Retrieval-Free Fact-Checking Techniques

• The paper introduces INTRA, a representation-based verifier that aggregates layer-wise internal signals to compute a truthfulness score from claims without external retrieval.
• It categorizes methods into logit-based and representation-based approaches, showing that representation methods outperform in generalizing over long-tail and out-of-domain data.
• Fact-checking without retrieval is significant for countering misinformation when explicit external evidence is scarce, addressing real-world limitations of retrieval-based systems.

Fact-checking without retrieval is the task of deciding whether a standalone natural-language claim is true or false using only what is encoded in a LLM’s parameters, with no web search, no database, no retrieval-augmented context, and no external tools. In this formulation, the verifier consumes only the claim text and computes a truthfulness score from the model’s internal processing of that sequence. The topic has acquired particular significance because retrieval-based pipelines are constrained by retrieval errors, external data availability, and the assumption that explicit counter-evidence already exists, whereas real-world misinformation often circulates precisely when such counter-evidence is absent or weak (Vazhentsev et al., 5 Mar 2026, Glockner et al., 2022).

1. Definition and formal task

In the retrieval-free formulation, a claim is a declarative sentence

$y = (y_1, \dots, y_n),$

and the goal is to compute a truthfulness score

$$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$

The verification function

$f : y \mapsto s$

must depend only on the claim text $y$, must not use the original prompt, surrounding context, or the model’s full output, and must not use any external knowledge source such as web search, knowledge bases, or RAG context. All evidence therefore comes from the model’s parametric knowledge and from internal signals such as hidden states $h_\ell(y_i)$, attention patterns, and output logits or probabilities (Vazhentsev et al., 5 Mar 2026).

This definition differs from retrieval-based fact-checking and from uncertainty estimation. Retrieval-based systems verify claims against retrieved documents and therefore focus on faithfulness to evidence rather than truth in isolation. Uncertainty estimation, by contrast, often measures how confident a model is in its own outputs. Retrieval-free fact-checking instead asks whether an arbitrary claim is factually correct, regardless of who produced it. In the evaluation described for this setting, the target label space is essentially binary—truthful versus hallucinated—even when some source datasets originally contained additional labels such as “neither,” which were filtered to binary (Vazhentsev et al., 5 Mar 2026).

A central implication is that the LLM is treated not only as a generator but also as a source of verification signals. This suggests a shift from evidence aggregation over external corpora to inspection of the model’s internal state as the locus of factual assessment.

2. Why retrieval-free conditions matter

Retrieval-free fact-checking is motivated not only by efficiency but also by a critique of standard NLP fact-checking task definitions. In the dominant FEVER-style pipeline, a claim $c$ is paired with retrieved evidence $E = \{e_1,\dots,e_n\}$, and a function $f(c,E)$ predicts a veracity label such as $\text{SUPPORTED}$, $\text{REFUTED}$, or $$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$0. This setup assumes that sufficiently strong counter-evidence is retrievable from a trusted corpus if the claim is false (Glockner et al., 2022).

The critique is that this assumption is often unrealistic for real misinformation. The analysis in (Glockner et al., 2022) argues that misinformation emerges “when the high demand for information cannot be met with a sufficient supply of credible answers,” producing data voids in which authoritative information is scarce or delayed. In such cases, explicit global counter-evidence may not pre-exist; professional fact-checkers often create the decisive refutation only after investigating the claim’s provenance, reasoning, or source material.

To characterize how professional verification proceeds under weak-evidence conditions, (Glockner et al., 2022) analyzes 100 misinformation claims from PolitiFact and Snopes, excludes 25/100 claims requiring multi-modal reasoning or scam/imposter webpage detection, and studies the remaining 75 textual claims. Four primary strategies are identified.

Strategy	Cases	Characterization
Global Counter-Evidence (GCE)	20	Direct textual contradiction; no source guarantee required
Local Counter-Evidence (LCE)	35	Refutation depends on showing the evidence is the same source or event
Non-Credible Source (NCS)	14	Refutation by establishing source non-credibility
No Evidence Assertion (NEA)	5	Verdict based on absence of trusted supporting evidence

For these 75 claims, LCE accounts for 46.7%, NCS for 18.7%, GCE for 26.7%, and NEA for 6.7%, with 1 additional “other” case. Aggregated differently, 49/75 claims, or approximately 65.3%, require a source guarantee, and 55/75 claims, or approximately 73.3%, do not rely on global counter-evidence. Only 26.7% are refuted in the way classical retrieval-based fact-checking applies cleanly (Glockner et al., 2022).

Two related notions are central in this critique. “Source Guarantee” is the guarantee that the evidence either is the claimant’s original source or refers to the same event, document, or dataset on which the claim relies. “Context Availability” is the broader information about who said the claim, where, when, and in response to what; it is described as a precondition for establishing a source guarantee. A plausible implication is that retrieval-free fact-checking, in the parametric sense, addresses one important failure mode of retrieval-based systems—absence of external counter-evidence—but does not by itself solve provenance tracing or contextual source verification.

3. Methodological families

The retrieval-free literature described in (Vazhentsev et al., 5 Mar 2026) divides methods into logit-based or probability-based approaches and representation-based approaches.

Logit-based methods treat the model as a conditional LLM and use token-level probabilities or uncertainty measures. Sequence probability (SP) is the log-likelihood of the claim under the model,

$$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$1

perplexity (PPL) is

$$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$2

and mean token entropy (MTE) is

$$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$3

Other uncertainty-oriented signals include CCP, RAUQ, Focus, and Attention Score. These methods are unsupervised: they do not require labels, but they rely on surface-level likelihood or confidence signals that are reported to be weakly aligned with truth, especially for long-tail facts and out-of-domain claims (Vazhentsev et al., 5 Mar 2026).

Representation-based methods instead treat the LLM as a feature extractor and train probes or heads on internal states. Examples include SAPLMA, MM, MIND, Sheeps, SATRMD, TAD, and UHead. The common pattern is to run the claim through the LLM, extract hidden states $$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$4, aggregate across tokens to a sequence-level embedding, and train a classifier or regressor to predict veracity. These methods are supervised and require labeled true/false claims (Vazhentsev et al., 5 Mar 2026).

The reported empirical conclusion is that representation-based methods outperform logit-based methods in the retrieval-free setting, especially for generalization across domains and long-tail knowledge. More specifically, (Vazhentsev et al., 5 Mar 2026) states that probability and entropy measures such as SP, PPL, and MTE are consistently weaker than representation-based methods on long-tail or out-of-domain datasets, while verbalized truthfulness is a strong unsupervised baseline but remains slower because it requires generation rather than a single forward pass.

4. INTRA and layer-interaction aggregation

The principal methodological contribution in (Vazhentsev et al., 5 Mar 2026) is INTRA, a representation-based verifier for “Intrinsic Truthfulness Assessment.” Its design is motivated by the claim that truthfulness is encoded in specific tokens and layers rather than uniformly across the network, that single-layer or single-token choices generalize poorly, and that internal truthfulness signals are spread across intermediate layers and tokens.

For a model with layers $$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$5, INTRA computes, at each layer, an attention distribution over token hidden states using a learnable vector $$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$6: $$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$7 The layer-specific sequence embedding is then

$$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$8

This is described as equivalent to a single attention head over tokens with a learned query $$s \in [0,1], \quad s \approx P(\text{Verified} \mid y).$$9, allowing the system to learn which tokens matter for truth assessment (Vazhentsev et al., 5 Mar 2026).

Each layer-specific embedding is mapped to a truthfulness probability by a linear classifier: $f : y \mapsto s$0 These layer-local classifiers are trained with binary cross-entropy. Because probabilities from different layers are not calibrated or on comparable scales, INTRA selects a band of middle layers, applies quantile normalization $f : y \mapsto s$1 to each layer’s probability distribution over the training set, and then fits a linear regressor across layers: $f : y \mapsto s$2 The aggregation stage uses an $f : y \mapsto s$3 squared-error regression objective, and training is two-stage on a train/validation split of PopQA-train: first fit $f : y \mapsto s$4 and $f : y \mapsto s$5, then fit $f : y \mapsto s$6, $f : y \mapsto s$7, and the quantile normalization on held-out data (Vazhentsev et al., 5 Mar 2026).

Several architectural choices are emphasized. INTRA uses all tokens in the claim but pools them with learnable attention rather than naïve averaging; it computes layer-wise truthfulness scores and aggregates them across a band of middle layers; it normalizes layer-wise probabilities before combining them; and it keeps the architecture shallow, using linear heads and a simple regressor, with the explicit goal of improving out-of-domain generalization. In contrast to UHead and TAD, it does not require the original prompt or the full generation; it operates on the claim alone (Vazhentsev et al., 5 Mar 2026).

Ablation results further identify the middle layers as the most informative. For Llama 3.1, using only early layers $f : y \mapsto s$8 or only late layers $f : y \mapsto s$9 yields clearly worse average ROC-AUC, approximately $y$0, whereas middle bands such as $y$1 or $y$2 yield the best results, with $y$3 slightly best overall at 77.9 average ROC-AUC. Using all layers does not improve over middle layers alone, and a single middle layer is substantially worse than aggregating multiple layers (Vazhentsev et al., 5 Mar 2026).

5. Evaluation framework and empirical findings

The benchmark introduced in (Vazhentsev et al., 5 Mar 2026) is explicitly designed for retrieval-free parametric fact-checking and emphasizes generalization. It evaluates 18 methods on 9 datasets and 3 open-source base LLMs—Llama 3.1-8B Instruct, Ministral 8B Instruct 2410, and Phi-4-mini Instruct—with GPT-4.1 used as a closed-source upper bound via prompting and verbalization. The 18 methods are grouped into one retrieval-based baseline (Verb+RAG), retrieval-free unsupervised methods (SP, PPL, MTE, Attention Score, Focus, RAUQ, CCP, and Verb), and retrieval-free supervised internal or probing methods (UHead, MM, CCS, TAD, SATRMD, MIND, Sheeps, SAPLMA, and INTRA) (Vazhentsev et al., 5 Mar 2026).

The dataset suite spans long-tail knowledge, long-form generation, human-authored claims, rule-generated claims, LLM-generated claims, cross-model variation, and multilinguality. The nine datasets are AC-PopQA, AC-WH, AVeriTeC, the UHead dataset, Cities, Companies, Common Claims, CounterFact, and X-Fact. All claims are filtered via Llama 3.3-72B to ensure they are self-contained atomic facts about a single entity and do not require external context. Only AC-PopQA has a train split; all other datasets are out-of-domain test-only, and all supervised methods are trained only on AC-PopQA-train (Vazhentsev et al., 5 Mar 2026).

The primary metrics are ROC-AUC and PR-AUC, both threshold-free. PR-AUC is computed with respect to hallucinations as the positive class. On average over all 9 datasets and 3 models, INTRA is reported as the best retrieval-free method overall, with average ROC-AUC 73.3 and average PR-AUC 69.0. It matches or exceeds Verb+RAG in ROC-AUC and surpasses it in PR-AUC by approximately 3 points on average, while being approximately 20 times faster. For Llama 3.1 specifically, INTRA achieves 77.7 average ROC-AUC across the 9 datasets; the second-best retrieval-free method is Sheeps at 75.0, while Verb+RAG also attains 77.7 ROC-AUC but lower PR-AUC. On Llama 3.1 PR-AUC, INTRA reaches 73.1, compared with 70.5 for Sheeps and 70.1 for Verb+RAG (Vazhentsev et al., 5 Mar 2026).

The computational profile is also reported explicitly. On Ministral-8B and PopQA, SP and PPL run in approximately 30 ms per claim, supervised probes in approximately 30–70 ms, INTRA in 56 ms, Verb in approximately 250 ms, and Verb+RAG in approximately 950 ms because of retrieval plus generation. This motivates the claim that INTRA is lightweight enough to run inline in generation loops or as a batch filter (Vazhentsev et al., 5 Mar 2026).

Robustness analyses refine these results. On PopQA divided by popularity bins, INTRA dominates across rare-to-popular entities with statistically significant gains ($y$4, Benjamini–Hochberg FDR correction), while the verbalized detector is best only in the rarest entity bucket, 0–100 views. On X-Fact, no single method wins in every language: CCP is best on Indonesian and Romanian, INTRA is best on Georgian, PPL is best on Tamil, and SP is slightly best on Portuguese. On long-form data, no strong monotonic increase in hallucination rates over normalized claim position is observed; for short generations INTRA significantly outperforms Verb at early and late positions, for medium-length generations Verb beats INTRA at early and late positions while INTRA is competitive in the middle, and for long-length claims the data is too sparse for significance (Vazhentsev et al., 5 Mar 2026).

These findings support several conceptual claims made in (Vazhentsev et al., 5 Mar 2026): internal representations encode truth signals; surface likelihood is not equivalent to truth; middle layers are more informative than early or late layers; and calibration plus aggregation of layer-wise scores improves robustness.

6. Dataset realism, leakage, and evidence sufficiency

A separate but closely related line of work argues that many standard fact-checking datasets do not model realistic low-retrieval conditions. The central proposal in (Glockner et al., 2022) is that evidence in datasets must satisfy two requirements for realistic fact-checking: it must be sufficient to refute the claim, and it must be unleaked from existing fact-checking articles.

The survey in (Glockner et al., 2022) reviews 16 fact-checking datasets and concludes that all fail to satisfy both criteria. Synthetic datasets based on perturbations of known truths guarantee the existence of global counter-evidence by construction, but do not reflect real-world misinformation. Datasets built from fact-checking websites often use fact-checking articles or quoted snippets as evidence; such evidence may be sufficient, but it is leaked because it is itself the output of the human verification process. Other datasets using web or Wikipedia evidence are criticized because evidence is often incomplete, insufficient, or contaminated by derivative reporting of fact-checker conclusions (Glockner et al., 2022).

The most detailed case study concerns MultiFC. Among 16,244 misinformation claims, automatically detected leakage appears in 8,999 claims (55.6%) by URL-based detection, 9,656 claims (59.7%) by phrase-based detection, and 11,267 claims (69.7%) by either criterion. Manual validation on 100 supposedly unleaked claims shows that about one-third still contain leaked evidence missed by the automatic patterns. Only 15 claims have unleaked refuting evidence, and in 10 of those cases the unleaked refuting evidence is overshadowed by leaked evidence in the same claim (Glockner et al., 2022).

Modeling experiments reinforce this critique. Using bert-base-uncased, four input configurations are evaluated on MultiFC: Sn-Text, Sn-Title, Snippets, and Full. Inputs are concatenated and truncated to 512 tokens, with standard cross-entropy on the $y$5 output, and results are reported in macro-F1 and micro-F1 averaged over three runs. On PolitiFact, the “Snippets” configuration without the claim reaches 59.7 macro-F1 and 59.2 micro-F1 on leaked test claims but only 22.1 macro-F1 and 23.0 micro-F1 on unleaked claims, a gap of approximately 36.2 micro-F1. Further analysis shows that when leaked snippets are removed, performance roughly halves on PolitiFact, and a claim-only BERT baseline performs far below the evidence-only model, indicating that the gains arise from patterns in the snippets rather than from semantic cues in the claims (Glockner et al., 2022).

This line of criticism does not invalidate retrieval-free parametric verification, but it changes the interpretation of benchmark performance. A plausible implication is that retrieval-free systems and retrieval-based systems should both be evaluated under explicit controls for leakage, sufficiency, provenance, and temporal availability of evidence.

7. Applications, limitations, and research directions

The application space proposed in (Vazhentsev et al., 5 Mar 2026) follows directly from the low-latency scalar score $y$6. The score can serve as a reward model in RLHF-style training or preference optimization, can be combined with other reward dimensions such as helpfulness, harmlessness, and style, and can be used in generation-time workflows for inline checking, filtering, regeneration, uncertainty marking, or triggering a retrieval-based cross-check when the score falls below a threshold. Because INTRA runs in approximately 50–60 ms per claim, it is presented as feasible for deployment alongside generation (Vazhentsev et al., 5 Mar 2026).

The limitations are equally explicit. Retrieval-free fact-checking is bounded by the model’s parametric coverage and training data biases; it struggles on truly unknown or recent facts outside pretraining data; it lacks explicit evidence and therefore provides weak interpretability; and multilingual performance drops substantially when methods are trained on English data alone. The paper also notes contamination or saturation in heavily reused rule-based datasets such as Cities and CounterFact, which can make them less informative about real-world performance (Vazhentsev et al., 5 Mar 2026).

The critique in (Glockner et al., 2022) adds a different class of limitation. Even when retrieval is allowed, current task formulations generally ignore source guarantees, context-dependent reasoning, and no-evidence assertions. The analysis therefore suggests that realistic fact-checking requires more than claim-to-evidence entailment: it may require provenance detection, reasoning about whether a claim follows from its cited source, logical fallacy or misinterpretation detection, and the ability to output uncertainty or “unsupported” judgments rather than forcing a binary verdict (Glockner et al., 2022).

The research directions proposed across these works are complementary rather than mutually exclusive. (Vazhentsev et al., 5 Mar 2026) suggests better layer-aware training, combination of parametric and retrieval-based fact-checkers, extension to long-form structured reward models incorporating claim position and discourse structure, and scaling to larger or more multilingual base models. (Glockner et al., 2022) suggests task redesign around source detection, evidence sufficiency, uncertainty classification, time-filtered dataset construction, and explicit separation of premise evidence from fact-checkers’ own explanatory articles. Taken together, these proposals suggest a research program in which retrieval-free parametric verification functions as a fast and scalable component, while provenance-aware and evidence-aware methods address cases where external knowledge, source tracing, or explicit justification is indispensable.