Prompt-Stable Claims as Clean Pseudo-Summaries

When a large language model writes summaries that are then used to train a smaller summarizer, every hallucinated fact in those machine-written summaries (pseudo-labels: model outputs treated as training targets) is learned by the student. The usual defense is to filter pseudo-labels with a factuality classifier. These classifiers mostly reward copying, so the student comes out faithful but dull and extractive. This proposal tests a different filter. We ask the teacher for the same summary under several instructions that mean the same thing ("Summarize this article in one sentence," "Write a one-line brief of this story," and so on). We then keep the claims that survive across those rewordings and down-weight the claims that appear only under some of them. The bet is that a fact grounded in the source keeps reappearing however the request is phrased, while a hallucination depends on the exact prompt. If so, a model's sensitivity to its instructions, usually treated as a defect, becomes a way to clean its own pseudo-labels.

What makes this more than a reuse of known parts is that the stability is measured over content, not wording. A teacher that paraphrases freely produces different sentences under every prompt, so surface disagreement would punish exactly the abstraction we want to keep. We therefore break each summary into atomic claims and match claims across prompt variants by meaning. A summary can then be lexically far from the source and still be fully stable. The central empirical question is whether this signal finds hallucinations at a fixed level of copying, or whether it is just one more proxy for extractiveness. The whole evaluation is built to answer that question.

Core Claim

Our claim is that claim-level agreement across semantically equivalent instructions predicts whether a teacher-written claim is unfaithful, even among claims with the same lexical overlap with the source. Training a student summarizer with each target token weighted by the stability of the claim it belongs to should then produce a student that is more faithful than one trained on unfiltered or classifier-filtered pseudo-labels, without moving it along the known trade-off curve toward copying. In the terms of Faithful or Extractive?, the student should show positive effective faithfulness: its faithfulness should exceed what a control model at the same extractiveness achieves.

Why This Gap Is Real

The trade-off literature motivates the problem and supplies the yardstick. Truth or Error? finds that more abstractive models make more factual errors. Training Dynamics for Text Summarization Models shows that models learn to copy early and pick up unsupported reference tokens late. Its fix, Loss Truncation, drops high-loss tokens and lowers sentence error rate on XSum from 0.67 to 0.59. FactPEGASUS cleans references by fixing or removing unsupported entities and cuts errors sharply, but it reports that its summaries become more extractive and choppy. The analysis that constrains our design most is Ladhak et al.'s. They show that factuality metrics correlate with copying and that Loss Truncation and similar methods mostly slide the model to a more extractive point on the same curve rather than improving on it. FRANK adds two further constraints: FactCC collapses on abstractive XSum, and the best metric correlates with human judgment at only about ρ ≈ 0.20. Any filter built on those classifiers inherits their bias toward copying. All of these methods act on human references or on a single model output; none uses agreement across instructions.

The prompt-sensitivity literature supplies the signal. On the Relation between Sensitivity and Accuracy in In-Context Learning shows that instability across prompt paraphrases and demonstration orders anticorrelates with accuracy, with Pearson r reaching -0.40. Its method, SENSEL, abstains on unstable predictions and beats probability-based confidence, but only for classification and only at test time. ProSA explains sensitivity as flat next-token distributions on instances the model does not grasp. Rethinking Hallucinations is a cautionary result: on QA benchmarks, common hallucination detectors capture consistency rather than correctness, and some errors are consistent across prompts. We take this seriously. Prompt stability cannot catch errors the teacher makes regardless of the prompt, and our evaluation separates those "prompt-agnostic" errors from prompt-dependent ones.

The pseudo-labeling literature supplies the training mechanism. Uncertainty-weighted losses come from Combating Noise. Its observation that noise concentrates near the decision threshold suggests that intermediate stability values deserve soft rather than hard treatment. Auto-Annotation Quality Prediction filters pseudo-labels by ensemble consensus at the level of local regions, and our claim-level masking is the text analogue of its pixel-level masking. Cross-Lingual Summarization with Pseudo-Label Regularization trains summarizers on soft teacher-ensemble targets and gains 2.2 ROUGE-L on WIKILINGUA, but never measures hallucination or extractiveness. That omission is precisely what we measure.

Research Program

The pipeline has four stages. In the first, an open instruction-tuned teacher (Llama 3-70B-Instruct, which ProSA found comparatively stable) summarizes unlabeled source articles from the XSum, Gigaword and WikiHow training sets, ignoring their references. Each article is summarized under $K=8$ instruction variants. We draw the variants from ProSA's four template families (plain input, role play, emotional framing, output requirements) and hold the length constraint fixed, so that variants differ only in phrasing.

In the second stage, each summary is split into atomic claims by an LLM-based decomposer, and claims are matched across variants by bidirectional entailment. For a claim $c$ in variant $k$, we define its stability as

$$s(c) = \frac{1}{K-1}\sum_{j \neq k} \mathbb{1}[\exists\, c' \in \text{variant } j \text{ that entails and is entailed by } c] - \lambda \cdot \mathbb{1}[\exists\, c' \text{ in some variant that contradicts } c],$$

where $\lambda$ adds an extra penalty when another variant actively contradicts the claim. A contradiction is a stronger warning than a mere absence, which may only reflect a difference in content selection.

In the third stage, we choose as the pseudo-label for each article the variant with the highest mean claim stability. During student training, each target token $t$ receives weight $w(t) = \sigma(\alpha(s(c_t)-\tau))$, where $c_t$ is the claim containing $t$, $\sigma$ is the logistic function, and $\alpha,\tau$ are tuned on a small development set. Connective tokens outside any claim get weight 1. The student loss is the weighted cross-entropy $\sum_t w(t)\,\mathcal{H}(p_\theta(\cdot\mid x, y_{<t}), y_t)$, which follows the form used in Combating Noise. Tokens are down-weighted rather than deleted, which avoids the fluency damage that FactPEGASUS's "remove" corrector causes.

In the fourth stage, the student (BART-base, matching the FactPEGASUS comparison point) is trained on these weighted pseudo-labels alone, and we compare it with students trained on the same documents under the alternative filters below.

The analysis component runs alongside the pipeline. We hire annotators to label about 600 teacher outputs (200 per dataset) at the claim level using the FRANK error typology. Each claim is also tagged with its lexical coverage of the source and with whether its error recurs across prompts or depends on the prompt. This yields the diagnostic on which the proposal turns: does stability separate faithful from unfaithful claims within coverage bins? We also fit a variant whose stability score is residualized on coverage. If the raw and residualized versions train equally good students, stability is carrying information beyond copying. If only the raw version works, the method is a disguised extractiveness filter, and we would report it as such.

Baselines

The naive reference point is a student trained on the teacher's output under one fixed instruction, with no filtering. It shows how much hallucination a modern teacher passes on and how abstractive the unfiltered student is. A second reference is the student trained on the original human references. On XSum, those references are themselves noisy, which makes the comparison informative.

The strongest existing competitor on the training side is Loss Truncation +Factuality from Training Dynamics, applied to the same single-prompt pseudo-labels. It is the closest prior idea: it down-weights suspect target tokens using the student's own loss, where we use teacher disagreement. Its reported result is SER 0.67 → 0.59 on XSum with human references. We run it on our pseudo-labels with the same burn-in and drop-rate grid as our $\alpha,\tau$ sweep. On the selection side, the strongest competitor is the Ladhak et al. selector. It reports faithfulness 86.00 at 68.38% coverage on Gigaword against a baseline of 83.33 at 76.12%, which is better faithfulness bought partly by more copying. We reproduce it on Gigaword with their quartile-trained candidates. We also build a variant that uses FactCC to choose among our eight prompt-variant teacher summaries, which isolates the choice of selector from the source of candidates. A third competitor filters pseudo-label tokens using teacher token entropy, in the style of BART-GBP, which reports a token-detection F1 of 0.2778 on FRANK.

The key ablation replaces the eight prompt variants with eight temperature samples under a single prompt and computes the same claim-level stability. This holds the number of teacher calls and the matching procedure constant and changes only where the variation comes from. All students share data, architecture, training steps and a 20-configuration tuning budget on the same development split. Beating the baselines means positive effective faithfulness relative to the Ladhak-style control curve on Gigaword and WikiHow, with human-judged faithfulness above Loss Truncation at equal or lower coverage. Matching a baseline's faithfulness only by raising coverage counts as a failure.

Why not just…?

Why not just filter pseudo-labels with FactCC or an entailment model? This is the obvious move, and the evidence in the file predicts it will make the student more extractive. Ladhak et al. show that these metrics reward coverage. FRANK shows that FactCC breaks down on XSum-style abstraction, with metric-human correlation capped near 0.20. We include the filter as a baseline. Our prediction is that it matches our faithfulness only at visibly higher coverage, which the effective-faithfulness analysis will show directly.

Why not just sample several outputs at temperature and use self-consistency? This is the real threat. It is cheaper to set up, and Rethinking Hallucinations warns that consistency-based detectors capture randomness rather than truth. If temperature samples give the same stability-error separation and the same student quality, then prompt variation contributes nothing beyond sampling noise, and we would report the simpler method. Our hypothesis is that rewording the instruction disturbs the teacher's reading of the task more than temperature does. ProSA's link between flat distributions and sensitivity points the same way, so prompt variation should expose more prompt-dependent hallucinations. The matched-budget ablation decides this.

Why not just use token entropy or the student's own loss? Single-pass confidence costs no extra teacher calls, and Loss Truncation already works on XSum. However, SENSEL's +4.1 AUC over MaxProb and entropy on GPT-J-6B suggests that agreement across perturbations carries information that a single output distribution lacks. Goyal et al. also note that a static loss threshold cannot tell a hard paraphrase from a hallucination. Our entropy and Loss Truncation baselines test whether that advantage carries over from classification to generated claims.

Why not just tell the teacher to be faithful, or use a more capable teacher? Instructions to stay faithful mostly push models toward copying, and ProSA's "unstable genius" finding means a stronger teacher may also vary more across prompts. We add a "be strictly faithful" single-prompt condition to the naive baseline. We also repeat the full pipeline with a second, more capable but more sensitive teacher, to see whether stability filtering helps more when the teacher is less stable.

Decisive Checks

Three results would settle the proposal. First, on the annotated claims, stability should reach higher AUROC for detecting unfaithful claims than entropy and FactCC within each coverage quartile, not just pooled across quartiles. Second, the stability-weighted student should lie above the Ladhak control curve on Gigaword and WikiHow, beating Loss Truncation in human-judged faithfulness at no higher coverage, and the coverage-residualized variant should retain most of that gain. Third, the prompt-variant signal should beat the matched temperature-sampling ablation, with the gain concentrated in errors the annotators labeled prompt-dependent. We expect none of the filters to help on prompt-agnostic errors, and we will report that limitation rather than hide it. If the first check holds and the third fails, the contribution narrows to claim-level self-consistency filtering for summarization. That result would still be useful, but it is a weaker claim, and we would state it that way.