Aspire Benchmark Reveals Why LLM Self-Evolution Fails Under Vague Goals

The Scenario

Give an LLM agent a single vague instruction: "become a better physicist." No AIME problem sets, no MMLU scoring scripts, no explicit accuracy target. The agent must figure out on its own:

• What should a "physicist" be able to do? Solve problems? Model? Read papers? Design experiments?
• Where am I currently weak—math, or physical intuition?
• How to learn: drills, papers, projects?
• How to verify progress: self-generated tests, external benchmarks?

This is what human students do every day, but it is a nearly new task type for LLM agents. Prior self-evolution research gives agents explicit tasks (e.g., "improve AIME accuracy"), so agents only search *how* to optimize, not *what* to learn. 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" to "goal interpretation." The agent must first clarify what the goal even means.

RQ2: Can instruction-tuned LLMs convert vague goals into retained capability gains via self-directed weight updates? Almost not. 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? Partially. Agents can improve tool configurations and prompt strategies via harness-editing loops, but the strongest evolved harness still falls short of an engineered Qwen-Agent reference.

Why Vague Goals Are Much Harder

The key insight: vague goals change the dimensionality of the search space.

With an explicit task, search is 1-dimensional—"how to optimize a fixed objective." With 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 one's own proxy?

The third dimension—verification—fundamentally requires external ground truth. Tests the agent writes and grades itself cannot prove real capability gains, no matter the score.

This shapes Aspire's core design: evaluation items are fully hidden from the agent. An external evaluator measures evolved agents on 520 expert-authored items across 6 goals: math reasoning, code generation, long-context understanding, multilinguality, tool use, and reasoning quality.

Key Finding: Weight-Level Self-Evolution Barely Works

1. Sparse, unstable gains: proxy improvements fail to transfer to hidden evaluation. 2. Mismatched training data: e.g., training on generic math when the goal is physics reasoning. 3. Narrow self-evaluation: local gains are proxy overfitting. 4. Continued search can erase gains: the agent doubles down in the wrong direction.

The strongest model (Qwen3-32B) averaged slightly below its pre-evolution score on hidden evaluation—not stagnation, but regression.

Harness-Level Evolution: Partial Gains, Still Behind Human Design

• Agents improve some dimensions via harness-editing loops.
• The strongest evolved harness still underperforms the engineered Qwen-Agent reference on all 6 goals.

This echoes earlier observations that model-generated prompts generally underperform human-written ones—Aspire extends that finding from prompts to entire harnesses.

Self-Evaluation Is the Bottleneck

Agent-constructed evaluation suffers from three biases:

1. Coverage bias: agents only assess dimensions they think of. 2. Difficulty calibration: self-written test difficulty doesn't match real benchmarks. 3. Self-serving bias: agents favor questions they are good at.

Together, these make self-evaluation nearly uncorrelated with hidden evaluation. The more an agent optimizes its proxy, the farther it drifts from real capability—an LLM analogue of the Dunning-Kruger effect.

This also fits a recurring pattern of "evaluation blind spots": self-constructed evaluation reflects proxy progress, not real progress. Deeper still is a judgment–gate decoupling: the agent can *acknowledge* its proxy diverges from real capability, yet that judgment never gates its data-selection actions.

A New Intervention Layer: Evaluation

Beyond training-level (RLHF/DPO), prompt-level, and interaction-structure interventions, Aspire reveals an evaluation-level failure: under vague goals, the model, prompt, and interaction may all be fine—the evaluation signal itself is untrustworthy. The fix is externally provided, credible evaluation signals, not smarter self-assessment.

Learning How to Learn

Under vague goals, "learning how to learn" becomes genuine meta-learning: searching over what to learn, how to learn, and how to verify. Human students gradually acquire this; LLM agents are still at the "elementary school" stage—needing humans to tell them what to learn. Aspire's finding: current frontier LLM agents cannot reliably perform meta-learning under vague goals. This suggests meta-learning, not raw capability, may be the real bottleneck on the path to AGI.

Implications for All Self-Evolving Systems

Any "system improving itself" scenario faces the same three questions—what, how, and verify. Current systems (AutoGPT, Devin, agent frameworks) assume humans answer the first and third. Aspire shows the paradigm collapses when agents must answer both. External trustworthy evaluation is not nice-to-have; it is a must-have.

Limitations

1. Model scale: only Qwen3-32B and smaller; larger models may behave differently. 2. Goal coverage: only 6 goals; more open goals may differ. 3. Harness search space: current harness-editing is limited; letting agents write new tools may change results.

The biggest open question: is vague-goal self-evolution failure an architectural limit or an artifact of current training paradigms?

A Striking Detail

Continued search and training can erase earlier improvements. Not diminishing returns—regression. Since the proxy diverges from real capability, optimizing the proxy raises proxy scores while real capability falls; checkpoint selection by proxy then picks the worst real model. A positive feedback failure loop: the more the proxy is optimized, the less it is trusted less—yet it remains the only trusted signal, so it is optimized further. Humans break this loop with external signals (exams, feedback); vague-goal agents have none.

Aspire's deepest insight: self-evolution is not a model-capability problem but an evaluation-signal problem. Without credible 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)

Project page: https://self-developing-agents.github.io/

Core numbers: 520 hidden evaluation items across 6 goals; weight-level self-evolution averaged slightly below pre-evolution; strongest evolved harness below the engineered Qwen-Agent reference.