Infinitely Further: Tool-Use Unlocks Length Generalization for State Space Models

State Space Models (SSMs) have emerged as an alternative to Transformers for long-context and long-form generation thanks to their fixed-size memory and linear computational complexity, but the authors first present a theoretical result showing that SSMs cannot exactly solve any “truly long-form generation” problem. In other words, they show that length scalability—widely regarded as SSMs' core strength—can be fundamentally limited.

This limitation is alleviated when external tool access is allowed. The authors argue that, given appropriate tool design and problem-specific training data, SSMs can solve any tractable problem and generalize even as problem length and complexity grow.

Building on this theoretical result, the authors also demonstrate experimentally that tool-augmented SSMs show strong length generalization across multiple tasks.

• Arithmetic tasks
• Reasoning tasks
• Coding tasks

The key message is that while SSMs alone have inherent constraints on long-form generation, they can become an efficient alternative in interactive tool-based or agentic settings. Rather than completely replacing Transformers, SSMs combined with external tools open up a new direction for tackling long-horizon problems.