Compute Aligned Training: Optimizing for Test Time Inference
What this work does
Treats inference strategies as operators on a base policy and derives strategy-dependent SFT/RL gradient weightings for pass@N, voting and best-of-N selection.
Connection to discovery
Motivates matching a discovery policy's training objective to the deployed search/selection harness instead of optimizing pass@1 in isolation.
Evidence
Inspected the effective-policy formulation, assumptions in the RL scaling derivation and reasoning/protein-generation evaluation setup.
Scope and limitations
Some derivations use explicit local/diagonal assumptions. Verifier access, selection cost and candidate dependence matter when extending the analysis to interactive experiments.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.