One strong recipe for reducing hallucination in long answers works like this. Have the model write documents. Break each document into small standalone facts. Keep only the facts the model itself "knows," as judged by a small classifier that reads the model's hidden activations. Then fine-tune on what is left. When this filtering goes through the model's own beliefs, it beats fine-tuning on human-written, fully correct text by 12–30 points of factual precision. The catch is that the model learns to write shorter answers and to abstain more often. Our hypothesis is that part of this cost, and some residual hallucination, comes from the filtering step itself. Deleting a fact from a document often leaves the text that depended on it in place: a later "he" with no antecedent, a "because of this" pointing at nothing, a character who was never introduced. A fine-tuning set full of such stitched-together documents teaches two bad habits. It teaches that answers are fragmentary lists. It also teaches the model to refer to things it never established, which is a form of fabrication.
We propose to treat the narrative coherence of the filtered training documents as a data-quality variable in its own right. When a fact is removed, everything in the document that presupposes it should be removed or repaired too, and no new facts may enter in the process. The claim to test is concrete. On matched factual content, training documents whose deletions have been repaired for coherence should yield models that are as precise as atomized epistemic training, write longer answers, and abstain less. Training documents with unrepaired deletions should measurably increase both hallucination and coherence errors downstream.
Core Claim
The work that motivates this proposal is The Curious Case of Factuality Finetuning. That paper fine-tunes on model-generated text filtered by internal-belief probes, which are logistic classifiers on hidden states that predict whether the model knows a claim. On Bios, the probe-filtered recipe moves Mistral-7B from 29.6% to 76.6% factuality, against 54.9% for gold-data fine-tuning. The paper also accepts a precision-versus-informativeness penalty and recombines atomized claims without regard to how they connect.
SNaC supplies the vocabulary for what goes wrong when claims are recombined. Its taxonomy of narrative errors includes unintroduced characters, missing referents, abrupt scene transitions and inconsistencies, and it shows these errors are invisible to overlap metrics. Our claim joins the two lines of work. Deletion-induced coherence defects are a specific kind of training-data noise, and they causally contribute both to the terseness and to some of the residual hallucination seen after factuality fine-tuning. Repairing them under a strict constraint that adds no new claims recovers informativeness without giving up precision.
The novelty is not a new factuality objective. It is the observation that fact filtering is a document-editing operation, and that how the edit is done is a data-quality choice nobody has measured.
Research Program
The core artifact is a claim-dependency graph built for each model-generated document. We decompose the document into atomic facts using the procedure of FActScore, which splits long text into short claims that can each be checked independently. Edges between claims come from two sources. Coreference supplies one: a later mention resolves to an entity introduced by an earlier claim. Discourse dependency supplies the other: a later claim is causally, temporally, or contrastively linked to an earlier one, in the sense of the discourse relations DiscoDVT uses as planning targets. A judge model labels presupposition edges ("would claim j be unintelligible or false-sounding if claim i were absent?"), and we audit a sample by hand.
The internal-belief probe from Newman et al. then marks unfamiliar claims. We remove those claims together with every claim that presupposes them and that the probe does not independently support. A repair model then rewrites only the sentence boundaries around each deletion: transitions, pronoun reintroductions, connectives.
The key constraint is a subset check. After the rewrite, we decompose the document again. We accept the rewrite only if every atomic fact it contains matches a retained fact. Otherwise we fall back to plain deletion. This keeps the repair step from slipping in new, possibly unfamiliar content, which would undo the point of filtering by the model's own beliefs.
The analysis side treats the deletion policy as the controlled variable in a data-quality study, in the spirit of The Dataset Multiplicity Problem. That paper shows small label errors flip many individual test predictions while aggregate metrics hide it. We hold the retained fact set fixed and vary only how coherently those facts are presented. We then measure what changes downstream, including the variance across resampled training sets rather than only the mean.
Baselines
The naive reference point is the base model with no fine-tuning, and next to it fine-tuning on gold human documents. Newman et al. report Mistral-7B on Bios at 29.6% (base) and 54.9% (gold), and we will reproduce both.
The strongest direct competitor is Newman et al.'s epistemic training: probe-filtered, atomized and recombined model generations. It reports 76.6% factuality for Mistral-7B on Bios, with 12–30 point gains over gold across Bios, Plots and Medical. It is the right competitor because it is the best training-data-only recipe in the file, and it shares our probe and our generations. We will run it using the same base generations, the same probe, the same probe threshold, the same number of fine-tuning tokens, and the same tuning sweep over learning rate and epochs. The fine-tuning-token match matters because deletion policies change document length. I found no reported number on Lacuna for the length or abstention rate of Newman et al.'s models, so we will measure those ourselves.
To beat this baseline, we must match its FActScore precision within noise across three seeds while producing more supported facts per answer, a lower abstention rate, and fewer SNaC-category errors.
A second baseline is FactAlign. It is a sentence-level preference method that rewards or penalizes individual sentence completions according to Wikipedia-verified atomic claims. It raised Gemma-2B's LongFact F1 from 61.74 to 86.51. We run it on the same base model and prompt pool to test whether a training-objective change already removes the informativeness penalty we attribute to data structure. A third, inference-time comparison is SHARS, which resamples sentences flagged by semantic-entropy uncertainty during decoding and reports roughly 20–26% precision improvement over greedy decoding on FActScore. It is orthogonal, so we also report it stacked on our model.
The ablations that isolate the key ingredient are:
- Plain deletion: remove unfamiliar claims with no dependency propagation and no repair.
- Propagation without repair: remove dependents but do not rewrite.
- Unconstrained repair: rewrite without the subset check.
Why not just…?
Why not just keep only documents in which every claim passes the probe? Whole-document filtering produces perfectly coherent training targets with no editing, and it is a genuine threat. Our worry is that it keeps mostly short, low-information documents and discards most of the long ones. That would itself push the model toward terseness and would make it difficult to supply a matched fine-tuning-token budget. We include it at a matched token budget, sampling more generations to fill the budget if needed. If it matches our method on precision, informativeness and coherence, the dependency graph is unnecessary and we will say so.
Why not just ask an LLM to rewrite the filtered text so it reads well? Generic rewriting is the obvious fix. HaluQuestQA shows generic feedback corrects long-form errors much worse than targeted span-level refinement (F1 0.55 versus 0.77). More importantly, a free rewriter will add connective claims of its own, which reintroduces exactly the unfamiliar content the probe removed. The unconstrained-repair ablation measures this directly as the fraction of rewritten facts not in the retained set, and tests whether those facts lower downstream precision.
Why not just filter training data with a coherence detector and a fact-checker jointly, or weight it by gradient agreement with a clean validation set? Two-score filtering, or a differentiable data selection scheme with a validation set judged on both factuality and coherence, discards defective documents rather than fixing them. It therefore faces the same data-loss problem as whole-document filtering. The validation-set approach additionally depends on a clean, representative validation set that does not exist for long-form narrative. We run joint two-score filtering at a matched budget as a further baseline.
Why not just use a bigger model? The file's evidence is that scale alone does not close either gap. ChatGPT scores 58.3% on FActScore biographies, and GPT-3-scale summarizers still make frequent character and scene errors in SNaC. We therefore treat scale as a reference point, not a solution.
Proposed Experiments
We train Mistral-7B and one 8B-class model on Newman et al.'s Bios, Plots and Medical setups. We evaluate on the FActScore biography benchmark, on LongFact using the SAFE-based F1 from Long-form factuality in large language models (which rewards both precision and the number of supported facts), and on held-out Plots prompts, where narrative coherence matters most.
Coherence is measured in two ways. The first is a detector trained on SNaC annotations; SNaC reports T5 character-error recall of 0.90 and scene-transition recall of 0.82. Because that detector was trained on book and movie summaries and may not transfer to biographies, the second is human annotation of 200 outputs per condition using SNaC's fine-grained schema. SNaC reports this schema reaches Krippendorff's α of 0.69 for character errors, compared with 0.19 for holistic judgment. We also report answer length, abstention rate and supported facts per answer.
The decisive check is a dose-response experiment. Starting from the repaired training set, we reintroduce unrepaired deletions at controlled rates (0%, 10%, 25%, 50% of documents) while holding the retained fact set fixed. If downstream FActScore precision and abstention do not move with the defect rate, and only coherence errors change, then the hallucination half of our claim is wrong. Coherence would then be a stylistic concern rather than a factuality one, and we would report it that way.
A second check counts how often the fine-tuned model's hallucinated facts attach to entities it never introduced. This tests whether dangling references in the training data teach fabrication specifically. Finally, we repeat training across five resampled generation pools to report variance, so that any gain is shown to be stable rather than a lucky draw.
References
- The Curious Case of Factuality Finetuning: Models' Internal Beliefs Can Improve Factuality
- SNaC: Coherence Error Detection for Narrative Summarization
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- DiscoDVT: Generating Long Text with Discourse-Aware Discrete Variational Transformer
- The Dataset Multiplicity Problem: How Unreliable Data Impacts Predictions
- FactAlign: Long-form Factuality Alignment of Large Language Models
- Building Reliable Long-Form Generation via Hallucination Rejection Sampling
- Fine-grained Hallucination Detection and Mitigation in Long-form Question Answering
- Optimizing Data Usage via Differentiable Rewards
- Long-form factuality in large language models