PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling Discovery Requires Forgetting

Shangyin Tan, Ben Zhou, Lakshya A Agrawal, Jialin Zhang, Yilin Chen, Qiuyang Mang, Donghyun Lee, Dennis Jacob, Koushik Sen, Alex Dimakis, Matei Zaharia

arXiv:2610.20563v1cs.AIcs.CL

TL;DR

As autoresearch agents run longer or work in communicating teams, accumulated history can make them conservative and cause progress to stall before the budget is spent. The paper periodically discards agent sessions and search state while keeping the best candidate, then evaluates this approach with communicating teams across four tasks. The teams with forgetting outperform parallel-agent baselines by up to 15% at the same budget.

  • Problem

    Longer single-agent, parallel-agent, and communicating-agent runs eventually plateau as accumulated history makes proposals conservative and stalls progress before the budget is exhausted.

    Source: §1 · p. 2

  • Method

    The method periodically discards every agent's session and search state, retains the best candidate, and restarts a communicating team from it.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    Across four tasks, a team with forgetting is best on every task and improves over parallel agents by up to 15% at the same budget.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The results suggest that forgetting can push autoresearch-agent performance further, with the next step being to decide when to forget and what to remember.

    Source: §6 · p. 9

Abstract

from arXiv · show

Automated research with LLM agents has shown promising results in various domains. Through iterative refinement and scored verifiers, LLM agents are capable of discovering new ideas. There are different ways to scale automated research: let one agent iterate for longer, run several agents in parallel, or further let the parallel agents communicate and build on one another's candidates. However, we find that all three strategies eventually plateau as the agents accumulate history. Each agent becomes anchored to its own past attempts, making fewer and smaller edits. Beyond this point, additional budget brings little further improvement. In this paper, we show that the key to scaling discovery further is to forget. Periodically discarding the search state, including every agent's session and the candidate history, and restarting from the best candidate found so far, unblocks the search and pushes the same budget further. We implement forgetting in a team of agents with communication and evaluate it on four tasks spanning heuristic programming, kernel optimization and model training. The results show up to 15% improvement compared to parallel agents, and up to 39% compared to a single agent. Our results suggest that, for open-ended discovery, what an agent is allowed to remember matters as much as how much compute it is given.

1 Introduction

Longer runs, parallel agents, and communication all eventually plateau as accumulated history makes agents more conservative. The paper proposes periodically forgetting that history while retaining the best candidate, and reports gains from this approach across four tasks.

1 Introduction

  • The three scaling strategies plateau as growing session and search history makes proposals increasingly conservative before the budget is exhausted.The paper attributes the stall to agents carrying accumulated state and session context into each new proposal.Source: §1 · p. 2
  • Forgetting discards sessions and candidate history while keeping the best candidate, so a fresh agent can continue without conditioning on failed attempts.The authors argue the best candidate often subsumes effective heuristics from earlier iterations.Source: §1 · p. 2
  • A communicating team periodically resets every agent from the best candidate found across the team, clearing accumulated history while retaining shared progress.Good ideas can propagate among agents between resets.Source: §1 · p. 2
  • Across four tasks, a four-agent team with forgetting is best on every task and improves over the strongest baseline by 1.9% to 15%.The tasks span heuristic programming, algorithmic optimization, VLIW kernel optimization, and small-language-model training.Source: §1 · p. 2
  • The paper reports that sequential and parallel agent scaling plateau under a fixed budget as sessions grow longer and agents become more conservative.It identifies periodically discarding agent sessions and candidate history, then restarting from the best candidate, as its proposed response.Source: §1 · p. 2

2 Related Work

The related work traces autoresearch from iterative LLM-guided optimization toward agents that run research loops in a continuous session. It distinguishes this work by resetting both session and workspace state, while situating communication among mixed findings on when agent collaboration helps.

2 Related Work

  • LLM-guided optimization includes evaluator-backed program search such as FunSearch and AlphaEvolve, prompt optimization, and coding agents running research loops.Prior approaches include population-level diversity and branching workspaces; studies also report that inherited failed attempts can reduce task solving and iterative loops can lose diversity and stall.Source: §2 · p. 3
  • Fork-and-Flush periodically creates fresh-context trajectories that inherit the whole workspace, whereas this work also removes workspace state and keeps only the best candidate [Cai et al., 2026].The paper also distinguishes its agents' communication within rounds from systems using an orchestrator.Source: §2 · p. 3
  • Prior multi-agent studies report benefits and failures: communication can propagate improvements but also homogenize proposals, with short-budget matched-compute teams often failing to beat a strong single agent.The passage also notes saturation from correlated outputs, compromise outcomes, and incorrect majority views.Source: §2 · p. 3

3 Agent Teams That Forget

The method periodically clears agents' sessions and search states while preserving the best candidate, and combines this reset with asynchronous sharing of candidates among parallel agents. A reducer controls what each agent sees and which session it continues.

3 Agent Teams That Forget

  • The search loop proposes a candidate, has a judge score it, and uses the outcome to guide the next proposal; the team variant periodically discards history and keeps the best candidate.Figure 2 contrasts continuous-history search with four communicating agents that restart from one retained candidate.Source: §3 · p. 3, Figure 2 · p. 4

3.1 Forgetting

  • Each forget ends a round by discarding the agent's session and search state, retaining the best candidate and its score for a new agent with empty context.The authors say the retained candidate contains most prior progress, while a fresh context avoids conditioning on earlier failed directions and produces more thinking tokens and larger edits.Source: §3.1 · p. 4

3.2 Teams

  • Agents work in separate sessions and states, see peers' candidates and scores before proposing, and communicate asynchronously without direct point-to-point messaging.At a team-wide forget, all agents restart from the best candidate found across the team.Source: §3.2 · p. 4

3.3 Implementation

  • The harness stores candidates as version-controlled commits with judge scores and feedback, and retains scripts for past sessions.Figure 3 depicts four concurrent loops, each with its own state, with each agent receiving peers' candidates.Source: §3.3 · p. 5, Figure 3 · p. 5
  • A reducer controls each agent's visible candidates and continuing session; forgetting clears both by starting a new state with only the best candidate and an empty context.Agents edit candidates through file tools, after which the harness records and judges each result.Source: §3.3 · p. 5

4 Experiments

Across four tasks, a communicating team with forgetting achieved the best result on every task. Forgetting continued to improve search after plateau, while parallelism and communication without resets did not consistently prevent stagnation.

4 Experiments

  • The experiments used three runs per task and compared a single agent, four parallel agents, and a four-agent communicating team that forgot three times.All settings used the same model and budget; parallel agents did not see one another’s work, while the team shared candidates.Source: §4 · p. 5

4.1 Parallel Scaling and Communication Still Plateau

  • Running four parallel agents helps on three of four tasks; on AHC014, they raise mean best score 21% over one agent, but gain only 1.2% in the budget’s second half.Communication alone does not remove the plateau: each team configuration’s progress flattened before each forget.Source: §4.1 · p. 6

4.2 Forgetting Unblocks the Plateau

  • Across four tasks, the team with forgetting achieved the best result on every task, averaging 10% better than parallel agents and 22% better than a single agent.On AHC014, its mean best score was 14% above parallel agents and 39% above a single agent.Source: §4.2 · p. 6, Table 1 · p. 6
  • The team’s mean best score improved after every forget on every task, and 32 of 36 forgets produced a new best candidate in the following round.On AHC014, the mean best score rose 11.2% after the first forget and 1.7% after the third, as progress slowed near task bounds.Source: Table 2 · p. 6, §4.2 · p. 7

4.3 Forgetting Outperforms Continuing at Lower Cost

  • From the same plateaued AHC014 state, forgetting gained 4.5% for one agent, 7.0% for parallel agents, and 9.2% for the team, at little more than half the cost.Continuing the sessions gained only 1.6%, 1.9%, and 1.4%, respectively; the team had the largest forgetting gain.Source: §4.3 · p. 7

5 Why Forgetting Helps

As sessions lengthened, agents made smaller edits and used less reasoning, sometimes ceasing to test ideas despite continued reasoning. Forgetting renewed larger edits, while communication helped agents turn a fresh but initially unsuccessful idea into an improvement.

Search Narrows as Sessions Grow

  • Before its first reset, the team’s edits fell from about 150 lines to 30 and reasoning from about 28k tokens to 7k, alongside a performance plateau.Across configurations, edits began around 100 lines and declined within the first few dollars; reasoning tokens fell to a third or less of their initial level.Source: §5 · pp. 7–8
  • On VLIW, 532 of 763 completed turns produced no candidate, consuming 66.2% of the single-agent model spend; 531 involved at most one file read and no edit.This occurred in all three replicates, and the passage distinguishes these empty turns from failures.Source: §5 · p. 8
  • In one VLIW session, eleven turns yielded no candidate even as the agent developed an untested selection-tree idea and declined to implement it as too complex and risky.The agent first rejected the idea using a mistaken cost comparison, later corrected that comparison, but still did not test the change.Source: §5 · pp. 8–9

Forgetting Renews Collective Search

  • After forgetting, median edits over the first ten proposals were 40 lines versus 6 under continuation, and reasoning-token use was 9.4k versus 2.9k.In a VLIW team, a fresh agent’s idea improved after it borrowed a peer’s code trick following two unsuccessful attempts; the team ended the round at 1,114 cycles.Source: §5 · p. 9

6 Conclusion

The paper concludes that periodically forgetting can address plateaus in long-running discovery agents by retaining the best candidate while discarding sessions and search state. Combined with communication, this approach gave the best result on all four evaluated tasks and improved on parallel agents by up to 15% at the same budget.

6 Conclusion

  • Combined with a communicating team, forgetting gave the best result on all four tasks and improved on parallel agents by up to 15% at the same budget.The authors identify deciding when to forget and what to remember as a natural next step.Source: §6 · p. 9

A Task Descriptions

Each task provides an editable candidate and a judge that scores it, with a per-replicate money budget for each treatment.

A Task Descriptions

  • Each task pairs an editable candidate with a scoring judge, and specifies the money spent per replicate of each treatment as its budget.Source: Table 4 · p. 13

B Prompts

The prompts keep search strategy unspecified while giving agents task and loop instructions. Session messages, workspace files, and an external scoring process structure each turn.

B Prompts

  • Every agent uses Pi’s default system prompt, and the prompts deliberately provide no search strategy.They describe the task and the loop.Source: B Prompts · p. 13
  • A session’s first turn receives the task, current status, and full instructions; later turns receive only status and a short reminder.After forgetting, a new session receives the full first-turn message again.Source: B Prompts · p. 13
  • In a team, peers’ candidates are workspace files that agents read with their own tools, and the prompt includes one line per peer.Source: B Prompts · p. 13
  • The `.context/` directory holds turn-prepared materials such as feedback, candidate lineage, other candidates, or the whole run; candidates are repository commits that can be inspected or compared.Source: B Prompts · p. 13
  • Agents end each session after improving a candidate; another process scores it, and the harness commits the working tree and resumes the process in a later turn.The instructions say not to evaluate the change or commit it themselves.Source: B Prompts · p. 14

Related papers

Loading 2610.20563v1…