Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
The standard diagnosis when an AI model gets a fact wrong is that the model never learned it. Engineering teams treat hallucinations as a knowledge gap, and the standard prescription follows logically: feed the model more data, scale it to more parameters, or bolt on a retrieval system to fetch facts from an external database. A new study from researchers at Google Research and Technion suggests this diagnosis is frequently wrong, and that the expensive remedies it leads to are often solving the wrong problem entirely.
The research, which tested 13 large language models across more than 4 million responses, finds that frontier models like GPT-5 and Gemini-3 successfully encode 95 to 98 percent of tested facts in their parameters. The knowledge is physically present inside the model. What breaks down is the model’s ability to surface that knowledge on demand. In other words, the problem is not empty shelves. It is lost keys.
The difference between storing a fact and being able to use it
The researchers introduce a framework they call knowledge profiling, which shifts the evaluation lens from simple question-level accuracy to something more granular. Rather than scoring whether a model answers a given prompt correctly or not, knowledge profiling tests a single underlying fact across multiple conditions: different phrasings, different query directions, and varying levels of computational effort. This allows the researchers to distinguish between two failure modes that standard accuracy benchmarks collapse into one indistinguishable error.
The first failure mode is an encoding failure, which the paper calls empty shelves. The model genuinely never learned the fact. It cannot reproduce it even when primed with its original training context. This is the failure mode most developers assume they are dealing with, and it genuinely calls for more pre-training data or a larger model.
The second failure mode is a recall failure, which the paper calls lost keys. The model has the fact encoded in its parameters and can reproduce it when given the original training context, but it cannot retrieve that fact in response to a naturally phrased question. The knowledge exists but is inaccessible through normal generation.
The paper illustrates this with a concrete example: the fact that Oasis played their first gig at the Boardwalk club. A model suffering a recall failure might complete a Wikipedia-style sentence about Oasis’s early days correctly, yet fail to answer the direct question “Where did Oasis play their first show?” even when given time to think. The same model, presented with that question in multiple-choice format, might select the correct answer immediately, because recognition is easier than generation.
“Encoding and recall failures are indistinguishable under accuracy metrics, yet they imply different limitations and solutions,” the researchers write. “Encoding failures call for pre-training interventions, such as scaling model size or data coverage. Recall failures suggest post-training interventions that often improve how models utilize what they already encode.”
Scaling solves storage but not access
One of the study’s most counterintuitive findings concerns what happens when you scale a model up. The researchers tracked the Gemma3 model family from 1 billion to 27 billion parameters. Scaling dramatically reduced encoding failures, dropping them from 85 percent to 23 percent. The model memorised vastly more facts. But the share of recall failures simultaneously increased, peaking at 40 percent without thinking enabled. As the model’s knowledge base expanded, a larger pool of encoded facts became trapped in an inaccessible state.
This means that scaling primarily solves the storage problem while leaving the access problem largely untouched, and in some respects making it worse. Companies that respond to hallucinations by fine-tuning larger models are, in many cases, spending heavily to fix something that was never broken while the actual bottleneck remains unaddressed.
“When facts come out wrong, the go-to move is to scale, meaning train a larger model or add more data,” Nitay Calderon, Research Scientist at Google, told VentureBeat. “Both are expensive, and if the facts are already encoded, neither helps.”
The same logic applies to Retrieval-Augmented Generation, the technique of connecting a model to an external vector database so it can look up facts rather than recall them from parameters. RAG is genuinely the right tool when the required information is recent, proprietary, or simply absent from the model’s training data. But deploying it as a blanket response to hallucinations adds latency and per-call cost for every fact the model could already recall if prompted correctly. “A lot of what teams solve with RAG are facts the model can already answer from memory,” Calderon said.
Thinking as a recovery mechanism, and its limits
The study’s most practically useful finding is that inference-time computation, specifically giving a model space to reason before answering, acts as a meaningful recovery mechanism for recall failures. When models were given the chance to generate intermediate reasoning steps before producing a final answer, they successfully retrieved 40 to 65 percent of the encoded facts they had initially failed to recall directly. The researchers compare this to the human tip-of-the-tongue experience, where deliberately retracing context eventually surfaces a memory that felt completely inaccessible moments earlier.
The mechanism works because generating relevant intermediate thoughts structurally primes the model to locate stored information. A model that fails to answer “Where did Oasis play their first gig?” directly might succeed after generating a reasoning chain about the band’s early history in Manchester, because that chain creates the contextual conditions under which the fact was originally encoded.
The practical catch is that only 10 to 20 percent of facts actually require this extended reasoning. Enabling thinking globally wastes compute budget on the 80 to 90 percent of queries the model handles fine without it. The challenge is routing selectively, and here the researchers are candid about a genuine limitation: today’s models lack reliable self-awareness about when they are about to fail. “To use the compute well, the model has to sense ahead of time that a plain answer is about to fail, so it can escalate before answering,” Calderon said. “That self-awareness is its own skill, and today’s models aren’t reliably good at it.” Google researchers are developing frameworks around what they call faithful uncertainty to address this metacognitive gap.
The study also surfaces two specific recall patterns that developers should know about. Rare facts are encoded at rates comparable to popular facts, but the recall gap between obscure and well-known facts exceeds 25 percent for frontier models, meaning long-tail knowledge is disproportionately locked away. Separately, models struggle badly with reverse queries. A model might easily answer that Oasis played their first gig at the Boardwalk, but fail to answer who played their first gig at the Boardwalk. The same model often selects the correct answer when given the reverse question in multiple-choice format, confirming the fact is encoded and recognisable but not generatable from that direction.
What this means for teams building on top of AI
For developers and product teams in Malaysia and Singapore building enterprise applications on top of frontier models, the practical implications are significant. The knowledge profiling framework is available as a benchmark called WikiProfile on Hugging Face, and the researchers have published all the prompts used to construct it, meaning teams can adapt the pipeline to evaluate their own internal corpora. Fully profiling a frontier model on the WikiProfile suite costs approximately USD 500, with meaningful cost reductions available by omitting multiple-choice variants or reducing response samples per question.
The researchers recommend four concrete shifts in how teams approach factual reliability. First, distinguish encoding failures from recall failures before reaching for RAG or a larger model. Second, use inference-time reasoning selectively rather than globally, and invest in routing logic that identifies which queries genuinely need it. Third, build generate-then-verify pipelines that exploit the asymmetry between generation and recognition, prompting the model to reflect on and check its own outputs after generating them. Fourth, evaluate models by probing the same underlying fact across multiple phrasings and directions rather than relying on single-prompt accuracy scores, which mask the actual shape of what a model can and cannot access.
The deeper implication of this research is a rebalancing of where AI investment creates returns. Pre-training at frontier scale is the exclusive domain of a handful of companies with capital measured in billions. But the levers this study highlights, including post-training adjustments, inference-time reasoning, and smarter evaluation, are accessible to any team working with commercially available models. “For companies that don’t build models from scratch, this is good news,” Calderon said. “Pre-training is hugely expensive and out of reach for most, but the levers that matter now are not.” The hallucination problem, it turns out, may be less about what models were taught and more about how well we have learned to ask.
