Aspire: Can Models Self-Evolve from Vague Goals?

Paper: Aspire: Can Models Self-Evolve from Vague Goals? (v1) https://arxiv.org/html/2608.31111v1

Aspire: operationalizing a goal before claiming progress

The previous checkpoint is an incomplete reference for self-improvement. Aspire makes the starting model, the agent’s proxy and the intended capability separate objects in the evaluation.

The managed interface supports data selection, synthesis, training, local validation and branching. Infrastructure mechanics remain controlled while strategic learning decisions belong to the agent. Weight evolution changes a descendant model; harness evolution changes supporting instructions, policy or workflow with runtime weights fixed. The reported weight searches keep the initial decision model fixed, so the evaluated process is not recursive replacement of that decision policy.

Our original mathematics example contrasts a broad reasoning goal with a local check that rewards a familiar answer format. A better proxy score can miss the intended capability. Aspire therefore keeps external evaluation separate. Adaptive feedback permits bounded aggregate queries on the same goal slice; final-only feedback withholds intermediate scores. Adaptive selection must not be described as an untouched confirmation test.

Of thirty adaptive configuration-goal cells, 28 produce an evaluated checkpoint and 21 produce an eligible checkpoint; only one cell retains an above-base improvement after rollback. Eligibility records a completed, admissible candidate; it is not itself a gain. One Qwen3.5-4B mathematics lineage scores 0.79, 1.75 and 2.38 after updates, while its base scored 17.86. An upward slope can be recovery from damage.

The fixed-runtime writing harness study separately freezes valid successors and executes them three times on twenty hidden writing items. Its reference task-macro score of 28.64 exceeds the closest valid successor’s 27.22. The useful chain is goal, proxy, candidate, external measurement and retained outcome. Each must remain visible before a learning loop can be called progress.

Watch the companion English animation with external English, Simplified Chinese or bilingual captions.

Paper, exact arXiv v1: https://arxiv.org/html/2608.31111v1

#SelfImprovement #AIAgents #AIResearch