Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Autonomous research should be judged by how fast it finds good answers, not only by the final score—and a bandit over parallel search chains can nearly match knowing the best method in advance.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 09:09 UTC pith:RBVT6BQ3

load-bearing objection Clean empirical case that AR search efficiency ≠ endpoint quality, plus a sensible adaptive harness—but fluid’s near-oracle number is partly fit to the same twelve tasks. the 3 major comments →

arxiv 2607.24647 v1 pith:RBVT6BQ3 submitted 2026-07-27 cs.AI cs.LG

Efficiency Matters in Autonomous Research

Haiqian Yang , Yuan Cao This is my paper
classification cs.AI cs.LG
keywords autonomous researchsearch efficiencyPareto frontier AUCanytime performanceportfolio banditfluid searchLLM-driven optimizationbudget allocation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that AI autonomous-research systems are usually scored only on the best solution they eventually find, which hides how wasteful the path was. As research moves into settings where each trial is a costly experiment, that path cost becomes decisive. The authors measure efficiency as the area under the curve of the best-so-far reward versus evaluation budget, and show on twelve systems-optimization tasks that hill climbing, beam search, tree search, and evolutionary search each win on some tasks and lose on others. Final quality and speed of improvement are not the same axis: a late breakthrough can look best at the endpoint while scoring poorly on efficiency. Because the right fixed structure is unknown beforehand, they introduce fluid search—a portfolio bandit that spreads a fixed budget across many hill-climbing chains and steers evaluations toward chains that are still improving. Fluid is the most efficient single policy they test and nearly matches an oracle that is told the best structure per task.

Core claim

No fixed classical search structure is uniformly most efficient for LLM-driven autonomous research, efficiency and final outcome are empirically distinct, and a budget-adaptive portfolio bandit over a forest of hill-climbing chains (fluid) achieves the highest overall efficiency—nearly matching a per-task oracle that already knows the best fixed structure for each task.

What carries the argument

Fluid search: a UCB portfolio bandit that allocates each next evaluation across a forest of restarting hill-climbing chains using recent improvement rate, incumbent quality, and an exploration bonus, plus levers (warm-start breadth ramp, archive reseeding, crashed-parent adoption) borrowed from fixed policies. Efficiency is scored by AUC of the Pareto frontier R(n) = max reward in the first n evaluations.

Load-bearing premise

That the efficiency rankings and fluid’s near-oracle gains will still hold when verification is truly expensive and the task landscape differs from twelve code-optimization tasks with one fixed LLM proposer and cheap automatic checks.

What would settle it

Rerun the same fixed policies and fluid on a domain with costly physical or lab verification (or a clearly different proposal model and crash structure): if a single fixed structure dominates AUC across tasks, or fluid falls well below the per-task oracle, the central efficiency and adaptivity claims fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • AR benchmarks should report AUC of best-so-far reward versus budget, not only terminal score.
  • Builders should treat budget allocation as an online decision rather than locking a search tree, beam, or population rule in advance.
  • When each evaluation is an expensive experiment, policies that improve early can be more usable than policies that only win at a large fixed budget.
  • Near-oracle fluid performance implies useful allocation signals already sit in the live search trajectory.
  • A natural next design is a heterogeneous portfolio that can switch search form, not only which chain gets the next trial.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Anytime AUC metrics from AutoML and molecular design are being imported as a first-class AR objective; agent leaderboards that only rank final medals may systematically mis-rank systems under tight real budgets.
  • If proposer quality or verifier delay changes the crash and improvement signals fluid relies on, the same bandit index may need retuning or richer state features.
  • Plateaus the paper attributes to vocabulary and verifier gaps suggest efficiency gains alone will not unlock open-ended discovery without expanding what can be proposed or recognized.
  • Cheap-verifier code tasks may understate the value of crash-repair and archive diversity that matter more when a failed trial wastes days or dollars.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper argues that autonomous-research (AR) systems should be evaluated not only by final outcome quality but by search efficiency — the shape of the budget-vs-reward curve. It formalizes this via the frontier reward R(n) = max_{i≤n} r_i and its AUC, plus a convergence ratio separating "how high" from "how fast." Holding proposer (Qwen3-Coder-480B), verifier, budget (500 evaluations), and frontier width (64) fixed, it compares greedy hill climbing, beam search, tree search, and evolutionary search on twelve AutoLab systems-optimization tasks, finding that no fixed structure is uniformly most efficient and that outcome and efficiency are only weakly correlated (Pearson r=0.40). It then introduces fluid, a UCB-style portfolio bandit over a forest of hill-climbing chains augmented with five levers (archive/reseeding, crash-repair, merge softening, warm-start breadth ramp), and reports that fluid achieves the best normalized AUC (0.780), essentially matching a per-task oracle over fixed policies (0.784). The authors conclude that trajectory signals suffice to allocate budget near-optimally online.

Significance. If the results hold, the paper makes two contributions of real value to the AR/agents literature: (1) a clean, well-motivated evaluation framing — AUC of the frontier reward R(n) alongside best reward, with a convergence-ratio decomposition (empirically only r=0.40 collinear with outcome, Fig. 3b) — that imports the anytime-performance perspective from AutoML/molecular design into LLM-driven research search; and (2) controlled evidence that four canonical search structures differ sharply in efficiency and that none dominates, with matched controls throughout (same proposer Qwen3-Coder-480B, same verifiers, 500-evaluation budget, frontier width 64, three seeds, twelve AutoLab tasks) and code released. The efficiency/outcome decoupling (e.g., aes128_ctr, where greedy wins on outcome but beam on AUC) is a crisp, useful demonstration. The fluid result is potentially the headline finding, but its current evidentiary basis (see major comments) is weaker than the framing experiments.

major comments (3)
  1. [§6, Fig. 4c] Fig. 4c and the concluding claim of §6: the near-oracle result (fluid 0.780 vs. oracle 0.784) is evaluated on the same twelve tasks from which fluid's design was reverse-engineered. The text is explicit that each lever is 'motivated by the fixed-policy results': the 16→64 warm-start ramp through evaluation 150 imports beam's early-depth advantage observed on gaussian_blur/flash_attention/levenshtein_distance; the archive and reseeding import evolutionary's crash robustness observed on bvh_raytracer/regex_engine; and the constants (w=1, c=1/2, improvement window of 5, Tw=150, repair bound, reseed ε) are hand-set. With roughly four task archetypes in the suite (saturated, fast-plateau, crash-prone, refinement-friendly), a portfolio with a beam-like opening and an evolutionary-like archive may be near-oracle on these tasks by construction. The inference drawn — 'the signals needed to alloca
  2. [§6, Fig. 4a and Appendix A] The five levers (archive+reseed, reseeding rule, crashed-parent adoption, merge softening, warm-start ramp) are said to be 'separately identifiable' because disabling all of them reduces fluid to the bare bandit forest — yet no ablation is reported. As a result the reader cannot attribute the 0.780 to the UCB core, to the imported levers, or to any single mechanism, and cannot tell whether the hand-set constants matter. At minimum: the bare-bandit baseline (all levers off), fluid-minus-warm-start, and fluid-minus-archive/reseed, plus a sensitivity sweep over w, c, the ρ window, and Tw=150. Given the in-sample design concern above, ablations are the cheapest available evidence that the trajectory (rather than the designer) is doing the allocation work.
  3. [Fig. 4b/c, §6] Two stacked sources of optimism affect the 0.780 vs. 0.784 comparison. (1) The oracle selects, per task, the fixed policy with the highest mean normalized AUC on the same three-seed runs — a winner's-curse inflation over 12 tasks × 4 policies with only 3 seeds each; a split-seed oracle (select on seeds 1–2, evaluate on seed 3) would quantify this. (2) The gap of 0.004 is far smaller than the ±1 SE bands shown in Fig. 4b/c, and no per-task paired comparison (fluid vs. oracle, or fluid vs. best fixed policy, with task-level error) is given; the 'essentially matching' conclusion currently rests on a difference well inside noise, which is fine for a non-inferiority reading but should be argued as such explicitly, with a per-task table and ideally more seeds on the tasks where the margins are decisive (gaussian_blur, z_order_range_scan, regex_engine, bvh_raytracer).
minor comments (6)
  1. [Table 1 / §4] Table 1 defines the task ceiling R⋆ as 'the best reward attained on the task by any policy and seed,' i.e., an observed-ceiling statistic computed on the same runs being normalized. This is acceptable as a descriptive scale, but it is itself a max over few samples and injects mild winner's-curse into every normalized number in Fig. 4. A sentence acknowledging this (or recomputing R⋆ from a disjoint seed) would suffice.
  2. [Fig. 3b] Pearson r = 0.40 for the ˆR vs. ¯R/ˆR decomposition is reported without an n, confidence interval, or p-value; with at most 48 task-policy points this should be stated so readers can judge how far from collinear the axes actually are.
  3. [§6, footnote 4] Footnote 4 notes the harness keeps up to 32 jobs in flight across 64 chains and that 'concurrency affects wall-clock throughput but not the evaluation budget.' Since the paper's framing is evaluation-count-as-budget, a brief note on whether asynchronous completion order interacts with the UCB selection (stale ρk for in-flight chains) would help reproducibility.
  4. [References] Reference [12] (Weng, 'LLM Powered Autonomous Agents') is a blog post cited alongside archival venues; consider citing a peer-reviewed survey of LLM agents instead or in addition.
  5. [Figures 2 and 4] Fig. 2: the per-panel x-axis range labels ('0 250 500') and the truncated panel titles in the rendered figure are hard to read at print size; consider consistent axis ticks and a shared legend. Also, Fig. 4c lists per-policy means (0.567–0.780) inside the caption rather than on the bars, which is easy to miss.
  6. [Abstract / §1] Abstract and §1: 'closely matching' / 'nearly matching' the oracle appears before the reader has any uncertainty quantification; consider stating the paired per-task gap or the SE once the statistical comment above is addressed.

Circularity Check

0 steps flagged

Empirical methods paper: AUC/oracle/fluid numbers are measured traces, not quantities forced equal to inputs by construction; design-on-same-suite is a validity concern, not derivation circularity.

full rationale

The load-bearing claims are empirical comparisons of search policies on twelve AutoLab tasks under a fixed proposer, budget, and frontier width. R(n)=max_{i≤n} r_i and AUC reward R̄=(1/N)Σ R(n) are descriptive summaries of observed evaluation traces; they are not fitted targets renamed as predictions, nor defined in terms of the quantity being ‘derived.’ The per-task oracle is explicitly a retrospective selector over the same fixed-policy runs and is labeled as such, so matching it is a benchmark comparison, not a self-definitional identity. fluid’s UCB allocation (ρ_k, q_k, exploration bonus) and its levers (warm-start ramp, archive, crash adoption) are hand-designed mechanisms whose performance is then measured by running the algorithm; the reported 0.780 vs 0.784 figures are experimental outcomes, not algebraic consequences of the design equations. Self-citation to Cao & Yang [42] (autonomy ladder) is contextual discussion only and does not underwrite the efficiency numbers. The skeptic’s point—that levers were motivated by fixed-policy results on the same twelve tasks, so near-oracle behavior may partly encode suite-specific design knowledge—is a real in-sample / generalization validity concern, but it does not reduce the central claim to an input-by-construction equality under the circularity taxonomy. No uniqueness theorem, ansatz-via-self-citation, or fitted-parameter-as-prediction chain is load-bearing. Score 1 only to mark that mild design-on-evaluation-suite dependence, not mathematical circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

Load-bearing content is experimental design plus classical search/bandit machinery, not new physics-style axioms. The claim rests on treating evaluation count as the budget, fixing a single-completion LLM proposer, AutoLab’s task-relative rewards, and several hand-chosen fluid knobs. Invented ‘entities’ are algorithmic constructs (fluid, AUC reward as AR objective), not ontological additions.

free parameters (6)
  • UCB incumbent weight w = 1
    Fixed to 1 in Uk = ρk + w qk + c sqrt(ln(n+1)/(nk+1)); controls quality vs recent improvement tradeoff in allocation.
  • UCB exploration constant c = 1/2
    Fixed to 1/2; sets exploration bonus strength across chains.
  • Warm-start pool M0 and threshold Tw = M0=16, Tw=150, M=64
    Active chains restricted to 16 until evaluation 150, then up to 64; chosen to imitate beam’s early depth compounding.
  • Evaluation budget B and frontier width = B=500, width=64
    All methods compared at 500 evaluations and matched width 64; shapes AUC and which policy looks best in late budget.
  • Improvement-rate window and ρ prior = window=5, prior=1/2
    ρk from five most recent evaluations; unevaluated/reseeded chains use ρ=1/2. Window length and prior are design choices affecting bandit scores.
  • Crash-repair bound, stagnation/reseed rules, merge-softening persistence, archive capacity, reseed ε = partially specified in Alg. 1
    Multiple discrete levers in §6/Alg. 1 (bounded crashed-parent adoption, reseeding on crash/stagnation, delayed duplicate collapse, archive sampling) are hand-specified; values are only partly quantified in text.
axioms (5)
  • domain assumption The dominant research budget is the number of verifier evaluations n, and policies are comparable at fixed n.
    Stated in §3.1 and footnote 1; excludes token/wall-clock-primary regimes except as secondary.
  • domain assumption A single LLM completion conditioned on task and history is an adequate proposal operator; multi-turn agentic loops can be stripped out to isolate search structure.
    §2 and §5 experimental protocol; core to attributing differences to search structure alone.
  • domain assumption AutoLab task-relative log-scaled rewards (baseline 0, reference 0.5, crashes gated) are a valid utility for ranking search efficiency.
    §5 and footnote 3; all AUC/best metrics inherit this scale before R★ normalization.
  • ad hoc to paper UCB-style indices on recent improvement and normalized incumbent quality are appropriate for allocating evaluations across asynchronous hill-climbing chains.
    §6 formula; classical UCB motivation, but the specific two-exploitation-term form and crash-regime qk=0 rule are design choices of fluid.
  • standard math Standard facts of multi-armed bandits, multi-start local search, and algorithm portfolios apply when arms are LLM-driven chains.
    Citations [26–31]; used as background, not re-proved.
invented entities (2)
  • fluid search (budget-adaptive portfolio bandit over a forest of LLM hill-climbing chains) no independent evidence
    purpose: Allocate a fixed evaluation budget online without committing to one classical search structure in advance.
    Primary methodological object of §6 and Algorithm 1; performance is the headline empirical result.
  • AUC reward ¯R as primary AR search-efficiency objective (with convergence ratio ¯R/ˆR) independent evidence
    purpose: Scalarize the Pareto frontier R(n) so anytime efficiency is optimized and reported alongside final ˆR.
    Table 1 and §4; AUC itself is prior art, but elevating it as a first-class AR metric is the paper’s evaluation proposal.

pith-pipeline@v1.2.0-grok45-kimik3 · 19195 in / 4157 out tokens · 80484 ms · 2026-07-31T09:09:57.671227+00:00 · methodology

0 comments
read the original abstract

AI-driven autonomous research (AR) systems are becoming increasingly effective across a broad range of tasks. Their performance, however, is still evaluated primarily by the quality of the final outcome. In this paper, we argue that the efficiency of the solution-search process is an equally important but often overlooked dimension of performance. A strong AR system should not only produce high-quality results, but also reach them with as small a budget as possible. Search efficiency will become increasingly important as AR expands from domains with inexpensive verification, such as mathematics and coding, to real-world scientific settings in which solution evaluation may require costly physical experiments. To capture this dimension, we propose evaluating AR systems using the area under the curve (AUC) of the Pareto frontier, alongside final outcome quality. We compare several families of search algorithms, including hill climbing, beam search, tree search, and evolutionary search, across twelve systems-optimization tasks. We find that no single search structure is consistently the most efficient. We also show that search efficiency and final outcome quality are distinct performance dimensions: a method that eventually achieves the best result may nevertheless improve slowly and consume substantially more evaluation budget before reaching that result. Because the most effective search policy is generally unknown in advance, we introduce an adaptive procedure called fluid search, which uses a portfolio bandit to dynamically allocate a fixed evaluation budget across a forest of search processes. Across the evaluated tasks, fluid search achieves the highest overall search efficiency, closely matching the performance of a per-task oracle that is given the best search structure for each task in advance.

Figures

Figures reproduced from arXiv: 2607.24647 by Haiqian Yang, Yuan Cao.

Figure 1
Figure 1. Figure 1: The two scalar summaries of a Pareto frontier of research quality, on two hypothetical systems that end at the same peak. (left) Frontier reward R(n): the systems reach the same final outcome at Rˆ = R(N), but with the same budget System A achieves superior solutions earlier than B. (right) The AUC reward R¯, the mean height of the area under each frontier (dashed lines), separates them: R¯A > R¯B because … view at source ↗
Figure 2
Figure 2. Figure 2: Frontier reward R(n) versus evaluation budget n for four fixed search policies across twelve AR tasks (mean over three seeds; the dotted line marks the reference level). Search strategy strongly shapes the Pareto frontier: policies that approach a similar endpoint can differ sharply in how quickly they reach it, and no single policy leads on every task. typically score partial reasoning states using a lear… view at source ↗
Figure 3
Figure 3. Figure 3: The four fixed policies across the twelve tasks (colors shared across panels). (a) AUC reward R¯ per task (mean over three seeds), rows grouped by the winning policy (black ring); overlapping dots are near-ties, and the long rows are the tasks where the choice of search structure matters most. (b) Best reward Rˆ versus convergence ratio R/¯ Rˆ. The markers indicate the metrics per task and policy. We evalu… view at source ↗
Figure 4
Figure 4. Figure 4: The mechanism and efficiency of the proposed fluid search policy. (a) Each row is a hill-climbing chain, a staircase rises when its incumbent improves (dots: evaluations; ×: crashes). The arrow on top indicates the shared budget. At each step the UCB bandit fills one open slot (faint rings), so favored chains become denser while neglected ones stay sparse. Two levers are imported from fixed policies: orang… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 2 canonical work pages

  1. [1]

    Cottereau, Ziwei Liu, Tat-Seng Chua, and Wei Tsang Ooi

    Lingdong Kong, Xian Sun, Wei Chow, Linfeng Li, Kevin Qinghong Lin, Xuan Billy Zhang, Song Wang, Rong Li, Qing Wu, Wei Gao, Yingshuo Wang, Shaoyuan Xie, Jiachen Liu, Leigang Qu, Shijie Li, Lai Xing Ng, Benoit R. Cottereau, Ziwei Liu, Tat-Seng Chua, and Wei Tsang Ooi. AI for Auto-Research: Roadmap & User Guide. arXiv preprint arXiv:2605.18661, 2026. URL htt...

    Pith/arXiv arXiv 2026

  2. [2]

    Efficient and Robust Automated Machine Learning.Advances in Neural Information Processing Systems, 28:2962–2970, 2015

    Matthias Feurer, Aaron Klein, Katharina Eggensperger, Jost Springenberg, Manuel Blum, and Frank Hutter. Efficient and Robust Automated Machine Learning.Advances in Neural Information Processing Systems, 28:2962–2970, 2015. URLhttps://proceedings.neurips. cc/paper/2015/hash/11d0e6287202fced83f79975ec59a3a6-Abstract.html

    2015

  3. [3]

    Using Anytime Algorithms in Intelligent Systems.AI Magazine, 17(3): 73–83, 1996

    Shlomo Zilberstein. Using Anytime Algorithms in Intelligent Systems.AI Magazine, 17(3): 73–83, 1996. URLhttp://rbr.cs.umass.edu/papers/Zaimag96.pdf

    1996

  4. [4]

    Wenhao Gao, Tianfan Fu, Jimeng Sun, and Connor W. Coley. Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization. arXiv preprint arXiv:2206.12411, 2022. URL https://arxiv.org/abs/2206.12411

    Pith/arXiv arXiv 2022

  5. [5]

    AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks? arXiv preprint arXiv:2606.05080, 2026

    Zhangchen Xu, Junda Chen, Yue Huang, Dongfu Jiang, Jiefeng Chen, Hang Hua, Zijian Wu, Zheyuan Liu, Zexue He, Lichi Li, Shizhe Diao, Jiaxin Pei, Jinsung Yoon, Hao Zhang, Mengdi Wang, Radha Poovendran, Misha Sra, Alex Pentland, and Zichen Chen. AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks? arXiv preprint arXiv:2606.050...

    Pith/arXiv arXiv 2026

  6. [6]

    Joseph Hoane, Jr., and Feng-hsiung Hsu

    Murray Campbell, A. Joseph Hoane, Jr., and Feng-hsiung Hsu. Deep Blue.Artificial Intelli- gence, 134(1-2):57–83, 2002. doi:10.1016/S0004-3702(01)00129-1

    doi:10.1016/s0004-3702(01)00129-1 2002

  7. [7]

    David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Master- ing the...

  8. [8]

    NeuralArchitectureSearch: ASurvey

    ThomasElsken, JanHendrikMetzen, andFrankHutter. NeuralArchitectureSearch: ASurvey. arXiv preprint arXiv:1808.05377, 2019. URLhttps://arxiv.org/abs/1808.05377

    Pith/arXiv arXiv 2019

  9. [9]

    John R. Koza. Genetic programming as a means for programming computers by natural selection.Statistics and Computing, 4(2):87–112, 1994. doi:10.1007/BF00175355

    doi:10.1007/bf00175355 1994

  10. [10]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475, 2024. doi:10.1038/ s41586-023-06924-6

    2024

  11. [11]

    Alexander Novikov, Ngân V˜ u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Ab- bas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent for scientific and al...

    Pith/arXiv arXiv 2025

  12. [12]

    LLM Powered Autonomous Agents

    Lilian Weng. LLM Powered Autonomous Agents. Blog post,https://lilianweng.github. io/posts/2023-06-23-agent/, 2023

    2023

  13. [13]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv preprint arXiv:2210.03629, 2023. URLhttps://arxiv.org/abs/2210.03629

    Pith/arXiv arXiv 2023

  14. [14]

    Zhe Ren, Yimeng Chen, Dandan Guo, Guowei Rong, Tonghui Li, R. B. Xiong, Qingfeng Lan, Wenyi Wang, Li Nanbo, Yibo Yang, Mingchen Zhuge, and Jürgen Schmidhuber. Self- Improvements in Modern Agentic Systems: A Survey. arXiv preprint arXiv:2607.13104, 2026. URLhttps://arxiv.org/abs/2607.13104

    Pith/arXiv arXiv 2026

  15. [15]

    Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan

    Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can Language Models Resolve Real-World GitHub Issues? arXiv preprint arXiv:2310.06770, 2024. URLhttps://arxiv.org/abs/2310.06770

    Pith/arXiv arXiv 2024

  16. [16]

    MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering

    Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Lilian Weng, and Aleksander Mądry. MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering. arXiv preprint arXiv:2410.07095, 2025. URLhttps://arxiv.org/abs/2410.07095

    Pith/arXiv arXiv 2025

  17. [17]

    Zhang, William Hu, Christopher Ré, and Azalia Mirhoseini

    Anne Ouyang, Simon Guo, Simran Arora, Alex L. Zhang, William Hu, Christopher Ré, and Azalia Mirhoseini. KernelBench: Can LLMs Write Efficient GPU Kernels? arXiv preprint arXiv:2502.10517, 2025. URLhttps://arxiv.org/abs/2502.10517. 13

    Pith/arXiv arXiv 2025

  18. [18]

    RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts

    Hjalmar Wijk, Tao Lin, Joel Becker, Sami Jawhar, Neev Parikh, Thomas Broadley, Lawrence Chan, Michael Chen, Josh Clymer, Jai Dhyani, Elena Ericheva, Katharyn Garcia, Brian Goodrich, Nikola Jurkovic, Holden Karnofsky, Megan Kinniment, Aron Lajko, Seraphina Nix, Lucas Sato, William Saunders, Maksym Taran, Ben West, and Elizabeth Barnes. RE-Bench: Evaluating...

    Pith/arXiv arXiv 2025

  19. [19]

    Miller, Abhishek Charnalia, Derek Dun- field, Carole-JeanWu, PontusStenetorp, NicolaCancedda, JakobNicolausFoerster, andYoram Bachrach

    Edan Toledo, Karen Hambardzumyan, Martin Josifoski, Rishi Hazra, Nicolas Baldwin, Alexis Audran-Reiss, Michael Kuchnik, Despoina Magka, Minqi Jiang, Alisia Maria Lupidi, An- drei Lupu, Roberta Raileanu, Kelvin Niu, Tatiana Shavrina, Jean-Christophe Gagnon-Audet, Michael Shvartsman, Shagun Sodhani, Alexander H. Miller, Abhishek Charnalia, Derek Dun- field,...

    arXiv 2025

  20. [20]

    AIDE: AI-Driven Exploration in the Space of Code

    Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, and Yuxiang Wu. AIDE: AI-Driven Exploration in the Space of Code. arXiv preprint arXiv:2502.13138, 2025. URLhttps://arxiv.org/abs/2502.13138

    Pith/arXiv arXiv 2025

  21. [21]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. arXiv preprint arXiv:2305.10601, 2023. URLhttps://arxiv.org/abs/2305.10601

    Pith/arXiv arXiv 2023

  22. [22]

    Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models. arXiv preprint arXiv:2310.04406, 2024. URLhttps://arxiv.org/abs/2310.04406

    Pith/arXiv arXiv 2024

  23. [23]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv preprint arXiv:2303.11366, 2023. URLhttps://arxiv.org/abs/2303.11366

    Pith/arXiv arXiv 2023

  24. [25]

    Scaling LLM Test-Time Com- pute Optimally can be More Effective than Scaling Model Parameters

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling LLM Test-Time Com- pute Optimally can be More Effective than Scaling Model Parameters. arXiv preprint arXiv:2408.03314, 2024. URLhttps://arxiv.org/abs/2408.03314

    Pith/arXiv arXiv 2024

  25. [26]

    Rafael Martí, Mauricio G. C. Resende, and Celso C. Ribeiro. Multi-start methods for com- binatorial optimization.European Journal of Operational Research, 226(1):1–8, 2013. doi: 10.1016/j.ejor.2012.10.012

    doi:10.1016/j.ejor.2012.10.012 2013

  26. [27]

    Lourenço, Olivier C

    Helena R. Lourenço, Olivier C. Martin, and Thomas Stützle. Iterated Local Search. In Fred W. Glover and Gary A. Kochenberger, editors,Handbook of Metaheuristics, volume 57 ofInterna- tional Series in Operations Research & Management Science, pages 320–353. Kluwer Academic Publishers, 2003. doi:10.1007/0-306-48056-5_11. 14

    doi:10.1007/0-306-48056-5_11 2003

  27. [28]

    Feo and Mauricio G

    Thomas A. Feo and Mauricio G. C. Resende. Greedy Randomized Adaptive Search Procedures. Journal of Global Optimization, 6(2):109–133, 1995. doi:10.1007/BF01096763

    doi:10.1007/bf01096763 1995

  28. [29]

    John R. Rice. The Algorithm Selection Problem. InAdvances in Computers, volume 15, pages 65–118. Elsevier, 1976. doi:10.1016/S0065-2458(08)60520-3

    doi:10.1016/s0065-2458(08)60520-3 1976

  29. [30]

    Gomes and Bart Selman

    Carla P. Gomes and Bart Selman. Algorithm portfolios.Artificial Intelligence, 126(1-2):43–62,

  30. [31]

    Finite-time Analysis of the Multiarmed Bandit Problem.Machine Learning, 47(2-3):235–256, 2002

    Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer. Finite-time Analysis of the Multiarmed Bandit Problem.Machine Learning, 47(2-3):235–256, 2002. doi:10.1023/A:1013689704352

    doi:10.1023/a:1013689704352 2002

  31. [32]

    Optimal Speedup of Las Vegas Algo- rithms.Information Processing Letters, 47(4):173–180, 1993

    Michael Luby, Alistair Sinclair, and David Zuckerman. Optimal Speedup of Las Vegas Algo- rithms.Information Processing Letters, 47(4):173–180, 1993. doi:10.1016/0020-0190(93) 90029-9

    work page doi:10.1016/0020-0190(93 1993

  32. [33]

    Hy- perband: A Novel Bandit-Based Approach to Hyperparameter Optimization

    Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hy- perband: A Novel Bandit-Based Approach to Hyperparameter Optimization. arXiv preprint arXiv:1603.06560, 2018. URLhttps://arxiv.org/abs/1603.06560

    Pith/arXiv arXiv 2018

  33. [34]

    Hoffman, and Nando de Freitas

    Eric Brochu, Matthew W. Hoffman, and Nando de Freitas. Portfolio Allocation for Bayesian Optimization. arXiv preprint arXiv:1009.5419, 2011. URLhttps://arxiv.org/abs/1009. 5419

    Pith/arXiv arXiv 2011

  34. [35]

    WiderorDeeper? ScalingLLMInference-TimeComputewithAdaptiveBranchingTreeSearch

    Yuichi Inoue, Kou Misaki, Yuki Imajuku, So Kuroki, Taishi Nakamura, and Takuya Akiba. WiderorDeeper? ScalingLLMInference-TimeComputewithAdaptiveBranchingTreeSearch. arXiv preprint arXiv:2503.04412, 2025. URLhttps://arxiv.org/abs/2503.04412

    arXiv 2025

  35. [36]

    Compute Allocation in Evolutionary Search: From Depth-Breadth to Multi-Armed Bandits

    Sixue Xing, Haoyu He, Kerui Wu, Zhuo Yang, Haozheng Luo, Tianfan Fu, and Aarthy Na- garajan. Compute Allocation in Evolutionary Search: From Depth-Breadth to Multi-Armed Bandits. arXiv preprint arXiv:2605.29268, 2026. URLhttps://arxiv.org/abs/2605.29268

    Pith/arXiv arXiv 2026

  36. [37]

    ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution

    Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin. ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution. arXiv preprint arXiv:2509.19349, 2025. URLhttps: //arxiv.org/abs/2509.19349

    Pith/arXiv arXiv 2025

  37. [38]

    Russell and Peter Norvig.Artificial Intelligence: A Modern Approach

    Stuart J. Russell and Peter Norvig.Artificial Intelligence: A Modern Approach. Pearson, Hoboken, NJ, 4 edition, 2020. ISBN 9780134610993

    2020

  38. [39]

    Conjecture Machines: AI Agents and the New Validation Bottleneck in Science

    Don Wallace, Conor Griffin, Sean O’Neill, Thang Luong, and Owen Larter. Conjecture Machines: AI Agents and the New Validation Bottleneck in Science. Google DeepMind,https://deepmind.google/public-policy/ conjecture-machines-ai-agents-and-the-new-validation-bottleneck-in-science/, 2026

    2026

  39. [40]

    The Harness Effect: How Orchestration Design Sets the 15 Token Economics of Enterprise Agentic AI

    Muayad Sayed Ali, Aliaksandra Novik, Anji Boddupally, Artem Yavorskyi, Chris Nickerson, Daniel Rica, Emily DuGranrut, Felix Leung, Garrett Prince, Grace Barnett, Heath Robin- son, Hosain Al Ahmad, Jesse Resnick, Juan Carlos Farah, Jyothi Swaroop Meruga, Leonid Kuznetsov, Luke Gorham, Marie Schmoll, Michael Paciullo, Saumya Das, Sharath Sheri- pally, Tommy...

    Pith/arXiv arXiv 2026

  40. [42]

    Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI

    Yuan Cao and Haiqian Yang. Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI. arXiv preprint arXiv:2607.09560, 2026. URLhttps://arxiv.org/ abs/2607.09560. 16 A ThefluidSearch Algorithm Algorithm 1fluid: Budget-Adaptive Portfolio Search Require:Baselinex 0, budgetB, chain countsM 0 ≤M, warm-start thresholdT w Ensure:Best valid...

    Pith/arXiv arXiv 2026

  41. [2001]

    doi:10.1016/S0004-3702(00)00081-3

    work page doi:10.1016/s0004-3702(00)00081-3

  42. [2016]

    doi:10.1038/nature16961

    doi:10.1038/nature16961

  43. [2023]

    URLhttps://arxiv.org/abs/2303.17651

    Pith/arXiv arXiv

  44. [2026]

    URLhttps://arxiv.org/abs/2607.18235

    Pith/arXiv arXiv