Real fake news is rarely false from start to finish. A typical deceptive article repeats verified facts and slips one distorted number, name, or claim into a sentence that is otherwise true. A detector that labels whole articles "fake" or "real" therefore cannot tell a reader or a fact-checker which words are wrong. Recent work shows that large language models (LLMs) handle this poorly. On MisSpans, a benchmark that pairs real and fake news and asks models to mark the exact false spans, fifteen LLMs mostly score below 0.5 F1. The same models still write plausible explanations of what is wrong (MisSpans). This gap suggests the models can judge a claim but cannot locate it. We propose to split the two. A team of narrowly prompted LLM calls first cuts an article into small candidate assertions, and then checks each assertion in isolation against evidence. The resulting per-span probabilities are distilled into a small extractive tagger. Because the tagger only marks tokens that already exist in the article, it cannot invent a false span that is not there.
In two sentences: LLMs fail at false-span detection mostly because finding the span is hard, not because judging it is hard, so recasting the task as "list every assertion, then verify each one separately" should give much better teacher labels. Distilling those labels as soft per-token probabilities into a compact encoder should give a cheap detector that works with very few human-labeled examples and transfers across news domains better than training on gold labels alone.
Core Claim
Our central hypothesis is that the MisSpans failure is mainly a localization failure. If it is, then a decomposed teacher should give much higher span-level precision and recall than the same LLM asked to find false spans in one pass. The pipeline has three roles. An enumerator segments the article into atomic assertions. A verifier checks one assertion at a time against the paired reference article or retrieved evidence, without seeing the rest of the article. A typer labels each flagged span as "Inconsistency" (it contradicts the evidence) or "No mention" (the evidence does not support it).
The second claim concerns distillation when labeled data is scarce. The verifier's confidence on each candidate span carries graded information that hard labels throw away. Distilling it should help most when human span labels are scarcest. This is the pattern Revisiting Knowledge Distillation: The Hidden Role of Dataset Size reports for image classifiers: gains of up to 25% over label training at small data fractions, which label smoothing does not reproduce. That result motivates the soft-label design, but it was shown on CIFAR and not on text spans, so whether it transfers here is part of what we test.
The "No mention" category is the news-domain form of an extrinsic hallucination, meaning content that has no support in the source. This makes the same student a natural candidate for localizing unsupported content in LLM outputs. We treat that as a secondary transfer test, not the main goal.
Research Program
Teacher. The decomposed teacher is built from separate LLM calls, each with its own instructions and its own limited context. This is the "context firewall" idea that the LLM-driven chemical synthesis platform uses to keep tool calls from confusing each other. We reuse two ideas from Falconer, which has an LLM decompose information extraction into atomic get_span and get_label operations and then trains a RoBERTa-large proxy on about 5% of a corpus. Falconer's setting is named-entity extraction. We apply the same move to adversarial misinformation, where the teacher's own hallucinations are the main danger. We add the isolated verification step because, unlike entity spans, false-claim spans can only be identified by comparison with evidence.
We deliberately do not use open-ended debate. MADR uses two debaters, a judge, and a refiner to raise fully faithful fact-checking explanations from 20% to 30% on PolitiHop. The low ceiling and the documented risk of debaters converging on a wrong premise lead us to prefer independent per-assertion checks, which cannot copy one another.
Student. The student is an encoder tagger, RoBERTa-large to match Falconer. For each token $t$ it predicts a distribution $p_\theta(\cdot \mid t)$ over three tags: true, inconsistent, and unsupported. We train it on the loss
$$\mathcal{L} = \sum_{t \in G} \mathrm{CE}\big(y_t,\, p_\theta(\cdot\mid t)\big) + \lambda \sum_{t \in U} \mathrm{KL}\big(q_t \,\|\, p_\theta(\cdot\mid t)\big).$$
Here $G$ is the set of tokens with human gold tags $y_t$, and $U$ is the set of tokens in teacher-annotated articles. The target $q_t$ is the verifier's tag distribution for the candidate assertion that contains $t$. It is estimated from the verifier's stated probability, or from agreement across a few independent verifier calls. The weight $\lambda$ is tuned on a development split. Tokens that the enumerator never places inside a candidate assertion receive a soft "true" target, not a hard one.
Relation to prior distillation work. This differs from attribution-based methods. AD-KD matches a teacher's Integrated Gradients token importances, gaining +3.2 on CoLA and +3.6 on RTE. FARD regularizes student attention toward a teacher's reasoning graph on math tasks. Both need white-box teachers and transfer why a teacher decided. We transfer where the teacher, after decomposition, believes the falsehood lies. That signal can be collected from black-box APIs, and it is directly the output we want.
Relation to prior distillation of multi-agent systems. Two papers already distill multi-agent reasoning into small models. Debate, Reflect, and Distill turns teacher–student debate trees into preference data and improves Mistral-7B from 23.98 to 38.16 on MMLU Pro and MATH. Thinking as Society simulates persona agents that react to multimodal news and distills them into a 7B model that reaches 67.15% on MFC-Bench. Both produce whole-answer or whole-document verdicts, and both students are generative models that can themselves hallucinate. Our student makes token-level decisions and is extractive by construction.
RAAR decomposes cross-domain misinformation detection into sentiment, semantic, and style sub-agents and trains 8B and 14B verifiers. It also stays at the document level. It motivates our cross-domain evaluation but is not a span-level competitor.
Baselines
Naive reference. The naive reference is the same RoBERTa-large tagger fine-tuned with cross-entropy on MisSpans gold spans alone, at each gold-data fraction. It uses identical architecture, epochs, and a hyperparameter search of the same size as every other variant. Any teacher-based method must beat it. Otherwise the whole teacher pipeline is wasted spend. We expect the gap to be largest at 5% gold and to shrink toward full data. The absence of that trend would itself be evidence against the scarcity argument.
Direct prompting. The direct-prompting baseline is the strongest available frontier LLM, zero-shot and one-shot, asked to output false spans and types in one pass. The MisSpans paper reports F1 "generally falling below 0.5" on MisSpansIdentity and MisSpansType across 15 LLMs. It gives no single best-model number that we can quote, so we will re-run the best model it names on our exact test split. Beating it means higher span F1 at a small fraction of its inference cost.
Monolithic teacher distilled (strongest competitor). This is a Falconer-style baseline: the same frontier LLM annotates the same unlabeled pool in one pass, and the same tagger is trained on its hard labels. It is the strongest competitor because it isolates whether decomposition adds anything beyond "an LLM labels data and a small model learns it." Falconer reports that this recipe gives 0.68 F1 against GPT-4o's zero-shot 0.53 on WikiNER, though on a different task. To make the comparison fair, the monolithic teacher gets the same total teacher-token budget as the decomposed pipeline. It spends the extra budget on repeated sampling with majority voting, so the comparison is not decomposition versus fewer calls. The decomposed student beats it if span F1 improves by a margin larger than the seed-to-seed spread at 5% and 25% gold, and if the improvement also appears on the held-out domain.
Ablations. Three ablations isolate the key ingredients. The first gives the verifier the whole article instead of one assertion, which tests the context firewall. The second replaces the soft targets $q_t$ with hard teacher labels, which tests the graded signal. The third replaces $q_t$ with uniform label smoothing, which tests whether the benefit is just regularization. Revisiting Knowledge Distillation found that label smoothing does not explain distillation gains in vision, and we test whether that holds for text spans.
Why not just…?
Why not just prompt a frontier model directly? MisSpans already shows that one-pass prompting stays below 0.5 F1 on span identification and typing, and that one-shot examples actually hurt smaller models. Direct prompting also costs a frontier call per article at deployment. It remains in our tables as a reference point, but the published evidence says it does not solve localization.
Why not just let one LLM label the data and distill that, skipping the agents? This is the real threat. If monolithic annotation with matched-budget voting produces a tagger as good as ours, the decomposition is unnecessary and we will say so. The deciding measurement comes before any student training. On the gold-labeled development set, we score the teachers themselves: span precision, span recall, and a teacher-hallucination rate, defined as the share of flagged spans that are actually true content. If the decomposed teacher does not beat the monolithic one on localization, we drop the student comparison and report the negative result about where LLMs fail.
Why not just use self-consistency, sampling one model many times to get soft labels? Agreement across samples does yield soft labels. However, every sample still attempts the combined find-and-judge task that MisSpans shows is weak, so the samples may agree on the same misplaced spans. We run exactly this as the matched-budget monolithic variant, with soft targets from vote fractions. A tie would mean the benefit comes from averaging, not from decomposition.
Why not just fine-tune on gold labels, or use a hallucination detector trained on weak labels? Gold-only fine-tuning is our naive baseline. OPDAI shows that small models trained on weak labels from prompt ensembles reach 0.836 accuracy on SemEval-2024 SHROOM. That result supports distillation in general, but it classifies whole outputs and does not localize spans. We test whether our span-level student is competitive on that binary task as a transfer check.
Proposed Experiments
Data and setup. The primary benchmark is MisSpans. It is built from FakeNewsAMT and entertainment news, with the fake article paired to a real one, which gives the verifier natural evidence. We report token-level and exact-span F1 for MisSpansIdentity and macro-F1 for MisSpansType. Gold training fractions are 5%, 10%, 25%, and 100%, with three seeds each. The unlabeled pool for teacher annotation is the AMTCele data used by RAAR.
Cross-domain test. We train on the FakeNewsAMT portion and test on the celebrity portion. No reported per-dataset RAAR number was found on Lacuna, so RAAR serves only as a qualitative document-level reference.
Transfer test. We apply the student's unsupported-tag score, aggregated to a binary decision, to SHROOM. We compare against OPDAI's reported 0.836 model-agnostic test accuracy, without expecting to exceed it given our task mismatch.
Explanation faithfulness. We will generate explanations for the student-flagged spans and apply the MADR human protocol on a sample. MADR's 30% fully faithful rate is the comparison point. The question is whether grounding explanations in correctly located spans raises that rate.
Cost. For every method we record cost and latency per article at inference and total teacher tokens at training.
Decisive checks. Four results decide the outcome. First, the decomposed teacher must beat the matched-budget monolithic teacher on development-set localization with a lower hallucination rate. Second, the soft-label student must beat hard-label and label-smoothing students, with a gap that widens as gold data shrinks. Third, the student must beat the Falconer-style student on the held-out domain. Fourth, the student must stay within a few F1 points of its own teacher at a small fraction of the teacher's cost. If the first check fails, the claim that localization is the bottleneck is wrong. If only the second fails, the distillation contribution reduces to cheap deployment of a better teacher.
References
- MisSpans: Fine-Grained False Span Identification in Cross-Domain Fake News
- Revisiting Knowledge Distillation: The Hidden Role of Dataset Size
- An automatic end-to-end chemical synthesis development platform powered by large language models
- A Tale of LLMs and Induced Small Proxies: Scalable Agents for Knowledge Mining
- Can LLMs Produce Faithful Explanations For Fact-checking? Towards Faithful Explainable Fact-Checking via Multi-Agent Debate
- AD-KD: Attribution-Driven Knowledge Distillation for Language Model Compression
- Towards Faithful Multi-step Reasoning through Fine-Grained Causal-aware Attribution Reasoning Distillation
- Debate, Reflect, and Distill: Multi-Agent Feedback with Tree-Structured Preference Optimization for Efficient Language Model Enhancement
- Thinking as Society: Multi-Social-Agent Self-Distillation for Multimodal Misinformation Detection
- RAAR: Retrieval Augmented Agentic Reasoning for Cross-Domain Misinformation Detection
- OPDAI at SemEval-2024 Task 6: Small LLMs can Accelerate Hallucination Detection with Weakly Supervised Data