Artificial intelligence has made significant progress through Large Language Models (LLM) and its advanced peers, Large Inference Models (LRMS), redefines how machines process and generate human-like texts. These models can write papers, answer questions, and even solve mathematical problems. However, despite its impressive capabilities, these models exhibit strange behavior: they often over-solve simple problems when dealing with complex problems. A recent study by Apple researchers provides valuable insight into the phenomenon. This article explores why LLM and LRMS act in this way and what it means for the future of AI.
Understand LLM and LRM
To understand why LLM and LRM behave, we first need to clarify what these models are. LLMS (such as GPT-3 or BERT) is trained on a huge text dataset to predict the next word of the sequence. This makes them excellent in tasks such as text generation, translation and summary. However, they are not inherent designs for reasoning, they involve logical inference or problem solving.
LRMS is a new model designed to address this gap. They combine techniques such as the Chain of Thinking (COT) prompts where intermediate reasoning steps are generated before providing the final answer. For example, when solving mathematical problems, LRM may break it down into steps like humans. As Apple Research shows, this approach can improve the performance of complex tasks, but faces challenges when dealing with various complexity issues.
Research and research
The Apple research team has adopted another approach to assess the reasoning capabilities of LLM and LRMS. Instead of relying on traditional benchmarks such as math or coding tests, which may be affected by data contamination (the model remembers the answer), they create a controlled puzzle environment. These include famous puzzles such as Hanoi Tower, pawn jumping, rivers and neighborhood worlds. For example, Hanoi Tower involves moving disks between nails by specific rules, and complexity increases as more disks are added. By systematically adjusting the complexity of these puzzles while maintaining a consistent logical structure, researchers can observe the performance of models in various difficulties. This approach allows them to analyze not only the final answer but also the reasoning process that provides a deeper understanding of the “thinking” of these models.
Discovery about overthinking and giving up
The study identified three different performance regimes based on problem complexity:
- At low complexity levels, standard LLMs generally perform better than LRMs because LRMS tend to overthink, resulting in additional steps, while standard LLMs are more efficient.
- For moderate complexity issues, LRMS shows excellent performance due to its ability to generate detailed inference traces, thus helping them effectively solve these challenges.
- Both LLM and LRM fail completely for high complexity issues. LRM in particular, despite the increased difficulty, the accuracy of LRM has completely collapsed and reduced the inference work.
For simple puzzles, such as Hanoi Tower with one or two disks, standard LLMS provides the right answer more efficiently. However, LRMS often overthink these issues, and even if the solution is simple, it can produce lengthy traces of reasoning. This suggests that LRM may mimic the exaggerated explanations in its training data, which may lead to inefficiency.
LRMS performs better in moderately complex situations. Their ability to generate detailed reasoning steps enables them to solve problems that require multiple logical steps. This allows them to outperform standard LLM, which is difficult to maintain coherence.
However, for highly complex puzzles, such as Hanoi Tower with many disks, both models completely failed. Surprisingly, despite sufficient computing resources, LRMS is more complex than a certain extent, thus reducing inference work. This “give up” behavior indicates that there are basic limitations in its ability to extend its reasoning ability.
Why does this happen
Overthinking simple puzzles may stem from the way LLM and LRM are trained. These models are learned from a large set of data that includes simplicity and detailed description. For simple questions, they may generate lengthy inference trajectories by default, even if the direct answer is sufficient, mimicking the lengthy examples in the training data. This behavior is not necessarily a flaw, but a reflection of their training, which prioritizes reasoning over efficiency.
The failure of complex problems reflects the inability of LLM and LRM to learn to generalize logical rules. As the complexity of the problem increases, their dependence on pattern matching can be broken, resulting in inference inconsistency and performance crashes. The study found that LRM cannot use clear algorithms and causes inconsistent among different puzzles. This emphasizes that these models can simulate reasoning, but they do not really understand the basic logic of human beings.
Diverse perspectives
The study sparked discussions in the AI community. Some experts believe that these findings may be misunderstood. They suggest that while LLM and LRM may not have the same reason as humans, they still show effective problem solving within certain complexity limitations. They stress that “reasoning” in AI does not need to reflect human cognition in order to make it valuable. Similarly, discussions about platforms such as Hacker News praise the rigorous approach to the study, but highlight the need for further research to improve AI reasoning. These views emphasize the ongoing debate about the reasoning that constitutes in AI and how we evaluate it.
Meaning and direction of the future
The findings of this study are of great significance to the development of AI. Although LRM represents progress in mimicking human reasoning, their limitations in dealing with complex problems and extending inference work suggest that current models are far from implementing generalizable inference. This highlights the need for a new evaluation method that focuses on the quality and adaptability of the reasoning process, not just the accuracy of the final answer.
Future research should aim to enhance the model’s ability to perform logical steps accurately and adjust the reasoning work according to the complexity of the problem. Developing benchmarks that reflect real-world reasoning tasks, such as medical diagnosis or legal argumentation, can provide more meaningful insights into AI capabilities. In addition, solving the over-reliance of the model on pattern recognition and improving its ability to generalize logical rules is crucial to advancing AI reasoning.
Bottom line
This study conducted a critical analysis of the reasoning ability of LLM and LRM. It shows that while these models over-analyze simple puzzles, they struggle with more complex puzzles, revealing their advantages and limitations. Although they perform well in some cases, they are unable to solve highly complex problems, highlighting the gap between simulated reasoning and real understanding. The study highlights the need to develop an AI system that can adapt to reasoning on a variety of complexities, so that it can solve problems of different complexities like humans.
