When a Large Language Model (LLM) suddenly masters complex mathematical reasoning through Reinforcement Learning (RL) or begins to "think" for longer periods at test time, it appears to be acquiring new intelligence. However, a growing body of research suggests that these post-training improvements are not creating new knowledge. Instead, they act as "distribution sharpeners" that help the model more efficiently navigate the latent knowledge it already acquired during pre-training. This creates a fundamental pre-training ceiling: a model's ultimate reasoning and factual capacity is strictly bounded by its initial training data.
Reinforcement Learning as a Sampling Optimizer
Recent breakthroughs in Reinforcement Learning with Verifiable Rewards (RLVR) have significantly boosted model performance on benchmarks like GSM8K and AIME. While this looks like the model is learning to reason, Yue et al. argue that RL primarily acts as a "sampling efficiency" booster. In this view, RL does not expand the model's fundamental reasoning boundary; rather, it makes the model much more likely to find a correct answer on the first try—an answer that the base model might have only found once in a thousand random samples.
This dependency means that if a base model lacks the latent capability to solve a task, no amount of reinforcement can produce a correct trajectory to reward. As Alam and Rastogi observe, RL often functions by re-weighting existing internal knowledge rather than teaching entirely new algorithms. Consequently, the model's performance is capped by the "reasoning potential" established during pre-training, a concept Zhang et al. define as the inverse of the attempts needed to solve a problem.
The Factual Floor of Latent Knowledge
The ceiling applies not just to logic, but to facts. Techniques designed to "steer" models toward specific semantic categories or to extract "latent knowledge" without supervision are ultimately tethered to the model's weights. Saeed and Papotti demonstrate that while type embeddings can force a model to predict a "City" instead of a "Year," they cannot induce the model to generate a correct city name if that fact was never learned during pre-training.
Similarly, unsupervised methods for discovering what a model "believes" to be true, such as those proposed by Burns et al., are constrained by the coherence of the model's internal representations. If the model lacks a factual foundation for a topic, consistency-based checks will fail. This suggests that the "truth" we extract from models is only as good as the data they were fed at the very beginning.
The Limits of "Thinking Harder"
If a model is struggling, can we simply let it deliberate longer? "Test-Time Scaling" (TTS) and "Chain-of-Thought" (CoT) reasoning allow models to explore multiple paths to a solution. However, Wang and Wang note that increased deliberation time cannot compensate for a fundamental lack of factual information or unknown theorems. If the necessary information is missing, extended reasoning processes will simply deliberate over incorrect premises.
This limitation extends to advanced decoding strategies. Methods like DoLa, which contrast different layers of a model to amplify factual signals, are still limited by the pre-existing knowledge base. As the authors of DoLa and Zhou et al. point out, while these methods can suppress "noise" from shallow layers to refine the selection of information, they cannot synthesize facts that were never present in the training corpus.
Can the Ceiling be Broken?
There is an ongoing debate about whether this ceiling is truly impenetrable. Liu et al. propose "Prolonged Reinforcement Learning" (ProRL), arguing that with periodic reference policy resets, models can eventually discover novel reasoning strategies that their base models could not crack even with thousands of attempts.
However, most current evidence suggests that for specialized domains—such as medical billing (Wang et al.) or psychological explanation (Zhu et al.)—the ability to reason effectively remains a function of the underlying knowledge base. As Yuan et al. argue, expanding the reasoning boundary likely requires external supervision or new data, as RL alone can only optimize what is already there. The "Aha!" moments we see in modern reasoning models may be less about the birth of new intelligence and more about the successful unveiling of latent potential.
Key Foundational Work
Discovering Latent Knowledge in Language Models Without Supervision (2023) — contributed 1 concept elements to this direction.
Skills-in-Context Prompting: Unlocking Compositionality in Large Language Models (2023) — contributed 1 concept elements to this direction.
You Are My Type! Type Embeddings for Pre-trained Language Models (2022) — contributed 1 concept elements to this direction.