Researchers from ByteDance have uncovered a surprising capability within large reasoning models: they often know exactly when they have finished solving a problem, yet our current systems force them to keep "thinking." This persistent internal dialogue, known as a Long Chain of Thought (CoT), frequently leads to redundant computations that delay answers without actually improving accuracy. In many cases, these extended reasoning paths can even introduce new errors, clouding a correct initial insight with unnecessary complexity.

To address this, the team introduced SAGE (Self-Aware Guided Efficient Reasoning), a sampling paradigm designed to unlock this latent self-awareness. By allowing a model to recognize its own success, SAGE eliminates the "chatter" that typically plagues complex inference tasks. This isn't just about saving time; it's about refining the logic of the model itself.

The researchers further enhanced this approach by integrating it into a reinforcement learning framework called SAGE-RL. This method trains the model to internalize these efficient reasoning patterns during standard inference (pass@1). The results are striking, showing significant gains in both speed and mathematical precision across several rigorous benchmarks. By teaching models to stop when they are ahead, ByteDance is paving the way for faster, more reliable AI assistants.

Researchers from ByteDance have uncovered a surprising capability within large reasoning models: they often know exactly when they have finished solving a problem, yet our current systems force them to keep "thinking." This persistent internal dialogue, known as a Long Chain of Thought (CoT), frequently leads to redundant computations that delay answers without actually improving accuracy. In many cases, these extended reasoning paths can even introduce new errors, clouding a correct initial insight with unnecessary complexity.

To address this, the team introduced SAGE (Self-Aware Guided Efficient Reasoning), a sampling paradigm designed to unlock this latent self-awareness. By allowing a model to recognize its own success, SAGE eliminates the "chatter" that typically plagues complex inference tasks. This isn't just about saving time; it's about refining the logic of the model itself.

The researchers further enhanced this approach by integrating it into a reinforcement learning framework called SAGE-RL. This method trains the model to internalize these efficient reasoning patterns during standard inference (pass@1). The results are striking, showing significant gains in both speed and mathematical precision across several rigorous benchmarks. By teaching models to stop when they are ahead, ByteDance is paving the way for faster, more reliable AI assistants.