Aspire: Can Models Self-Evolve from Vague Goals?
Yuhao Wu, Jingyuan Zhang, Jiajun Shi, Yuxuan Zhang, Xinping Lei, Junting Zhou, Zexuan Wang, Yuchen Wu, Huan Zhou, Duo Wang, Yinzhu Piao, Yongchang Peng, Yunfeng Shi, Jin Chen, Zuo Wang, Jinkai Liu, Jiaheng Liu, Wenxuan Zhang, Shen Yan, Wenhao Huang, Ge Zhang
cs.CL
2026-09-01
Aspire hides 520 items behind a vague goal. One of 12 final-only pairs beats base; adaptive search keeps one 4B math gain, 17.86 to 20.10. Best harness trails Qwen-Agent at 28.64.
People learn physics or research from a vague aim, not from a prepared exam. They decide what is missing, what to practice on, and how to tell that they actually improved. Most LLM self-evolution work starts where PostTrainBench starts: the task, the script, and the metric are already chosen. The search is over how to optimize, not over what to optimize.
Aspire isolates that missing step. The agent gets a natural-language capability goal. Evaluation items stay sealed. It must pick data, pick SFT or GRPO, build a validation signal, and decide when to stop. The test is whether the intended capability moved, not whether a self-made proxy went up.
Six goals sit on 520 expert-authored hidden items: science 75, humanities 110, medicine 100, math 126, logic and instruction following 89, academic writing 20. GPQA, MMLU-Pro, and MedQA are format and difficulty references only; no source item is copied. Training data is overlap-audited. The protocol returns either a bounded number of aggregate scores or a single terminal score, never items or per-item labels.
The agent sees one typed tool: search and register data, launch LoRA SFT or GRPO, poll jobs, score home-made validation sets, branch or stop. The controller owns credentials, scheduling, and checkpoint checks. Weight evolution mutates the model with the harness frozen. Harness evolution mutates prompts, tool policy, and workflow with weights frozen. Under score gating, a selected checkpoint is kept only if it beats the incoming score; otherwise the controller rolls back. Nonnegative retained gain is a selection rule, not evidence that every update helped.
RQ1 replaces PostTrainBench’s benchmark name with a broad capability phrase and keeps the original evaluator. RQ2 starts from instruction-tuned checkpoints and searches over weights. RQ3 freezes Qwen3.5-4B and asks different decision models to emit one successor harness.
RQ1: Claude Opus 4.8 scores 27.07 under vague goals versus 32.90 on the official explicit-task reference; GPT-5.6 scores 29.58 versus 36.23. In matched Claude traces, vague goals add 2,109 seconds of decision thinking and 0.61 GPU-idle hours, and cut 1.27 hours of active train/eval. Accesses to task materials are about 2.98× denser.
RQ2 final-only: 24 runs, 12 model–goal pairs. Only one Avg@2 beats base: Qwen3.5-9B science, 45.33 to 48.00. Three of 24 terminal checkpoints beat base; 21 roll back. Adaptive feedback: 28/30 cells produce a checkpoint, 21 are eligible, two have a best score above base, and only Terra’s 4B math run is retained, 17.86 to 20.10. Later descendants fall to 2.78 then recover to 18.17. Sol searches hardest, 33 checkpoints and 76.56 GPU-hours, and never clears base.
A concrete collapse: five terminal checkpoints trained on numeric-label MMLU SFT. All 21,000 targets are single digits, all 279 eval outputs are single digits, scores 0, 0, 0, 6.141, 0.
RQ3: original Qwen-Agent task-macro 28.64. Sol’s successor 27.22, Terra 20.76, Luna 19.32. The 4B Creator emits no valid harness. Luna stops after pushing an 8-item homemade checklist from 7/8 to 8/8, then scores 19.32 on the hidden set.
For recursive self-improvement, finishing a training loop is not closing a capability loop. Agents already download data, launch LoRA, and submit checkpoints. Given a vague goal, most updates damage behavior the instruction-tuned model already had. Slope versus the previous checkpoint misleads: many rises are recovery from training injury.
The practical rule is blunt. Without hidden evaluation and rollback against the base score, self-training will collapse format and capability together. Harness edits overfit narrow homemade checklists in the same way.
Six goals and 520 items are a narrow slice; writing has 20 tasks. Adaptive cells are one run each. Terra’s 20.10 was selected with repeated aggregate queries on the same hidden slice, with no confirmation split, so it is not a replicated gain. Reported runs never promote a trained descendant into the next-round decision model, so this is not recursive replacement.
RQ1’s official references and vague-goal traces are not fully matched on metadata; the paper treats the gap as observational, not a pure prompt effect. Content-level traces sit behind controlled access, so failure-mode attribution cannot be fully rechecked in public. Unrelated capabilities are not measured.