ByteDance Seed's Self-Developing Agents: 3 benchmarks show AI self-improvement fails at goal validation
机器之心 · wechat · 2026-09-16
ByteDance Seed, TokenWave and collaborators released the Self-Developing Agents project, arguing that common setups relying on a reliable GoldenVerifier ("half-loop RSI") miss the full recursive self-improvement loop, and introduced three benchmarks:
- ASPIRE: tests learning from vague goals like "improve math reasoning" with 6 capability targets and 520 expert-written hidden eval questions. Vaguer goals push agents to spend more time interpreting objectives (LoRA usage rose from 24.1% to 89.8%), yet only 1 of 30 units produced a retained capability gain.
- S³Gym: in seven text games, agents must self-judge and self-improve. Judgment accuracy correlated weakly with later performance; no universal memory format worked across environments, and weight updates could regress some games while improving others.
- HarnessDev: models create and iteratively modify their own harness. Positive feedback during development often didn't transfer to hidden tasks, and some changes broke existing functionality and had to be rolled back.
The project concludes the bottleneck isn't whether agents can change themselves, but building reliable mechanisms for goal formation, verification, and retaining generalizable improvements.
More from coding & agent
- Dev ditches Cursor, says he switched back to VS Code and loves it — Arindam_1729 · 2026-09-21
- Agent system layers are mostly transitional, engineer argues — rakyll · 2026-09-21
- Don't feed full HTML: how an agent fixes internal links without hallucination — cneuralnetwork · 2026-09-21
- Engineer quips: AGI is inevitable once we figure out the YAML stuff — rakyll · 2026-09-21
- One RTX 3090 ran Qwen 27B autonomously for 3 weeks — it shipped working CUDA kernels — skeole · 2026-09-21
- Dev buys a Meta coding subscription for its 'excellent model, crazy quota, low price' — intellectronica · 2026-09-21