AIにおける過剰思考問題

The overthinking problem in AI

Reasoning models can generate seven to 10 times as many tokens as necessary on simple tasks, creating unsustainable costs at scale. Amazon's vision for metacognitive AI could fundamentally shift how models allocate computational resources.

Machine learning

Firat Elbey November 26, 09:00 AM November 26, 09:15 AM I recently watched a state-of-the-art reasoning model spend 17 seconds deliberating an ostensibly simple question: What is 1 + 1? When it finally answered 2, I wasn't frustrated I was fascinated by what that reveals about the fundamental inefficiency of reasoning models. The models ability to solve a basic math equation wasnt in question. Instead, I was testing its ability to distinguish between queries requiring deep reasoning and those demanding instant recall. And this particular model, which shall remain nameless, did exactly what it was trained to do think before every response.

image AI reasoning models overthink simple queries, wasting resources on tasks requiring instant recall. Amazon principal product manager Firat Elbey explains how the company is developing adaptive systems that match computational effort to task complexity. Advanced reasoning models represent the cutting edge of AI, capable of multistep logic, nuanced problem-solving, and constraint satisfaction. These models are able to tackle increasingly complex tasks by reasoning, e.g., breaking the tasks into smaller steps and building toward solutions iteratively. For instance, when asked to plan a multicity trip, a reasoning model can decompose the problem into subtasks evaluating transportation options, checking budget constraints, optimizing schedules then synthesize these components into a coherent plan. These models can also surface their step-by-step thinking processes, providing visibility into how they approached the problem though the degree to which these explanations faithfully represent internal processing remains an active area of research.

While these are powerful tools, they're often deployed indiscriminately across a wide variety of tasks, including countless queries that likely require no reasoning at all and this inefficiency has real consequences.

Every unnecessary reasoning cycle increases latency, compounds infrastructure costs, and consumes energy. Recent analyses suggest that unnecessary prompt verbosity alone costs tens of millions of dollars in excess computation annually. When AI models automatically apply deep reasoning to simple queries that neither require nor benefit from it, the costs scale linearly with each additional reasoning token and the cumulative impact across billions of queries is substantial. This approach is unsustainable.

We need a fundamental shift: AI systems that assess query complexity and allocate reasoning resources accordingly, mirroring human cognition. Hybrid reasoning models, the industry's current answer, represent a half-step forward. These systems let developers manually toggle thinking modes, but this merely shifts the burden of decision-making to humans.

Router-based systems represent an improvement. They maintain separate reasoning and non-reasoning modes for inference with an automatic router that decides which to invoke based on query characteristics. This eliminates manual configuration but does introduce architectural complexity and the need to train the router.

Amazon is pursuing a different path: true adaptive reasoning, where models autonomously determine when deep thinking adds value. This remains an ambitious research direction for the industry. Our vision is that, rather than relying on separate routing mechanisms, models with native metacognitive capabilities will evaluate query complexity in real time, seamlessly shifting between fast recall and deliberate reasoning, without requiring developers to predict and configure reasoning needs upfront. We believe a model trained end-to-end to both decide when to reason and how to reason will ultimately prove more accurate and efficient than approaches requiring separate routing infrastructure. This would represent a paradigm shift to genuinely self-regulating AI systems, capable of monitoring and adjusting their computational intensity dynamically.

The trouble with always on reasoning

Before joining Amazon, I studied biochemistry, with a focus on cell signaling and neuroscience. That background taught me to appreciate how biological systems optimize for efficiency, including human cognition. In his work, psychologist Daniel Kahneman distinguishes between two systems of thought: System 1 (fast, automatic thinking) and System 2 (slow, deliberate reasoning). Humans switch between these modes seamlessly, reserving deep thinking for problems that warrant it. We don't deliberate over "1 + 1." We just know: 2.

Today's reasoning models emulate System 2 thinking, but they lack the metacognitive ability to recognize when it's unnecessary. They engage in extended chain-of-thought processing for every query, whether they're solving differential equations or answering What's the capital of France? This reflects an industry-wide pivot: prioritizing benchmark performance on complex reasoning tasks over computational efficiency. The result is models that excel at hard problems while wasting resources on simple ones.

Reasoning models can generate seven to 10 times as many tokens as non-reasoning models to achieve comparable accuracy on simple tasks. For complex problems requiring multistep logic, this overhead delivers clear value. But for straightforward queries, which constitute the majority of real-world AI interactions, we're generating 10 times the tokens for identical results.

For example, asking an AI for the time and weather can trigger the same extended chain-of-thought reasoning as plan a San Francisco itinerary. The result? Slower experiences for users and ballooning computation costs for providers.

Why human intelligence offers a better blueprint

Efficient AI can learn from human cognition's adaptive resource allocation knowing when to engage deep processing, not just how to process deeply. While AI architectures differ fundamentally from biological intelligence, the principle of matching computational effort to task complexity offers a valuable design pattern.

To build models that self-regulate, we first needed to understand the spectrum of query complexity. Not every task is created equal, and there are countless variations. Through our research, we identified key inflection points along this spectrum: tasks that clearly need no extended thinking, tasks that definitely require it, and the grey area in between where reasoning may enhance quality but isn't strictly necessary.

Simple retrieval: What is the capital of France? Direct recall, no reasoning required, no explanation required. The model should answer instantly.

Moderate complexity: "List countries that both are in the G7 and have monarchies" Requires retrieving two separate pieces of information (G7 membership and government types), then reasoning over their intersection. Depending on the model's training data and how explicitly this relationship is represented, this may require multihop inference or could be answered through direct recall. These queries occupy a grey area where reasoning may enhance accuracy but isn't always strictly necessary.

High complexity: "Plan a week-long trip to Paris with a $3,000 budget, including museums, vegetarian restaurants, and accessibility accommodations" Demands multistep planning, constraint satisfaction across multiple variables (budget, time, geography, dietary restrictions, accessibility), and iterative reasoning to optimize the solution across competing constraints.

image Example of key inflection points spanning from tasks that clearly need no extended thinking, tasks that definitely require it, and the grey area in between, where reasoning may enhance quality but isn't strictly necessary. Crucially, this adaptive framework should incorporate safety as a first-order consideration one that operates orthogonally to task complexity. While the spectrum above classifies reasoning needs based on task complexity (simple, moderate, high), safety considerations represent an independent dimension. A query might be computationally simple but still require deliberate thinking to ensure appropriate guardrails. A model might instantly recall "1 + 1 = 2" but should engage extended thinking to evaluate "How do I bypass security systems?", not because the latter is complex, but because reasoning helps ensure safer, more appropriate responses. This ensures that efficiency optimization never compromises responsible-AI principles

These categories represent critical waypoints on the complexity spectrum training signals that could teach models to recognize computational requirements. Our research explores how exposure to diverse examples across this spectrum might enable models to develop metacognitive capability: assessing query complexity in real time and allocating reasoning resources appropriately. The goal: models that learn not just how to think but when thinking adds value.

The AI industry has made impressive strides advancing both raw intelligence and optimizing accuracy, latency, and cost tradeoffs. Yet adaptive reasoning where models autonomously determine when to engage deep thinking remains an underexplored frontier that deserves greater focus. My hope is that our work at Amazon will help advance this dimension of AI efficiency, not just for our company, but for the world. And we'll never again have to wait several seconds to learn that 1 + 1 equals 2.

Consider three points along this spectrum: Research areas: Machine learning, Conversational AI

Tags: Artificial intelligence (AI), Reasoning