pith. machine review for the scientific record. sign in

arxiv: 2605.05703 · v2 · submitted 2026-05-07 · 💻 cs.MA · cs.AI· cs.LG

Recognition: no theorem link

Active Learning for Communication Structure Optimization in LLM-Based Multi-Agent Systems

Huchen Yang , Xinghao Dong , Dan Negrut , Jin-Long Wu
Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:42 UTC · model grok-4.3

classification 💻 cs.MA cs.AIcs.LG
keywords active learningmulti-agent systemslarge language modelscommunication structure optimizationensemble Kalman inversioninformation-theoretic selectiontask selectionBayesian approximation
0
0 comments X

The pith

Task selection by expected shifts in graph parameter distributions optimizes communication structures in LLM multi-agent systems more reliably than random sampling.

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

The paper establishes that random task sampling creates unstable results when optimizing the communication graph of LLM-based multi-agent systems, because tasks differ sharply in how much they reveal about the best graph structure. It introduces an information-theoretic selection rule that scores each candidate task by the magnitude of the change it would produce in the distribution over graph parameters. This change is approximated through ensemble Kalman inversion, which supplies a derivative-free Bayesian update suitable for noisy black-box evaluations. The resulting active-learning loop, paired with an embedding-based candidate pool and surrogate-assisted batch sampling, produces better final communication structures under tight computational limits. A sympathetic reader cares because real deployments cannot afford to evaluate every possible task, and instability from poor training sets has been a practical barrier.

Core claim

The paper claims that task informativeness equals the expected change a task induces in the posterior distribution over communication-graph parameters; ensemble Kalman inversion supplies an efficient, derivative-free approximation to this quantity, and selecting tasks according to the resulting scores yields more effective communication-structure optimization than random sampling, both in standard settings and when some agents are attacked.

What carries the argument

ensemble Kalman inversion approximation to the Bayesian update that quantifies how much a candidate task shifts the distribution over communication-graph parameters

If this is right

  • Active task selection reduces sensitivity of the final communication graph to the particular training set chosen.
  • The same framework remains effective when some agents behave adversarially.
  • Embedding-based candidate-pool construction plus surrogate modeling and batch Thompson sampling together keep the method computationally tractable.
  • The approach applies under constrained training budgets where exhaustive evaluation of all tasks is impossible.

Where Pith is reading between the lines

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

  • If the approximation holds, the same selection principle could be used to optimize other black-box structural choices such as agent role assignments or shared prompt templates.
  • The method points toward derivative-free active-learning techniques for any multi-agent system whose performance is measured only through noisy end-to-end evaluations.
  • Extending the candidate pool construction beyond embeddings to include domain-specific similarity metrics might further improve selection quality in specialized task domains.

Load-bearing premise

That ensemble Kalman inversion supplies a sufficiently accurate approximation to the true Bayesian update for measuring task informativeness in noisy black-box multi-agent LLM systems, and that the embedding-based candidate pool remains representative of the full task distribution.

What would settle it

A controlled experiment in which tasks ranked highest by the informativeness estimator produce no greater improvement in final system performance than randomly chosen tasks of equal number would directly falsify the claim that the estimator selects valuable tasks.

Figures

Figures reproduced from arXiv: 2605.05703 by Dan Negrut, Huchen Yang, Jin-Long Wu, Xinghao Dong.

Figure 1
Figure 1. Figure 1: illustrates this benefit: under a matched 8× total MAS-inference token budget, active learning (1× training + 7× selection) yields higher average downstream accuracy than the 8× random training. 3 Related work view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed active learning framework. The framework first selects rep view at source ↗
Figure 3
Figure 3. Figure 3: Data distribution of downstream accuracy on MMLU In the benign setting without agent attacks ( view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between different meth￾ods. EKI and Fisher coreset (det) perform best overall, while EKI does not use a coreset. 5.4 Comparison with other informativeness-based active learning methods We compare our method with several informativeness-based active learning baselines on MMLU under agent attacks. All methods share the same 1000-to-50 representative-selection stage and differ only in how they sele… view at source ↗
Figure 6
Figure 6. Figure 6: A representative example of graph change on MMLU under agent attack. The left panel view at source ↗
Figure 7
Figure 7. Figure 7: Representative MMLU tasks at different EKI ranking positions from the same 50-task view at source ↗
Figure 7
Figure 7. Figure 7: Representative MMLU tasks at different EKI ranking positions from the same 50-task [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity to ensemble size: Top-10 overlap with the ensemble-size-20 ranking (set as view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity to ensemble size: Top-10 overlap with the ensemble-size-20 ranking (set as [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Top-10 overlap with the final three-step ranking as a function of the number of EKI iterations view at source ↗
Figure 9
Figure 9. Figure 9: Top-10 overlap with the final three-step ranking as a function of the number of EKI iterations [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
read the original abstract

Optimizing the communication structure of large language model based multi-agent systems (LLM-MAS) has been shown to improve downstream performance and reduce token usage. Existing methods typically rely on randomly sampled training tasks. However, tasks may differ substantially in difficulty and domain, and thus they are not equally informative for updating communication structure, making optimization under limited training budgets often unstable and highly sensitive to the particular training set. To actively identify the most valuable tasks for communication-structure optimization, we propose an ensemble-based information-theoretic task selection framework. The proposed method estimates task informativeness by how much a candidate task changes the distribution over graph parameters, using ensemble Kalman inversion as an efficient and derivative-free approximation of the corresponding Bayesian update. The resulting estimator is especially suitable for black-box and noisy multi-agent systems. To enhance scalability, we construct a compact candidate pool through embedding-based representative selection and combine the informative selection with surrogate modeling and batch Thompson sampling. We validate our method in both benign settings and settings with agent attacks, demonstrating its effectiveness for communication-structure optimization under constrained computational budgets.

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 / 2 minor

Summary. The manuscript proposes an ensemble-based information-theoretic active learning framework for selecting training tasks to optimize communication structures in LLM-based multi-agent systems. Task informativeness is quantified by the shift it induces in the posterior over graph parameters, with ensemble Kalman inversion (EnKI) used as a derivative-free approximation to the Bayesian update. The approach augments this with embedding-based candidate pool construction, surrogate modeling, and batch Thompson sampling, and reports validation in both benign and agent-attack settings under constrained computational budgets.

Significance. If the empirical claims hold, the work provides a targeted method to reduce the instability of communication-structure optimization that arises from random task sampling, which is especially relevant for black-box, noisy LLM-MAS under tight token or compute limits. The use of EnKI to handle derivative-free, stochastic forward models is a pragmatic adaptation, and the explicit treatment of agent-attack scenarios adds practical value. Credit is due for framing the problem around information gain on graph parameters rather than downstream task performance alone.

major comments (2)
  1. [§3.2] §3.2 (EnKI approximation): The central claim that the ensemble shift under EnKI accurately ranks tasks by true information gain for communication-graph parameters rests on the assumption that the Gaussian ensemble update sufficiently approximates the posterior change. In LLM-MAS the forward model is highly non-linear, stochastic, and the graph parameters are typically discrete (adjacency or topology choices), so the ensemble can collapse or bias the estimated informativeness. A load-bearing validation—e.g., a controlled comparison of EnKI scores against exact posterior shifts on a small discrete graph model—is required before the superiority over random sampling can be accepted.
  2. [Experimental section] Experimental section (results tables): The abstract asserts effectiveness under constrained budgets in both benign and attack settings, yet the provided text contains no quantitative metrics, error bars, ablation on the EnKI component, or protocol details (number of runs, budget levels, graph sizes). Without these, the data support for the claim that the method outperforms random sampling cannot be assessed, undermining the central empirical contribution.
minor comments (2)
  1. [§2] Notation: The mapping from communication structure to the parameter vector θ is introduced without an explicit equation or example in the early sections; a small illustrative diagram or definition would improve readability.
  2. [§3.3] Candidate pool construction: The embedding-based representative selection is described at a high level; the precise embedding model, distance metric, and selection algorithm (k-means, greedy, etc.) should be stated explicitly with a reference or pseudocode.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and describe the revisions we will incorporate to address the concerns raised.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (EnKI approximation): The central claim that the ensemble shift under EnKI accurately ranks tasks by true information gain for communication-graph parameters rests on the assumption that the Gaussian ensemble update sufficiently approximates the posterior change. In LLM-MAS the forward model is highly non-linear, stochastic, and the graph parameters are typically discrete (adjacency or topology choices), so the ensemble can collapse or bias the estimated informativeness. A load-bearing validation—e.g., a controlled comparison of EnKI scores against exact posterior shifts on a small discrete graph model—is required before the superiority over random sampling can be accepted.

    Authors: We acknowledge that the reliability of the EnKI-based informativeness ranking depends on how well the Gaussian ensemble update approximates the true posterior shift, especially given the non-linear, stochastic forward models and discrete graph parameters in LLM-MAS. While EnKI is a standard derivative-free technique for such intractable settings and has been validated in other non-linear domains, we agree that a direct comparison to exact posterior shifts would strengthen the central claim. In the revised manuscript, we will add a controlled validation study on a small-scale discrete graph model (e.g., using exact enumeration or MCMC on toy instances) to compare EnKI scores against ground-truth information gain. This will be included in an expanded §3.2 or a new appendix. revision: yes

  2. Referee: [Experimental section] Experimental section (results tables): The abstract asserts effectiveness under constrained budgets in both benign and attack settings, yet the provided text contains no quantitative metrics, error bars, ablation on the EnKI component, or protocol details (number of runs, budget levels, graph sizes). Without these, the data support for the claim that the method outperforms random sampling cannot be assessed, undermining the central empirical contribution.

    Authors: We appreciate the referee highlighting this gap in the submitted version. The experimental section will be substantially expanded in the revision to include full quantitative results: performance metrics (e.g., communication-structure stability, downstream accuracy, token savings) with error bars from multiple runs (we will specify 5–10 random seeds), ablation studies isolating the EnKI component, and complete protocol details (budget levels, graph sizes, number of tasks, statistical tests). These additions will be presented in revised tables and text to rigorously support the superiority over random sampling in both benign and agent-attack scenarios. revision: yes

Circularity Check

0 steps flagged

No circularity: framework combines external approximation techniques without self-referential reduction

full rationale

The abstract and described framework present an active learning procedure that estimates task informativeness via the change in graph-parameter distribution, approximated by ensemble Kalman inversion (EnKI) as a derivative-free Bayesian update. This is combined with embedding-based candidate selection, surrogate modeling, and batch Thompson sampling. No equations or claims reduce the informativeness score, the selection criterion, or the final optimization result to a fitted parameter, self-definition, or load-bearing self-citation by construction. EnKI is invoked as an established external method suitable for black-box systems rather than derived from the paper's own inputs. The central claim therefore remains independent of tautological re-labeling or circular fitting, consistent with the reader's assessment of score 2.0 (minor at most).

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions from ensemble Kalman methods and active learning; no new physical entities are introduced. Free parameters such as ensemble size or embedding dimension are implicit but not enumerated in the abstract.

axioms (1)
  • domain assumption Ensemble Kalman inversion provides a usable derivative-free approximation to the Bayesian posterior update for graph-parameter distributions
    Invoked to justify the informativeness estimator in the proposed framework

pith-pipeline@v0.9.0 · 5492 in / 1222 out tokens · 49028 ms · 2026-05-11T01:42:43.406222+00:00 · methodology

Review history (2 revisions) →

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

48 extracted references · 48 canonical work pages · 5 internal anchors

  1. [1]

    arXiv preprint arXiv:2405.02957 , year =

    Junkai Li, Yunghwei Lai, Weitao Li, Jingyi Ren, Meng Zhang, Xinhui Kang, Siyu Wang, Peng Li, Ya-Qin Zhang, Weizhi Ma, et al. Agent hospital: A simulacrum of hospital with evolvable medical agents.arXiv preprint arXiv:2405.02957, 2024

    work page arXiv 2024

  2. [2]

    A survey of llm-based agents in medicine: How far are we from baymax? Findings of the Association for Computational Linguistics: ACL 2025, pages 10345–10359, 2025

    Wenxuan Wang, Zizhan Ma, Zheng Wang, Chenghan Wu, Jiaming Ji, Wenting Chen, Xiang Li, and Yixuan Yuan. A survey of llm-based agents in medicine: How far are we from baymax? Findings of the Association for Computational Linguistics: ACL 2025, pages 10345–10359, 2025

    work page 2025

  3. [3]

    Chronollm: a framework for customizing large language model for digital twins generalization based on pychrono.arXiv preprint arXiv:2501.04062, 2025

    Jingquan Wang, Harry Zhang, Khailanii Slaton, Shu Wang, Radu Serban, Jinlong Wu, and Dan Negrut. Chronollm: a framework for customizing large language model for digital twins generalization based on pychrono.arXiv preprint arXiv:2501.04062, 2025

    work page arXiv 2025

  4. [4]

    SciML Agents: Write the Solver, Not the Solution

    Saarth Gaonkar, Xiang Zheng, Haocheng Xi, Rishabh Tiwari, Kurt Keutzer, Dmitriy Morozov, Michael W Mahoney, and Amir Gholami. SciML agents: Write the solver, not the solution. arXiv preprint arXiv:2509.09936, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    work page 2024

  6. [6]

    LLM-based multi-agent systems for software engineering: Literature review, vision, and the road ahead.ACM Transactions on Software Engineering and Methodology, 34(5):1–30, 2025

    Junda He, Christoph Treude, and David Lo. LLM-based multi-agent systems for software engineering: Literature review, vision, and the road ahead.ACM Transactions on Software Engineering and Methodology, 34(5):1–30, 2025

    work page 2025

  7. [7]

    MetaGPT: Meta programming for a multi-agent collaborative framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, et al. MetaGPT: Meta programming for a multi-agent collaborative framework. InThe twelfth international conference on learning representations, 2024

    work page 2024

  8. [8]

    Chatdev: Communicative agents for software development

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development. InProceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers), pages 15174–15186, 2024

    work page 2024

  9. [9]

    S2-mad: Breaking the token barrier to enhance multi-agent debate efficiency

    Yuting Zeng, Weizhe Huang, Lei Jiang, Tongxuan Liu, Xitai Jin, Chen Tianying Tiana, Jing Li, and Xiaohua Xu. S2-mad: Breaking the token barrier to enhance multi-agent debate efficiency. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long P...

    work page 2025

  10. [10]

    Understanding the information propagation effects of communication topologies in LLM-based multi-agent systems

    Xu Shen, Yixin Liu, Yiwei Dai, Yili Wang, Rui Miao, Yue Tan, Shirui Pan, and Xin Wang. Understanding the information propagation effects of communication topologies in LLM-based multi-agent systems. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 12347–12361, 2025

    work page 2025

  11. [11]

    AFlow: Automating Agentic Workflow Generation

    Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. Aflow: Automating agentic workflow generation.arXiv preprint arXiv:2410.10762, 2024

    work page internal anchor Pith review arXiv 2024

  12. [12]

    Topological structure learning should be a research priority for LLM-based multi-agent systems.arXiv preprint arXiv:2505.22467, 2025

    Jiaxi Yang, Mengqi Zhang, Yiqiao Jin, Hao Chen, Qingsong Wen, Lu Lin, Yi He, Srijan Kumar, Weijie Xu, James Evans, et al. Topological structure learning should be a research priority for LLM-based multi-agent systems.arXiv preprint arXiv:2505.22467, 2025. 10

    work page arXiv 2025

  13. [13]

    Assemble your crew: Automatic multi-agent communication topology design via autoregressive graph generation

    Shiyuan Li, Yixin Liu, Qingsong Wen, Chengqi Zhang, and Shirui Pan. Assemble your crew: Automatic multi-agent communication topology design via autoregressive graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 23142–23150, 2026

    work page 2026

  14. [14]

    Multi-agent architecture search via agentic supernet.arXiv preprint arXiv:2502.04180, 2025

    Guibin Zhang, Luyang Niu, Junfeng Fang, Kun Wang, Lei Bai, and Xiang Wang. Multi-agent architecture search via agentic supernet.arXiv preprint arXiv:2502.04180, 2025

    work page arXiv 2025

  15. [15]

    Cut the crap: An economical communication pipeline for LLM-based multi-agent systems

    Guibin Zhang, Yanwei Yue, Zhixun Li, Sukwon Yun, Guancheng Wan, Kun Wang, Dawei Cheng, Jeffrey Xu Yu, and Tianlong Chen. Cut the crap: An economical communication pipeline for LLM-based multi-agent systems. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=LkzuPorQ5L

    work page 2025

  16. [16]

    Amas: Adaptively determining communication topology for LLM-based multi-agent system

    Hui Yi Leong, Yuheng Li, Yuqing Wu, Wenwen Ouyang, Wei Zhu, Jiechao Gao, and Wei Han. Amas: Adaptively determining communication topology for LLM-based multi-agent system. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 2061–2070, 2025

    work page 2025

  17. [17]

    G-designer: Architecting multi-agent communication topologies via graph neural networks.arXiv preprint arXiv:2410.11782, 2024

    Guibin Zhang, Yanwei Yue, Xiangguo Sun, Guancheng Wan, Miao Yu, Junfeng Fang, Kun Wang, Tianlong Chen, and Dawei Cheng. G-designer: Architecting multi-agent communication topologies via graph neural networks.arXiv preprint arXiv:2410.11782, 2024

    work page arXiv 2024

  18. [18]

    Agentdropout: Dynamic agent elimination for token-efficient and high-performance LLM-based multi-agent collaboration

    Zhexuan Wang, Yutong Wang, Xuebo Liu, Liang Ding, Miao Zhang, Jie Liu, and Min Zhang. Agentdropout: Dynamic agent elimination for token-efficient and high-performance LLM-based multi-agent collaboration. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 24013–24035, 2025

    work page 2025

  19. [19]

    OPTAGENT: Optimizing multi-agent LLM interactions through verbal reinforcement learning for enhanced reasoning

    Zhenyu Bi, Meng Lu, Yang Li, Swastik Roy, Weijie Guan, Morteza Ziyadi, and Xuan Wang. OPTAGENT: Optimizing multi-agent LLM interactions through verbal reinforcement learning for enhanced reasoning. InProceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association f...

    work page 2025

  20. [20]

    Ensemble Kalman methods for inverse problems.Inverse Problems, 29(4):045001, 2013

    Marco A Iglesias, Kody JH Law, and Andrew M Stuart. Ensemble Kalman methods for inverse problems.Inverse Problems, 29(4):045001, 2013

    work page 2013

  21. [21]

    Ensemble Kalman inversion: a derivative-free technique for machine learning tasks.Inverse Problems, 35(9):095005, 2019

    Nikola B Kovachki and Andrew M Stuart. Ensemble Kalman inversion: a derivative-free technique for machine learning tasks.Inverse Problems, 35(9):095005, 2019

    work page 2019

  22. [22]

    Active learning literature survey

    Burr Settles. Active learning literature survey. 2009

    work page 2009

  23. [23]

    A survey on active learning: State-of-the-art, practical challenges and research directions.Mathematics, 11(4):820, 2023

    Alaa Tharwat and Wolfram Schenck. A survey on active learning: State-of-the-art, practical challenges and research directions.Mathematics, 11(4):820, 2023

    work page 2023

  24. [24]

    Bayesian approaches to associative learning: From passive to active learning

    John K Kruschke. Bayesian approaches to associative learning: From passive to active learning. Learning & behavior, 36(3):210–226, 2008

    work page 2008

  25. [25]

    Batchbald: Efficient and diverse batch acquisition for deep Bayesian active learning.Advances in neural information processing systems, 32, 2019

    Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep Bayesian active learning.Advances in neural information processing systems, 32, 2019

    work page 2019

  26. [26]

    A survey of deep active learning.ACM computing surveys (CSUR), 54 (9):1–40, 2021

    Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning.ACM computing surveys (CSUR), 54 (9):1–40, 2021

    work page 2021

  27. [27]

    Active learning with real annotation costs

    Burr Settles, Mark Craven, and Lewis Friedland. Active learning with real annotation costs. In Proceedings of the NIPS workshop on cost-sensitive learning, volume 1. Vancouver, CA:, 2008

    work page 2008

  28. [28]

    Deep Bayesian active learning for natural language processing: Results of a large-scale empirical study

    Aditya Siddhant and Zachary C Lipton. Deep Bayesian active learning for natural language processing: Results of a large-scale empirical study. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2904–2909, 2018

    work page 2018

  29. [29]

    Bayesian active learning for classification and preferenc e learning,

    Neil Houlsby, Ferenc Huszár, Zoubin Ghahramani, and Máté Lengyel. Bayesian active learning for classification and preference learning.arXiv preprint arXiv:1112.5745, 2011. 11

    work page arXiv 2011

  30. [30]

    Deep Bayesian active learning with image data

    Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. InInternational conference on machine learning, pages 1183–1192. PMLR, 2017

    work page 2017

  31. [31]

    Improving generalization with active learning

    David Cohn, Les Atlas, and Richard Ladner. Improving generalization with active learning. Machine learning, 15(2):201–221, 1994

    work page 1994

  32. [32]

    Maximizing expected model change for active learning in regression

    Wenbin Cai, Ya Zhang, and Jun Zhou. Maximizing expected model change for active learning in regression. In2013 IEEE 13th international conference on data mining, pages 51–60. IEEE, 2013

    work page 2013

  33. [33]

    Bayesian optimal experimental design with Wasserstein information criteria.arXiv preprint arXiv:2504.10092, 2025

    Tapio Helin, Youssef Marzouk, and Jose Rodrigo Rojo-Garcia. Bayesian optimal experimental design with Wasserstein information criteria.arXiv preprint arXiv:2504.10092, 2025

    work page arXiv 2025

  34. [34]

    Bayesian experimental design for model discrepancy calibration: A rivalry between Kullback–Leibler divergence and Wasserstein distance.arXiv preprint arXiv:2601.16425, 2026

    Huchen Yang, Xinghao Dong, and Jin-Long Wu. Bayesian experimental design for model discrepancy calibration: A rivalry between Kullback–Leibler divergence and Wasserstein distance.arXiv preprint arXiv:2601.16425, 2026

    work page arXiv 2026

  35. [35]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  36. [36]

    Diffusion Posterior Sampling for General Noisy Inverse Problems

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffu- sion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

    work page internal anchor Pith review arXiv 2022

  37. [37]

    Exponential convergence of Langevin distributions and their discrete approximations

    Gareth O Roberts and Richard L Tweedie. Exponential convergence of Langevin distributions and their discrete approximations. 1996

    work page 1996

  38. [38]

    Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    work page 2019

  39. [39]

    Bayesian experimental design for model discrepancy calibration: An auto-differentiable ensemble Kalman inversion approach.Journal of Computational Physics, 545:114469, 2026

    Huchen Yang, Xinghao Dong, and Jin-Long Wu. Bayesian experimental design for model discrepancy calibration: An auto-differentiable ensemble Kalman inversion approach.Journal of Computational Physics, 545:114469, 2026

    work page 2026

  40. [40]

    Reverse-annealed sequential Monte Carlo for efficient Bayesian optimal experiment design

    Jake Callahan, Andrew Chin, Jason Pacheco, and Tommie Catanach. Reverse-annealed sequential Monte Carlo for efficient Bayesian optimal experiment design. InThe Thirty- ninth Annual Conference on Neural Information Processing Systems, 2025. URL https: //openreview.net/forum?id=jut5q3UYRz

    work page 2025

  41. [41]

    Sentence-BERT: Sentence embeddings using siamese BERT- networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT- networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 3982–3992. Association for Computational Linguistics, 11 2019

    work page 2019

  42. [42]

    MiniLM: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Advances in neural information processing systems, 33:5776–5788, 2020

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. MiniLM: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Advances in neural information processing systems, 33:5776–5788, 2020

    work page 2020

  43. [43]

    Instruction embedding: Latent representations of instructions towards task identification.Advances in Neural Information Processing Systems, 37:87683–87711, 2024

    Yiwei Li, Jiayi Shi, Shaoxiong Feng, Peiwen Yuan, Xinglin Wang, Boyuan Pan, Heda Wang, Yao Hu, and Kan Li. Instruction embedding: Latent representations of instructions towards task identification.Advances in Neural Information Processing Systems, 37:87683–87711, 2024

    work page 2024

  44. [44]

    Measuring massive multitask language understanding.Proceedings of the International Conference on Learning Representations (ICLR), 2021

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.Proceedings of the International Conference on Learning Representations (ICLR), 2021

    work page 2021

  45. [45]

    Training Verifiers to Solve Math Word Problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021. 12 A Methodology details and extensions A.1 EKI and score-based posterior sampling...

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  46. [46]

    (23) Similarly, it can be shown that C ll ≈G qC zz G⊤ q

    Justification of the first approximation Consider C zl = 1 J−1 JX j=1 (z(j) − ¯z)(l(j) − ¯l)⊤ ≈ 1 J−1 JX j=1 (z(j) − ¯z) h Gq(z(j) − ¯z) i⊤ = 1 J−1 JX j=1 (z(j) − ¯z)(z(j) − ¯z)⊤G⊤ q =C zz G⊤ q . (23) Similarly, it can be shown that C ll ≈G qC zz G⊤ q . Therefore, the first approximation holds, and the approximation will be exact when the forward model is linear

    work page

  47. [47]

    task usage

    Proof of the final equality Here we show C zz G⊤ q GqC zz G⊤ q + Γ −1 =C locG⊤ q Γ−1.(24) We first defineS=G qC zz G⊤ q + Γfor notational simplicity. Starting from the right-hand side, ClocG⊤ q Γ−1 =C locG⊤ q Γ−1SS −1 =C locG⊤ q Γ−1(GqC zz G⊤ q + Γ)S−1 =C loc G⊤ q Γ−1GqC zz G⊤ q S−1 +G ⊤ q Γ−1ΓS−1 =C loc G⊤ q Γ−1GqC zz G⊤ q S−1 +G ⊤ q S−1 =C loc G⊤ q Γ−1G...

    work page

  48. [48]

    The answer is

    Overall, across the two commonly used ways to allocate an equal training budget, we consistently observe that active learning outperforms random task training. In addition, on MMLU we performed a sanity check by swapping the random-training repetition schedule from 20 tasks × 1 to 10 tasks × 2 while keeping the total task usages fixed. The random baseline...

    work page 1990