TTCS: Test-Time Curriculum Synthesis for Self-Evolving
What this work does
Co-optimizes a test-question-conditioned curriculum synthesizer and solver using self-consistency and a difficulty signal favoring partially solvable generated questions.
Connection to discovery
A candidate for generating locally useful training problems when a target research problem is too difficult to produce positive rollouts directly.
Evidence
Inspected the variance-based curriculum reward and math-benchmark comparisons against TTRL and sampling baselines.
Scope and limitations
Answer agreement is not independent correctness. A scientific version needs external experimental/verifier signals and checks against a synthesizer/solver reinforcing shared misconceptions.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.