AI Models Can Recall Forgotten Facts by Thinking Longer
Research shows large language models can recover up to 65% of unrecalled facts through extended reasoning processes.
1The Recall Problem
When large language models (LLMs) make errors, developers typically assume missing knowledge. However, research from Google and Technion reveals models encode 95-98% of facts but fail to access them properly. This indicates recall mechanisms, not storage, are the primary bottleneck. Improving access to existing knowledge may be more effective than simply increasing model size or training data.
2Knowledge Profiling Approach
Researchers developed a novel evaluation method called 'knowledge profiling' that goes beyond traditional accuracy metrics. This framework measures whether facts are encoded, their accessibility through different phrasings, and the computational effort required for retrieval. It distinguishes between encoding failures (needing more training) and recall failures (needing better access methods).
3Five Knowledge Profiles
The study categorizes model knowledge processing into five types: Direct recall, encoding failure (empty shelves), recall failure (lost keys), recall with thinking, and inference without encoding. For example, a model failing to directly answer where Oasis played their first gig could deduce it by chaining facts about Manchester music venues - similar to human tip-of-the-tongue experiences.
4Limits of Model Scaling
Findings challenge the assumption that bigger models always perform better. When scaling Gemma3 from 1B to 27B parameters, encoding failures dropped from 85% to 23%, but recall failures peaked at 40%. This shows scaling mainly addresses storage, not access problems. Larger models accumulate more knowledge but struggle to retrieve it efficiently.
5Recovery Mechanisms
The study tested inference-time computation techniques like chain-of-thought. Frontier models (GPT-5, Gemini-3) recovered 40-65% of initially unrecalled facts through these methods. This mirrors human memory retrieval where deliberate effort (like retracing context) helps recall seemingly forgotten information.
6Long-Tail Knowledge Gap
A significant recall disparity exists between popular and obscure facts, exceeding 25% in frontier models. While encoding rates are similar, rare facts face substantially worse recall performance. This reflects training data distribution biases and highlights the need for specialized query optimization techniques.
7Query Style Impact
How users phrase questions directly affects model performance. Recall degrades when queries diverge from training patterns. For example, reverse questions ('What has Y as X?' vs 'What is X of Y?') significantly reduce accuracy. This has crucial implications for user interface design and query formulation.
8Engineering Implications
These findings suggest shifting engineering priorities from model size to inference-time techniques that improve access to existing knowledge. Methods like chain-of-thought, self-questioning, and contextual priming may unlock models' full potential more efficiently than simply adding parameters or training data.
Original article: VentureBeat — Yapay Zeka Modelleri Unuttuklarını Düşünerek Hatırlayabilir