When a zero-shot text classifier advises a person, it usually hands over one thing: a label. That label is fragile. The classifier chose it by testing how strongly the input text "entails" sentences built from class names, and those scores shift when the class names are merely reworded. People under time pressure tend to accept whatever label they are shown, including wrong ones. The main fix in current practice is to hide the label when the model seems unsure. Hiding it also throws away everything the model learned about the input, and it hurts in exactly the cases where the person could have done better with partial help.
Communication theory has a clean answer to this all-or-nothing problem. Hybrid digital-analog coding sends a coarse discrete message, which is either received exactly or lost entirely, together with a continuous "analog" residual that gets noisier as conditions worsen but never vanishes. We propose treating classifier advice the same way. The digital layer is the label, and it is shown only when the prediction holds up when the class names are paraphrased. The analog layer is always shown: graded entailment scores over short, checkable statements that define each candidate class. The claim is that this display gives better human-AI team accuracy than showing the label always, hiding it when uncertain, or showing a confidence number. It should also degrade gracefully when the model is wrong or the person is rushed, where current designs fail sharply.
Core Claim
Entailment-based classification reframes classification as natural language inference: the input is the premise, "This text is about {label}" is the hypothesis, and labels are ranked by entailment probability (Indirectly Supervised Natural Language Processing; BTZSC). Because the hypothesis is built from a class name, the output inherits semantic label dependency, meaning that accuracy is tied to how descriptive and how exactly worded the class names are. In vision-language few-shot learning, the gains from initializing with class names disappear completely when names are replaced by identifiers like "Class 001" (Exploiting Category Names for Few-Shot Classification with Vision-Language Models). Entailment classifiers are also known to swing with template wording and to exploit lexical artifacts in hypotheses (Issues with Entailment-based Zero-shot Text Classification).
We turn this weakness into two signals. The first is whether the top label survives paraphrases of the class names and templates. This acts as a measure of "channel quality" that decides whether the digital layer is sent. The second is a set of decomposed class definitions, such as "the writer reports a card payment that did not go through," which give the classifier a second, finer-grained vocabulary. Scoring the input against these statements yields a continuous profile that a human can check against the text in front of them.
The specific, testable claim has three parts. Compared with uncertainty-gated suppression, the hybrid display should reduce "negative consultations," where a person reverses a correct judgment to match wrong advice. It should also increase team accuracy on the items where the human alone beats the model. And it should do both without lowering accuracy when the model is right.
Why the coding analogy is more than a metaphor
In hybrid coding, a quantized digital layer is superimposed on a scaled analog error signal. A receiver in good conditions decodes both, while one in poor conditions still recovers the analog baseline (Joint Source-Channel Coding With Time-Varying Channel and Side-Information). Two results from this literature motivate the design, and one constrains it.
The motivating results concern side information. When the receiver holds correlated side information, purely digital separation is suboptimal, and hybrid schemes reach the optimal distortion region (The Achievable Distortion Region of Bivariate Gaussian Source on Gaussian Broadcast Channel). The human reader is exactly such a receiver: they see the raw text and bring their own knowledge, which the auditor study found to matter most when humans outperform the model.
The constraining result is that uncoded analog transmission can beat layered digital schemes when receivers have little side information (Wyner-Ziv Coding over Broadcast Channels: Digital Schemes). This warns that the digital layer may add nothing, so an analog-only display must be a first-class condition rather than a strawman.
Deep hybrid systems (Hybrid Digital-Analog Semantic Communications) split neural image features into an encrypted digital core and an analog auxiliary layer, and they evaluate image reconstruction on Kodak. We borrow the split-and-superimpose structure, not the physical-layer machinery. Our "channel" is a noisy classifier together with a human under load, and our distortion measure is decision error.
Research Program
Building the display. We first build the two layers from off-the-shelf zero-shot scorers evaluated in BTZSC: a reranker (Qwen3-Reranker-8B), an embedding model (GTE-large-en-v1.5), and a classic NLI cross-encoder.
For the digital layer, each class name is paraphrased into several variants, crossed with several templates. The top label is transmitted only if it is stable across a fixed fraction of these variants. Otherwise the label is withheld.
For the analog layer, a language model writes three to five short defining statements per class from the class name alone. This keeps the setting zero-shot. For the top few candidate classes, the scorer then computes graded entailment and contradiction for each statement against the input. These are shown as bars beside the text. This layer is always sent, including when the label is withheld.
Offline analysis. Before running any people, we ask whether paraphrase stability predicts errors better than the scorer's own maximum probability. We also ask whether the statement profile carries information about the true class beyond the top label. We measure the second question as how much a small probe trained on the profile improves over the label alone.
Human study. We then test the display under time pressure, using the manipulation shown to induce automation bias in computational pathology (Automation Bias in AI-Assisted Medical Decision-Making under Time Pressure in Computational Pathology). Every participant first gives an unaided answer, then sees the advice and gives a final answer. This lets us count negative and positive consultations directly.
Baselines
The naive reference is the human alone, with no advice. Every assisted condition must be compared against it, because the auditor study showed that assistance can end up worse than no help in human-superior regimes.
The second reference is "always show the label," which is the default deployment of an entailment classifier. The label comes from the strongest reported zero-shot scorer, Qwen3-Reranker-8B. BTZSC reports a benchmark-wide macro F1 of 0.72 for it, against up to 0.67 for instruction-tuned LLMs such as Mistral-Nemo. The file reports no per-dataset scores for our chosen tasks: there is no reported result found on Lacuna for Banking77 or EmpatheticDialogues individually, so we will measure these ourselves.
The strongest existing mitigation is auditor-style suppression (Auditor models to suppress poor artificial intelligence predictions can improve human-artificial intelligence collaborative performance). In that work, a second model hides uncertain predictions, and the result is significantly higher team AUROC than humans alone (P < 8.2 × 10⁻⁶) when the model beats the human. The same paper reports no fairness benefit, and sometimes harm, when humans beat the model. We will run it with the same base scorer. The auditor is trained on held-out model errors, with the same labeled calibration budget our paraphrase gate uses to set its threshold. Its suppression rate is matched to ours, so that differences reflect what is shown rather than how often something is hidden.
We also include a confidence display: the top-three labels with calibrated probabilities, using temperature scaling on the same calibration set. This is the simplest continuous signal.
Three ablations isolate the proposal's ingredients:
- Analog-only: statement profile, never a label. This tests the Nayak-style prediction that the digital layer adds nothing.
- Digital-only with paraphrase gating: label shown when stable, nothing otherwise. This tests whether the gain comes from the gate or from the residual.
- Shuffled-profile control: bars permuted across statements. This mirrors the shuffled-face control in Learning Representations by Humans, for Humans and tests whether people read the content or are merely reassured by richer visuals.
The hybrid "beats" suppression if it has higher final team accuracy overall and a lower negative-consultation rate. It must also show a significant team-accuracy gain on the stratum of items where humans alone outperform the model, pre-registered with a mixed-effects model over participants and items.
Why not just…?
Why not just show calibrated confidence or the top three labels? This is the most serious threat, because it is cheap and continuous. Our argument is that a single probability tells the person how much to trust the model but not what to check. The statement profile, by contrast, points to verifiable claims about the text, which lets the reader combine it with what they see. This is the role side information plays in the coding results. The claim could be false, and the confidence-display condition settles it directly. If confidence display matches the hybrid on negative consultations and on the human-superior stratum, we will report that the analog layer adds nothing beyond calibration.
Why not just suppress uncertain predictions with an auditor? The auditor paper itself reports the failure mode we target: suppression does not help, and can hurt, when humans outperform the model. Suppression leaves the person with nothing on exactly the hard items. Our matched-suppression-rate comparison tests whether keeping the residual on withheld items recovers that loss.
Why not just use a better classifier? If the reranker were accurate enough, bias would matter less. But BTZSC's best benchmark-wide macro F1 is 0.72, and it reports that LLMs lag on high-cardinality intent and nuanced emotion. That leaves many errors in exactly the tasks we study. We run all three scorers. If the hybrid's advantage shrinks toward zero as scorer accuracy rises, that is a legitimate, reportable limit on the idea.
Why not just add explanations like highlights? Eye-tracking studies in colonoscopy show that highlighting markers change how people search and cause healthy tissue to be flagged (The influence of computer-aided polyp detection systems on reaction time for polyp detection and eye gaze). The statement profile could anchor people in the same way. The shuffled-profile control and the negative-consultation counts are designed to detect this.
Proposed Experiments
Tasks. We use three BTZSC datasets chosen to vary difficulty and the human-versus-model balance:
- Banking77: 77 intents, where cardinality stresses the scorer.
- EmpatheticDialogues emotion: a nuanced task where humans may beat the model.
- AG News: broad topics, where the model is likely strong.
Offline checks. First, we compare error-detection AUROC for paraphrase stability versus maximum probability, per dataset and per scorer. Second, we measure the information gain of the statement profile over the label.
Human study. Crowdworkers are assigned to the conditions between subjects, with a within-subject time-pressure factor.
Metrics. We measure final team accuracy, negative and positive consultation rates, decision time, and accuracy on the human-superior stratum. We also measure unaided accuracy on a final block of items with no advice, as a short-term check for deskilling.
Degradation test. The decisive "cliff" test plots team accuracy against two factors: model correctness (right versus wrong advice) and time pressure. The label-always condition should drop sharply on wrong-advice items, and the more so under pressure. The hybrid should fall less, without losing ground on right-advice items.
Falsification. The idea fails if any of the following holds:
- The analog-only condition matches the hybrid, meaning the gate is useless.
- The confidence display matches it, meaning the residual is unnecessary.
- The shuffled profile helps as much as the real one, meaning people are responding to reassurance rather than content.
Each outcome is informative, and we will report whichever we find.
References
- Indirectly Supervised Natural Language Processing
- BTZSC: A Benchmark for Zero-Shot Text Classification Across Cross-Encoders, Embedding Models, and Rerankers
- Exploiting Category Names for Few-Shot Classification with Vision-Language Models
- Issues with Entailment-based Zero-shot Text Classification
- Joint Source-Channel Coding with Time-Varying Channel and Side-Information
- The achievable distortion region of bivariate Gaussian source on Gaussian broadcast channel
- Wyner-Ziv coding over broadcast channels: digital schemes
- Hybrid Digital-Analog Semantic Communications
- Automation Bias in AI-assisted Medical Decision-making under Time Pressure in Computational Pathology
- Auditor models to suppress poor artificial intelligence predictions can improve human-artificial intelligence collaborative performance
- Learning Representations by Humans, for Humans
- The influence of computer-aided polyp detection systems on reaction time for polyp detection and eye gaze