Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Competition and Attraction Improve Model Fusion

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Model merging can evolve neural networks from random weights, a new evolutionary algorithm demonstrates.

desk verdict A genuinely new evolutionary-merging recipe whose strongest claim outruns its evidence: no control isolates whether SLERP crossover helps from scratch. read the letter →

arxiv 2508.16204 v1 pith:6RKLX7VI submitted 2025-08-22 cs.AI cs.NE

João Abrantes , Robert Tjarko Lange , Yujin Tang This is my paper
classification cs.AIcs.NE
keywords modelmergingevolutionaryalgorithmimplicitfitnesssharingdiversitypreservationSLERPMNISTlargelanguagemodelsdiffusion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes Model Merging of Natural Niches (M2N2), an evolutionary algorithm that treats model merging as an optimization over both mixing coefficients and merging boundaries. It claims, for the first time, that model merging can evolve a working neural network entirely from random initialization, matching CMA-ES on MNIST while using far fewer compute resources. When applied to large language models and text-to-image diffusion models, M2N2 merges specialized models into one that retains both skills, surpassing existing merging baselines. The reason to care: if true, gradient-free model fusion becomes a general tool for combining and even from-scratch training networks without backpropagation or access to training data.

What carries the argument

The central object is the split-point merging operator h_M2N2 (Eq. 2): it concatenates SLERP-interpolated parameter blocks before and after a random split point, so the merge boundary itself is part of the search. It is paired with a capacity-limited fitness (Eq. 3) that implements implicit fitness sharing without a hand-defined diversity metric, and an attraction score (Eq. 4) that scores a potential mate by how much it improves the first parent on contested resources. Together they convert model merging into an evolutionary search over both coefficients and boundaries.

What would settle it

On a small network, initialize two parents from very different random seeds, run M2N2 for many generations, and record whether merged offspring ever outlive the best parent; if the archive's merged models collapse into non-functional networks because SLERP crosses incompatible weight spaces, the central claim is refuted. Alternatively, replace SLERP with a non-linear interpolation that provably breaks mergeability and show M2N2's diversity mechanisms cannot compensate.

Watch

Extended reading notes

Core claim

M2N2 maintains an archive of models and repeatedly merges pairs using spherical linear interpolation (SLERP) at a randomly sampled split point. Two mechanisms make this work: limited-resource competition, where the fitness a model earns from a training example is discounted by how much the whole archive already solves that example, and an attraction heuristic that selects the second parent as the model that most improves the first parent's weak points. The paper reports that this setup evolves MNIST classifiers from random weights to test accuracy comparable to CMA-ES in about one fifteenth of the compute, and that on LLM and diffusion-model merging it beats fixed-boundary baselines, includi

Load-bearing premise

The load-bearing premise is that the SLERP split-point merge of two archive models keeps producing viable offspring; the paper offers no compatibility check, so the from-scratch and scale-up results rest on the unverified assumption that evolutionary pressure alone keeps models merge-compatible.

Editorial extensions

If this is right

  • Model merging is no longer only a post-hoc combination of pretrained weights; it can serve as an optimization loop that produces functional models from random initialization.
  • Dynamic merging boundaries matter more than coefficient search alone: on the math+agentic LLM merge, M2N2 without split-point drops from 63.49 to 61.15 average, and CMA-ES's per-layer coefficients score 49.06.
  • The resource-competition fitness can replace hand-designed diversity metrics, which the paper shows with MAP-Elites retaining unhelpful low performers on MNIST.
  • Because the diffusion merged model keeps English image understanding despite being evolved solely on Japanese fitness, merging here avoids catastrophic forgetting without access to original training data.
  • The attraction heuristic is cheap and transferable, suggesting mate-selection strategies become important as crossover cost grows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If evolutionary pressure alone keeps co-evolving models merge-compatible, M2N2 could be used to co-adapt populations of large models without adding a compatibility regularizer — a hypothesis the paper states as untested.
  • The competition mechanism is defined per training example, so it should transfer to any differentiable or non-differentiable per-sample scorer; a natural next test is applying it to reinforcement-learning or black-box tasks where per-sample rewards are available.
  • Because split points can be drawn per parameter block, M2N2 might extend to architectures with heterogeneous components (e.g., separate heads, adapters) where layer-aligned merging is meaningless.
  • The from-scratch result on a 19k-parameter network does not establish that merging can evolve very large models from scratch; that is an extrapolation, not a demonstrated claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces M2N2 (Model Merging of Natural Niches), an evolutionary model-merging method that removes fixed parameter-group boundaries by sampling a random split point and mixing ratio per merge, augments the fitness function with resource competition (Eq. 3/5), and pairs parents via an attraction heuristic (Eq. 4). The authors claim (a) the first demonstration that model merging can evolve models entirely from scratch, based on MNIST experiments; (b) that M2N2 scales to merging LLMs and diffusion-based image generation models, achieving state-of-the-art performance; and (c) that it preserves capabilities beyond the optimized fitness, such as bilingual image generation. The paper includes ablations for attraction, split-point, archive size, and competition intensity, and releases code.

Significance. If the claims hold, the paper makes a useful contribution to evolutionary model merging: a gradient-free, boundary-free fusion method with a principled diversity mechanism, applicable to models too large for gradient-based optimization. The experimental scope is broad (MNIST, 7B LLMs, SDXL-based diffusion models), the ablations are informative, and the code release is a concrete asset. The main limitations are not in the method's design but in the support for two headline claims: the from-scratch result lacks a mutation-only control, and the diffusion comparison optimizes a different objective from the CMA-ES baseline. These are fixable with additional experiments, and the core mechanism (competition plus pair selection plus flexible split points) appears plausible and worth publishing once the evidence is tightened.

major comments (3)
  1. [§4.1, Eq. 5 and Eq. 2] The from-scratch MNIST experiment is the sole support for the abstract's 'for the first time, model merging can be used to evolve models entirely from scratch.' The text states that all model-merging methods use the same Gaussian mutation and the same SLERP-with-split-point crossover, and that 'the split-point and attraction score have a minimal impact' from scratch. This leaves competition (Eq. 5) as the only identified differentiator between M2N2 and the GA baselines. Without a control that uses the same competition archive and mutation but disables the merge/crossover operator, the result may show only that a niching evolutionary algorithm with Gaussian mutation and a competition fitness can optimize a small MLP; it does not isolate the contribution of model merging. Please add a mutation-only control under M2N2's competition regime (or otherwise demonstrate that offspring produced by
  2. [§4.3, 'Baseline' paragraph and Table 2] The diffusion comparison is not apples-to-apples. The paper states that CMA-ES minimizes FID while M2N2 maximizes NCS: 'our method maximizes NCS instead of minimizing the Fréchet Inception Distance (FID) as they did.' Table 2 then reports both FID and NCS and presents M2N2 as superior on both. But the CMA-ES baseline was optimized for FID, not NCS, so its lower NCS is expected, and M2N2's slightly better FID despite not optimizing it is interesting but, without repeated runs or error bars, not conclusive. The 'state-of-the-art performance' claim in the abstract requires an equivalent-objective comparison: either run CMA-ES on the same per-sample NCS objective, or run M2N2 with an FID-based fitness, and report variances. As written, the scale-up claim for diffusion is not established.
  3. [§5, Limitations, and Eq. 2] The paper candidly notes that merging fails when fine-tuned models deviate significantly from their base, and hypothesizes that 'models with divergent state representations are incompatible for merging,' but no compatibility metric or regularization is introduced. Every call to Eq. 2 assumes that a SLERP of two archive members yields a functional offspring. The from-scratch and large-scale results rely on the untested assumption that evolutionary pressure alone keeps the archive merge-compatible. This is a load-bearing correctness risk for the method's general applicability. Please provide evidence that archive models remain merge-compatible over generations (e.g., track offspring viability or merge-fitness correlation as a function of generation), or discuss how the current experiments constrain this risk. A sentence in the limitations is not sufficient for a claim of general scalabilit
minor comments (5)
  1. [§3.1] The paper says M2N2 'samples two parameters (w_m, w_s)' but does not specify the distribution or whether w_s is uniform over parameter indices. It also claims the method 'progressively explores a broader set of boundaries,' but with independent random sampling there is no explicit progress schedule. Please clarify the sampling process and what 'progressively' means precisely.
  2. [Eq. 3] Equation 3 is ambiguous: z_j is defined as a sum over the archive, and the argmax over theta does not specify whether the candidate theta is included in z_j. Define the fitness of a candidate relative to a fixed archive, or state explicitly that z_j includes the candidate. This matters for implementation and for interpreting Eq. 5.
  3. [§4.3, references] The text says 'Our work builds on [34] approach using CMA-ES for merging.' Reference [34] is the STAIR Captions dataset paper, not a CMA-ES merging method; this appears to be a citation error, likely intended to refer to the EvoSDXL-JP work (Ref. [21]) or Ref. [1]. Please correct.
  4. [§4.1 and Table 2] The MNIST curves and the diffusion table would benefit from statistical significance statements or effect sizes. Ten runs are reported for MNIST, but the paper does not test whether differences (e.g., M2N2 vs. GA) are significant. Table 2 reports no variance or number of runs; please add this information.
  5. [Throughout] Several typographical issues: 'Llamma-2-7b' should be 'Llama-2-7B'; 'enviroments' -> 'environments'; 'test split consistent of' -> 'consisted of'; 'split' vs 'split-point' is sometimes inconsistent. These do not affect the science but should be cleaned up.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: M2N2 is an empirical method paper whose claims are supported by ablations and external benchmarks; the only same-lab citation is non-load-bearing.

full rationale

This paper is an empirical method paper, not a derivation. M2N2's three components are defined heuristically in Eq. 2 (split-point SLERP), Eq. 3/5 (competition fitness), and Eq. 4 (attraction); none of these are fitted to the reported test metrics, and the paper reports ablations (Table 1, Fig. 4) rather than reverse-engineering hyperparameters. The from-scratch MNIST result compares M2N2 with GA (using the same merge operator), MAP-Elites, and CMA-ES. The absence of a mutation-only/competition-only control is a genuine experimental-confound weakness, but it is not a circular reduction: the claim that merging can evolve models from scratch is an empirical claim, not a definitional one. The LLM and diffusion claims are evaluated on held-out splits (GSM8k/WebShop, COCO NCS/FID) against external baselines, so they do not reduce to the fitness inputs. The only same-lab citation, EvoSDXL-JP [21], is used to justify a peripheral design choice—merging attention layers independently—and is not invoked as a uniqueness theorem or as the evidence for the claimed state-of-the-art result. Section 5's limitation on model compatibility is an honest caveat, not a tautology. The training-coverage plot (Fig. 3) tracks the objective that the competition fitness was intentionally designed to promote, but this is a sanity check of the mechanism, not a 'prediction' equivalent to the input by construction. No load-bearing step reduces Eq. X to Eq. Y by definition. The score of 1 reflects one minor same-lab citation with otherwise self-contained empirical support.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims rest on the empirical assumption that SLERP weight interpolation produces functional offspring, on the task-specific design of capacity c_j, and on hand-set hyperparameters (archive size, alpha) that are ablated but not derived. There are no new physical or structural entities.

free parameters (4)
  • alpha (competition intensity) = 1 (default); swept in Figure 4 (right)
    Controls how much fitness is reduced by competition (Eq. 5); the paper shows smaller alpha converges earlier to worse solutions.
  • archive size P = 20 (MNIST), 15 (LLM)
    Hand-set; Figure 4 (left) shows an archive-size trade-off between early speed and final quality.
  • capacity c_j = 1 for binary tasks, max_i s(x_j|theta_i) for continuous scores
    Task-dependent design choice in Eq. 3 that determines how fitness is distributed per data point.
  • epsilon = small positive
    Zero-division guard in the denominator of Eq. 3 and Eq. 4; not otherwise specified.
assumptions (3)
  • domain assumption SLERP interpolation between parameter vectors of two models yields a functional model
    Used in Eq. 2 for every merge; if interpolation produces broken models, the whole algorithm fails. The paper's Limitations section acknowledges that divergent models are incompatible for merging.
  • domain assumption Per-example scores s(x_j|theta) are available and meaningful as resources
    Eq. 3 requires per-sample rewards; the diffusion experiment switches from FID to NCS for this reason, so the fitness surrogate may not align with the reported FID metric.
  • ad hoc to paper The archive's evolutionary pressure keeps models compatible for merging
    Stated in Section 5 as a hypothesis: 'We hypothesize that models with divergent state representations are incompatible for merging'; not tested. The from-scratch result depends on this remaining true.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Competition and Attraction Improve Model Fusion." pith.science (2026). https://pith.science/paper/6RKLX7VI

@misc{pith2026250816204,
  author       = {Pith},
  title        = {Pith review of: Competition and Attraction Improve Model Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6RKLX7VI}},
  note         = {Machine review of arXiv:2508.16204}
}
read the original abstract

Model merging is a powerful technique for integrating the specialized knowledge of multiple machine learning models into a single model. However, existing methods require manually partitioning model parameters into fixed groups for merging, which restricts the exploration of potential combinations and limits performance. To overcome these limitations, we propose Model Merging of Natural Niches (M2N2), an evolutionary algorithm with three key features: (1) dynamic adjustment of merging boundaries to progressively explore a broader range of parameter combinations; (2) a diversity preservation mechanism inspired by the competition for resources in nature, to maintain a population of diverse, high-performing models that are particularly well-suited for merging; and (3) a heuristicbased attraction metric to identify the most promising pairs of models for fusion. Our experimental results demonstrate, for the first time, that model merging can be used to evolve models entirely from scratch. Specifically, we apply M2N2 to evolve MNIST classifiers from scratch and achieve performance comparable to CMA-ES, while being computationally more efficient. Furthermore, M2N2 scales to merge specialized language and image generation models, achieving state-of-the-art performance. Notably, it preserves crucial model capabilities beyond those explicitly optimized by the fitness function, highlighting its robustness and versatility. Our code is available at https://github.com/SakanaAI/natural_niches

Figures

Figures reproduced from arXiv: 2508.16204 by the authors.

Figure 1
Figure 1. Left, previous methods group the parameters of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The plots show the accuracy on the test split vs the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Left: The percentage of training data points that can [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Left, the percentage of training data points that can [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of generated images across seed models [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Images generated by each model when receiving [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection

    cs.LG 2026-03 conditional novelty 6.0 of 10

    Predicting multi-head attention queries from history and scoring cosine mismatch against an EMA target, combined with reconstruction error, improves unsupervised multivariate anomaly ranking and localization.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. 2024. Evolu- tionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187 (2024)

    arXiv 2024

  2. [2]

    AUTOMATIC1111. 2022. Stable Diffusion WebUI. https://github.com/ AUTOMATIC1111/stable-diffusion-webui

    work page 2022

  3. [3]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 (2021)

    arXiv 2021

  4. [4]

    Paul Darwen and Xin Yao. 1996. Every niching method has its niche: Fitness shar- ing and implicit sharing compared. InParallel Problem Solving from Nature—PPSN IV: International Conference on Evolutionary Computation—The 4th International Conference on Parallel Problem Solving from Nature Berlin, Germany, September 22–26, 1996 Proceedings 4 . Springer, 398–407

    work page 1996

  5. [5]

    Kenneth Alan De Jong. 1975. An analysis of the behavior of a class of genetic adaptive systems. University of Michigan

    work page 1975

  6. [6]

    Kalyanmoy Deb and David E Goldberg. 1989. An investigation of niche and species formation in genetic function optimization. In Proceedings of the third international conference on Genetic algorithms . 42–50

    work page 1989

  7. [7]

    David E Goldberg, Kalyanmoy Deb, and Jeffrey Horn. 1992. Massive multimodal- ity, deception, and genetic algorithms.. In PPSN, Vol. 2

    work page 1992

  8. [8]

    David E Goldberg, Jon Richardson, et al. 1987. Genetic algorithms with sharing for multimodal function optimization. InGenetic algorithms and their applications: Proceedings of the Second International Conference on Genetic Algorithms, Vol. 4149. Cambridge, MA, 414–425

    work page 1987

Show all 35 references
  1. [9]

    Nikolaus Hansen and Andreas Ostermeier. 2001. Completely derandomized self-adaptation in evolution strategies. Evolutionary computation 9, 2 (2001), 159–195

    2001

  2. [10]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  3. [11]

    HuggingFace. 2023. Open LLM Leaderboard. https://huggingface.co/spaces/ HuggingFaceH4/open_llm_leaderboard. HuggingFace

    2023

  4. [12]

    So Kuroki, Taishi Nakamura, Takuya Akiba, and Yujin Tang. 2024. Agent Skill Ac- quisition for Large Language Models via CycleQD.arXiv preprint arXiv:2410.14735 (2024)

    2024 arXiv

  5. [13]

    Maxime Labonne. 2024. Merge Large Language Models with mergekit. Hugging Face Blog. https://huggingface.co/blog/mlabonne/merge-models

    2024

  6. [14]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proce...

    2014

  7. [15]

    Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 (2023)

    2023 arXiv

  8. [16]

    Jean-Baptiste Mouret and Jeff Clune. 2015. Illuminating search spaces by mapping elites. arXiv preprint arXiv:1504.04909 (2015)

    2015 arXiv

  9. [17]

    Alain Pétrowski. 1996. A clearing procedure as a niching method for genetic algorithms. In Proceedings of IEEE international conference on evolutionary com- putation. IEEE, 798–803

    1996

  10. [18]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  11. [19]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  12. [20]

    RunDiffusion. 2024. Juggernaut-XL-v9. https://huggingface.co/RunDiffusion/ Juggernaut-XL-v9. Accessed: January 2025

    2024

  13. [21]

    Sakana AI. 2025. EvoSDXL-JP: Evolutionary Model Merging for Japanese-English Bilingual Text-to-Image Generation. https://sakana.ai/evosdxl-jp/ Accessed: 2025-01

    2025

  14. [22]

    Jimmy Secretan, Nicholas Beato, David B D Ambrosio, Adelein Rodriguez, Adam Campbell, and Kenneth O Stanley. 2008. Picbreeder: evolving pictures collab- oratively online. In Proceedings of the SIGCHI conference on human factors in computing systems. 1759–1768

    2008

  15. [23]

    Makoto Shing, Takuya Akiba, and Jerry Chi. [n. d.]. Japanese Stable Diffusion XL. [https://huggingface.co/stabilityai/japanese-stable-diffusion-xl](https:// huggingface.co/stabilityai/japanese-stable-diffusion-xl)

  16. [24]

    Robert E Smith, Stephanie Forrest, and Alan S Perelson. 1993. Searching for di- verse, cooperative populations with genetic algorithms.Evolutionary computation 1, 2 (1993), 127–149

    1993

  17. [25]

    Kenneth O Stanley and Risto Miikkulainen. 2002. Evolving neural networks through augmenting topologies. Evolutionary computation 10, 2 (2002), 99–127

    2002

  18. [26]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  19. [27]

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. 2024. Diffusion model alignment using direct preference optimization. InProceedings of the IEEE/CVF Conference on Computer Visio...

    2024

  20. [28]

    Tom White. 2016. Sampling generative networks. arXiv preprint arXiv:1609.04468 (2016)

    2016 arXiv

  21. [29]

    Ka-Chun Wong. 2015. Evolutionary multimodal optimization: A short survey. arXiv preprint arXiv:1508.00457 (2015)

    2015 arXiv

  22. [30]

    Ka-Chun Wong, Chun-Ho Wu, Ricky KP Mok, Chengbin Peng, and Zhaolei Zhang. 2012. Evolutionary multimodal optimization using the principle of locality. Information Sciences 194 (2012), 138–170

    2012

  23. [31]

    Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, et al. 2024. AgentGym: Evolving Large Language Model-based Agents across Diverse Environments. arXiv preprint arXiv:2406.04151 (2024)

    2024 arXiv

  24. [32]

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. 2023. TIES-Merging: Resolving Interference When Merging Models. arXiv:2306.01708 [cs.LG] https://arxiv.org/abs/2306.01708

    2023 arXiv

  25. [33]

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems 35 (2022), 20744–20757

    2022

  26. [34]

    Yuya Yoshikawa, Yutaro Shigeto, and Akikazu Takeuchi. 2017. STAIR cap- tions: Constructing a large-scale Japanese image caption dataset. arXiv preprint arXiv:1705.00823 (2017)

    2017 arXiv

  27. [35]

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Forty-first International Conference on Machine Learning

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.