Does Retrieval Cure Hallucination? — What RAG Fixes, and What It Does Not

When We Checked All 8,050 References

There is a study in which three top-tier language models in 2026 were made to write 270 cardiology reviews with web search turned on. The researchers cross-checked every single one of the 8,050 references attached to those texts against PubMed, a database of medical literature. The proportion of problematic references ranged from 11.5% to 29.6% depending on the model. It would be a mistake to read this range as a median. Two of the three models were in the 29% range, and 11.5% is the value for just the lowest single model.

What deserves closer attention is the shape of the error. 77% of the errors were misattributions. Misattribution is an error where the cited paper number actually exists, but points to a different paper. Cases of entirely fabricating a nonexistent literature were rare at 0.6%. The abstract does not separately state the denominator for this ratio (DOI: 10.64898/2026.08.31.26361806, preprint).

Following the definition in "How AI Gets Things Wrong — The Test-Taker Who Stumbles Outside the Exam Hall", a hallucination is when AI produces content that sounds plausible but is factually incorrect or inconsistent with the given material. Both numbers pointing to the wrong paper and nonexistent papers fall under this definition. Put those two numbers together and you get a single sentence: Even with web search turned on, hallucinations remained, and the vast majority of the remaining hallucinations were not fabricated nonexistent literature, but rather real literature attached in the wrong places. However, since the condition described in the abstract is only the one with search turned on, we cannot go so far as to say that search changed the shape of the error.

The fix we will deal with here is retrieval-augmented generation. Exactly as defined in "Why Do LLMs 'Lie'? — Where Hallucinations Come From, and How Line-by-Line Fixes Go", retrieval-augmented generation (RAG) is a method that searches for external documents and places them in front of the model before answering, having it answer based on those documents. Taking an internal company policy chatbot as an example, when a question comes in, it pulls out a few relevant excerpts from the rulebook and hands them to the model along with the question, and the model in principle looks only at those excerpts to answer. As "How Generative AI Answers — A Sentence Built One Piece at a Time" pointed out, the basic procedure of a language model does not have a search step, and search is a separately attached device. "How AI Gets Things Wrong" noted that this device reduced hallucinations but did not eliminate them, and "Why Do LLMs 'Lie'?" passed the problems that remain even with search attached over to this article. This article looks at what that device fixes and what it fails to fix.

A review paper notes both the reason this fix is used and its limitation. It was proposed as a practical fix because it supplements knowledge contained in the numbers determined through learning inside the model—in other words, parameters—with external evidence fetched at the moment of answering. And it adds immediately: RAG fundamentally guarantees neither truthfulness nor source attribution (DOI: 10.54254/2753-8818/2026.dl34055).

Three things are clarified before reading. First, this article does not evaluate specific products or models. It does not even carry over the names of the models revealed by the earlier study. What this article uses is the shape of the error. Second, the grades of the cited studies are uneven and skewed to one side. There are many preprints on this topic that have not undergone peer review, and the studies this article relies on most heavily are largely preprints. Peer-reviewed pages were generally used as a supplement, and preprints were indicated each time. Third, the issue of searched-in documents acting like instructions to manipulate the model is covered in "When Someone Else's Sentence Becomes Your Instruction — Prompt Injection". What we look at here is what happens when the searched-in documents are wrong or insufficient.

Prescriptions Actually Work

The evidence that attaching search makes things better comes from the clinic. A systematic review gathered studies that used retrieval-augmented generation in systems assisting doctors' decisions. A systematic review is a study that collects and reviews related research comprehensively using a set procedure, and this review followed its reporting guideline, PRISMA. Thirty studies meeting the conditions among those published from January 2023 to December 2025 were gathered. In the compared studies, the approach with search attached was associated with answers sticking better to factual evidence and fewer unsupported claims compared to language models without search attached. This was especially true in tasks relying on guidelines and protocols. The authors call this practical progress under controlled conditions (DOI: 10.21203/rs.3.rs-9741159/v1, preprint). This means search actually works under controlled conditions.

The first notch for measuring where the improvement comes from lies on the search side. Recall is the proportion of documents that should have been found that were actually brought back. If three clauses needed for an answer are in the rulebook but only two are brought out, the recall is two-thirds. If recall is low, no matter how good the next step is, the answer cannot be filled.

There are also values measured in live services. This is a study reporting a retrieval-augmented platform actually operated by a global-scale digital infrastructure provider, covering 334,818 search units gathered from eight sources (11.7% of which are images). The benefit of search measured by the authors was that the relevance of the answers was +0.25 higher than when the model was made to answer using only memory without documents (DOI: 10.2139/ssrn.7360668, preprint). The caveat attached to this figure is revisited later.

Two more caveats are attached to the conclusion of the same clinical review, and the rest of this article follows those caveats. The first caveat is the evaluation method. Evaluations were mostly retrospective methods looking back at past records, metric-focused, and had little clinician participation. There was not a single prospective clinical outcome study that tracked results forward while actually using it on patients. The second caveat is the kind of failure. The failures that appeared repeatedly were retrieval errors, integration that reflected evidence incompletely or selectively, and hallucinations that remained even when a retrieved context was present. The authors summarize that clinical validation has not kept pace side-by-side with structural innovation (DOI: 10.21203/rs.3.rs-9741159/v1). There is still an unverified distance between working under controlled conditions and being ready for use in the medical practice setting.

The Problem Statements Papers Themselves Write Down

In papers on this topic, reports that "our method reduced hallucinations by N%" are very common. Because these have the same character as a manufacturer reviewing its own product, performance figures must be read with caution, and how to read them will be covered in a later section.

Even in such papers, there is one honest passage. It is the problem statement—the passage written first to explain why their method is necessary by showing why existing approaches fall short. The five sentences below are all written at the very beginning of the abstracts of those papers, and they are motivations, not performance outcomes.

First, standard pipelines have no checking component. A pipeline is the processing path that runs from receiving a question to finding documents and writing an answer. A paper proposing to attach a real-time hallucination detection and correction device here writes this at the beginning of its abstract: because retrieval and generation are functionally disconnected, there is no device within the standard pipeline to verify whether what the model wrote actually follows from what was retrieved (DOI: 10.2139/ssrn.7242858, preprint; the authors note that this is a proposed architecture and that empirical validation is left as future work).

Second, even citations are written by the model. This is the first sentence of a paper proposing deterministic citation verification, which checks citations against the same rules every time. Retrieval-augmented generation reduces unsupported assertions but does not eliminate them, because in common implementations the model writes the answer and the sources of that answer together. The author proposes separating model-fabricated citations from citations attached to retrieved records into distinct categories (DOI: 10.2139/ssrn.7322378, preprint; the author notes themselves that the formal content is more definitional than deep and the contribution is structural).

Third, hallucinations remain even when relevant documents are properly retrieved. A paper proposing sentence-level retrieval writes that mitigation is not the same as removal. The existing implementations examined by that paper allow the model to still pull in knowledge it learned during pretraining when assembling the answer. Pretraining is the first-stage learning done on a massive amount of text. Therefore, hallucinations can remain even when relevant documents are properly retrieved (DOI: 10.2139/ssrn.7333058, preprint).

Fourth, when evidence and memory conflict, the memory may be chosen. A paper proposing a method (low-rank adaptation) that further trains only a tiny fraction of the model to align it toward the evidence divides the remaining hallucinations into two types: fabricating plausible but unsupported assertions, and following the memorized bias by ignoring the evidence when the retrieved evidence conflicts with what was memorized through learning (DOI: 10.66917/ijaeic.a000015).

Fifth, the evidence itself is not clean. A paper proposing a dynamic memory approach lists three conditions under which hallucinations remain: when noise is mixed into the retrieved evidence, when pieces of evidence contradict one another, and when the evidence is insufficient to begin with. It also adds that existing adaptive systems, which decide whether to retrieve on the fly, have neither a device to gauge retrieval quality before generation nor a device to catch contradictions among pieces of evidence (DOI: 10.2139/ssrn.6329665, preprint).

These five sentences were written by different teams presenting different methods, yet they point to the same place. They also overlap with the four places where the earlier review paper pointed out that errors arise: when crafting the query, when finding documents, when gathering evidence, and when attaching the answer to the evidence. That review concludes that RAG should not be treated as a complete solution to hallucinations. The value of RAG lies in making generation tied to external evidence possible, not in guaranteeing it (DOI: 10.54254/2753-8818/2026.dl34055).

Faithful but Wrong Answers

Suppose the company rulebook says, "Apply for leave two weeks in advance," but that rule changed to one week last month and only the rulebook hasn't been updated. A chatbot copies the rulebook verbatim and answers, "Two weeks in advance." To evaluate this answer, we need two axes.

Grounding faithfulness is the degree to which sentences written by a model actually follow from the document placed in front of it. Factuality is the degree to which those sentences are actually true in the world.

That answer we just looked at is perfectly faithful to the grounds and factually wrong. Conversely, if a chatbot guesses from its own memory, "It's one week these days," the fact is correct, but it does not follow from the given grounds. Unless we check which axis every subsequent score is measuring, we will misread the numbers.

The creators of metrics also split into the two axes. A team proposing a tool to measure grounding coverage writes their design goal as follows: measuring whether each factual claim is explicitly supported by the retrieved context, independently of whether it matches the correct answer on the surface (DOI: 10.36227/techrxiv.176784505.56675782/v1, preprint · authors' own method). According to that design, an answer that faithfully transcribes a wrong document is faithful but wrong, and gets a perfect score on this metric.

That is why the preceding review divided the evaluation axes into four: factuality, grounding faithfulness, citation quality, and retrieval relevance. Having the axes separate means that fixing one can leave the other three intact (DOI: 10.54254/2753-8818/2026.dl34055).

The Scale Runs Backward

The same problem catches scoring. When there are too many answers for humans to grade, grading is delegated to another language model. Just as defined in "The Score Went Up. But What Went Up?", a language model delegated to grade like this is called a model judge. It is also called an LLM judge. The approach asks a model judge for a single answer, "Is this answer attached to the grounds?", and records the score.

The central finding picked by the preceding live-service study is also about this measurement by model judges. This paper calls the degree to which an answer is attached to the grounds "groundedness," which is the same axis as the grounding faithfulness mentioned earlier.

This is the first finding. Model judges that grade based only on the answer were systematically wrong in their groundedness judgments. They gave higher scores to questions that could not be answered in the first place than to questions that could be answered.

So, the retrieved context was also shown to the model judge. That eliminated one shortcut, but created another. Systems unable to retrieve on this platform refused to answer instead of making things up, and to a model judge looking at the context, a refusal is easily judged as an answer well-attached to the grounds. This is because a refusal contains no claims to conflict with the grounds. As a result, the groundedness scored by the model judge ran backward from the actual retrieval quality measured by correct answer labels. Across various settings where Top-1 recall varied up to 11-fold, the Spearman correlation measuring how much the two rankings moved in the same direction was −0.89. This is a value meaning the rankings were almost completely opposite, and it means the better the retrieval setting, the lower the groundedness score.

The authors' prescription is simple. Because groundedness scores cannot distinguish correct refusals from answers made right by attaching to grounds, groundedness scores must always be reported together with the refusal rate. Also, when the model judge was switched to a different family, the systems' rankings themselves shook. Kendall's tau, which measures how much rankings match depending on the judge family, was 0.10, showing almost no agreement. Therefore, the authors pin down even their measured retrieval gain of +0.25 as "a value limited to that model judge" (DOI: 10.2139/ssrn.7360668, preprint).

The reliability of model judges was also tested on the task of pointing out the causes of failure in agents—programs that work through multiple steps on their own. This is a study that tested ten model judges across two agent tasks with intentionally planted flaws. Even in the condition showing the main experiment's logs in the greatest detail, with standard prompts no judge exceeded a margin of 0.16 (the difference in accuracy between the two causes) separating the two causes, and a single prompt revision raised it to 0.34. Because deterministic rules not using models separated the two causes from the same logs, the information to separate the causes was inside the logs. The authors read this result as a reliability warning within the scope of models, prompts, and tasks tested (DOI: 10.3390/make8090275). This test is covered in detail by "Multiplication Won't Tell You — Where Agents Break on Long Tasks".

Where It Breaks

What Was Not Found vs. What Was Found But Not Used

Search failures fall into two categories. There are cases where the necessary evidence was ultimately not found, and cases where relevant evidence was found but not properly used. A study that used this distinction dissected six search agents—which repeatedly search to find answers on their own—at each search step against human-assigned relevance judgments for each document (arXiv:2608.01913, preprint).

In the same study, search effort and answer quality correlated only weakly. Answer accuracy tracked better with the quality of the retrieved evidence than with the number of searches or the amount of context consumed. In particular, it correlated well with the recall measured by combining all documents found up to that point, in other words, cumulative recall. Furthermore, even though useful evidence often appeared early, agents tended to keep searching, leading to long stretches of search steps with few useful yields (arXiv:2608.01913, preprint).

Separating these two failures helps gauge where adjustments are needed. The directions proposed by the authors also break down into crafting better queries, better selecting evidence and managing context, and deciding when to stop based on whether sufficient evidence has been gathered (arXiv:2608.01913). A study that created a benchmark to separately measure search errors and generation errors targets the same spot. A benchmark is a collection of test questions used to compare scores by having multiple AI models solve the same problems. However, the authors limit those results to open-weight models between 2B and 8B in size, meaning small-to-medium models. Open-weight models are models whose parameters are publicly released (DOI: 10.1007/s10994-026-07121-y).

Methods to predict in advance whether a search will go well have also been attempted. A study published in a peer-reviewed journal proposed a framework that predicts query difficulty using only features from after the search is completed. The correlation between those predicted values and actual generation quality ranged from 0.22 to 0.48. This means there is still a gap between knowing if a search will go well and knowing if the answer will be good. Still, when search depth was adjusted according to that prediction, generation quality improved consistently, with the greatest gains seen in queries where searches previously tended to go poorly. It also surpassed existing methods that adjust searches without training (DOI: 10.1145/3827605).

There are also limits to increasing the retrieval volume. In one small-scale experiment, blindly increasing the candidate documents for dense retrieval—which finds documents with similar meanings—caused the recall to hit a ceiling, yielding diminishing returns. The accuracy of the best configuration proposed in that experiment was 62% (DOI: 10.66245/jyi.v1.i1.002, a small-scale experiment published in a student journal). Feeding in long contexts is not free either. A survey reviewing biomedical RAG notes the phenomenon where information placed in the middle of a long context gets buried as a technical difficulty (DOI: 10.70121/001c.158711).

Documents Are Not Shaped That Way to Begin With

There is a study in Japan that fed four institutional manuals unchanged into a search system designed to assist with ethical reviews of clinical research. When a model judge graded the quality of the 594 retrieved snippets using five metrics, the medians of all five metrics were 2.0 or lower. Fewer than 20% of the snippets exceeded the practical threshold (4 points or higher). Even when filtering for only the top-ranking results, less than half passed the threshold. There was an exception with a single metric, an item the authors named 'faithfulness'. Whether this means the same thing as the evidence faithfulness mentioned earlier cannot be verified from the abstract. This scoring yielded almost identical results even when repeated five times (the intraclass correlation, which measures agreement between repeated scores, was 0.936). The median length of the retrieved snippets was 3,861 characters, greatly exceeding the configured 500 tokens. Tokens are units into which language models chop up text to process it. The authors interpret this as a signal where core information is diluted within long snippets.

The same study tried re-feeding a single paragraph after cleaning up only its structure and formatting. The rewritten side received a perfect score on every metric. However, the rewritten side had only six snippets. We must not forget the asymmetry of 594 versus 6. The authors' conclusion is that the quality of the original documents entering the search targets constrains search performance. This is a challenge to the assumption that you just need to upload existing documents (DOI: 10.64898/2026.01.01.26343326, preprint · Tier 1 basic research · the authors note that the inter-rater agreement in the primary evaluation was low at Fleiss' kappa 0.269, indicating that the framework itself needs refining).

There are also measurements taken from a different angle. A team that built a benchmark evaluating medicine, law, and finance under a common procedure reports that in the legal domain, switching the collection of search target documents—namely, the corpus—from synthetic data to an actual Australian legal corpus raised the citation accuracy score from near 0 to 0.678. In the same benchmark, dense retrieval scored 1.6 percentage points higher in overall score than BM25, the traditional method of finding documents with overlapping words. The two values measure different things. One is the increment of the overall score and the other is the absolute level of a single citation accuracy score, so their magnitudes cannot be directly compared. Yet the sentence the authors attach to this result points clearly in one direction. Real data is not an item you can compromise on (DOI: 10.21203/rs.3.rs-9859942/v1, preprint).

There is also no single correct answer for what size to cut documents into for indexing. In a study controlling and comparing four sizes, cutting text finely into propositional units—which are short statements containing a single fact—yielded up to 6.8 percentage points higher top-20 recall on one benchmark compared to paragraph-level units. The accuracy rate for questions asking for a single fact was also the highest. However, that advantage diminished for questions requiring multiple documents to be pieced together, where paragraph-level units performed similarly or slightly better. The authors' conclusion is that there is no single size that pulls ahead under all conditions (DOI: 10.66372/jger.v3i1.6).

How to Read "Our Method Reduced It by N%"

This is a sentence format common in papers on this topic, which we put off earlier. The method of attaching a knowledge graph is the first example. As defined in "Why Do LLMs 'Lie'?", a knowledge graph is a structured knowledge storage where facts are explicitly contained, like Wikipedia. This method reports that exact match increased by 11.4% and the hallucination rate decreased by 18.2% compared to a strong RAG baseline (DOI: 10.54097/e4991d70). A baseline is the comparison method set up to test the new method against, and exact match is the rate at which the answer is identical to the correct answer. The low-rank adaptation method reports that the factual hallucination rate decreased by 43.7% compared to the standard RAG baseline. The faithfulness score went from 0.642 to 0.891, and F1, a score measuring the degree of overlap with the correct answer, increased by 11.8% (DOI: 10.66917/ijaeic.a000015). The risk control method for technical documents reports a 43% reduction in the hallucination rate compared to the standard RAG baseline (DOI: 10.66238/ijcbs29).

All three sentences are values where the authors compared their method to their baseline. Therefore, three things must be asked together. Who measured it, what was it compared against, and what is left over. The answer to the third question is especially easy to leave out. Among the three reports above, only one shows the remaining absolute amount. The faithfulness score of 0.891 is not 1. The other two only have the reduced percentage, and the baseline figures are not written in the abstract.

There are also cases where the remaining amount is written down. In a study applying twelve retrieval-augmented methods to 250 patient case descriptions (vignettes), a self-reflective design that makes the model recheck its own answer lowered the hallucination rate down to 5.8%. Before this value is a number of achievement, it is a number for the amount left over in that experiment. In the same experiment, the setting with the best retrieval accuracy had a ratio of relevant documents among the top 5 (precision) of 0.68 or higher. This means that among the top five documents, there were on average a little over three relevant documents or more (DOI: 10.3390/electronics14214227).

When received as absolute performance, the position of the scale becomes clear. A 2025 preprint used a method of finding empty spots in evidence and searching back through them, obtaining an F1 of 0.453 on HotpotQA, a question benchmark where multiple documents must be stitched together to get an answer. This value is 8.3 points higher than the strongest iterative baseline, and the authors write that this is the highest score achieved by this family of methods on those benchmarks (arXiv:2510.22344, preprint). This is precisely the result introduced by "Why Do LLMs 'Lie'?" as an upward jump. Read that same result as an absolute value, and it means that the highest score for this family fell short of half in its overlap with the correct answer.

Another team using the same HotpotQA under fair corpus retrieval conditions writes that the exact match of the highest setting was 11.5% and F1 was 16.8%. This is a value improved by 0.5 percentage points over the baseline. Because the experimental conditions differ, it cannot be directly compared to the previous result. Success and failure were split by component. The component that catches contradictions between pieces of evidence found contradictions in 97.5% of the retrieval results. Conversely, another module that was theoretically plausible actually degraded performance in a small generative model (DOI: 10.2139/ssrn.6329665, preprint). This is a case where a paper presenting its own method even records the failures of its own modules.

In a cross-three-domain benchmark, the highest overall score among six settings was 0.585 (DOI: 10.21203/rs.3.rs-9859942/v1).

"Reduced by 43%" and "absolute value 0.453" do not contradict each other. They measure the same world with different rulers, and both are needed to complete the sentence. If you only copy the reduced percentage and leave out the remaining amount, the sentence gets cut off. For two of the three reports seen in this article, the remaining amount was left out starting right from the abstract.

The Value of Refusing to Answer

One path remains. Withholding judgment means that when an answer cannot be provided from the given evidence, the system stops and says it does not know instead of making up an answer. It is like finishing by saying "There is no relevant provision in the rulebook" when you cannot find the relevant clause in the rulebook.

Withholding actually works. When the system was made to refuse answers when evidence was insufficient, baseless claims decreased and reliability went up. This is the result of testing across three types of embedding models, which turn text into bundles of numbers to find similar text, and three stages of search depth. However, a measurable trade-off between precision and refusal arose together. This means that as wrong answers decrease, the questions for which no answer is received at all increase. The authors summarize that preventing hallucinations is not just a matter of the retrieval process, but also a matter of response control and evaluation (DOI: 10.2139/ssrn.6702638, preprint).

Yet in a benchmark that measured withholding separately, the tested systems did not withhold. This is a study comparing six memory structures on a controlled benchmark handling loan eligibility screening and insurance claim adjudication. All six structures gave answers for every single case. This happened even though the paper intentionally created and measured withholding as one of four evaluation axes. The authors call this an axis the field is not aiming at yet. The same paper also notes that retrieval-based structures collapsed on the factual accuracy axis, and that one of their pre-registered predictions was heavily contradicted by the data (arXiv:2604.19457, preprint).

The danger on the opposite side is also documented. A team that built a clinical question-answering prototype for resource-constrained environments notes that while standard RAG achieved high accuracy, it showed a dangerous "yes" bias. When asked, it tended to answer yes almost no matter what. That team's method chose to give up some accuracy in order to reduce that overconfidence (DOI: 10.21203/rs.3.rs-8163345/v1, preprint · prototype · author's own method).

For withholding to do its job, one more thing is needed. Distinguishing between what the search failed to find and what is absent from the corpus in the first place. The two require different things to be said to the user. There is research that attempted this distinction. By diagnosing the structure of the network formed by concepts within documents, it tried to separate retrieval deficits from evidence scarcity in the corpus itself (DOI: 10.64898/2026.02.18.26346545, preprint · single-case proof of concept).

A logical extreme has also been proposed. If generation is not done at all, hallucinations are zero. There is a design that indexes documents at the sentence level to return the original source sentences as they are, leaving generation as an option. The authors write that this design does not just reduce hallucinations, it eliminates them (DOI: 10.2139/ssrn.7333058, preprint · quantitative comparison not in the abstract). The price is clear. What you get back is a search result, not an answer.

Here the problem from the previous section returns. To a model judge looking at context, a refusal is easily judged as an answer well-anchored to the evidence. Increasing refusals improves the faithfulness metric, while decreasing the number of answers the user receives. A single metric cannot tell these two apart. Therefore, the demand to report the refusal rate together is a minimum condition for reading the grounding score.

So What Must Be Asked

"Does this system hallucinate?" cannot be answered in a single word. What can be answered is the following list:

① Is what is needed for the answer inside the corpus? If not, there is no need to ask ② through ④, and the only remaining question is ⑤. The quality of the original document limits retrieval performance (DOI: 10.64898/2026.01.01.26343326, preprint).

② Did it go and fetch it? This is the place of recall. Answer accuracy aligns better with cumulative recall than with the number of retrieval attempts (arXiv:2608.01913).

③ Did it actually use what it brought back? This is the place of utilization, and the reality of the failure where it fetches something and then fails to use it has been measured empirically using human-rated relevance (arXiv:2608.01913, preprint). The explanation that when evidence contradicts memory, it chooses memory is a statement written by that paper as a motivation for its method, not a measurement (DOI: 10.66917/ijaeic.a000015).

④ Is what it brought back correct? This is the place of factuality, not evidence faithfulness. Because a flawed document makes even a faithful answer flawed—just like the vacation policy example—this item cannot be solved by retrieval performance alone, and requires document-side management as well.

⑤ What does it do when it fails to find it? This is the place of withholding, and unless withholding rates are looked at together, the metrics are read backward (DOI: 10.2139/ssrn.7360668, preprint).

There is a sixth as well. Who is checking? A paper that models citation practices into a five-step staircase points to this spot. Step 1 is reading the original text directly, Step 3 is a human checking what the AI found, and Step 5 is fully automated citation without human oversight. The critical transition the authors point out is not between humans and AI, but between Step 3 and Step 4. It is between the point where a human's ability to spot something unusual still operates and the point where it stops. And that transition is invisible from the outside. No feature of a submitted manuscript reveals how much the author checked (DOI: 10.2139/ssrn.6318660, preprint · this is a behavioral model, not an empirical study).

The reason this point is structurally difficult is already present in the second problem statement. The model authors the answer and the source for that answer together. Therefore, one proposal suggests outputting only the citations attached to the retrieved records on the screen instead of citations made up by the model. And it states that when that inspection gate is ambiguous, it should fail toward closing rather than passing (DOI: 10.2139/ssrn.7322378, preprint).

There is no reason to lean into cynicism here. The study that opened this article also attached a verification procedure. Out of 270 references judged by experts, a verification procedure built with a language model caught 60 out of 62 problem references. The sensitivity—the rate of catching problem references as problems—was 96.8%, the specificity—the rate of leaving untroubled references alone—was 98.6%, and misattribution and fabrication were caught entirely (DOI: 10.64898/2026.08.31.26361806, preprint). This means that once what constitutes a problem is clearly defined, automated inspection can also be built.

There is no reason to lean into optimism either. That inspection is at the level of citation numbers. The proposition that a number points to the correct paper and the proposition that the paper actually supports that sentence are different things. Sitting in the seat that automatically judges the latter is ultimately a language model, and the verification procedure that just scored 96.8% was a language model to begin with. In the earlier live-service study, the model judge gave higher scores to unanswerable questions and read refusals as answers well-grounded in evidence. Whether this verification procedure takes the same shortcut does not appear in the abstract, but if the grader is a language model, that point must be checked first.

Search gives the model something to look at. It does not make it speak what it saw.

There are things search fixes as well. In the earlier clinical review (preprint), the approach coupled with search was associated with a reduction in unsupported claims under controlled conditions. However, there are five items between seeing and speaking. Is it in the corpus? Did it fetch it? Did it use it? Is it correct? Does it know how to stop when it isn't there? Not a single one of those items becomes true automatically just by attaching search. Therefore, this prescription is an unfinished prescription. What must be asked is not whether search was attached, but what the state of each of the five items is, and only when asked that way does it become a question with an answer.

Source Papers

Preprints are manuscripts that have not undergone peer review. Each instance is marked accordingly in the text.

• "Citation reliability of frontier large language models in medical writing and its automated verification" (2026, preprint) — DOI: 10.64898/2026.08.31.26361806 — Complete collation of 8,050 references across 270 reviews. Problem rate 11.5–29.6% (two out of three in the 29% range), misattribution 77%, fabrication 0.6%, language model verifier detected 60 out of 62 items.
• "Hallucination in Large Language Models and Retrieval-Augmented Generation" (2026) — DOI: 10.54254/2753-8818/2026.dl34055 — RAG does not guarantee truthfulness or citation, and errors occur across four places: query, retrieval, integration, and grounding.
• "Retrieval-Augmented Large Language Models for Clinical Decision Support: A Systematic Review" (2026, preprint) — DOI: 10.21203/rs.3.rs-9741159/v1 — PRISMA 30 papers. Associated with improved grounding, but this is progress under controlled conditions with 0 prospective outcome studies.
• "Production-scale multimodal retrieval-augmented generation for network operations" (2026, preprint) — DOI: 10.2139/ssrn.7360668 — An answer-only judge rates unanswerable questions higher, and even when viewing context, refusal receives a perfect score, causing evidence-basedness to regress by −0.89.
• "FaithGate: A Proposed Architecture for Real-Time Hallucination Detection and Correction" (2026, preprint) — DOI: 10.2139/ssrn.7242858 — Lacks a mechanism to verify whether generated output follows from retrieved material (proposed architecture; empirical validation is future work).
• "Deterministic Citation Validation for Grounded Language-Model Systems" (2026, preprint) — DOI: 10.2139/ssrn.7322378 — Groundless assertions remain because the model writes the answer and source together.
• "No Pre-Trained Knowledge, No Hallucination" (2026, preprint) — DOI: 10.2139/ssrn.7333058 — Mitigation is not elimination. A design that returns original text via sentence-level indexing (no quantitative comparison).
• "Parameter-Efficient Contextual Calibration for Hallucination Mitigation in Domain-Specific LLM RAG" (2026) — DOI: 10.66917/ijaeic.a000015 — Problem statement that when evidence contradicts memory, memory is chosen. Own method's faithfulness 0.642→0.891.
• "DynamiRAG: Dynamic Memory RAG with Entropy-Aware Hallucination Prediction" (2026, preprint) — DOI: 10.2139/ssrn.6329665 — Hallucinations persist when evidence is noisy, contradictory, or insufficient. Exact match 11.5% (+0.5%p), contradiction detection 97.5%, one module drops in a small generator.
• "Evidence Coverage Evaluator" (2026, preprint) — DOI: 10.36227/techrxiv.176784505.56675782/v1 — A design declaration measuring whether each claim is supported independently of surface accuracy.
• "One Reflex, Two Answers: Model Judges Weakly Discriminate Localized from Distributed Causes" (2026) — DOI: 10.3390/make8090275 — Ten judges' margins are near 0 (0.16→0.34), but model-free deterministic rules separated the two causes.
• "Diagnosing Search Behavior and Failure Modes in Long-Horizon Search Agents" (2026, preprint) — arXiv:2608.01913 — Separates failures into search gaps and utilization gaps. Answer quality aligns with cumulative recall rather than search effort.
• "Disentangling Faithfulness Hallucinations in Retrieval-Augmented Generation" (2026) — DOI: 10.1007/s10994-026-07121-y — A benchmark separating retrieval errors from generation errors (limited to 2B–8B open-weight models).
• "Adaptive Query Performance Prediction for Retrieval-Augmented Generation" (2026) — DOI: 10.1145/3827605 — The correlation between predicted values and generation quality is 0.22–0.48, but adjusting search depth using that prediction improved quality.
• "The Impact of Query Decomposition and Cross-Encoder Reranking in Multi-Hop RAG" (2026) — DOI: 10.66245/jyi.v1.i1.002 — Diminishing returns in recall ceiling even when increasing candidates, peak accuracy 62% (small-scale student journal experiment).
• "The Role of Retrieval-Augmented Generation in Improving Factual Accuracy for Medical Large Language Models" (2026) — DOI: 10.70121/001c.158711 — Biomedical RAG survey. Notes the "lost in the middle" phenomenon within long contexts as a difficulty.
• "Measurement of retrieved chunk quality from real-world knowledge in RAG: A Phase 1 foundational study" (2026, preprint) — DOI: 10.64898/2026.01.01.26343326 — 594 original chunks showed a median across five metrics below 2.0 and under 20% passing threshold (intraclass correlation 0.936), while 6 rewritten chunks achieved perfect scores.
• "TriDomRAG: A Tri-Domain Benchmark for Evaluating RAG in Healthcare, Legal, and Finance" (2026, preprint) — DOI: 10.21203/rs.3.rs-9859942/v1 — Replacing synthetic legal data with a real corpus raised citation rates from near 0 to 0.678. Changing the retrieval method yielded +1.6%p.
• "Passage, Sentence, or Proposition? An Empirical Comparison of Retrieval Granularity Effects" (2025) — DOI: 10.66372/jger.v3i1.6 — Proposition granularity wins on recall by up to 6.8%p, but its advantage shrinks in multi-document questions, leaving no universal size.
• "Dynamic Knowledge Graph Augmentation Enhances Factual Accuracy in Retrieval Based Generation" (2026) — DOI: 10.54097/e4991d70 — Authors' own method. 18.2% reduction in hallucination rate compared to strong baseline, no absolute remainder provided.
• "Scalable Retrieval-Augmented Generation for Technical Documents with Hallucination Risk Control" (2025) — DOI: 10.66238/ijcbs29 — Authors' own method. 43% reduction in hallucination rate compared to standard baseline, no absolute baseline value provided.
• "Evaluating Retrieval-Augmented Generation Variants for Clinical Decision Support" (2025) — DOI: 10.3390/electronics14214227 — Twelve variants across 250 patient vignettes. Lowest hallucination rate 5.8%, top 5 relevance above 0.68.
• "FAIR-RAG: Faithful Adaptive Iterative Refinement for Retrieval-Augmented Generation" (2025, preprint) — arXiv:2510.22344 — Multi-document question F1 0.453 (+8.3 points). The authors note this as the highest performance among methods in this family.
• "Evaluating hallucination Mitigation in RAG through Guarded Response Strategies" (2026, preprint) — DOI: 10.2139/ssrn.6702638 — Forcing refusal when evidence is insufficient reduces claims, but creates a precision-refusal tradeoff.
• "Four-Axis Decision Alignment for Long-Horizon Enterprise AI Agents" (2026, preprint) — arXiv:2604.19457 — Measured withholding as one of four axes, but all six architectures answered every single case.
• "Causal-RAG: Causally-Augmented Retrieval for Clinical Decision Support in Low-Resource Settings" (2025, preprint) — DOI: 10.21203/rs.3.rs-8163345/v1 — Standard RAG showed a dangerous "yes" bias alongside high accuracy (prototype, own method).
• "Graph-Augmented Retrieval for Digital Evidence-Based Medical Synthesis" (2026, preprint) — DOI: 10.64898/2026.02.18.26346545 — An attempt to distinguish between retrieval deficits and the scarcity of evidence within the corpus itself (single-case proof of concept).
• "The Verification Depth Gradient: A Five-Stage Model of Citation Practice in the Age of AI" (2026, preprint) — DOI: 10.2139/ssrn.6318660 — The deterministic transition lies between stages 3 and 4, and that transition is invisible from the outside.