pith. sign in

arxiv: 2606.08032 · v1 · pith:C2M7M5UInew · submitted 2026-06-06 · 📊 stat.ML · cs.LG

Variational Proximal Policy Optimization

Ousmane Amadou Dia This is my paper

Pith reviewed 2026-06-27 19:20 UTC · model grok-4.3

classification 📊 stat.ML cs.LG
keywords variational policy optimizationstein variational gradient descentmixture of expertsreinforcement learning from human feedbackproximal policy optimizationpolicy optimizationreasoning benchmarksfunctional kernels
0
0 comments X

The pith

A particle-based variational framework maps policy optimization to Stein variational gradients inside Mixture-of-Experts models to supply geometry-based proximal control.

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

The paper introduces Variational Proximal Policy Optimization to address policy mode collapse, brittle exploration, and distribution drift in reinforcement learning from human feedback. It recasts policy updates as a particle-based variational inference problem solved by Stein Variational Gradient Descent within a Mixture-of-Experts architecture. Functional kernels defined over localized expert prototypes, together with an expert orthogonalization loss, create a geometry-based mechanism that limits how far policy updates can stray. Experiments on a 33B/4B sparse Mixture-of-Experts model report concrete gains on complex reasoning tasks, including a 179-point ELO increase on Codeforces and a 32 percent drop in tokens required on AIME problems. A sympathetic reader would care because the approach aims to make large-model RLHF training both more stable and more efficient without heavy dependence on hand-tuned clipping or KL penalties.

Core claim

VP₂O frames policy optimization as particle-based variational inference solved by Stein Variational Gradient Descent in a Mixture-of-Experts architecture. Functional kernels over localized expert prototypes and an expert orthogonalization loss supply a geometry-based proximal-control mechanism. This mechanism is designed to reduce reliance on fixed clipping or KL schedules while preserving training stability. On a 33B/4B sparse Mixture-of-Experts model the method yields gains across complex reasoning benchmarks, including a +179 ELO gain on Codeforces and a 32 percent reduction in token count on AIME mathematical reasoning tasks.

What carries the argument

Variational Proximal Policy Optimization (VP₂O) as a mapping of policy optimization to Stein Variational Gradient Descent within a Mixture-of-Experts architecture, using functional kernels over expert prototypes together with an expert orthogonalization loss to enforce geometry-based proximal control.

If this is right

  • Policy optimization becomes less sensitive to the precise values of clipping thresholds or KL coefficients.
  • Large Mixture-of-Experts models trained for reasoning achieve both higher benchmark scores and lower token usage.
  • Stability is maintained through geometry-based constraints rather than through explicit proximal penalties.
  • The same particle-based variational approach produces measurable lifts on both competitive programming and mathematical reasoning tasks.

Where Pith is reading between the lines

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

  • The same kernels and orthogonalization step could be tested on policy optimization methods other than PPO.
  • The geometry-based control might reduce the amount of hyperparameter search needed when scaling to larger expert counts.
  • If the mechanism generalizes, it could be applied to non-reasoning domains where mode collapse is also observed.

Load-bearing premise

The geometry-based proximal-control mechanism supplied by functional kernels and expert orthogonalization actually reduces reliance on fixed clipping or KL schedules while preserving stability.

What would settle it

An ablation experiment on the same 33B/4B model and benchmarks in which the functional kernels and expert orthogonalization loss are removed, followed by measurement of whether the reported ELO gain and token reduction disappear or whether training instability increases.

Figures

Figures reproduced from arXiv: 2606.08032 by Ousmane Amadou Dia.

Figure 1
Figure 1. Figure 1: Performance curves on academic benchmarks under the 8K generation context. [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance curves under the 16K generation length. Advantages are larger and more [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Thought and solution token counts per task throughout training ( [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
read the original abstract

Reinforcement Learning from Human Feedback via Proximal Policy Optimization often suffers from policy mode collapse, brittle exploration loops, and distribution drift. This paper introduces Variational Proximal Policy Optimization (\(\textsc{VP}_2\textsc{O}\)), a particle-based variational inference framework that maps policy optimization to Stein Variational Gradient Descent within a Mixture-of-Experts architecture. By leveraging functional kernels over localized expert prototypes alongside an expert orthogonalization loss, \(\textsc{VP}_2\textsc{O}\) introduces a geometry-based proximal-control mechanism that can reduce reliance on fixed clipping or KL schedules. Our results on a 33B/4B sparse Mixture-of-Experts model show several improvements across complex reasoning benchmarks, establishing a \(+\mathbf{179}\) ELO gain on Codeforces and a \(\mathbf{32\%}\) reduction in token count on AIME mathematical reasoning tasks.

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

2 major / 0 minor

Summary. The manuscript introduces Variational Proximal Policy Optimization (VP₂O), a particle-based variational inference framework that reformulates PPO-style policy optimization as Stein Variational Gradient Descent inside a Mixture-of-Experts architecture. Functional kernels defined over localized expert prototypes together with an expert orthogonalization loss are claimed to supply a geometry-based proximal-control mechanism that can reduce reliance on fixed clipping ratios or KL penalties. On a 33B/4B sparse MoE model the method is reported to deliver a +179 ELO gain on Codeforces and a 32% reduction in token count on AIME mathematical-reasoning tasks.

Significance. If the geometry-based proximal mechanism were shown to substitute for or relax clipping/KL schedules while preserving stability, the work would supply a principled alternative to standard PPO for RLHF on large sparse models and could improve exploration and reduce mode collapse. The reported benchmark deltas are large enough to be practically interesting, but the absence of any derivation, ablation, or statistical protocol prevents the result from being evaluated at present.

major comments (2)
  1. [Abstract] Abstract, paragraph describing the framework: the central claim that functional kernels over expert prototypes plus expert orthogonalization supply a geometry-based proximal-control mechanism that 'can reduce reliance on fixed clipping or KL schedules' is unsupported; no comparison of clipping ratios, KL coefficients, or stability metrics (e.g., policy entropy, gradient norms) between VP₂O and baseline PPO is provided, nor any ablation isolating the kernels/orthogonalization as the source of any observed stability.
  2. [Abstract] Abstract, results paragraph: the reported +179 ELO gain on Codeforces and 32% token reduction on AIME rest on a single 33B/4B sparse MoE experiment whose baseline implementation, hyperparameter schedule, data-selection procedure, number of runs, and variance are not described; without these details the attribution of gains to the proposed proximal mechanism cannot be assessed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for stronger empirical grounding of our claims. We address each major comment below and will revise the manuscript to incorporate additional details, ablations, and clarifications where feasible.

read point-by-point responses

  1. Referee: [Abstract] Abstract, paragraph describing the framework: the central claim that functional kernels over expert prototypes plus expert orthogonalization supply a geometry-based proximal-control mechanism that 'can reduce reliance on fixed clipping or KL schedules' is unsupported; no comparison of clipping ratios, KL coefficients, or stability metrics (e.g., policy entropy, gradient norms) between VP₂O and baseline PPO is provided, nor any ablation isolating the kernels/orthogonalization as the source of any observed stability.

    Authors: We acknowledge that the abstract presents the geometry-based proximal mechanism as a central contribution without accompanying empirical comparisons or ablations. The full manuscript contains a theoretical derivation mapping the functional kernels and orthogonalization loss to a proximal operator within the SVGD framework, but we agree this does not substitute for direct evidence. In revision we will add ablations that isolate the kernels and orthogonalization (by removing each component) together with side-by-side plots of policy entropy, gradient norms, and performance under matched versus relaxed clipping/KL schedules. revision: yes

  2. Referee: [Abstract] Abstract, results paragraph: the reported +179 ELO gain on Codeforces and 32% token reduction on AIME rest on a single 33B/4B sparse MoE experiment whose baseline implementation, hyperparameter schedule, data-selection procedure, number of runs, and variance are not described; without these details the attribution of gains to the proposed proximal mechanism cannot be assessed.

    Authors: The reported numbers come from a single training run of the 33B/4B model, which was the only feasible scale given compute limits. We will expand the experimental section to document the exact baseline PPO implementation, hyperparameter schedules, data-selection pipeline, and training protocol. We will also state explicitly that variance across independent seeds was not measured and treat this as a limitation of the current evaluation. revision: partial

Circularity Check

0 steps flagged

No circularity detected; abstract contains no equations or derivation chain

full rationale

The abstract and description introduce VP₂O as mapping PPO to SVGD in MoE with functional kernels and orthogonalization loss for proximal control, claiming benchmark gains. No equations, fitting procedures, self-citations, or 'predictions' of derived quantities are present. The central claim that the mechanism 'can reduce reliance' on clipping/KL is stated as a possibility without reduction to inputs by construction. No load-bearing steps reduce to self-definition or fitted inputs. This is the expected non-finding when no derivation chain is supplied.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no equations, so free parameters, axioms, and invented entities cannot be enumerated.

pith-pipeline@v0.9.1-grok · 5664 in / 1064 out tokens · 17374 ms · 2026-06-27T19:20:45.943570+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

219 extracted references · 21 canonical work pages · 5 internal anchors

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    IEEE transactions on medical imaging , volume=

    Uncertainty driven probabilistic voxel selection for image registration , author=. IEEE transactions on medical imaging , volume=. 2013 , publisher=

    2013

  5. [5]

    2026 , eprint=

    Reinforcement Learning via Self-Distillation , author=. 2026 , eprint=

    2026

  6. [6]

    2026 , eprint=

    Advancing Expert Specialization for Better MoE , author=. 2026 , eprint=

    2026

  7. [7]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  8. [8]

    2025 , eprint=

    OpenThoughts: Data Recipes for Reasoning Models , author=. 2025 , eprint=

    2025

  9. [9]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Ziyi and Liu, Aixin and Xue, Bing and Wang, Bingxuan and Wu, Bochao and Feng, Bei ...

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

  10. [10]

    2025 , eprint=

    Group Sequence Policy Optimization , author=. 2025 , eprint=

    2025

  11. [11]

    2025 , eprint=

    Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs , author=. 2025 , eprint=

    2025

  12. [12]

    Structured Pruning of Neural Networks with Budget-Aware Regularization , journal =

    Carl Lemaire and Andrew Achkar and Pierre. Structured Pruning of Neural Networks with Budget-Aware Regularization , journal =. 2018 , url =

    2018

  13. [13]

    2023 , eprint=

    Object-Centric Slot Diffusion , author=. 2023 , eprint=

    2023

  14. [14]

    2017 , eprint=

    Stein Variational Policy Gradient , author=. 2017 , eprint=

    2017

  15. [15]

    2023 , eprint=

    How to Exploit Hyperspherical Embeddings for Out-of-Distribution Detection? , author=. 2023 , eprint=

    2023

  16. [16]

    2024 , eprint=

    Conditional Vendi Score: An Information-Theoretic Approach to Diversity Evaluation of Prompt-based Generative Models , author=. 2024 , eprint=

    2024

  17. [17]

    2025 , eprint=

    No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping , author=. 2025 , eprint=

    2025

  18. [18]

    2017 , eprint=

    Proximal Policy Optimization Algorithms , author=. 2017 , eprint=

    2017

  19. [19]

    2017 , eprint=

    Trust Region Policy Optimization , author=. 2017 , eprint=

    2017

  20. [20]

    2016 , eprint=

    Asynchronous Methods for Deep Reinforcement Learning , author=. 2016 , eprint=

    2016

  21. [21]

    2025 , eprint=

    Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts , author=. 2025 , eprint=

    2025

  22. [22]

    2018 , eprint=

    High-Dimensional Continuous Control Using Generalized Advantage Estimation , author=. 2018 , eprint=

    2018

  23. [23]

    Top- n : Eliminating Noise in Logit Space for Robust Token Sampling of LLM

    Tang, Chenxia and Liu, Jianchun and Xu, Hongli and Huang, Liusheng. Top- n : Eliminating Noise in Logit Space for Robust Token Sampling of LLM. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.528

    work page doi:10.18653/v1/2025.acl-long.528 2025

  24. [24]

    2023 , eprint=

    The Vendi Score: A Diversity Evaluation Metric for Machine Learning , author=. 2023 , eprint=

    2023

  25. [25]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  26. [26]

    Proceedings of the 40th International Conference on Machine Learning , articleno =

    Gao, Leo and Schulman, John and Hilton, Jacob , title =. Proceedings of the 40th International Conference on Machine Learning , articleno =. 2023 , publisher =

    2023

  27. [27]

    2016 , eprint=

    Concrete Problems in AI Safety , author=. 2016 , eprint=

    2016

  28. [28]

    2024 , eprint=

    Secrets of RLHF in Large Language Models Part II: Reward Modeling , author=. 2024 , eprint=

    2024

  29. [29]

    2023 , eprint=

    Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback , author=. 2023 , eprint=

    2023

  30. [30]

    2021 , eprint=

    Variational Bayesian Optimistic Sampling , author=. 2021 , eprint=

    2021

  31. [31]

    Sixteenth European Workshop on Reinforcement Learning , year=

    Overcoming Policy Collapse in Deep Reinforcement Learning , author=. Sixteenth European Workshop on Reinforcement Learning , year=

  32. [32]

    Proceedings of the 30th International Conference on Neural Information Processing Systems , pages =

    Liu, Qiang and Wang, Dilin , title =. Proceedings of the 30th International Conference on Neural Information Processing Systems , pages =. 2016 , isbn =

    2016

  33. [33]

    Advances in Neural Information Processing Systems , volume =

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  34. [34]

    , booktitle =

    Korbak, Tomasz and Perez, Ethan and Buckley, Christopher L. , booktitle =. 2022 , url =

    2022

  35. [35]

    Advancing Expert Specialization for Better

    Guo, Hongcan and Lu, Haolang and Nan, Guoshun and Chu, Bolun and Zhuang, Jialin and Yang, Yuan and Che, Wenhao and Cao, Xinye and Leng, Sicong and Cui, Qimei and Jiang, Xudong , journal =. Advancing Expert Specialization for Better. 2025 , url =

    2025

  36. [36]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Weighted importance sampling for off-policy learning with linear function approximation , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2014 , publisher =

    2014

  37. [37]

    2023 , eprint=

    Click: Controllable Text Generation with Sequence Likelihood Contrastive Learning , author=. 2023 , eprint=

    2023

  38. [38]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    High-Confidence Off-Policy Evaluation , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2015 , doi =

    2015

  39. [39]

    2024 , eprint=

    Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs , author=. 2024 , eprint=

    2024

  40. [40]

    2024 , eprint=

    From r to Q^* : Your Language Model is Secretly a Q-Function , author=. 2024 , eprint=

    2024

  41. [41]

    2024 , eprint=

    Slot State Space Models , author=. 2024 , eprint=

    2024

  42. [42]

    2022 , eprint=

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback , author=. 2022 , eprint=

    2022

  43. [43]

    CoRR , volume =

    Francesco Locatello and Dirk Weissenborn and Thomas Unterthiner and Aravindh Mahendran and Georg Heigold and Jakob Uszkoreit and Alexey Dosovitskiy and Thomas Kipf , title =. CoRR , volume =. 2020 , url =. 2006.15055 , timestamp =

    arXiv 2020

  44. [44]

    CoRR , volume =

    Jonathan Ho and Ajay Jain and Pieter Abbeel , title =. CoRR , volume =. 2020 , url =. 2006.11239 , timestamp =

    Pith/arXiv arXiv 2020

  45. [45]

    2022 , eprint=

    Object Scene Representation Transformer , author=. 2022 , eprint=

    2022

  46. [46]

    Multi-Object Representation Learning with Iterative Variational Inference , journal =

    Klaus Greff and Rapha. Multi-Object Representation Learning with Iterative Variational Inference , journal =. 2019 , url =. 1903.00450 , timestamp =

    arXiv 2019

  47. [47]

    CoRR , volume =

    Shaolei Zhang and Yang Feng , title =. CoRR , volume =. 2021 , url =. 2109.05244 , timestamp =

    arXiv 2021

  48. [48]

    Burgess and Lo

    Christopher P. Burgess and Lo. MONet: Unsupervised Scene Decomposition and Representation , journal =. 2019 , url =. 1901.11390 , timestamp =

    Pith/arXiv arXiv 2019

  49. [49]

    2022 , eprint=

    Illiterate DALL-E Learns to Compose , author=. 2022 , eprint=

    2022

  50. [50]

    CoRR , volume =

    Olaf Ronneberger and Philipp Fischer and Thomas Brox , title =. CoRR , volume =. 2015 , url =. 1505.04597 , timestamp =

    Pith/arXiv arXiv 2015

  51. [51]

    2022 , eprint=

    VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance , author=. 2022 , eprint=

    2022

  52. [52]

    Taming Transformers for High-Resolution Image Synthesis , journal =

    Patrick Esser and Robin Rombach and Bj. Taming Transformers for High-Resolution Image Synthesis , journal =. 2020 , url =. 2012.09841 , timestamp =

    arXiv 2020

  53. [53]

    High-Resolution Image Synthesis with Latent Diffusion Models , journal =

    Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Bj. High-Resolution Image Synthesis with Latent Diffusion Models , journal =. 2021 , url =. 2112.10752 , timestamp =

    Pith/arXiv arXiv 2021

  54. [54]

    CoRR , volume =

    Gautam Singh and Fei Deng and Sungjin Ahn , title =. CoRR , volume =. 2021 , url =. 2110.11405 , timestamp =

    arXiv 2021

  55. [55]

    2023 , eprint=

    Neural Systematic Binder , author=. 2023 , eprint=

    2023

  56. [56]

    CoRR , volume =

    Najmeh Abiri and Mattias Ohlsson , title =. CoRR , volume =. 2020 , url =. 2004.02581 , timestamp =

    arXiv 2020

  57. [57]

    CoRR , volume =

    Prafulla Dhariwal and Alex Nichol , title =. CoRR , volume =. 2021 , url =. 2105.05233 , timestamp =

    Pith/arXiv arXiv 2021

  58. [58]

    Information Pursuit:

    Ehsan Jahangiri and Erdem Y. Information Pursuit:. CoRR , volume =. 2017 , url =. 1701.02343 , timestamp =

    Pith/arXiv arXiv 2017

  59. [59]

    2023 , eprint=

    Variational Information Pursuit for Interpretable Predictions , author=. 2023 , eprint=

    2023

  60. [60]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Energy-based Out-of-distribution Detection , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  61. [61]

    2023 , eprint=

    Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints , author=. 2023 , eprint=

    2023

  62. [62]

    CoRR , volume =

    Pang Wei Koh and Thao Nguyen and Yew Siang Tang and Stephen Mussmann and Emma Pierson and Been Kim and Percy Liang , title =. CoRR , volume =. 2020 , url =. 2007.04612 , timestamp =

    arXiv 2020

  63. [63]

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

    Simonyan, Karen and Vedaldi, Andrea and Zisserman, Andrew , keywords =. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps , publisher =. 2013 , copyright =. doi:10.48550/ARXIV.1312.6034 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1312.6034 2013

  64. [64]

    Balasubramanian , title =

    Aditya Chattopadhyay and Piyushi Manupriya and Anirban Sarkar and Vineeth N. Balasubramanian , title =. CoRR , volume =. 2019 , url =. 1902.02302 , timestamp =

    Pith/arXiv arXiv 2019

  65. [65]

    "Why Should

    Marco T. "Why Should. CoRR , volume =. 2016 , url =. 1602.04938 , timestamp =

    Pith/arXiv arXiv 2016

  66. [66]

    Selvaraju and Abhishek Das and Ramakrishna Vedantam and Michael Cogswell and Devi Parikh and Dhruv Batra , title =

    Ramprasaath R. Selvaraju and Abhishek Das and Ramakrishna Vedantam and Michael Cogswell and Devi Parikh and Dhruv Batra , title =. CoRR , volume =. 2016 , url =. 1610.02391 , timestamp =

    arXiv 2016

  67. [67]

    Goodfellow and Moritz Hardt and Been Kim , title =

    Julius Adebayo and Justin Gilmer and Michael Muelly and Ian J. Goodfellow and Moritz Hardt and Been Kim , title =. CoRR , volume =. 2018 , url =. 1810.03292 , timestamp =

    arXiv 2018

  68. [68]

    CoRR , volume =

    Tom Zahavy and Shie Mannor , title =. CoRR , volume =. 2019 , url =. 1901.08612 , timestamp =

    arXiv 2019

  69. [69]

    The Exp3-IX Algorithm , DOI=

    Lattimore, Tor and Szepesvári, Csaba , year=. The Exp3-IX Algorithm , DOI=. Bandit Algorithms , publisher=

  70. [70]

    CoRR , volume =

    Aadirupa Saha and Pierre Gaillard and Michal Valko , title =. CoRR , volume =. 2020 , url =. 2004.06248 , timestamp =

    arXiv 2020

  71. [71]

    Zhang and Michael Ruan and Eric Wang and So Hasegawa and Jimmy Ba and Roger Baker Grosse , booktitle=

    Juhan Bae and Michael R. Zhang and Michael Ruan and Eric Wang and So Hasegawa and Jimmy Ba and Roger Baker Grosse , booktitle=. Multi-Rate. 2023 , url=

    2023

  72. [72]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead , url =

    Rudin, Cynthia , biburl =. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead , url =. Nature Machine Intelligence , keywords =. doi:10.1038/s42256-019-0048-x , interhash =

    work page doi:10.1038/s42256-019-0048-x

  73. [73]

    Haeffele and Rene Vidal and Donald Geman , title =

    Aditya Chattopadhyay and Stewart Slocum and Benjamin D. Haeffele and Rene Vidal and Donald Geman , title =. doi:10.1109/tpami.2022.3225162 , url =

    work page doi:10.1109/tpami.2022.3225162 2022

  74. [74]

    Improved Algorithms for Linear Stochastic Bandits , url =

    Abbasi-yadkori, Yasin and P\'. Improved Algorithms for Linear Stochastic Bandits , url =. Advances in Neural Information Processing Systems , editor =

  75. [75]

    Supervised Topic Models , url =

    Mcauliffe, Jon and Blei, David , booktitle =. Supervised Topic Models , url =

  76. [76]

    and Ng, Andrew Y

    Blei, David M. and Ng, Andrew Y. and Jordan, Michael I. , title =. J. Mach. Learn. Res. , month =. 2003 , issue_date =

    2003

  77. [77]

    1994 , Journal =

    Primary biliary cirrhosis: prediction of short-term survival based on repeated patient visits , Author =. 1994 , Journal =. doi:10.1016/0270-9139(94)90144-9 , Number =

    work page doi:10.1016/0270-9139(94)90144-9 1994

  78. [78]

    Fader and Bruce G.S

    Peter S. Fader and Bruce G.S. Hardie , abstract =. How to project customer retention , journal =. 2007 , issn =. doi:https://doi.org/10.1002/dir.20074 , url =

    work page doi:10.1002/dir.20074 2007

  79. [79]

    Survival and event history analysis: A process point of view , author=

  80. [80]

    Kalbfleisch, J. D. and Prentice, R. L. , biburl =. The

Showing first 80 references.