AIM treats research ideas as explicit search objects: 67.0% and 55.8% average scores across 10 AutoLab tasks, matching the strongest baseline up to 3.1x faster in wall-clock time

The paper offers a first categorization of automated research by its primary unit of search: solution-driven search over executable artifacts versus idea-driven search that explicitly selects research directions before delegating implementation, and it identifies three central challenges of the latter: idea management scaffold, idea selection, and idea-solution integrity. Prior systems organize ideas with fixed scaffolds such as lists, search trees, or beam-style retention, and select with fixed rules such as UCB or MCTS or with direct LLM judgment; this work treats idea management itself as an autonomous, evidence-conditioned design problem and names idea-solution integrity as a distinct risk. A conceptual categorization and problem definition, supported by Figure 1 and the related-work comparison in Section 2, without quantitative metrics.

AIM's Agentic Surrogate uses Organize and Estimate operators to rebuild the growing idea pool into semantic clusters and to produce ordinal promisingness rankings at cluster and idea level, after which Agentic Acquisition's Dispatch uses two sequential LLM calls to assign each branch a joint cluster-level and idea-level explore/exploit action. Unlike fixed rules or direct LLM candidate selection, the dispatch actions are conditioned on explicit promisingness estimates and accumulated experimental evidence; ablations show that removing the whole Agentic Surrogate drops Flash Attention from 90.5 to 85.9, removing Organize gives 89.0 and removing Estimate 87.9, while embedding-based K-means clustering reaches only 83.4 and direct raw-score prediction 89.3. Single-task (Flash Attention) ablation reporting means and standard errors; the paper also reports that cluster-level rank estimates correlate positively with observed ranks, at roughly the 0.5 level on average (given as a range in the text).

Before a result enters the search state, the Solution Auditor checks whether the implementation follows the intended idea, satisfies task requirements, and obtains its score without exploiting the verifier; trivial, task_mismatch, and reward_hacking results are discarded, while idea_mismatch results have their idea reconstructed and the score and lessons re-attributed. This directly addresses the attribution risk specific to idea-driven search, where an unfaithful implementation would misattribute scores and derived lessons; the paper reports that the ratio of idea-mismatch flags drops substantially after reconstruction and gives qualitative cases such as code relying on the very technique the idea set out to avoid. Ablation shows removing the Solution Auditor lowers Flash Attention from 90.5 to a value in the high 80s (given parenthetically in the text), supported by before-and-after flag-ratio figures and qualitative cases.

The Resource Planner dynamically decides how many parallel branches to run per iteration under fixed total branch and execution budgets, trading parallel breadth against sequential adaptivity; the theoretical analysis adds an attainable-semantic-coverage proposition and a sufficient condition showing that broader coverage matters more when many directions are plausible but few are competitive. The paper formalizes when idea-level search helps: Proposition 1 states that an extreme idea-driven procedure attains semantic coverage at least as large as any procedure evaluating at most the same number of executable solutions, and Proposition 2 gives a sufficient condition in which the required coverage grows linearly with effective semantic breadth. The theory section states assumptions, definitions, and proofs (Appendix F.2); empirically, convex-hull area over Gemini-Embedding-001 embeddings projected by two-dimensional PCA and mean pairwise cosine distance serve as rough proxies for explored breadth, with idea-driven pools wider on Flash Attention (AIM 0.1150, ScientistOne 0.1199, versus EvoX at 0.0153) and a smaller gap on Radix Sort.

The results target automated research agents operating under a fixed experimental budget with executable verifiers, in tasks such as system optimization, model development, and CUDA kernels that can be repeatedly compiled, executed, and scored; for researchers and engineering teams building or evaluating research agents, the work offers a reusable component split: idea organization and promisingness estimation, cluster-level and idea-level exploration-exploitation dispatch, implementation-integrity auditing, and parallel-branch budget allocation. The qualitative examples and ablations also indicate the scaffold can pair with different solvers, for instance remaining at or above ScientistOne across five benchmarks when the default solver is replaced by Claude Code.

The theoretical analysis relies on assumptions such as faithful realization, which the paper itself calls somewhat idealized: in practice solver agents may produce incomplete or incorrect implementations, and some tasks inherently require substantial solution-level refinement before a direction's value can be assessed. The optimal balance between semantic exploration and implementation refinement may therefore vary by task, and adapting it dynamically to implementation difficulty, solver fidelity, and task semantic structure is left as future work. Empirically, explored breadth is proxied roughly by embedding convex-hull area and mean pairwise cosine distance, and the Estimate operator's ranking correlation is not strictly monotonic on AES128-CTR and FFT, which the authors conjecture may stem from multiple clusters with similar expected performance or from implementation noise. AIM also tends to need more executions to reach its best score and consumes more tokens, which the paper frames as a trade-off against final performance; on Data Select IFEval, AdaEvolve still scores higher, which the authors conjecture reflects a task better suited to smooth local search. This summary is based on the full paper text and the homepage evidence bundle and does not include complete numeric tables that appear only in appendices.