pith. sign in

arxiv: 2606.21891 · v1 · pith:HY5JINA4new · submitted 2026-06-20 · 💻 cs.AI · cs.CL

Learning the ARTS of Search for Automated Discovery

Gurusha Juneja , Arnav Kumar Jain , Deepak Nathani , William Yang Wang , Xin Eric Wang This is my paper

Pith reviewed 2026-06-26 12:01 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords automated discoverytree searchreasoning language modelstest-time traininghypothesis searchagentic reasoningMLGymMLEBench
0
0 comments X

The pith

A reasoning language model improves hypothesis search by diagnosing execution logs to separate faulty implementations from bad ideas.

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

Current tree-search methods for automated discovery rank hypotheses partly by how well they were executed, so a strong idea with a buggy run can lose to a weak idea with polished code. ARTS instead has the language model read the full execution history, decide whether each past failure came from the hypothesis itself or from its implementation, and pick the next step accordingly. Test-time training embeds the growing search tree directly into the model's weights so the full history never exceeds the context limit. On 22 tasks the approach raises normalized scores by more than 15 percent over prior leaders, and a 4-billion-parameter open model reaches parity with much larger closed models at up to five times lower inference cost while recovering solutions that standard heuristics discard.

Core claim

ARTS deploys a reasoning language model to inspect prior execution logs, diagnose whether earlier failures arose from faulty implementations or bad hypotheses, and select the hypothesis to build on next. Test-time training instills the knowledge of the search tree in the model weights to handle growing context. Across 22 tasks from MLGym and MLEBench, ARTS outperforms leading algorithms with over 15.3 percent relative improvement in the normalized score. With test-time training a Qwen3-4B agent matches closed-source frontier models like Gemini-3 Pro and GPT o3-reasoning with up to 5x lower inference cost. On partially observable RL tasks the test-time trained Qwen3-4B scientist surpasses ART

What carries the argument

Agentic Reasoning for Tree Search (ARTS): a reasoning language model that reads full execution logs to diagnose whether failures stem from bad hypotheses or faulty implementations, then chooses the next hypothesis, with test-time training used to encode the search tree into the model weights.

If this is right

  • ARTS raises normalized scores by more than 15.3 percent over leading algorithms on 22 tasks from MLGym and MLEBench.
  • A 4B-parameter open model with test-time training reaches the performance of closed-source frontier models at up to 5x lower inference cost.
  • On partially observable RL tasks the same small model rediscovers the human-best recurrent-memory solution that standard heuristics discard.

Where Pith is reading between the lines

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

  • The diagnosis step could be added to other iterative search procedures that currently rely only on reward signals.
  • Test-time training of search history may prove useful for any long-horizon reasoning task that outgrows fixed context windows.
  • Similar log-based diagnosis might reduce wasted experiments in non-ML domains such as automated chemistry or materials design.

Load-bearing premise

A reasoning language model can accurately tell from execution logs alone whether a previous failure was caused by a bad hypothesis or merely by a faulty implementation.

What would settle it

On the same 22 tasks, replace the model's diagnosis step with random selection of the next hypothesis and measure whether the 15.3 percent normalized-score gain disappears.

Figures

Figures reproduced from arXiv: 2606.21891 by Arnav Kumar Jain, Deepak Nathani, Gurusha Juneja, William Yang Wang, Xin Eric Wang.

Figure 1
Figure 1. Figure 1: A single search step of ARTS. A node is one validated experiment with a hypothesis, code, logs, and score. Starting from the current search tree (Tree to expand), the scientist inspects the candidate nodes and reasons about why each scored as it did, then selects the most promising node rather than the one with the highest score. In this figure, it picks the node with score 0.10 since the low score comes f… view at source ↗
Figure 2
Figure 2. Figure 2: Aggregate normalized performance. We report the reliable metrics [1] and observe that ARTS significantly outperforms baselines at Mean and IQM while achieving lower optimality gap. 40 GB A100 GPU to train any downstream model during the search process. We provide details about implementation, prompts, finetuning with RL, and compute in Appendix I. To compute the normalized score for rliable [1] we use the … view at source ↗
Figure 3
Figure 3. Figure 3: Hourly progression of normalized score on randomly selected tasks under the same 8-hour [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Swap ablations. (a) Executor swap with the scientist fixed to o3. (b) Scientist swap with [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: HMS Brain Activity search trees. ARTS (top) drafts a pretrained ResNet-50 over 3-channel log-mel spectrograms at depth 1 and reaches KL 0.557. The next two children regress (EfficientNet￾B0, 0.702; SpecAugment, 0.565). ARTS re-reads the regressed nodes’ training logs, classifies them as implementation-wrong (under-trained under heavy augmentation), retains the ResNet-50 + spectrogram family, and the next s… view at source ↗
Figure 6
Figure 6. Figure 6: Vesuvius Ink Detection search trees. ARTS (top) keeps five distinct axes as siblings of root, deepens the dominant lever (root_3, both fragments, 0.479) with augmentation to 0.575. AIRA (bottom) commits to the 3D-U-Net branch via UCB, abandons a stronger 2.5D-pretrained sibling at depth 1, and only rediscovers that family 12 nodes deeper at 0.510. margin notes give the scientist’s tool call (INSPECT/READ) … view at source ↗
Figure 7
Figure 7. Figure 7: MetaMaze search trees. ARTS (top, 7 nodes) recovers from a config-truncation crash via failure attribution, then climbs 15.73 → 48.57 on a pure-MLP path – the scientist explicitly identifies recurrent memory as the missing axis at steps 2, 3 and 5 but the verbalized sampler always selects a cheaper alternative. AIRA (bottom, 13 nodes) proposes a PPO+LSTM recurrent policy at root_0 and reaches 48.04 on firs… view at source ↗
Figure 8
Figure 8. Figure 8: MountainCar search trees. ARTS (top) makes three sequential moves — tune PPO h-params, add state-dependent action σ, add regularization — and reaches reward 95.73 at depth 4. AIRA (bottom) drafts three roots; one diverges catastrophically, UCB commits to the 38.16 draft, chains 17 further levels to 94.82. The table margin reflects baseline runs that frequently diverge. 22 [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 10
Figure 10. Figure 10: Token and component ablations. (a) Token usage estimated from saved trajectory text. (b) [PITH_FULL_IMAGE:figures/full_fig_p035_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: MetaMaze: AIRA, ARTS-o3, and ARTS+TTT search trees on the same task and budget. Nodes are coloured blue (baseline), purple (progress), teal (new best, ⋆), and red (regression / abandoned). AIRA picks LSTM at depth 1 (48.04) and never refines it. ARTS-o3 recovers from a crash via failure attribution and climbs 15.7 → 48.57 on an MLP path, but never lands LSTM. ARTS+TTT commits to the recurrent direction and… view at source ↗
Figure 11
Figure 11. Figure 11: Reward-function ablation for test-time training. Scores are normalized by the task baseline [PITH_FULL_IMAGE:figures/full_fig_p037_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Diversity after TTT. Left: number of strategy categories used out of 11. Right: early (blue) [PITH_FULL_IMAGE:figures/full_fig_p038_12.png] view at source ↗
read the original abstract

Scientific discovery can be formulated as an iterative search process over the space of hypotheses and experiments. Contemporary methods navigate this space using heuristics such as MCTS. These algorithms conflate the merit of a hypothesis with the quality of its experimental execution. A promising hypothesis with preliminary execution is therefore ranked below a modest hypothesis whose execution is refined. Moreover, prior methods prune the search logs as the search progresses because the accumulated history outgrows the context window. We propose Agentic Reasoning for Tree Search (ARTS), where we deploy a reasoning language model to navigate this space. The model inspects prior execution logs, diagnoses whether earlier failures arose from faulty implementations or bad hypotheses, and selects the hypothesis to build on next. To mitigate challenges with context length, ARTS uses test-time training to instill the knowledge of search tree in the model weights. Across 22 tasks from MLGym and MLEBench, we show that ARTS outperforms leading algorithms, with over 15.3% relative improvement in the normalized score. With test-time training we show that a Qwen3-4B agent can match performance with closed-source frontier models like Gemini-3 Pro and GPT o3-reasoning with upto 5x lower inference cost. We further observe that on partially observable RL tasks, the test-time trained Qwen3-4B scientist surpasses ARTS with the o3 scientist by rediscovering the human-best recurrent-memory solution that heuristic methods prune away.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes Agentic Reasoning for Tree Search (ARTS), in which a reasoning LM inspects execution logs to diagnose whether prior failures stem from faulty implementations or flawed hypotheses and then selects the next hypothesis to pursue. Test-time training is used to embed search-tree knowledge into model weights and mitigate context-length limits. On 22 tasks drawn from MLGym and MLEBench the method is reported to deliver >15.3% relative improvement in normalized score over leading baselines; a test-time-trained Qwen3-4B agent is claimed to match closed-source frontier models (Gemini-3 Pro, o3) at up to 5× lower inference cost and, on partially observable RL tasks, to rediscover human-best recurrent-memory solutions that heuristic search prunes.

Significance. If the performance claims prove robust, the separation of hypothesis quality from execution quality via explicit diagnosis would constitute a substantive advance over MCTS-style search for automated discovery. The demonstration that test-time training enables a 4B open model to match much larger closed models is potentially impactful for cost and accessibility. The absence of any quantitative validation of the diagnosis step, however, leaves the source of the reported gains unclear and therefore limits the assessed significance at present.

major comments (3)
  1. [Experimental results] Experimental results (abstract and §4): the central claims of 15.3% relative improvement and frontier-model parity rest on point estimates without error bars, statistical significance tests, or any description of how the 22 tasks were selected or how scores were normalized.
  2. [Experimental results] Experimental results (abstract and §4): no ablation isolates the contribution of the LM diagnosis step from the test-time training component, so it is impossible to determine whether the reported gains derive from the core innovation or from test-time training alone.
  3. [Method description] Method description (abstract and §3): the paper provides no quantitative evaluation of diagnosis quality (human agreement, accuracy on held-out logs, or correlation between diagnosis correctness and downstream success), which is the load-bearing assumption distinguishing ARTS from heuristic or MCTS baselines.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting needs for greater statistical rigor and validation of key components. We address each major comment below and will incorporate the suggested improvements in the revised manuscript.

read point-by-point responses

  1. Referee: Experimental results (abstract and §4): the central claims of 15.3% relative improvement and frontier-model parity rest on point estimates without error bars, statistical significance tests, or any description of how the 22 tasks were selected or how scores were normalized.

    Authors: We agree that the current results would benefit from additional statistical detail. In the revision we will report means and standard deviations across three independent runs per task (varying random seeds for both the search and test-time training), include Wilcoxon signed-rank tests for significance against baselines, describe the task selection (all 22 tasks from MLGym and MLEBench that provide executable environments and ground-truth metrics), and explicitly state the normalization procedure (per-task min-max scaling to [0,1] using the range of scores observed across all evaluated methods). revision: yes

  2. Referee: Experimental results (abstract and §4): no ablation isolates the contribution of the LM diagnosis step from the test-time training component, so it is impossible to determine whether the reported gains derive from the core innovation or from test-time training alone.

    Authors: This is a fair criticism. The revised §4 will contain two new ablation studies: (i) ARTS with the diagnosis step replaced by standard MCTS value estimation, and (ii) a non-ARTS baseline agent equipped with the same test-time training procedure. These comparisons will quantify the incremental benefit of the explicit diagnosis mechanism over test-time training alone. revision: yes

  3. Referee: Method description (abstract and §3): the paper provides no quantitative evaluation of diagnosis quality (human agreement, accuracy on held-out logs, or correlation between diagnosis correctness and downstream success), which is the load-bearing assumption distinguishing ARTS from heuristic or MCTS baselines.

    Authors: We accept that direct validation of the diagnosis step strengthens the central claim. The revision will add a new subsection in §3 reporting: (a) human evaluation on a random sample of 100 execution logs with two annotators, including Cohen’s kappa for inter-annotator agreement, (b) accuracy of the model’s diagnoses against the human labels, and (c) correlation between diagnosis correctness and subsequent improvement in normalized task score. These metrics will be presented alongside the end-to-end results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method evaluated on external fixed benchmarks

full rationale

The paper proposes the ARTS algorithm as an agentic search procedure that deploys a reasoning LM to inspect execution logs and diagnose failure causes, then reports normalized-score gains on 22 fixed tasks from MLGym and MLEBench against external baselines. No equations, derivations, or first-principles predictions appear; performance is measured by direct empirical comparison on held-out benchmarks whose success criteria are independent of the method's internal diagnosis step. No self-citation chains, fitted parameters renamed as predictions, or self-definitional loops are present. The central claim therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities. The central claim rests on the unstated assumption that execution logs contain sufficient signal for the LM to distinguish implementation faults from hypothesis quality.

pith-pipeline@v0.9.1-grok · 5798 in / 1324 out tokens · 26740 ms · 2026-06-26T12:01:37.515144+00:00 · methodology

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

71 extracted references · 33 canonical work pages · 18 internal anchors

  1. [1]

    Deep reinforcement learning at the edge of the statistical precipice

    Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, and Marc G Belle- mare. Deep reinforcement learning at the edge of the statistical precipice. Advances in Neural Information Processing Systems, 2021

    2021

  2. [2]

    The surprising effectiveness of test-time training for few-shot learning

    Ekin Akyürek, Mehul Damani, Adam Zweiger, Linlu Qiu, Han Guo, Jyothish Pari, Yoon Kim, and Jacob Andreas. The surprising effectiveness of test-time training for few-shot learning. In Forty-second International Conference on Machine Learning, 2025. URL https: //openreview.net/forum?id=asgBo3FNdg

    2025

  3. [3]

    Seungju Back, Dongwoo Lee, Naun Kang, Taehee Lee, S. K. Hong, Youngjune Gwon, and Sungjin Ahn. Understanding lora as knowledge memory: An empirical analysis, 2026. URL https://arxiv.org/abs/2603.01097

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  4. [4]

    Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes

    Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chemical research with large language models. Nature, 624:570–578, 2023

    2023

  5. [5]

    OpenAI Gym

    Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. URLhttps://arxiv.org/abs/1606.01540

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  6. [6]

    Toward understanding in-context vs

    Bryan Chan, Xinyi Chen, András György, and Dale Schuurmans. Toward understanding in-context vs. in-weight learning. In The Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=aKJr5NnN8U

    2025

  7. [7]

    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 M ˛ adry. Mle-bench: Evaluating machine learning agents on machine learning engineering, 2024. URL https://arxiv.org/abs/2410.07095

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    MARS: Modular Agent with Reflective Search for Automated AI Research

    Jiefeng Chen, Bhavana Dalvi Mishra, Jaehyun Nam, Rui Meng, Tomas Pfister, and Jinsung Yoon. Mars: Modular agent with reflective search for automated ai research, 2026. URL https://arxiv.org/abs/2602.02660

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Sela: Tree-search enhanced llm agents for automated machine learning, 2024

    Yizhou Chi, Yizhang Lin, Sirui Hong, Duyi Pan, Yaying Fei, Guanghao Mei, Bangbang Liu, Tianqi Pang, Jacky Kwok, Ceyao Zhang, Bang Liu, and Chenglin Wu. Sela: Tree-search enhanced llm agents for automated machine learning, 2024. URL https://arxiv.org/abs/ 2410.17238

    work page arXiv 2024

  10. [10]

    Arc prize 2024: Technical report, 2025

    Francois Chollet, Mike Knoop, Gregory Kamradt, and Bryan Landers. Arc prize 2024: Technical report, 2025. URLhttps://arxiv.org/abs/2412.04604

    work page arXiv 2024

  11. [11]

    Friedman, J

    Dean De Cock. Ames, iowa: Alternative to the boston housing data as an end of semester regression project. Journal of Statistics Education, 19(3), 2011. doi: 10.1080/10691898.2011. 11889627. URLhttps://doi.org/10.1080/10691898.2011.11889627

    work page doi:10.1080/10691898.2011 2011

  12. [12]

    Gemini 3 flash, 2025

    Google DeepMind. Gemini 3 flash, 2025. URL https://deepmind.google/models/ gemini/flash/

    2025

  13. [13]

    MLEvolve: An autonomous system for end-to-end machine learning algorithm design and optimization.https: //github.com/InternScience/MLEvolve, 2026

    Shangheng Du, Xiangchao Yan, Shiyang Feng, Bo Zhang, Lei Bai, et al. MLEvolve: An autonomous system for end-to-end machine learning algorithm design and optimization.https: //github.com/InternScience/MLEvolve, 2026. GitHub repository

    2026

  14. [14]

    FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI

    Elliot Glazer, Ege Erdil, Tamay Besiroglu, Diego Chicharro, Evan Chen, Alex Gunning, Caroline Falkman Olsson, Jean-Stanislas Denain, Anson Ho, Emily de Oliveira Santos, Olli Järviniemi, Matthew Barnett, Robert Sandler, Matej Vrzala, Jaime Sevilla, Qiuyu Ren, Elizabeth Pratt, Lionel Levine, Grant Barkley, Natalie Stewart, Bogdan Grechuk, Tetiana Grechuk, S...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Towards an AI co-scientist

    Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno, Khaled Saab, Dan Popovici, Jacob Blum, Fan Zhang, Katherine Chou, Avinatan Hassidim, Burak Gokturk, Amin Vahdat, Pushmeet Kohli, Yossi Matias, Andrew Carroll, Kavita Kulkarni, Nenad Tomasev, Yuan Guan, 12...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  16. [16]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, ...

    work page doi:10.1038/s41586-025-09422-z 2025

  17. [17]

    Llm-first search: Self-guided exploration of the solution space, 2025

    Nathan Herr, Tim Rocktäschel, and Roberta Raileanu. Llm-first search: Self-guided exploration of the solution space, 2025. URLhttps://arxiv.org/abs/2506.05213

    work page arXiv 2025

  18. [18]

    RULER: What's the Real Context Size of Your Long-Context Language Models?

    Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What’s the real context size of your long-context language models?, 2024. URLhttps://arxiv.org/abs/2404.06654

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=nZeVKeeFYf9

    2022

  20. [20]

    Mlagentbench: Evaluating language agents on machine learning experimentation, 2024

    Qian Huang, Jian V ora, Percy Liang, and Jure Leskovec. Mlagentbench: Evaluating language agents on machine learning experimentation, 2024. URL https://arxiv.org/abs/2310. 03302

    2024

  21. [21]

    Multi-turn code generation through single-step re- wards

    Arnav Kumar Jain, Gonzalo Gonzalez-Pumariega, Wayne Chen, Alexander M Rush, Went- ing Zhao, and Sanjiban Choudhury. Multi-turn code generation through single-step re- wards. In Forty-second International Conference on Machine Learning, 2025. URL https: //openreview.net/forum?id=aJeLhLcsh0

    2025

  22. [22]

    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, 2025. URL https: //arxiv.org/abs/2502.13138

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Titanic - machine learning from disaster, 2012

    Kaggle. Titanic - machine learning from disaster, 2012. URL https://www.kaggle.com/c/ titanic. Kaggle competition. Accessed: 2026-05-07. 13

    2012

  24. [24]

    Histopathologic cancer detection, 2018

    Kaggle. Histopathologic cancer detection, 2018. URL https://www.kaggle.com/ competitions/histopathologic-cancer-detection. Kaggle competition. Accessed: 2026-05-07

    2018

  25. [25]

    Toxic comment classification challenge, 2018

    Kaggle. Toxic comment classification challenge, 2018. URL https://www.kaggle.com/ competitions/jigsaw-toxic-comment-classification-challenge . Kaggle competi- tion. Accessed: 2026-05-07

    2018

  26. [26]

    APTOS 2019 blindness detection, 2019

    Kaggle. APTOS 2019 blindness detection, 2019. URL https://www.kaggle.com/ competitions/aptos2019-blindness-detection. Kaggle competition. Accessed: 2026- 05-07

    2019

  27. [27]

    RSNA-MICCAI brain tumor radiogenomic classifica- tion, 2021

    Kaggle. RSNA-MICCAI brain tumor radiogenomic classifica- tion, 2021. URL https://www.kaggle.com/competitions/ rsna-miccai-brain-tumor-radiogenomic-classification . Kaggle competition. Accessed: 2026-05-07

    2021

  28. [28]

    Spaceship titanic, 2022

    Kaggle. Spaceship titanic, 2022. URL https://www.kaggle.com/competitions/ spaceship-titanic. Kaggle competition. Accessed: 2026-05-07

    2022

  29. [29]

    Vesuvius challenge - ink detection, 2023

    Kaggle. Vesuvius challenge - ink detection, 2023. URL https://www.kaggle.com/ competitions/vesuvius-challenge-ink-detection . Kaggle competition. Accessed: 2026-05-07

    2023

  30. [30]

    HMS - harmful brain activity classification, 2024

    Kaggle. HMS - harmful brain activity classification, 2024. URL https://www.kaggle.com/ competitions/hms-harmful-brain-activity-classification . Kaggle competition. Accessed: 2026-05-07

    2024

  31. [31]

    autoresearch: AI agents running research on single-GPU nanochat train- ing automatically

    Andrej Karpathy. autoresearch: AI agents running research on single-GPU nanochat train- ing automatically. https://github.com/karpathy/autoresearch, March 2026. GitHub repository

    2026

  32. [32]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Techni- cal report, University of Toronto, 2009. URL https://www.cs.toronto.edu/~kriz/ learning-features-2009-TR.pdf

    2009

  33. [33]

    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, 2025. URL https://arxiv.org/abs/2509.19349

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    I-mcts: Enhancing agentic automl via introspective monte carlo tree search, 2026

    Zujie Liang, Feng Wei, Wujiang Xu, Lin Chen, Yuxi Qian, and Xinhui Wu. I-mcts: Enhancing agentic automl via introspective monte carlo tree search, 2026. URL https://arxiv.org/ abs/2502.14693

    work page arXiv 2026

  35. [35]

    and Lin, Kevin and Hewitt, John and Paranjape, Ashwin and Bevilacqua, Michele and Petroni, Fabio and Liang, Percy

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173, 2024. doi: 10.1162/tacl_a_00638. URLhttps://aclanthology.org/2024.tacl-1.9/

    work page doi:10.1162/tacl_a_00638 2024

  36. [36]

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist: Towards fully automated open-ended scientific discovery, 2024. URL https: //arxiv.org/abs/2408.06292

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    Discoverybench: Towards data-driven discovery with large language mod- els

    Bodhisattwa Prasad Majumder, Harshit Surana, Dhruv Agarwal, Bhavana Dalvi Mishra, Abhijeetsingh Meena, Aryan Prakhar, Tirth V ora, Tushar Khot, Ashish Sabharwal, and Peter Clark. Discoverybench: Towards data-driven discovery with large language mod- els. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.n...

    2025

  38. [38]

    MLGym: A new framework and benchmark for advancing AI research agents

    Deepak Nathani, Lovish Madaan, Nicholas Roberts, Nikolay Bashlykov, Ajay Menon, Vincent Moens, Mikhail Plekhanov, Amar Budhiraja, Despoina Magka, Vladislav V orotilov, Gaurav Chaurasia, Dieuwke Hupkes, Ricardo Silveira Cabral, Tatiana Shavrina, Jakob Nicolaus Foerster, Yoram Bachrach, William Yang Wang, and Roberta Raileanu. MLGym: A new framework and ben...

  39. [39]

    URLhttps://openreview.net/forum?id=ryTr83DxRq. 14

  40. [40]

    Alexander Novikov, Ngân V˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas 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 algor...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Introducing openai o3 and o4-mini, Apr 2025

    OpenAI. Introducing openai o3 and o4-mini, Apr 2025. URL https://openai.com/index/ introducing-o3-and-o4-mini/

    2025

  42. [42]

    2024 , booktitle =

    Guilherme Penedo, Hynek Kydlíˇcek, Loubna Ben allal, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro V on Werra, and Thomas Wolf. The fineweb datasets: Decanting the web for the finest text data at scale. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems...

    work page doi:10.52202/079017-0970 2024

  43. [43]

    To backtrack or not to backtrack: When sequential search limits model reasoning, 2025

    Tian Qin, David Alvarez-Melis, Samy Jelassi, and Eran Malach. To backtrack or not to backtrack: When sequential search limits model reasoning, 2025. URL https://arxiv.org/ abs/2504.07052

    work page arXiv 2025

  44. [44]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. Gpqa: A graduate-level google-proof q&a benchmark, 2023. URLhttps://arxiv.org/abs/2311.12022

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  45. [45]

    Openevolve: an open-source evolutionary coding agent, 2025

    Asankhaya Sharma. Openevolve: an open-source evolutionary coding agent, 2025. URL https://github.com/algorithmicsuperintelligence/openevolve

    2025

  46. [46]

    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, 2023. URL https://arxiv.org/abs/2303.11366

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  47. [47]

    Towards execution-grounded automated ai research, 2026

    Chenglei Si, Zitong Yang, Yejin Choi, Emmanuel Candès, Diyi Yang, and Tatsunori Hashimoto. Towards execution-grounded automated ai research, 2026. URL https://arxiv.org/abs/ 2601.14525

    work page arXiv 2026

  48. [48]

    Learning to (Learn at Test Time): RNNs with Expressive Hidden States

    Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, and Carlos Guestrin. Learning to (learn at test time): Rnns with expressive hidden states, 2025. URL https: //arxiv.org/abs/2407.04620

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    Ghiringhelli, Takenori Yamamoto, Yury Lysogorskiy, Lars Blu- menthal, Thomas Hammerschmidt, Jacek R

    Christopher Sutton, Luca M. Ghiringhelli, Takenori Yamamoto, Yury Lysogorskiy, Lars Blu- menthal, Thomas Hammerschmidt, Jacek R. Golebiowski, Xiangyue Liu, Angelo Ziletti, and Matthias Scheffler. Crowd-sourcing materials-science challenges with the NOMAD 2018 Kag- gle competition. npj Computational Materials, 5:111, 2019. doi: 10.1038/s41524-019-0239-3. U...

    work page doi:10.1038/s41524-019-0239-3 2018

  50. [50]

    The plant pathology challenge 2020 data set to classify foliar disease of apples

    Ranjita Thapa, Kai Zhang, Noah Snavely, Serge Belongie, and Awais Khan. The plant pathology challenge 2020 data set to classify foliar disease of apples. Applications in Plant Sciences, 8(9):e11390, 2020. doi: https://doi.org/10.1002/aps3.11390. URL https: //bsapubs.onlinelibrary.wiley.com/doi/abs/10.1002/aps3.11390

    work page doi:10.1002/aps3.11390 2020

  51. [51]

    AI research agents for machine learning: Search, exploration, and generalization in MLE-bench

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

    2026

  52. [52]

    Amplifying human performance in combinatorial competitive programming, 2024

    Petar Veliˇckovi´c, Alex Vitvitskyi, Larisa Markeeva, Borja Ibarz, Lars Buesing, Matej Ba- log, and Alexander Novikov. Amplifying human performance in combinatorial competitive programming, 2024. URLhttps://arxiv.org/abs/2411.19744

    work page arXiv 2024

  53. [53]

    Group-evolving agents: Open-ended self-improvement via experience sharing

    Zhaotian Weng, Antonis Antoniades, Deepak Nathani, Zhen Zhang, Xiao Pu, and Xin Eric Wang. Group-evolving agents: Open-ended self-improvement via experience sharing, 2026. URLhttps://arxiv.org/abs/2602.04837

    work page arXiv 2026

  54. [54]

    Re-bench: Evaluating frontier ai r&d capabilities of language model agents against human experts, 2025

    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...

    work page arXiv 2025

  55. [55]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017

    Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. URL https://arxiv.org/abs/1708. 07747

    2017

  56. [56]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  57. [57]

    Tree of thoughts: Deliberate problem solving with large language models

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, volume 36, pages 11809–11822. Curran Associates, Inc.,...

    2023

  58. [58]

    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, 2023. URL https: //arxiv.org/abs/2210.03629

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  59. [59]

    MinAtar: An Atari-Inspired Testbed for Thorough and Reproducible Reinforcement Learning Experiments

    Kenny Young and Tian Tian. Minatar: An atari-inspired testbed for thorough and reproducible reinforcement learning experiments, 2019. URLhttps://arxiv.org/abs/1903.03176

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  60. [60]

    Learning to Discover at Test Time

    Mert Yuksekgonul, Daniel Koceja, Xinhao Li, Federico Bianchi, Jed McCaleb, Xiaolong Wang, Jan Kautz, Yejin Choi, James Zou, Carlos Guestrin, and Yu Sun. Learning to discover at test time, 2026. URLhttps://arxiv.org/abs/2601.16175

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  61. [61]

    Verbalized sampling: How to mitigate mode collapse and unlock LLM diversity, 2026

    Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael Tomz, Christopher D Manning, and Weiyan Shi. Verbalized sampling: How to mitigate mode collapse and unlock LLM diversity, 2026. URLhttps://openreview.net/forum?id=9jQkmGunGo

    2026

  62. [62]

    the next logical leap

    Stephen Zhang, Mustafa Khan, and Vardan Papyan. Attention sinks: A ’catch, tag, release’ mechanism for embeddings. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URLhttps://openreview.net/forum?id=r8UWp9JeJi. 16 Appendix Contents AHMS Search Trajectory:ARTSvs. AIRA . . . . . . . . . . . . . . . . . . . . . . . . . . ...

    2026

  63. [63]

    CanARTSseparate a promising hypothesis from a bad execution?

  64. [64]

    Does auditing identify the root cause of failed or regressed experiments?

  65. [65]

    Does the scientist reason from logs rather than only follow final scores?

  66. [66]

    Does reasoning-based parent selection spend budget better than heuristic selection?

  67. [67]

    Can the scientist propose directions that are not already in the tree?

  68. [68]

    Does verbalized sampling preserve enough diversity?

  69. [69]

    why” with “that

    How do baselines fail differently under the same budget? We answer each question separately below. F.1 CanARTSseparate a promising hypothesis from a bad execution? Answer: yes.The logs show that ARTS often preserves a promising hypothesis after a flawed implementation. On HMS Brain Activity, the task is to predict seizure-related activity from EEG spectro...

    2019

  70. [70]

    Lets you understand EXACTLY what was tried and why it succeeded or failed

    INSPECT nodes -- see the actual commands and output the executor ran for any node. Lets you understand EXACTLY what was tried and why it succeeded or failed

  71. [71]

    CatBoost (0.91) + LightGBM (0.90) plateau -- try feature engineering next

    READ files -- see the contents of any workspace file (e.g. target.py, evaluate.py, baseline.py, strategy.py). Lets you understand the task code, opponent strategy, evaluation logic, or data format before proposing an experiment. Respond in EXACTLY this format: INSPECT: node_id_1, node_id_2 [OR] INSPECT: NONE READ: filename1.py, filename2.py [OR] READ: NON...