tau- and α\alpha-dependent terms guide hyperparameter search. Measurements at the default LTA temperature τ=0.1\tau=0.1 further show that the actual mean smoothing bias is much smaller than its worst-case bound, helping explain the observed gains when the conservative sufficient condition is unmet. (4) In the paper's finite hypothesis example, we verify that the sufficient condition holds and certifies the transfer from empirical PEAL optimality to empirical edit-distance optimality, aligning with the strong empirical LTA performance observed at the same small temperature.

Theoretical Extension to Unequal Sequence Lengths

Theorem 1 in the initial submission adopted equal predicted and target sequence lengths (n=mn = m) as a simplified case, and the result extends directly to general unequal lengths. Let the predicted length be n≥1n \geq 1 and the target length be m≥1m \geq 1. Keeping all other assumptions unchanged, consider the general setting where all samples and candidate models share a fixed length pair (n,m)(n, m) with n≠mn \neq m.

Extending the approximation bound to this (n+1)×(m+1)(n+1) \times (m+1) rectangular grid requires only two adjustments to the equal-length derivation. First, because PEAL and the discrete edit distance assign identical fixed costs to insertions and deletions, their discrepancy arises solely from aligned matching/substitution steps; since any monotone edit path aligns at most min⁡(n,m)\min(n, m) element pairs, the probabilistic substitution error accumulates over at most min⁡(n,m)\min(n, m) positions. Second, the total number of monotone paths aggregated by the soft-min operator becomes Kn,mK_{n, m}.

Normalizing both PEAL and the discrete edit distance by the target length mm, the per-sample approximation error bound takes the general form:

=\frac{\min(n,m)}{m}\bigl[1-(1-\xi)^\alpha+\xi^\alpha\bigr] +\frac{\tau\log K_{n,m}}{m}.$$ For a training set of $N$ equally weighted samples, whenever $\beta_{n,m} < 1/(Nm)$, optimizing empirical PEAL to near-optimality within the hypothesis class guarantees near-optimal average normalized edit distance. Setting $n = m$ directly recovers the expression in Theorem 1, confirming that the conditional consistency guarantee holds generally for unequal sequence lengths. (The full derivation on the rectangular grid is provided in Appendix K.4, Eq. (73), and the corresponding excess-risk transfer is proven in Appendix L.3 of the revised manuscript.)

Part 2

Error Bound Sources and Theoretical Motivation for PEAL-FT

The generalized approximation error bound βn,m\beta_{n,m} in Appendix K.4, Eq. (73), provides direct theoretical motivation for our two-stage training strategy (PEAL-FT). In βn,m\beta_{n,m}, the error accumulated from discrete-to-continuous probabilistic substitution is governed by the confidence-dependent term min⁡(n,m)m[1−(1−ξ)α+ξα]\frac{\min(n,m)}{m}[1-(1-\xi)^\alpha+\xi^\alpha]. Under α=1\alpha = 1 used in SlowFast, this term simplifies to min⁡(n,m)m⋅2ξ\frac{\min(n,m)}{m}\cdot 2\xi. As prediction probabilities concentrate on the model's argmax classes and ξ\xi decreases, this confidence-dependent term shrinks linearly with ξ\xi. This directly motivates PEAL-FT: cross-entropy (CE) initialization first builds initial representations and prediction confidence (reducing ξ\xi), followed by pure PEAL fine-tuning for sequence-level alignment. Our SlowFast ablation below evaluates the equal-length case n=m=20n=m=20.

To verify transition timing experimentally, under identical data, learning rate schedule, and total budget (50 epochs) sharing the same initial CE trajectory, we ablated introducing PEAL at epochs 10, 15, 20, 25, and 30 (each using a 10-epoch linear ramp-up transition). In this single-seed ablation, all groups switching between epochs 15–30 outperformed continuous CE on both verb and noun at the final checkpoint under both argmax and best-of-5 evaluation, whereas switching too early at epoch 10 led to noun degradation. These results support pre-training followed by fine-tuning (transition curves are reported in Appendix D.6, Figures 5 and 6).

Part 3

Guidance of Error Bounds on Hyperparameter Search Ranges

The approximation width βn,m\beta_{n,m} in Appendix K.4, Eq. (73), also guides the initial hyperparameter search through its confidence-dependent term min⁡(n,m)m[1−(1−ξ)α+ξα]\frac{\min(n,m)}{m}[1-(1-\xi)^\alpha+\xi^\alpha] and its soft-min smoothing term τlog⁡Kn,m/m\tau\log K_{n,m}/m, which characterize the effects of temperature τ\tau and cost exponent α\alpha. The SlowFast results below evaluate the equal-length case n=m=20n=m=20.

Temperature τ\tau and Empirical Approximation Error. The smoothing term τlog⁡Kn,m/m\tau\log K_{n,m}/m increases linearly with τ\tau (reducing to τlog⁡Kn,n/n\tau\log K_{n,n}/n when n=mn=m), motivating an initial search over smaller positive temperatures such as 0.01, 0.05, and 0.1. Empirically, Table 3(b) reports Verb/Noun edit distances of 0.6878/0.7622 at τ=0.01\tau=0.01. Increasing τ\tau to 0.05 lowers Noun ED to its minimum of 0.7551 while raising Verb ED to 0.6941. At the default τ=0.1\tau=0.1, the scores are 0.6875/0.7628, and Verb ED reaches its minimum of 0.6845 at τ=0.2\tau=0.2 (where Noun ED is 0.7642). Increasing τ\tau further to 0.5 degrades both heads to 0.6910/0.7653 relative to τ=0.2\tau=0.2. Thus, while the worst-case smoothing bound grows monotonically with τ\tau, empirical performance varies nonmonotonically and favors different temperatures across heads, requiring final selection on validation data.

To understand why the default setting τ=0.1\tau=0.1 performs well even though its worst-case smoothing bound falls outside the conservative sufficient condition, we directly measure how closely PEAL approximates discrete edit distance in actual LTA runs at τ=0.1\tau=0.1. For sequences of length 20, let ℓraw=Lτ/20\ell_{\mathrm{raw}}=L_\tau/20, ℓE=dED/20\ell_{\mathrm E}=d_{\mathrm{ED}}/20, and ℓS=(Lτ+τlog⁡K20,20)/20\ell_{\mathrm S}=(L_\tau+\tau\log K_{20,20})/20 denote the normalized raw PEAL objective, normalized argmax edit distance, and path-count-corrected theoretical surrogate, respectively. The table below reports the signed difference ℓ−ℓE\ell-\ell_{\mathrm E}, mean absolute difference, and median absolute difference across the evaluated LTA runs:

ObjectiveHeadMean signed differenceMean absolute differenceMedian absolute difference
Raw PEAL / 20Verb0.0410420.0454740.033773
Raw PEAL / 20Noun0.0119810.0243200.007270
Path-count-corrected surrogateVerb0.2070110.2070110.198578
Path-count-corrected surrogateNoun0.1779500.1781490.165732

The raw objective exhibits mean absolute discrepancies of 0.045474 for Verb and 0.024320 for Noun, with smaller median absolute discrepancies of 0.033773 and 0.007270 (reflecting a long tail in the sample-wise error distribution). Moreover, let LhardL_{\mathrm{hard}} denote the minimum path cost under the same probabilistic edit costs. The normalized soft-min smoothing bias is (Lhard−Lτ)/20(L_{\mathrm{hard}}-L_\tau)/20, whose worst-case path-count bound at τ=0.1\tau=0.1 is τlog⁡K20,20/20≈0.165969\tau\log K_{20,20}/20 \approx 0.165969. In practice, the measured mean smoothing bias is approximately 0.00812 for Verb and 0.00265 for Noun—more than an order of magnitude below this worst-case bound. (The larger discrepancy of the path-count-corrected surrogate in the table simply reflects adding the full constant 0.165969 to every loss, which leaves gradients and within-class minimizers unchanged for fixed sequence lengths.) These measurements show that while the worst-case bound overestimates average smoothing bias and the uniform sufficient condition is not satisfied at τ=0.1\tau=0.1, the trained objective still tracks edit distance closely on average, explaining its strong empirical performance at τ=0.1\tau=0.1 (0.6875/0.7628).

Cost Exponent α\alpha. For fixed ξ\xi, increasing α\alpha increases 1−(1−ξ)α1-(1-\xi)^\alpha while decreasing ξα\xi^\alpha, so the combined confidence-dependent term depends on their trade-off. Table 3(a) illustrates this sensitivity on LTA: α=1\alpha=1 yields Verb/Noun edit distances of 0.6875/0.7628 (achieving the lowest Noun ED among tested exponents); α=1.5\alpha=1.5 achieves the lowest Verb ED at 0.6865 while Noun ED rises to 0.7705; and α=3\alpha=3 degrades both heads to 0.7045/0.7703. Because the error decomposition alone does not prescribe an empirically optimal exponent, these results support anchoring the search around α=1\alpha=1 and selecting the final value on validation data. A formal consistency certificate additionally requires βn,m<g\beta_{n,m}<g (Appendices L.2–L.3), where gg is the minimum positive empirical edit-distance excess risk within the hypothesis class.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.