New papers say scaffolds explain only 1.5% of agent performance variance
gerardsans · x · 2026-07-21
A thread about two papers argues that “agent” scaffolds may contribute far less than people assume.
- MORPHEUS (NeurIPS 2026) is cited as finding that scaffolds explain only 1.5% of performance variance, while the base next-token model explains 41.4%—about 28× more.
- The implication: multi-agent swarms, chain-of-thought wrappers, and similar orchestration layers often just reshape prompts around the same frozen weights.
- A reply cites CORRAL (arXiv:2604.18805), which reportedly ran 25k+ trajectories across 8 scientific domains and found that 68% ignored evidence, 71% never updated beliefs, and only 26% showed refutation-driven revision.
- The combined argument is that many “agentic” behaviors may be an illusion created by scaffolding plus narrative, not genuine adaptation or reasoning.
More from AGI Musings
- François Fleuret: Only Two Long-Term Futures — No Super AI, or Staying Fully Human With It — francoisfleuret · 2026-09-11
- IG reel debunking the 'winning the AI race against China' fallacy hits 500k likes — louisvarge · 2026-09-11
- Post-AI World Leaves No Room for Learning on the Job — rachittshah · 2026-09-11
- Researcher questions AI safety eval firm, citing 'blatantly sloppy' security and monitoring — Kyrannio · 2026-09-11
- AI researcher memes agent-swarm tinkering with He Jiankui's embryo-editing quote — dejavucoder · 2026-09-11
- nabla_theta: happy to be wrong if the AI utopia arrives with little ex ante risk — nabla_theta · 2026-09-11