Read-Head Disagreement as a Free Committee for Active Learning in Simultaneous Translation

A simultaneous translation model starts writing the translation before the speaker has finished the sentence. One common design, monotonic multihead attention, gives each attention head its own read cursor. The cursor only moves forward, and each head decides for itself when it has read enough source words to let the next target word be written. These heads already act like a small committee voting on when to commit. On sentences the model has mastered they tend to agree. On sentences it has not, some heads stop early and others keep reading. This proposal treats that disagreement as an uncertainty signal and uses it to choose which source sentences to send to human translators when the translation budget is small.

The appeal is cost. The classic way to find uncertain examples is query-by-committee: train several models and query the examples on which they disagree. Scoring a large unlabeled pool then costs one full decode per committee member, and that scoring step is often the bottleneck in active learning (Understanding the Training Speedup from Sampling with Approximate Losses shows uncertainty scoring can cost more wall-clock time than it saves). Reading disagreement out of one model's own heads costs a single greedy decode. Whether this signal is worth anything beyond the model's ordinary word-level confidence is the open empirical question, and the proposal is built to answer it.

Core Claim

Disagreement among monotonic read heads about where to stop reading measures something token entropy does not: whether the model knows how much source context a given output word needs. We claim this signal has three properties. First, it is computable during one greedy simultaneous decode of an unlabeled source sentence, at no extra cost. Second, ranked moderately rather than maximally, it selects sentences whose translations improve the quality-versus-delay trade-off more per labeled sentence than random or entropy-based selection. Third, its extreme tail is dominated by sentences with heavy word reordering. A monotonic student cannot learn much from those sentences, so selecting by maximum disagreement should backfire.

For a source sentence $x$ decoded into $U$ target tokens, let $t^h_i$ be the source position at which head $h$ (of $H$ heads) stops before writing target token $i$, and let $T$ be the source length. The score is

$$ D(x) = \frac{1}{U}\sum_{i=1}^{U} \frac{\mathrm{Var}_{h}\left(t^h_i\right)}{T^2}, $$

which is the average spread of the heads' read positions, normalized so that long and short sentences are comparable. During training we can replace $t^h_i$ with the expected stop position under the model's soft stopping probabilities. At pool-scoring time the hard positions come for free from decoding.

Relation to Prior Work

The mechanism comes from Monotonic Multihead Attention, which gives each Transformer head an independent stop-or-advance policy. In its MMA-IL variant, a head that stops attends softly to everything read so far. That paper treats head disagreement purely as a latency problem and suppresses it with a head-divergence loss. This creates a real tension, because that loss shrinks exactly the variance we want to read. We address it directly: the divergence-loss weight is an experimental factor, and we test whether the ranking signal survives under the latency-optimal setting.

The earlier monotonic attention papers, Online and Linear-Time Attention by Enforcing Monotonic Alignments and Monotonic Chunkwise Attention, constrain the proposal. They document that forward-only reading fails under strong word reordering, and that failure is why we expect the most-disagreeing sentences to be poor training material.

Two lines of work motivate the "moderate, not maximal" selection rule. Robust Active Distillation shows on image and text classification that the most uncertain points coincide with unreliable labels, and it spreads its budget over moderately hard points. Privacy-preserving and Uncertainty-aware Federated Trajectory Prediction for Connected Autonomous Vehicles finds that selecting at median rather than maximal uncertainty avoids corrupt outliers. We carry this idea to sequence alignment, where the "unlearnable" tail has a concrete cause: the source word order cannot be followed left to right.

The closest data-centric work is Better Simultaneous Translation with Monotonic Knowledge Distillation. It rewrites training targets to be more monotonic using an offline teacher, but it curates the whole corpus statically and uses no uncertainty. Active Data Curation Effectively Distills Large-Scale Multimodal Models also selects data by uncertainty, but it needs a large external reference model and works only on vision-language models. Our signal needs neither a teacher nor extra models.

On the ensemble side, Collaborative Learning for Deep Neural Networks trains several heads on a shared backbone to get ensemble benefits at single-model test cost, and LoRA-Ensemble ensembles lightweight adapters for uncertainty. Both build the committee on purpose. We ask whether a committee the architecture already contains is good enough.

Research Program

The setting is pool-based active learning for simultaneous translation. A modest seed set of parallel sentences is available. A large pool of source sentences has its reference translations hidden, and the learner chooses which to reveal.

We train an MMA-IL model on the seed set and decode the whole pool greedily in simultaneous mode, recording each head's stop positions and each token's predictive entropy from the same pass. We then rank the pool, reveal a fixed batch of references, retrain from scratch on the enlarged set, and repeat for several rounds.

The selection rule is a quantile band on $D(x)$, for example the 50th to 90th percentile. The band is chosen on a held-out acquisition run on a separate split, not on test data. The top-$k$ rule is reported alongside it.

We also run an analysis track to explain why the signal works or fails. We correlate $D(x)$ with token entropy, with per-sentence quality gain after labeling, and with a reordering measure: the number of crossings in the argmax attention alignment of a separately trained offline soft-attention model. If the extreme tail of $D(x)$ is mostly high-crossing sentences, the band rule is justified mechanistically rather than by tuning.

A secondary track varies the head-divergence loss weight, including a setting where it is switched off only for the scoring model. This measures how much latency regularization erodes the uncertainty signal.

Baselines

The naive reference is random selection. At each round we reveal a uniformly random batch of the same size and retrain with identical architecture, hyperparameters, and step budget. Any method that does not beat random at equal labeled-sentence counts is not worth its complexity.

The strongest cheap competitor is single-model token entropy: the mean predictive entropy of the same MMA-IL model over its own greedy simultaneous decode. It is computed in exactly the same pass as $D(x)$, so its scoring cost is identical. This baseline isolates the key ingredient, because the only difference between it and our method is whether uncertainty is read from what the model writes or from when its heads stop. It gets the same band-versus-top selection options and the same held-out tuning of the band.

The strongest uncertainty estimator is true query-by-committee with $M=4$ independently trained MMA-IL models. Disagreement is computed both on stop positions and on output tokens, at four times the scoring cost. A Monte Carlo dropout committee of four stochastic decodes from one model is a middle point. Both are run on the same seed sets and selection batch sizes, and their scoring wall-clock is reported. Beating these is not required for success: matching the four-model committee's learning curve at one quarter of the scoring cost would itself be a positive result.

The full-data reference is the published MMA-IL result. Monotonic Multihead Attention reports about 28.5 BLEU at an average lag of about 3.5 source tokens on IWSLT 2015 English–Vietnamese, and about 30.5 BLEU on WMT15 German–English against an offline Transformer ceiling of about 32.3 BLEU. We found no reported result on Lacuna for active-learning baselines (random, entropy, or committee) for simultaneous translation on these benchmarks, so we produce those curves ourselves.

"Beating" a baseline means a higher area under the curve of BLEU versus number of labeled sentences, evaluated at a fixed latency operating point, across three seeds. The confidence interval of the difference must exclude zero. We also report the labeled fraction at which each method reaches 95% of the full-data BLEU.

Why not just…?

Why not just use token entropy, which is equally free? This is the real threat. Head disagreement and token entropy may be highly correlated, and if so our signal adds nothing. The proposal does not assume otherwise. The head-to-head comparison at identical scoring cost, plus a combined score, settles it. We commit in advance: if disagreement neither beats entropy nor improves on it when combined, we report alignment uncertainty as redundant for selection. The reason to expect a difference is that entropy is high both when the model lacks source context and when the target is genuinely ambiguous, while read-position disagreement responds specifically to the first case. That is the case most relevant to a model whose errors come from writing too early, as shown by the hallucination findings in Better Simultaneous Translation with Monotonic Knowledge Distillation.

Why not just train a real ensemble and do query-by-committee? It is the gold standard for disagreement, and we run it. Its cost scales with committee size at every scoring round over the whole pool, and Understanding the Training Speedup from Sampling with Approximate Losses shows that scoring overhead can erase the benefit of selection. Our claim is about the curve of quality against total compute, labeling plus scoring, and that comparison is reported directly.

Why not just pick the most uncertain sentences, the textbook rule? Robust Active Distillation and the federated trajectory study both find that extreme uncertainty selects noise. In our setting the analogous "noise" is reordering that a forward-only reader cannot resolve. We run the top-$k$ rule anyway and correlate the tail with alignment crossings, so the band rule has to earn its place.

Why not just distill a monotonic student on all the data, skipping active learning? When the full corpus is labeled, that is the right tool, and our method is not meant to compete with it there. The question here is the low-budget regime, where references are the scarce resource. The full-data MMA-IL numbers serve as the ceiling we measure distance to.

Proposed Experiments

The primary benchmark is IWSLT 2015 English–Vietnamese. Its size makes multi-round retraining affordable, and its mostly monotonic word order makes it the easy case. WMT15 German–English is the stress test, because clause-final verbs create exactly the reordering tail we predict to be harmful. We seed with a small fraction of each training set, acquire in equal batches over several rounds, and evaluate at several latency settings of the model's lag control. We report BLEU, average lag, the hallucination rate used in the monotonic distillation work, and scoring wall-clock.

The first decisive check is the matched-cost comparison with token entropy. If $D(x)$ loses or ties on both language pairs, the central claim fails.

The second check is the German–English tail analysis. Top-$k$ disagreement should underperform the band rule there by more than on English–Vietnamese, and the top tail should be enriched for high-crossing sentences. If neither happens, the reordering explanation is wrong, even if the signal still helps.

The third check is the divergence-loss sweep. If the ranking signal vanishes at the latency-optimal weight, practitioners face a real trade-off between low-latency training and useful uncertainty. The workaround of training the scoring model without the loss would then need to be shown to transfer.