Scaling Discovery Requires Forgetting — Periodic restart from the best candidate — What does it mean?
The news. In October 2026 an arXiv paper, Scaling Discovery Requires Forgetting, reported that three ways of scaling automated research (one agent iterating longer, several agents in parallel, parallel agents that build on one another's candidates) all plateau as history builds up. The fix it proposes is to periodically discard the search state and restart from the best candidate. It evaluates this on four tasks across heuristic programming, kernel optimization and model training.
Picture a team working at one whiteboard that never gets erased. Every failed idea, every crossed-out sketch and every note about "we already tried that" stays on it. At first the board looks like a help, because it records what was tried. After enough rounds it can be a hindrance: the board is so full that each new idea is squeezed into the gaps around old ones, and the team makes small touch-ups instead of bold changes.
The paper describes this in agent terms. As an agent accumulates history, it becomes anchored to its own past attempts and makes fewer and smaller edits. The context failure modes in the curriculum describe the general version: context that keeps growing can steer the model instead of informing it. Here the harm is specific. The history is not wrong, it is narrowing: it records what was tried and how each attempt went, so new edits stay close to old ones.
| Scaling strategy | What the budget buys | What the paper reports |
|---|---|---|
| One agent, more iterations | Longer history in one session | Plateaus as history accumulates |
| Parallel agents | More independent sessions | Plateaus as history accumulates |
| Parallel agents that share candidates | More sessions plus cross-pollination | Plateaus as history accumulates |
| Any of the above plus periodic forgetting | The same budget, with history cleared at checkpoints | Up to 15% over parallel agents, up to 39% over a single agent (4 tasks) |
The remedy is the board-wiping move. Periodically, the system discards the search state, including every agent's session and the whole candidate history, and restarts from the best candidate found so far. Only that one candidate crosses over. In the metaphor, someone photographs the best sketch, wipes the board clean, and redraws just that sketch as the starting point. Objective progress is kept; the narrative around it is dropped.
The paper implements this in a team of agents that communicate, and reports it on four tasks. The architectural details of the restart schedule are not in the abstract, so treat the exact interval as something to tune rather than a fixed recipe. The following contrast is my inference, not a finding of the paper. It differs from compaction: the fixes for long context summarize history to make it smaller, while forgetting removes it and keeps only the winning artifact. A restart has a plausible cost, because discarded history can contain lessons the fresh agents must rediscover, and coordination costs apply to a team that restarts together.
Worked example (illustrative, not from the paper). Suppose a team has a budget of 100 rounds and a verifier score that starts at 100. For the first 40 rounds each round adds about 1.0, so the best score reaches 140. Then history anchoring sets in and each round adds only about 0.1. Without forgetting, the last 60 rounds add 60 × 0.1 = 6, ending at 146.
Now restart at round 40: keep the score-140 candidate, clear every session and the candidate list. Assume the fresh agents recover about 0.4 per round for the remaining 60 rounds. That adds 60 × 0.4 = 24, ending at 164 versus 146, about 12% higher for the same budget. The numbers are chosen to show the mechanism; the measured gains are the paper's 15% and 39% figures, and they depend on the task.
Goes deeper in: AI Agents → Context Engineering → The 4 Failure Modes
Related explainers
- CliffCompaction — Re-compacting from the original context — a different answer to the same question: what to carry forward when context grows.