A Concrete Scenario
Give an LLM agent a task: "become a better physicist."
No AIME question bank, no MMLU scoring script, no explicit metric like "accuracy improves X% after 100 training steps." Just one sentence. The agent must figure out on its own:
- What should a "physicist" be able to do? Solve problems? Model systems? Read papers? Design experiments?
- Where am I weak now—math or physical intuition?
- How to learn: drill problems, read papers, do projects?
- How to verify progress: self-generated tests? External benchmarks?
- Agents improve some dimensions via harness-editing loops
- But the strongest evolved harness still underperforms the engineered Qwen-Agent reference across all 6 goals
Human students do this daily. For LLM agents, it is an almost entirely new task type.
Existing LLM self-evolution research gives agents explicit tasks—"improve accuracy on AIME." The agent only searches for "how to optimize" (data, training methods), not "what to learn." This setting is called explicit-task optimization.
Aspire asks: what happens when "what to learn" is also left to the agent?
Three Research Questions
RQ1: What changes when an explicit post-training task is replaced by a vague capability goal?
Search effort shifts from "how to optimize" toward "interpreting the goal." The agent must first clarify what the target even means.
RQ2: Can an instruction-tuned LLM convert a vague goal into retained capability gains via self-directed weight updates?
Almost no. Weight-level gains are sparse and unstable. Agents often train on mismatched data, trust narrow self-evaluations, and local gains fail to transfer to hidden evaluation. Continued search and training can even erase earlier progress.
RQ3: Can the agent harness (supporting scaffolding) evolve, beyond model weights?
Partially. Agents improve tool configurations and prompt strategies via harness-editing loops. But the strongest evolved harness still underperforms the engineered Qwen-Agent reference—human-designed scaffolding wins.
Why Vague Goals Are Much Harder Than Explicit Tasks
The paper's deepest insight: vague goals change the dimensionality of the search space.
Under an explicit task, the search space is 1-dimensional: "how to optimize a fixed objective." Under a vague goal, it becomes 3-dimensional:
1. What to optimize: how to decompose "become a better physicist" into subgoals? 2. How to optimize: which data and methods? 3. How to verify: how to know progress is real, not overfitting to your own proxy?
The third dimension—verification—inherently requires external ground truth. Tests the agent writes and grades itself cannot prove real capability gains, no matter the score.
This is Aspire's core design: evaluation items are fully hidden from the agent. An external evaluator tests the evolved agent with 520 hidden expert-authored items across 6 goals, covering math reasoning, code generation, long-context understanding, multilingual ability, tool use, and reasoning quality.
Key Finding: Weight-Level Self-Evolution Barely Works
1. Sparse, unstable weight-level gains: proxy improvements fail to transfer to hidden evaluation 2. Mismatched training data: e.g., goal is "physics reasoning," agent trains on generic math problems 3. Narrow self-evaluation: agents trust their own limited evaluations, making local gains proxy overfitting 4. Continued search can erase gains: training in the wrong direction actively removes progress
The strongest model (Qwen3-32B) averaged slightly below its pre-evolution baseline on hidden evaluation across 6 goals—not stagnation, regression.
Harness-Level Evolution: Partially Effective, Still Losing to Humans
Why Self-Evaluation Is the Bottleneck
Under explicit tasks, evaluation is given (AIME accuracy, MMLU scores)—a trustworthy external signal. Under vague goals, agents must construct their own evaluation, which suffers from:
1. Coverage bias: agents only evaluate dimensions they think of, missing ones they don't 2. Difficulty calibration: agent-written items don't match real benchmark difficulty distributions 3. Self-serving bias: agents favor questions they're good at
Together, these make self-evaluation nearly uncorrelated with hidden evaluation—the more the agent optimizes its proxy, the further it drifts from real capability. This mirrors the Dunning-Kruger effect in human learning.
Another Case for the "Evaluation Blind Spot Law"
This paper precisely hits the "evaluation blind spot law": evaluation systems systematically miss one class of failure—and it's exactly the most important class.
Prior cases: QuoteBench (matching scores masking pipeline damage), Omission Blindness (LLM judges detect commission but not omission), Calibrated Enough to Know (models know when outcomes are unpredictable, but the action gate ignores that judgment).
Aspire adds: agent-constructed evaluations reflect proxy progress but not real capability progress. Deeper isomorphism: judgment-gate decoupling—the agent "knows" its proxy diverges from real capability if asked, but that judgment never reaches the action gate selecting training data.
A New Intervention Layer: Evaluation
Existing alignment intervention levels: training (RLHF/DPO), inference-time prompts, interaction structures, output-level skills. Aspire reveals another: the evaluation layer. Under vague goals, the failure isn't model capability, prompts, or interaction—it's that the evaluation signal itself is untrustworthy. The fix is not to teach agents to self-evaluate, but to outsource evaluation to trusted external systems.
A Deeper Philosophical Question: What Is "Learning How to Learn"?
Under explicit tasks, "learning how to learn" is 1-D: search better optimization methods. Under vague goals it's 3-D: what to learn, how to learn, how to verify—genuinely meta-learning. Human students gradually acquire this: elementary students need teachers to tell them what to learn; university students plan their own direction. Current LLM agents are at the "elementary school" stage. Aspire's finding: even the strongest current agents cannot reliably perform meta-learning under vague goals. This suggests meta-learning may be the true bottleneck toward AGI—not model capability or search algorithms, but not knowing what to learn.
Implications for All Self-Evolving Systems
Any "system improving itself" scenario faces the same three questions: what to optimize, how to optimize, how to verify. Current self-evolving systems (AutoGPT, Devin, various agent frameworks) assume humans answer the first and third. Aspire shows: when these are handed to the agent, the whole self-evolution paradigm collapses. External trusted evaluation is not "nice to have"—it is "must have."
Limitations and Open Questions
1. Model scale: only Qwen3-32B and smaller models tested; larger models may behave differently 2. Goal coverage: 6 goals; more open goals ("become a better programmer") may differ 3. Harness search space: more radical harness evolution (agents writing new tools) is untested
The biggest open question: is vague-goal self-evolution failure an architectural limit or a training-paradigm artifact?
A Fascinating Detail
One detail worth re-reading: continued search and training can erase earlier improvements. This is not diminishing returns—it is regression. Because proxy and real capability diverge, the agent keeps training on the proxy, selects the "best checkpoint" by proxy score—which is the checkpoint with the worst real capability. A positive feedback failure loop: the more you optimize the proxy, the further from real capability; the further from real capability, the more you rely on the proxy. Humans have external signals (exams, feedback, social input) to break this loop. LLM agents under vague goals do not.
Perhaps Aspire's deepest insight: self-evolution is not a model-capability problem but an evaluation-signal problem. Without trusted external evaluation, self-evolution necessarily degenerates into proxy overfitting.
---
Paper: Aspire: Can Models Self-Evolve from Vague Goals?
Authors: Yuhao Wu, Jingyuan Zhang, Jiajun Shi, et al. (ByteDance Seed, Singapore University of Technology and Design, M-A-P, TokenWave.AI)
Published: 2026-08-31
Project page: https://self-developing-agents.github.io/
Key data: 520 hidden evaluation items × 6 goals; weight-level self-evolution averaged slightly below baseline; strongest evolved harness below the engineered Qwen-Agent reference