TTSR: Test-Time Self-Evolving via Reflection
A new method lets language models improve their own reasoning at test time through self-reflection, raising concerns about self-improving AI capabilities.
A new method lets language models improve their own reasoning at test time through self-reflection, raising concerns about self-improving AI capabilities.
Auto-scored: Self-improving reasoning at test time represents a real capability escalation with potential for misuse or loss of control.