Language models that summarize or answer from a document often get the facts wrong in small ways: a slightly different number, a swapped name, a confident generic phrase where the source said something specific. Two separate fixes exist. The first asks the same question several ways and averages the answers, because any single phrasing gives a noisy reading of what the model knows. The second is a "pointer-generator," a decoder that at each step chooses between writing a word from its vocabulary and copying a token directly out of the source. This note proposes joining them in one specific way. Run a few paraphrased prompts side by side, and at each step ask two questions. Do the prompts agree on which source position the model is looking at? Do they agree on which vocabulary word it wants to write? When the prompts agree on where to look but disagree on what to say, copy from the source. When they agree on what to say, let the model write freely.
The point of this design is that disagreement between prompts is a cheap, training-free warning sign. Paraphrasing an instruction should not change a fact that is grounded in the document, but it does change a fact the model is inventing from memory. Existing methods either average vocabulary predictions across prompts without any copy route, or compute a copy decision from a single prompt whose attention may be pointing at the wrong entity. Neither uses the relation between the two kinds of agreement to decide when to copy.
Core Claim
A copy gate computed from cross-prompt agreement will reduce entity- and number-level hallucination in document-grounded generation, compared with multi-prompt averaging alone and with single-prompt pointer copying. It will do this at the compute cost of two or three batched prompts, and it will not collapse into the over-extractive "fragment stitching" that pointer models are known for. The claim is falsifiable. If naively averaging each prompt's own pointer-generator output matches the agreement gate, the idea reduces to a trivial combination and is not worth a paper.
Where This Sits
The multi-prompt side supplies the method. M-Ped packs $n$ prompt variants into one batch with left-padding and averages their next-token distributions. It reports that two to three prompts suffice and gives about +1.5 LENS on simplification and +1.5 d-BLEU on some translation directions. It has no copy route and was never evaluated for factuality. MPD shows that single-prompt accuracy swings by over 8% with phrasing and that aggregating hidden states across templates stabilizes classification. It only classifies and never generates. Together these motivate the proposal: prompt variance is real and averaging helps. The proposal differs by treating variance as a signal rather than only as noise to be averaged away.
The pointer side also supplies method, and it contains the closest prior work. LargePiG shows that a frozen decoder-only model already contains a pointer-generator. It takes the attention from the last token to the input document as a copy distribution. It sets the copy probability from the divergence between intermediate-layer and final-layer predictions, and it reports gains over DoLa on TruthfulDQG with about 6% latency on LLaMA2-7B. We reuse its attention-to-token construction but replace its gate. LargePiG decides whether to copy from layer disagreement inside one prompt, and it has no defense when that prompt's attention lands on the wrong span. FiD-PGN adds a supervised copy gate to Fusion-in-Decoder and reports 51.4 vs 48.5 exact match on NaturalQuestions, with +9.6 on questions whose answers never appeared in training. That is evidence that copying matters most for facts the model cannot have memorized, but it requires fine-tuning. See et al. introduced the mixture gate and also documented its failure mode: summaries that stitch source fragments together. Don't Copy That! shows copying hurts when the output must depart from the source wording. These last two papers constrain the design, and they are why the gate must stay off when the prompts agree on what to write.
The hallucination side supplies the problem, the metrics, and the strongest competitor. Context-aware decoding (CAD) contrasts predictions with and without the document. This is itself a two-prompt ensemble. It lifts LLaMA-7B's XSUM FactKB score from 38.9 to 53.8, but at about 83% extra latency, and it loses ROUGE on abstractive data when its weighting $\alpha$ is large. Agrawal et al. find that answers to varied indirect queries are inconsistent exactly when a reference is fabricated. This is our motivating evidence that cross-prompt inconsistency tracks invention, although they use it only after generation, as a detector. FaithBench and HaluEval show that after-the-fact LLM judges perform near chance on hard cases, with a best balanced accuracy of 62.31% and ChatGPT at 58.53% on summarization. That argues for intervening during decoding rather than checking afterwards.
Method
Take $n$ prompt templates (we use $n \in \{2,3\}$) that wrap the same source document and differ only in instruction wording. Batch them as M-Ped does. Because the document text is identical in each prompt, source positions can be aligned by character offset.
At step $t$, template $k$ gives two things: a vocabulary distribution $P^k_{vocab}$ and a pointer distribution $a^k$ over the aligned source positions $i$, built as in LargePiG. We form a consensus pointer from the geometric mean, so that a source position survives only if every prompt attends to it:
$$\tilde a_i = \frac{\prod_k (a^k_i)^{1/n}}{G}, \qquad G = \sum_i \prod_k (a^k_i)^{1/n} \in [0,1].$$
$G$ is pointer agreement. It equals 1 only when all prompts point identically. Generator disagreement $D \in [0,1]$ is the generalized Jensen–Shannon divergence among the $P^k_{vocab}$, divided by $\log n$ so that it lies between 0 and 1. The copy gate and the output distribution are
$$p_{copy} = G \cdot D, \qquad P(w) = (1-p_{copy})\,\bar P_{vocab}(w) + p_{copy}\sum_{i:\,w_i = w} \tilde a_i,$$
where $\bar P_{vocab}$ is the M-Ped average. The same generated token is appended to all $n$ streams.
The gate behaves as follows in each case:
- Agree on what to write. The gate is near zero, and the method reduces to M-Ped. This covers function words and paraphrase, which protects abstraction.
- Agree on where to look, disagree on what to write. The method copies. This is the case of a grounded entity that the vocabulary route renders inconsistently.
- Disagree on both. The gate stays low, and we record the step as high-risk. This gives a free token-level hallucination flag, which we evaluate separately.
We deliberately include no learned parameters. An optional sharpening exponent on $p_{copy}$ is tuned on the validation split, with the same budget given to the baselines' knobs.
Research Program
The work has four parts.
The decoder. We build it on LLaMA2-7B and Qwen1.5-7B, the models LargePiG used. Open weights are needed for attention access.
Where the gate fires. We measure the joint distribution of $G$ and $D$ on tokens labeled as grounded, hallucinated, or paraphrased. The hypothesis is testable before any decoding comparison: hallucinated entity tokens should show high $D$. If they do not, the project stops there.
Template choice. Following Voronov et al.'s warning that templates don't transfer (Mind Your Format), we fix a template pool per task on validation data. We then report variance over random draws of $n$ templates from the pool, not just the best draw.
The risk flag. We test the "disagree on both" flag as a detector, as a secondary contribution.
Baselines
The naive reference is single-prompt greedy decoding with the median-performing template from the pool. Reporting the median rather than the best template avoids flattering the baseline.
M-Ped with the same $n$ templates. This isolates whether a copy route adds anything beyond vocabulary averaging. It runs on identical batches, so compute is exactly matched. No factuality number for M-Ped exists; no reported result was found on Lacuna for M-Ped on XSUM, CNN/DM, or any hallucination benchmark.
LargePiG with a single prompt. This is the direct pointer competitor, run with each template separately. We report the mean and the best template. LargePiG reports gains over DoLa on TruthfulDQG MC1–MC3, but the file quotes no values; no reported MC number was found on Lacuna.
CAD. This is the strongest existing method. Its reported result is XSUM FactKB rising from 38.9 to 53.8 for LLaMA-7B (CAD). CAD costs two passes per token, so our $n=2$ setting is exactly compute-matched to it, and $n=3$ is reported separately. We tune its $\alpha$ over the same grid size as our sharpening exponent. The paper's reported optimum is 0.3. We also re-run CAD ourselves on LLaMA2-7B rather than relying on the reported LLaMA-7B figure. Beating CAD means higher FactKB than our CAD re-run at equal passes per token, with ROUGE and the novel-n-gram ratio no worse.
Naive combination. This ablation isolates the key ingredient. Run LargePiG inside each prompt stream, then average the $n$ final mixtures M-Ped style. It uses the same attention, the same prompts, and the same compute. Only the gate differs.
Same-prompt ensemble. A second ablation uses $n$ copies of one template with shuffled demonstration order instead of different wordings. It tests whether instruction diversity is what produces the useful disagreement.
Why not just…?
Why not just find the single best prompt? MPD reports that single-prompt accuracy moves by over 8% with wording, and Mind Your Format reports that winning templates don't transfer across models. The median-template baseline, the best-template LargePiG baseline, and a cross-model check will show directly whether a tuned single prompt closes the gap. If the best single template on LLaMA2 also wins on Qwen and matches our method, the ensemble is unnecessary, and we will report that.
Why not just run CAD? CAD is a serious threat. Its XSUM gain is large, and it is already a two-prompt contrast. Its weaknesses are cost, its lack of any route for exact rare tokens, and ROUGE loss at high $\alpha$ on Debatepedia. The compute-matched $n=2$ comparison settles this question. We also stratify by entity rarity, following FiD-PGN's no-overlap analysis, because a copy route should win specifically on rare names and numbers. If CAD wins overall and on rare entities, our contribution shrinks to latency and the risk flag, and we would say so.
Why not just average LargePiG across prompts? This is the obvious combination and the real test of novelty. An arithmetic average of attention spreads mass over every prompt's distortions, whereas the geometric consensus keeps only shared positions. Averaging gates also discards the what-versus-where contrast. The naive-combination ablation decides this. If it matches the agreement gate within seed variance, we abandon the gate claim.
Why not just use a post-hoc judge? FaithBench reports that the best detector reaches 62.31% balanced accuracy on hard cases and that chain-of-thought judging does worse than zero-shot. HaluEval reports 58.53% for ChatGPT on summarization. Judging after generation also cannot repair an entity during decoding. We compare our risk flag against a zero-shot judge on the same outputs.
Proposed Experiments and Decisive Checks
Benchmarks. We use XSUM and CNN/Daily Mail, evaluated with FactKB as in CAD. Alongside FactKB we report ROUGE and the novel-n-gram ratio, so that over-copying is visible. We use Debatepedia as the abstractive stress case where CAD lost ROUGE, and TruthfulDQG with MC1–MC3 for comparison with LargePiG. For entity fidelity, we measure precision of named entities and numbers against the source, stratified by corpus frequency.
Human evaluation. On 300 XSUM outputs per system, annotators apply FaithBench's four-way taxonomy: consistent, benign, questionable, unwanted. For the risk flag, token-level human labels on a subset give an AUC to compare against the zero-shot judge.
Latency. Measured in seconds per token alongside CAD's reported overhead.
Decisive checks.
- On validation data, hallucinated entity tokens show markedly higher $D$ than grounded ones. If not, stop.
- The agreement gate beats the naive-combination ablation on FactKB and entity precision at equal compute.
- At $n=2$ it matches or beats our CAD re-run on FactKB while keeping ROUGE and the novel-n-gram ratio within one point of M-Ped.
- On Debatepedia it does not show CAD's ROUGE drop.
Passing the first two checks justifies the idea. Passing all four makes it the preferred training-free decoder for grounded generation.
References
- M-Ped: Multi-Prompt Ensemble Decoding for Large Language Models
- Multi-Prompting Decoder Helps Better Language Understanding
- LargePiG: Your Large Language Model is Secretly a Pointer Generator
- A Copy-Augmented Generative Model for Open-Domain Question Answering
- Get To The Point: Summarization with Pointer-Generator Networks
- Pointer-Generator Networks for Low-Resource Machine Translation: Don’t Copy That!
- Context-aware Decoding Reduces Hallucination in Query-focused Summarization
- Do Language Models Know When They're Hallucinating References?
- FaithBench: A Diverse Hallucination Benchmark for Summarization by Modern LLMs
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- Mind Your Format: Towards Consistent Evaluation of In-Context Learning Improvements