An incorrect LLM answer tells you that output failed, but not whether behavioral encoding or recall failed. Google's WikiProfile study separates facts classified as not encoded from facts that show behavioral encoding evidence but remain hard to retrieve. The paper does not test a production diagnosis ladder; this article derives one for choosing among access probes, inference-time reasoning, external retrieval, and new training data.
Reading time: 9 minutes · About 1,800 words
TL;DR
- WikiProfile evaluates 2,150 Wikipedia-derived facts with ten tasks per fact across 13 LLMs and roughly 4.5 million responses.
- On this benchmark, GPT-5 and Gemini 3 Pro showed encoding evidence for 95–98% of scorable model-fact pairs; 26–34 percentage points were encoded but not directly recalled, and 11–12 points remained recall failures under the tested thinking condition.
- Recognition and encoding probes are behavioral evidence, not proof that a fact occupies a clean, readable location inside model weights.
- Among previously not-known facts, thinking recovered about 40–65% of those classified as encoded, but only 5–15% of those classified as not encoded.
- Treat a wrong answer as a routing problem: diagnose the bottleneck before paying for more context, more compute, retrieval, or retraining.
One wrong answer hides several failure modes
Conventional factuality benchmarks score a question as right or wrong. That is useful for comparing end performance, but weak for diagnosis. The same wrong answer can come from at least four different conditions:
- The relevant fact was not behaviorally encoded.
- The fact is accessible only when the query resembles its Wikipedia source-text context.
- The fact is recognizable when shown among choices but cannot be generated unaided.
- The fact becomes accessible only after additional inference-time computation.
These states demand different interventions. More pretraining data might help the first. A better query or thinking budget may help the second and fourth. External retrieval can bypass all four, but adds latency, cost, source quality, and citation work.
The bottleneck matters more than the size of the available toolbox. Optimizing a non-bottleneck creates expense without improving factual reliability.
How WikiProfile measures a fact instead of a question
The Google Research overview and paper define knowledge profiling at the level of facts. The released WikiProfile dataset contains 2,150 English Wikipedia facts that passed automated filtering and targeted manual quality validation. Each fact is paired with ten tasks:
- two encoding probes: proposition completion and a contextual question in a pre-training-like source-text setting;
- four closed-book knowledge questions: direct, direct-natural, reverse, and reverse-natural;
- four multiple-choice variants to probe recognition.
Candidate facts and questions were generated through an automated pipeline using Gemini 2.5 Pro with thinking, grounded in web search, and filtered for ambiguity. Targeted manual checks removed low-quality items. The researchers then collected roughly 4.5 million responses from 13 LLMs.
For each model, every question was sampled eight times at temperature 1. A model encodes a fact when either encoding probe scores above 0.5. It knows a fact only when all four closed-book knowledge questions score above 0.5. Recognition is analyzed separately and does not define knows.
The paper reports five main behavioral profiles. Rare cases that do not fit them are excluded, including fewer than 0.5% classified as not encoded but known without thinking.
| Profile | Operational meaning | Engineering implication |
|---|---|---|
| Encoding failure | Not encoded by the two probes and not known under the tested conditions | Investigate data coverage, model scale, or probe false negatives |
| Recall failure | Encoded, but still not known with thinking | Test other elicitation or post-training methods, or use external evidence |
| Direct recall | Encoded and known without thinking | Verify the resulting claim as usual |
| Recall with thinking | Encoded, not known without thinking, known with thinking | Test selective use of higher inference compute |
| Inference without encoding | Not encoded by the probes, but known with thinking | Check reasoning, guessing, and encoding-probe false negatives |
This taxonomy is more actionable than accuracy alone. It is still behavioral. A probe success does not identify a literal database record inside the weights, and a failed probe cannot prove metaphysical absence. It shows what a model exposes under controlled prompts.
Recall is the measured bottleneck on this benchmark
Across the tested frontier models, behavioral encoding approached saturation while recall remained incomplete. In Google's reported results, GPT-5 and Gemini 3 Pro showed encoding evidence for 95–98% of scorable model-fact pairs. Among all scorable pairs, about 26–34 percentage points were classified as encoded yet not directly recalled. After the tested thinking condition, 11–12 percentage points remained in the recall-failure profile.
Two systematic associations emerged: fact popularity and query direction.
Rare facts had encoding rates closer to popular facts than ordinary accuracy would suggest, while their recall rates were much worse. Reverse questions also produced a revealing split: models often failed to generate the reversed relationship, yet recognized the correct answer when it appeared among distractors.
That pattern reframes two familiar problems. Some long-tail failures may reflect poor access rather than absent storage. The reversal curse can also appear as an asymmetric retrieval problem: recognition survives even when free generation fails.
The conclusion should stay inside the study's boundary. WikiProfile uses validated English Wikipedia facts and a particular family of behavioral probes. It does not establish that all domain knowledge, recent events, private enterprise data, or procedural skills are already encoded in frontier models.
Thinking can act as recall assistance
Among facts that were previously not known, thinking-optimized models recovered roughly 40–65% of those classified as encoded. The same analysis recovered only 5–15% of previously not-known facts classified as not encoded.
This difference supports a useful mechanism-level hypothesis: extra inference often helps search for access paths to existing knowledge, instead of manufacturing missing facts from nothing. It also explains why indiscriminately enabling maximum reasoning is inefficient. Thinking costs tokens and latency, and a residual failure remains.
These findings motivate testing a router instead of using one universal setting. The paper does not establish a reliable trigger policy; deciding when to invoke thinking remains an open metacognitive problem.
A proposed four-level diagnosis for factual failures
The paper does not evaluate this production ladder. The following protocol is an engineering proposal derived from its behavioral distinctions.
Level 1: change the access path
Start with a low-cost paraphrase while keeping the target fact constant. WikiProfile found no significant overall effect between natural and highly rephrased wording after false-discovery-rate correction, so paraphrasing is an individual diagnostic probe rather than a proven general recovery method.
Then test the relationship in the opposite direction. This changes which entity must be generated, so it can reveal directional access asymmetry but does not itself recover or verify the original target answer.
Level 2: test recognition without treating it as proof
Provide plausible alternatives and ask the model to select one, then request a reason. Recognition after generation failure is evidence of an access gap. It is not enough for high-stakes use because distractor design can cue the answer and a model can guess.
Level 3: add bounded thinking
Allow a larger reasoning budget, decomposition, or self-check. Compare whether the answer changes and whether independent evidence supports the change. A longer chain is a recovery attempt, not a citation.
Level 4: retrieve and verify external evidence
For changing, consequential, or disputed facts, query an authoritative source and preserve the URL, date, and relevant passage. Retrieval moves the task from parametric recall to evidence-backed generation. It also creates a new failure surface: stale documents, poor ranking, prompt injection, and incorrect synthesis.
If retrieval still fails, classify the answer as unresolved. A reliable system needs an abstention state.
Build an evidence trace, not a confidence theater
A factual response should carry an observable route:
claim: "..."
initial_answer: "..."
probe_results:
paraphrase_consistent: false
reverse_query_consistent: false
recognition_success: true
thinking_recovered: true
external_evidence:
required: true
sources: ["authoritative URL"]
final_status: verified | conflicting | unresolved
This trace is an engineering proposal for debugging the access path; it does not verify a claim by itself. Model access and claim truth are separate layers. The AI advice confidence trap covers human confidence and abstention, while latent workspace versus external memory covers the state-lifetime boundary.
What to measure in an LLM factuality evaluation
Aggregate accuracy remains useful, but a diagnostic evaluation should add:
- paraphrase and query-direction consistency;
- free recall versus recognition;
- recovery rate from bounded thinking;
- cost and latency per recovered fact;
- source-backed verification rate;
- false recovery rate, where thinking changes a wrong answer into another wrong answer;
- abstention quality when evidence remains unavailable.
The important unit is the complete route from question to verified claim. A benchmark that stops at first-token correctness cannot tell you which intervention will improve the deployed system.
FAQ
Does WikiProfile prove that LLMs store almost every fact they see?
No. GPT-5 and Gemini 3 Pro met the paper's encoding criterion for 95–98% of scorable model-fact pairs: either proposition completion or contextual questioning scored above the threshold. The result is specific to the benchmark, models, sampling, and probes; it is not a claim about every fact in training data.
Is recognition evidence that a model knows a fact?
Recognition is one useful access probe. It can distinguish some generation failures from total inaccessibility, but answer choices can cue or enable guessing. High-stakes claims still need independent evidence.
Should every factual query use maximum reasoning?
No. Thinking recovered many encoded-but-inaccessible facts in the study, but it adds compute and did not eliminate failures. Use routing rules based on task risk and cheaper diagnostic signals.
Does retrieval-augmented generation solve the recall bottleneck?
It bypasses parametric recall by supplying external text. Reliability then depends on source selection, freshness, injection resistance, synthesis, and citation verification.
The next action
Take a factual evaluation set you already use and add three variants for each claim: a paraphrase, a reversed query, and a recognition probe. Measure what bounded thinking recovers, then require authoritative retrieval for the remaining high-risk failures. This turns a generic wrong-answer rate into a map of the actual bottlenecks.
References and further reading
- Google Research, Empty shelves or lost keys? Recall is the bottleneck for parametric factuality
- Calderon et al., Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality
- Google, WikiProfile dataset