How longer thinking boosts factual recall in frontier LLMs

Longer generation can recover up to 65% of facts that a model initially fails to recall, because the model continues to search its internal knowledge during extra token steps. This effect, demonstrated on GPT-5 and Gemini-3, shows that many hallucinations are not due to missing data but to premature stopping.

What is "longer thinking" in large language models and how does it work?

In the context of LLMs, "longer thinking" refers to allowing the model to generate more tokens before producing a final answer. Instead of stopping after a short response, the model keeps producing text, effectively giving itself more internal steps to retrieve and reason over stored knowledge.

The mechanism relies on the transformer architecture’s attention layers, which can re‑examine earlier context as new tokens are added. Each additional token provides a fresh opportunity for the model to surface a fact that was encoded parametrically but not yet expressed.

Researchers at Google Research and Technion demonstrated this by prompting frontier models to continue generating after an initial answer. The models often corrected themselves or added missing details, indicating that the information was already present in the weights.

Because the process does not require external retrieval systems, it is a purely architectural advantage: the model’s own latent memory is tapped more thoroughly when given time.

Why does longer thinking matter for factual accuracy?

Hallucinations—confidently wrong statements—are a major obstacle for deploying LLMs in critical applications. Traditionally, developers assume the model lacks the necessary fact and respond by scaling up data or adding retrieval‑augmented pipelines.

The new findings suggest a different diagnosis: the fact may already be stored, but the model stops too early. By extending the generation window, a large proportion of these “missing‑knowledge” errors can be corrected without any architectural changes.

This matters for enterprises that need reliable outputs without incurring the cost of additional retrieval infrastructure. A simple prompting tweak—asking the model to think longer—can raise factual recall rates dramatically, as shown by the up‑to‑65% recovery rate reported in the VentureBeat article.

Moreover, the approach aligns with emerging best practices like chain‑of‑thought prompting, where models are guided to reason step‑by‑step. Longer thinking is a natural extension of that philosophy, giving the model more room to articulate its reasoning.

What evidence shows longer thinking improves recall?

The evidence comes from a study highlighted by VentureBeat, which examined frontier models such as GPT‑5 and Gemini‑3. The researchers first asked the models factual questions and recorded which answers were correct.

When the same questions were re‑asked with a directive to continue generating for additional tokens, the models recovered up to 65% of the facts they previously missed. The study also measured that these models internally encode 95‑98% of tested facts, confirming that the knowledge is present but not always accessed.

Crucially, the recovery was achieved without any changes to model weights, data, or external tools—only by extending the generation length. This demonstrates that the limitation is procedural rather than architectural.

These results contrast with the conventional response of increasing model size or adding retrieval modules, which can be costly and complex. The study’s authors argue that developers should first explore longer thinking as a low‑cost mitigation strategy.

What are the implications and future directions?

If longer thinking can routinely retrieve hidden facts, it may reshape how developers design prompts and evaluation pipelines. Prompt engineers might include explicit “think longer” cues or set higher token limits for critical queries.

However, longer generation also raises concerns about latency and token cost, especially for commercial APIs. Balancing accuracy gains against increased compute will be a key operational decision.

Future research is likely to explore adaptive stopping criteria, where the model decides in real time whether additional tokens are needed to improve confidence. Integrating dynamic token budgeting could deliver the best of both worlds: higher accuracy without unnecessary expense.

Finally, the findings encourage a re‑examination of hallucination diagnostics. Instead of assuming missing knowledge, practitioners may first test whether a longer generation resolves the error, reserving more heavyweight solutions for truly unknown facts.

Frequently asked questions

How many extra tokens should I add to improve factual recall?

There is no one‑size‑fits‑all answer; experiments often start with an additional 20‑30 tokens and adjust based on response quality and cost considerations.

Does longer thinking work for all types of questions?

It is most effective for factual queries where the answer exists in the model’s parameters. Complex reasoning or opinion‑based prompts may see diminishing returns.

Will longer generation increase the risk of new hallucinations?

Generally, extending generation reduces early‑stop hallucinations, but it can also introduce new errors if the model wanders off topic. Careful prompt design mitigates this risk.

Can I use longer thinking with retrieval‑augmented models?

Yes; combining retrieval with extended generation can further boost accuracy, as the model can cross‑verify retrieved facts while still leveraging its internal memory.

The bottom line

• Longer thinking can recover up to 65% of missed facts without extra data or retrieval systems.
• Frontier models already encode 95‑98% of factual knowledge; the bottleneck is often premature stopping.
• Implementing a simple token‑extension or “think longer” prompt is a low‑cost way to improve accuracy.
• Developers must balance the accuracy gains against higher latency and token costs.
• Future work will focus on adaptive stopping and hybrid approaches that combine internal recall with external retrieval.