Many decisions must be made before a sequence has finished arriving. A cardiologist's software reads an electrocardiogram (ECG) one heartbeat at a time. A trading system watches an option's price path day by day. In both cases the system must decide when it has seen enough to commit to a label. The best current recipe, Classifier-Induced Stopping from A Policy for Early Sequence Classification, first trains a classifier with dense temporal supervision: the sequence's final label is attached to every prefix, so the model learns to guess the answer at every point in time. It then trains a small stopping policy to imitate the best stopping time, which can be computed in hindsight on the training data. This works well. The weak point is that the policy trusts one network's view of its own certainty. That view is least reliable exactly where the decision matters most, in the earliest prefixes, where the authors themselves report degraded behavior in roughly the first 0.3 seconds of an ECG.
We propose to train two architectures with different built-in assumptions on the same dense per-prefix objective. One is a causal convolutional network that only sees local windows; the other is a causal transformer that attends over the whole prefix. We then give the stopping policy one extra input at each timestep: how much the two models disagree. The claim is that, on early and ambiguous prefixes, disagreement between structurally different models detects confident-but-wrong guesses that a single model's softmax confidence cannot see. A homogeneous ensemble of the same total size misses many of these as well. The expected payoff is fewer premature wrong stops at the same average delay.
Core Claim
Dense supervision on early prefixes forces a network to output a confident label even when the evidence is insufficient. A single network therefore produces early confidence that partly reflects which shortcut it latched onto, not how much evidence it has seen. Two networks with different inductive biases latch onto different shortcuts, so their disagreement rises on exactly those prefixes. Two copies of the same architecture, differing only in random seed, share shortcuts and stay in agreement.
The stopping policy therefore receives two signals at each timestep $t$. The first is the averaged within-model predictive entropy, a proxy for "the evidence is not in yet." The second is the cross-architecture disagreement $D_t = \mathrm{JS}(p^{\text{cnn}}_t, p^{\text{tf}}_t)$, the Jensen–Shannon divergence between the two models' class distributions on prefix $x_{1:t}$, a proxy for "the models are confidently guessing differently." This split mirrors the aleatoric-plus-epistemic decomposition that Complementing Self-Consistency with Cross-Model Disagreement for Uncertainty Quantification found effective for catching confident language-model hallucinations. We move it into the streaming, per-prefix setting, where it controls an action rather than a post-hoc score.
Relation to Prior Work
The proposal reuses its method from two sources. From A Policy for Early Sequence Classification we take the training pipeline unchanged: a classifier supervised at every prefix, oracle stopping times computed in hindsight, and a policy that imitates them. Only the policy's input features change, so any gain is attributable to the new signal. From Caser+ and CosRec+ we take the architecture trick that makes a convolutional network produce a prediction at every step: left-padding plus cumulative pooling. That paper showed densely supervised convolutional networks can match or beat a transformer recommender, with CosRec+ beating SASRec by 53% NDCG@10 on Foursquare Tokyo. It evaluated the two architectures only as competitors. We use the gap between them as a signal.
Three papers motivate the proposal by showing that heterogeneous disagreement carries information that homogeneous ensembles lack. Entropy-Based Uncertainty Modeling for Trajectory Prediction found that mixed ensembles of different trajectory forecasters track displacement error and out-of-distribution corruption better than homogeneous ones. It applied this only post hoc, to finished forecasts, not per timestep to drive decisions. Transformer-CNN Cohort found that a convolutional and a transformer student avoid reinforcing each other's mistaken pseudo-labels, unlike same-architecture pairs. That result is purely spatial. Asymmetric Duos showed that a cheap sidekick costing 10–20% of the base model's compute can match a 5-model deep ensemble on uncertainty. It used static images and fixed fusion weights. We borrow its asymmetric sizing so the duo costs little more than one model.
Two papers constrain the design. Ensembling Tabular Foundation Models shows that heterogeneous ensembles can hit a diversity ceiling and that naively averaging their probabilities can damage calibration. We respond in two ways. First, we use disagreement as an input feature to a learned policy rather than averaging probabilities into the final answer. Second, we measure error correlation between the two members directly; if they are nearly redundant, the claim fails. Language Models, Graph Searching, and Supervision Adulteration shows that dense step-wise targets can push a model toward local shortcuts. This is the failure we hope disagreement exposes. It also warns that both architectures might find the same shortcut, which is again testable through error correlation.
Research Program
We will build, for each benchmark, a duo trained jointly on the shared dense loss. The two members do not distill into each other, so their errors stay independent. The base member is a causal transformer. The sidekick is a causal one-dimensional convolutional network with cumulative pooling, sized at roughly 10–20% of the transformer's compute, following the Asymmetric Duos sizing. At every prefix we record each member's class distribution, the mean entropy, and $D_t$. The stopping policy is the same small network used by Classifier-Induced Stopping. It sees a short history of these features and is trained to imitate the hindsight-optimal stopping time under each accuracy/delay trade-off.
The analysis has three parts. The first is a per-prefix study of when $D_t$ is high. We test whether it concentrates on prefixes where at least one member is confident and wrong. The second is a measurement of error correlation between members as a function of prefix length, for the heterogeneous duo and for homogeneous duos, to check whether we are below the diversity ceiling. The third is a calibration audit of the averaged prediction per prefix length, to check whether the calibration trap appears in the streaming setting.
Baselines
The naive reference is a fixed confidence threshold. One densely supervised transformer stops as soon as its maximum class probability exceeds a threshold, and sweeping the threshold traces a Pareto curve. This shows what any learned policy must beat.
The strongest existing method is Classifier-Induced Stopping with a single densely supervised classifier, as in A Policy for Early Sequence Classification. That paper reports an average 11.8% improvement in Pareto-frontier AUC over standard methods and a 35.3% gain over PPO on PTB-XL ECG. The file reports no absolute Pareto-AUC value, so we will reproduce this baseline ourselves. It gets the same splits, the same total parameter and compute budget as the duo (we enlarge the single model to match), and the same hyperparameter search budget. Beating it means a higher Pareto AUC on all three tasks, with a lower premature-wrong-stop rate at matched average delay.
The control that most directly tests our claim is a homogeneous duo. It consists of two transformers, or a large plus a small transformer, matched to the heterogeneous duo's compute, trained from different seeds, and given the identical disagreement feature. A 5-seed deep ensemble serves as an upper reference at roughly 5× cost.
Two ablations isolate the key ingredient. The first gives the policy the heterogeneous duo's entropies but removes $D_t$. The second gives the policy the two models' concatenated hidden states instead of $D_t$, which tests whether an explicit disagreement scalar adds anything over raw access to both views. For calibration context, the PTB-XL benchmark paper Deep Learning for ECG Analysis reports a full-sequence CNN ensemble macro AUC of about 0.939 on diagnostic statements. This bounds the accuracy any early policy can approach as it observes the whole signal.
Why not just…?
Why not just calibrate the single model's confidence per prefix length? Temperature scaling fitted separately at each prefix length fixes average overconfidence but leaves the ranking of inputs unchanged. A confident wrong early guess stays confident relative to its peers. We will include per-prefix temperature scaling on the single-model baseline. If it closes most of the gap, then the problem was miscalibration rather than missing epistemic information, and the claim is weakened.
Why not just use disagreement within a homogeneous ensemble? This is the real threat, because seed diversity alone sometimes suffices. The evidence against it is indirect. Mixed forecaster ensembles beat homogeneous ones in Entropy-Based Uncertainty Modeling for Trajectory Prediction. Cross-family disagreement beat same-model sampling in Complementing Self-Consistency with Cross-Model Disagreement, with +0.15 AUROC on HotpotQA. Neither study tested streaming prefixes. The compute-matched homogeneous duo settles it. If it matches the heterogeneous duo, we will report the idea as reducing to ensemble disagreement and drop the heterogeneity claim.
Why not just learn the stopping rule end-to-end with reinforcement learning? The cited early-classification work shows that an agent which stops early never sees the rest of the sequence, which makes learning noisy. Imitating hindsight-optimal stopping beat PPO by 35.3% on PTB-XL. We therefore keep the imitation pipeline and change only the policy's inputs.
Proposed Experiments
We use the three tasks from the early-classification work. The first is PTB-XL (dataset paper) at 100 Hz on the five diagnostic superclasses; the benchmark paper reports no significant gain from 500 Hz for diagnosis. The second is word-by-word IMDB sentiment. The third is 30-day S&P 500 option profitability. The primary metric is Pareto-frontier AUC of accuracy against fraction of the sequence observed. Secondary metrics are the premature-wrong-stop rate at matched mean delay, the gap between policy and hindsight-optimal stopping times, and per-prefix expected calibration error of the averaged prediction.
Three outcomes are decisive. First, at matched compute, the heterogeneous duo must beat the homogeneous duo on Pareto AUC on at least two of three tasks, and its advantage must concentrate in the first 10–20% of each sequence. Second, member error correlation must be measurably lower for the heterogeneous pair at early prefixes. If it is not, any gain is unexplained, and the diversity ceiling has been reached. Third, removing $D_t$ while keeping everything else must erase most of the gain. If the concatenated-hidden-state ablation matches the full method, the contribution is "give the policy two architectures' views," a weaker but still useful finding that we will state as such.
References
- A Policy for Early Sequence Classification
- Complementing Self-Consistency with Cross-Model Disagreement for Uncertainty Quantification
- Caser+ and CosRec+: Closing the Gap Between CNNs and Attention Models in SRS
- Entropy-Based Uncertainty Modeling for Trajectory Prediction in Autonomous Driving
- Transformer-CNN Cohort: Semi-supervised Semantic Segmentation by the Best of Both Students
- Asymmetric Duos: Sidekicks Improve Uncertainty
- Ensembling Tabular Foundation Models: A Diversity Ceiling and a Calibration Trap
- Language Models, Graph Searching, and Supervision Adulteration: When More Supervision is Less and How to Make More More
- Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL
- PTB-XL, a large publicly available electrocardiography dataset