Self-Developing Agents: Three Benchmarks Reveal Why AI Struggles to Self-Improve

The article introduces the Self-Developing Agents project from ByteDance Seed and TokenWave, which proposes three benchmarks to study whether AI agents can achieve recursive self-improvement (RSI) by forming goals from vague objectives, learning from experience, and retaining system improvements. The authors argue that current "semi-loop RSI" relies on an external golden verifier, whereas a full closed loop requires the agent to decide what to learn, how to verify, and how to preserve verified changes.

ASPIRE: Learning from Vague Goals

Paper: https://arxiv.org/abs/2608.31111 | Project & Blog: https://self-developing-agents.github.io/ ASPIRE tests an agent's ability to pursue a natural-language capability direction (e.g., "improve mathematical reasoning") without a predefined dataset or reward function. The benchmark includes 6 capability goals, 520 expert-written hidden evaluation questions, and a minimal interactive environment that supports weight and harness updates. It also provides 48 paired "vague goal / explicit task" runs with auditable traces.

Key finding: Agents spend more time interpreting the goal and selecting proxy metrics, leaving less budget for actual training and evaluation. LoRA usage rises from 24.1% to 89.8% under vague goals, yet only 1 out of 30 units produced a retained gain after filtering. Completing an update does not guarantee the target capability improved.

S³Gym: Self-Testing, Self-Judging, Self-Improving

Paper: https://arxiv.org/abs/2608.31100 | Project & Blog: https://self-developing-agents.github.io/ S³Gym places agents in seven text-based games where they must explore, self-evaluate, and self-improve without access to the true step rewards or final scores. After exploration, agents are tested under stricter conditions with different random seeds; exploration logs are not fed back as training material.

Key findings:

Self-judgment accuracy does not reliably translate into subsequent performance gains. Agents can judge how well they did but often fail to adjust the next action accordingly.

No single experience-retention method works across all games. Raw history and summary memory each have strengths; summarization preserves reusable strategies but may lose critical state-action details.

Model-weight updates can harm existing capabilities. In one trajectory the model improved on a trust game but regressed continuously on Plants vs. Zombies, while other games showed no change. The cause of regression is not yet determined.

The benchmark therefore evaluates experience value by its effect on future actions, not by the plausibility of the summary.

HarnessDev: Persisting System Improvements

Paper: https://arxiv.org/abs/2609.01437 | Project & Blog: https://self-developing-agents.github.io/ HarnessDev lets a model create and iteratively modify its own execution system (code, ML experiments, writing, search). The development model and the execution model are separated; the execution model's weights are frozen so that only the harness changes are measured. The benchmark tracks whether new mechanisms are actually invoked during runs (e.g., whether written memory and state-saving code are called).

Key finding: Systems that improve on development feedback often overfit and degrade on held-out tasks. In fixed-Gemini-execution runs with Qwen and DeepSeek developers, selected versions improved on seen feedback but both declined on hidden SWE-Pro tasks. Modifications can also break existing functionality (e.g., message cleanup and context compression disrupted valid tool-message sequences, leading to rollbacks). HarnessDev therefore emphasizes verifying that new mechanisms are exercised at runtime, that improvements transfer to unseen tasks, and that rollback is possible.

Conclusion

The three benchmarks collectively show that current agents possess basic self-training, self-summarization, and self-modification abilities, but the real bottleneck is not "can they change" but "can they accurately judge what to change, which experience to absorb, and whether the change actually helps." The main problems are proxy-goal mismatch and overfitting to local feedback. Future self-developing agents will need more reliable goal-formation, verification, and retention mechanisms to continuously filter effective experience and distill truly generalizable improvements into stable capabilities.