pith. machine review for the scientific record. sign in

arxiv: 2604.18026 · v1 · submitted 2026-04-20 · 💻 cs.LG · cs.AI

Recognition: unknown

RASP-Tuner: Retrieval-Augmented Soft Prompts for Context-Aware Black-Box Optimization in Non-Stationary Environments

Enze Pan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 04:41 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords black-box optimizationnon-stationary environmentscontext-aware tuningretrieval-augmented promptsmixture of expertsdynamic regretonline adaptation
0
0 comments X

The pith

RASP-Tuner retrieves similar past contexts to supply soft prompts that let a mixture-of-experts surrogate adapt black-box tuning to non-stationary regimes at low per-step cost.

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

The paper casts online black-box tuning as regret minimization conditioned on an observed context that switches among a small number of latent regimes. It decomposes the task into three steps: retrieve a proxy for the current regime from history, feed the retrieved soft prompt together with the current parameters and context into a mixture-of-experts loss predictor, and perform most adaptation inside the low-dimensional prompt subspace while triggering full surrogate updates only on error spikes. A RealErrorComposer converts heterogeneous streaming metrics into a single [0,1] target via EMA-stabilized logistic scores. The resulting algorithm is shown to match or beat GP-UCB and CMA-ES on seven of nine synthetic non-stationary tasks at horizon 100 while running eight to twelve times faster than sliding-window GP-UCB, and to supply sufficient conditions for bounded dynamic regret under an idealized cluster-separated convex model.

Core claim

RASP-Tuner instantiates a first-principles decomposition for context-aware black-box optimization: (i) identify the active regime by nearest-neighbor retrieval of past contexts, (ii) predict short-horizon loss with a mixture-of-experts surrogate whose input is the concatenation of parameters, context, and a retrieved soft prompt, and (iii) adapt chiefly inside the prompt subspace, invoking expensive full surrogate refits only when scalarized prediction error or expert disagreement exceeds a threshold. The RealErrorComposer supplies a differentiable training target by mapping heterogeneous metrics to [0,1] via EMA-stabilized logistic scores. Under an idealized RA-GD analysis in a cluster-sepa

What carries the argument

Retrieval-augmented soft prompt: a low-dimensional vector retrieved from similar past contexts that is concatenated to the surrogate input and becomes the primary subspace for online adaptation.

If this is right

  • On nine synthetic benchmarks at T=100 the method improves or matches paired-test regret against GP-UCB and CMA-ES on seven tasks.
  • Wall-clock time per step is 8-12 times lower than sliding-window GP-UCB on the same hardware.
  • The RA-GD analysis gives sufficient conditions for bounded dynamic regret when regimes are cluster-separated and strongly convex.
  • The RealErrorComposer produces a single differentiable target from any collection of heterogeneous streaming metrics.

Where Pith is reading between the lines

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

  • The same retrieval-plus-soft-prompt pattern could be inserted into other online optimizers that already maintain a low-dimensional adaptation subspace.
  • If context similarity can be computed incrementally, the approach may extend to continuous or slowly drifting context spaces without requiring an explicit regime count.
  • The error-spike trigger for full updates offers a general way to trade off surrogate fidelity against compute in any streaming surrogate-based optimizer.

Load-bearing premise

Contexts revisit a small set of latent regimes so that retrieving similar past contexts reliably supplies useful soft prompts.

What would settle it

Run the method on a stream whose contexts are all distinct and never repeat; if cumulative regret then equals that of a non-retrieval baseline, the retrieval premise does not hold.

Figures

Figures reproduced from arXiv: 2604.18026 by Enze Pan.

Figure 1
Figure 1. Figure 1: Architecture of RASP-Tuner. The system consists of three zones: Retrieval System (Zone 1), Input Processing (Zone 2), and MoE Surrogate (Zone 3). Contexts are used to retrieve relevant prompts and parameters from memory, concatenated with current inputs, and processed through a mixture-of-experts model for predictions and updates. Algorithm 1 RASP-Tuner (one-step update) 1: Input: Bounds, context fn, metri… view at source ↗
Figure 2
Figure 2. Figure 2: Main results across 9 non-stationary benchmarks. Each row corresponds to one environment, with loss (left) and cumulative regret (right) vs. steps. Color convention (all plots): RASP-Tuner (red), CMA-ES (green), Bayesian Optimization (blue), Random Search (gray). In drifting and regime-switching tasks (Wafer Etching, Switching LQR, Regime Switch, Server Flash Crowd, Real Trace Replay, LLM Inference), RASP-… view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of RASP-Tuner. Left: Ablation comparing full RASP (with PromptMemory) vs. a NoMemory variant where contexts are masked to zero. Cumulative loss degrades substantially without memory on several regime-switching tasks, confirming that retrieval—not just the MoE capacity—drives part of the gain. Middle: Latency (ms/step) for RASP, CMA-ES, and BO. BO is roughly 8–12× slower per step under our timing s… view at source ↗
Figure 4
Figure 4. Figure 4: Adversarial Context failure case. Cumulative regret for RASP-Tuner (red), CMA-ES (green), Bayesian Optimization (blue), and Random Search (gray) (same color convention as [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-world streams: validation loss (left) and cumulative regret (right) per dataset; shaded regions are 95% CI over seeds. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
read the original abstract

Many deployed systems expose black-box objectives whose minimizing configuration shifts with an externally observed context. When contexts revisit a small set of latent regimes, an optimizer that discards history pays repeated adaptation cost; when each step must remain inexpensive, full Gaussian-process (GP) refits at high observation counts are difficult to sustain. We cast online tuning as context-conditioned regret minimization and present RASP-Tuner, which instantiates a decomposition motivated by first principles: (i) identify a regime proxy by retrieving similar past contexts; (ii) predict short-horizon loss with a mixture-of-experts surrogate whose input concatenates parameters, context, and a retrieved soft prompt; (iii) adapt chiefly in a low-dimensional prompt subspace, invoking full surrogate updates only when scalarized error or disagreement spikes. A RealErrorComposer maps heterogeneous streaming metrics to [0,1] via EMA-stabilized logistic scores, supplying a single differentiable training target. On nine synthetic non-stationary benchmarks, an adversarial-context sanity check, and three tabular real-world streams (Section on real-world experiments), RASP-Tuner improves or matches cumulative regret relative to our GP-UCB and CMA-ES implementations on seven of nine synthetic tasks under paired tests at horizon T=100, while recording 8-12 times lower wall-clock per step than sliding-window GP-UCB on identical hardware. Idealized analysis in a cluster-separated, strongly convex regime model (RA-GD) supplies sufficient conditions for bounded dynamic regret; the deployed pipeline violates several of these premises, and we articulate which gaps remain open.

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

1 major / 2 minor

Summary. The manuscript introduces RASP-Tuner for context-aware black-box optimization in non-stationary environments. It retrieves similar past contexts to supply soft prompts to a mixture-of-experts surrogate (input concatenates parameters, context, and prompt), adapts primarily in the low-dimensional prompt subspace, triggers full surrogate updates only on error or disagreement spikes, and uses a RealErrorComposer (EMA-stabilized logistic scores) to map heterogeneous metrics to a [0,1] differentiable target. On nine synthetic non-stationary benchmarks plus an adversarial-context check and three real tabular streams, it improves or matches cumulative regret versus GP-UCB and CMA-ES on seven of nine tasks under paired tests at horizon T=100 while achieving 8-12x lower wall-clock time per step than sliding-window GP-UCB. An idealized RA-GD analysis in a cluster-separated strongly convex regime supplies sufficient conditions for bounded dynamic regret, but the authors explicitly note that the deployed pipeline violates several of those premises and articulate the remaining gaps.

Significance. If the empirical improvements are robust and the method can be placed on firmer theoretical footing, RASP-Tuner would offer a practical, low-cost alternative to repeated full GP refits for online tuning tasks where contexts recur across a small number of latent regimes. The computational savings and the honest acknowledgment of analysis gaps are strengths, but the absence of a regret bound that covers the implemented algorithm limits the significance of the theoretical contribution.

major comments (1)
  1. [Abstract and RA-GD analysis] Abstract and RA-GD analysis section: the idealized RA-GD model supplies sufficient conditions for bounded dynamic regret under cluster-separated, strongly convex regimes, yet the deployed pipeline (retrieval + mixture-of-experts surrogate + prompt-subspace adaptation + RealErrorComposer) is stated to violate several of those premises. No alternative regret analysis is supplied for the actual algorithm, so the central claim that RASP-Tuner improves or matches regret on seven of nine tasks lacks theoretical grounding that would explain why the method succeeds when the analyzed conditions fail. This gap is load-bearing for the paper's positioning as a principled context-aware optimizer.
minor comments (2)
  1. [Method description] The EMA decay rate in RealErrorComposer and the disagreement threshold for triggering full surrogate updates are listed as free parameters; their sensitivity to performance and any default values or tuning protocol should be reported explicitly.
  2. [Experiments] The abstract reports results at horizon T=100; longer-horizon experiments or scaling behavior with observation count would strengthen the practical claims.

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for the detailed and constructive feedback. We address the major comment below, with an honest assessment of what can and cannot be revised.

read point-by-point responses

  1. Referee: [Abstract and RA-GD analysis] Abstract and RA-GD analysis section: the idealized RA-GD model supplies sufficient conditions for bounded dynamic regret under cluster-separated, strongly convex regimes, yet the deployed pipeline (retrieval + mixture-of-experts surrogate + prompt-subspace adaptation + RealErrorComposer) is stated to violate several of those premises. No alternative regret analysis is supplied for the actual algorithm, so the central claim that RASP-Tuner improves or matches regret on seven of nine tasks lacks theoretical grounding that would explain why the method succeeds when the analyzed conditions fail. This gap is load-bearing for the paper's positioning as a principled context-aware optimizer.

    Authors: We agree that the RA-GD analysis applies strictly to an idealized retrieval-augmented gradient descent procedure under cluster-separated strongly convex assumptions and does not bound the full deployed pipeline. The manuscript already states that the implemented system violates several of those premises. The RA-GD section is offered as a source of intuition for the retrieval-plus-prompt-adaptation decomposition rather than a guarantee for the mixture-of-experts surrogate, RealErrorComposer, or approximate retrieval. No alternative regret bound for the complete algorithm is supplied because extending the analysis to cover these components appears technically difficult and is left open. The central empirical claim (competitive regret on seven of nine tasks with 8-12x lower per-step cost) is presented as an experimental result, not as a consequence of the idealized bound. We will revise the abstract and the RA-GD section to state more explicitly that the theoretical results apply only to the simplified model and do not extend to the deployed RASP-Tuner, thereby clarifying the paper's positioning as a practical method with partial theoretical support. revision: partial

standing simulated objections not resolved
  • A regret bound that covers the full deployed RASP-Tuner (including retrieval approximation, mixture-of-experts dynamics, and RealErrorComposer) is not provided and cannot be derived within the scope of this work.

Circularity Check

0 steps flagged

No circularity: empirical claims independent of idealized RA-GD analysis

full rationale

The paper states its performance claims solely via empirical paired tests on nine synthetic benchmarks and real-world streams at T=100, without deriving those outcomes from the RA-GD model. It explicitly notes that the deployed pipeline (retrieval + mixture-of-experts + prompt adaptation + RealErrorComposer) violates several premises of the cluster-separated strongly-convex RA-GD regime used only for sufficient conditions on bounded dynamic regret. No equations reduce a prediction or first-principles result to its own inputs by construction, no self-citation supplies a load-bearing uniqueness theorem, and no fitted parameter is relabeled as a prediction. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

Review performed on abstract only; free parameters and axioms are inferred from stated components rather than explicit equations.

free parameters (2)
  • EMA decay rate in RealErrorComposer
    Stabilizes logistic scores for heterogeneous streaming metrics; value not reported in abstract.
  • Disagreement threshold for full surrogate update
    Controls when expensive refits occur; not numerically specified.
axioms (2)
  • domain assumption Contexts revisit a small set of latent regimes
    Explicitly stated as the condition under which retrieval supplies useful history.
  • domain assumption Short-horizon loss is predictable from parameter-context-prompt concatenation
    Underlies the mixture-of-experts surrogate design.
invented entities (1)
  • RealErrorComposer no independent evidence
    purpose: Maps heterogeneous streaming metrics to a single differentiable [0,1] target via EMA-stabilized logistic scores
    New component introduced to supply a unified training signal; no independent evidence outside the paper.

pith-pipeline@v0.9.0 · 5585 in / 1633 out tokens · 43114 ms · 2026-05-10T04:41:55.762331+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

61 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    Evolutionary computation , volume=

    Completely derandomized self-adaptation in evolution strategies , author=. Evolutionary computation , volume=. 2001 , publisher=

    2001

  2. [2]

    Advances in neural information processing systems , volume=

    Contextual gaussian process bandit optimization , author=. Advances in neural information processing systems , volume=

  3. [4]

    Advances in Neural Information Processing Systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in Neural Information Processing Systems , volume=

  4. [5]

    International conference on machine learning , pages=

    Model-agnostic meta-learning for fast adaptation of deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  5. [6]

    International Conference on Learning Representations , year=

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer , author=. International Conference on Learning Representations , year=

  6. [7]

    Advances in neural information processing systems , volume=

    Practical bayesian optimization of machine learning algorithms , author=. Advances in neural information processing systems , volume=

  7. [8]

    2023 , eprint=

    ISP meets Deep Learning: A Survey on Deep Learning Methods for Image Signal Processing , author=. 2023 , eprint=

    2023

  8. [9]

    2014 IEEE international conference on cloud engineering , pages=

    Autoscaling web applications in heterogeneous cloud infrastructures , author=. 2014 IEEE international conference on cloud engineering , pages=. 2014 , organization=

    2014

  9. [10]

    Data Drift Quantification Metrics for Month-on-Month Semiconductor Data , year=

    Bhattacharya, Nandita and Barari, Adrita and Pedapudi, Lakshmi Narayana and Kim, Hoyeon and Kim, Kyuhwan and Seo, Jihye , booktitle=. Data Drift Quantification Metrics for Month-on-Month Semiconductor Data , year=

  10. [11]

    Artificial intelligence and statistics , pages=

    Deep kernel learning , author=. Artificial intelligence and statistics , pages=. 2016 , organization=

    2016

  11. [12]

    Advances in neural information processing systems , volume=

    Sparse Gaussian processes using pseudo-inputs , author=. Advances in neural information processing systems , volume=

  12. [13]

    International Conference on Machine Learning , pages=

    Adaptive and safe Bayesian optimization in high dimensions via one-dimensional subspaces , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  13. [14]

    Advances in neural information processing systems , volume=

    Learning to learn by gradient descent by gradient descent , author=. Advances in neural information processing systems , volume=

  14. [15]

    Proceedings of the IEEE , volume=

    Taking the human out of the loop: A review of Bayesian optimization , author=. Proceedings of the IEEE , volume=. 2015 , publisher=

    2015

  15. [16]

    The journal of machine learning research , volume=

    Random search for hyper-parameter optimization , author=. The journal of machine learning research , volume=. 2012 , publisher=

    2012

  16. [17]

    Automated Machine Learning: Methods, Systems, Challenges , editor =

    Hyperparameter Optimization , author =. Automated Machine Learning: Methods, Systems, Challenges , editor =

  17. [18]

    International Conference on Learning Representations , year =

    Generalization through Memorization: Nearest Neighbor Language Models , author =. International Conference on Learning Representations , year =

  18. [19]

    Artificial Intelligence and Statistics , pages=

    Time-varying Gaussian process bandit optimization , author=. Artificial Intelligence and Statistics , pages=. 2016 , organization=

    2016

  19. [20]

    Proceedings of the Twentieth International Conference on Machine Learning (ICML'03) , pages =

    Online Convex Programming and Generalized Infinitesimal Gradient Ascent , author =. Proceedings of the Twentieth International Conference on Machine Learning (ICML'03) , pages =

  20. [21]

    Operations Research , volume =

    Non-stationary Stochastic Optimization , author =. Operations Research , volume =. 2015 , doi =

    2015

  21. [22]

    IEEE Journal of Selected Topics in Signal Processing , volume =

    Online Convex Optimization in Dynamic Environments , author =. IEEE Journal of Selected Topics in Signal Processing , volume =. 2015 , doi =

    2015

  22. [23]

    Foundations and Trends

    Introduction to online convex optimization , author=. Foundations and Trends. 2016 , publisher=

    2016

  23. [24]

    Proceedings of the 34th Conference on Learning Theory (COLT) , pages=

    Non-stationary Reinforcement Learning without Prior Knowledge: An Optimal Black-box Approach , author=. Proceedings of the 34th Conference on Learning Theory (COLT) , pages=. 2021 , organization=

    2021

  24. [26]

    Shi, Weijia and Min, Sewon and Yasunaga, Michihiro and Seo, Minjoon and James, Rich and Lewis, Mike and Zettlemoyer, Luke and Yih, Wen-tau , journal=

  25. [28]

    Proceedings of the 29th International Conference on Computational Linguistics (COLING) , pages=

    Context-Tuning: Learning Contextualized Prompts for Natural Language Generation , author=. Proceedings of the 29th International Conference on Computational Linguistics (COLING) , pages=

  26. [29]

    Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    Time-aware Prompting for Text Generation , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    2022

  27. [30]

    Anupam, Sagnik and Shypula, Alexander and Bastani, Osbert and others , journal=

  28. [31]

    Gupta, Taneesh and Shandilya, Shivam and Zhang, Xuchao and Madhavan, Rahul and Ghosh, Supriyo and Bansal, Chetan and Yao, Huaxiu and Rajmohan, Saravan , booktitle=

  29. [33]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  30. [34]

    , journal=

    Chu, Jaewon and Lee, Seunghun and Kim, Hyunwoo J. , journal=

  31. [35]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , pages=

    PromptBoosting: Black-Box Text Classification with Ten Forward Passes , author=. Proceedings of the 40th International Conference on Machine Learning (ICML) , pages=

  32. [37]

    Context tuning for retrieval augmented generation

    Anantha, R., Bethi, T., Vodianik, D., and Chappidi, S. Context tuning for retrieval augmented generation. arXiv preprint arXiv:2312.05708, 2023

    work page arXiv 2023

  33. [38]

    W., Pfau, D., Schaul, T., Shillingford, B., and De Freitas, N

    Andrychowicz, M., Denil, M., Gomez, S., Hoffman, M. W., Pfau, D., Schaul, T., Shillingford, B., and De Freitas, N. Learning to learn by gradient descent by gradient descent. In Advances in neural information processing systems, volume 29, 2016

    2016

  34. [39]

    LLM Program Optimization via Retrieval Augmented Search

    Anupam, S., Shypula, A., Bastani, O., et al. LLM Program Optimization via Retrieval Augmented Search . arXiv preprint arXiv:2501.18916, 2025

    work page arXiv 2025

  35. [40]

    and Bengio, Y

    Bergstra, J. and Bengio, Y. Random search for hyper-parameter optimization. The journal of machine learning research, 13 0 (1): 0 281--305, 2012

    2012

  36. [41]

    Non-stationary stochastic optimization.Operations Research, 63(5):1227–1244, 2015

    Besbes, O., Gur, Y., and Zeevi, A. Non-stationary stochastic optimization. Operations Research, 63 0 (5): 0 1227--1244, 2015. doi:10.1287/opre.2015.1408

    work page doi:10.1287/opre.2015.1408 2015

  37. [42]

    Time-varying gaussian process bandit optimization

    Bogunovic, I., Scarlett, J., and Cevher, V. Time-varying gaussian process bandit optimization. In Artificial Intelligence and Statistics, pp.\ 314--323. PMLR, 2016

    2016

  38. [43]

    and Wang, L

    Cao, S. and Wang, L. Time-aware prompting for text generation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pp.\ 7231--7246, 2022

    2022

  39. [44]

    Chu, J., Lee, S., and Kim, H. J. PRESTO: Preimage-Informed Instruction Optimization for Prompting Black-Box LLMs . arXiv preprint arXiv:2510.25808, 2025

    work page arXiv 2025

  40. [45]

    Model-agnostic meta-learning for fast adaptation of deep networks

    Finn, C., Abbeel, P., and Levine, S. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pp.\ 1126--1135. PMLR, 2017

    2017

  41. [46]

    CARMO: Dynamic Criteria Generation for Context Aware Reward Modelling

    Gupta, T., Shandilya, S., Zhang, X., Madhavan, R., Ghosh, S., Bansal, C., Yao, H., and Rajmohan, S. CARMO: Dynamic Criteria Generation for Context Aware Reward Modelling . In Findings of the Association for Computational Linguistics: ACL 2025, pp.\ 2202--2261, 2025

    2025

  42. [47]

    Hall, E. C. and Willett, R. M. Online convex optimization in dynamic environments. IEEE Journal of Selected Topics in Signal Processing, 9 0 (4): 0 647--662, 2015. doi:10.1109/JSTSP.2015.2404790

    work page doi:10.1109/jstsp.2015.2404790 2015

  43. [48]

    Online reinforcement learning in non-stationary context-driven environments

    Hamadanian, P., Nasr-Esfahany, A., Schwarzkopf, M., Sen, S., and Alizadeh, M. Online reinforcement learning in non-stationary context-driven environments. arXiv preprint arXiv:2302.02182, 2023

    work page arXiv 2023

  44. [49]

    and Ostermeier, A

    Hansen, N. and Ostermeier, A. Completely derandomized self-adaptation in evolution strategies. Evolutionary computation, 9 0 (2): 0 159--195, 2001

    2001

  45. [50]

    Hazan, E. et al. Introduction to online convex optimization. Foundations and Trends in Optimization , 2 0 (3-4): 0 157--325, 2016

    2016

  46. [51]

    Promptboosting: Black-box text classification with ten forward passes

    Hou, B., O'Connor, J., Andreas, J., Chang, S., and Zhang, Y. Promptboosting: Black-box text classification with ten forward passes. In Proceedings of the 40th International Conference on Machine Learning (ICML), pp.\ 13309--13324, 2023

    2023

  47. [52]

    Generalization through memorization: Nearest neighbor language models

    Khandelwal, U., Fan, A., Jurafsky, D., Zettlemoyer, L., and Lewis, M. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020

    2020

  48. [53]

    and Ong, C

    Krause, A. and Ong, C. Contextual gaussian process bandit optimization. Advances in neural information processing systems, 24, 2011

    2011

  49. [54]

    The Power of Scale for Parameter-Efficient Prompt Tuning

    Lester, B., Al-Rfou, R., and Constant, N. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021

    work page internal anchor Pith review arXiv 2021

  50. [55]

    u ttler, H., Lewis, M., Yih, W.-t., Rockt \

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K \"u ttler, H., Lewis, M., Yih, W.-t., Rockt \"a schel, T., et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems, volume 33, pp.\ 9459--9474, 2020

    2020

  51. [56]

    Few-shot recognition via stage-wise retrieval-augmented finetuning

    Liu, T., Zhang, H., Parashar, S., and Kong, S. Few-shot recognition via stage-wise retrieval-augmented finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  52. [57]

    P., and De Freitas, N

    Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., and De Freitas, N. Taking the human out of the loop: A review of bayesian optimization. Proceedings of the IEEE, 104 0 (1): 0 148--175, 2015

    2015

  53. [58]

    arXiv preprint arXiv:2301.12652 , year=

    Shi, W., Min, S., Yasunaga, M., Seo, M., James, R., Lewis, M., Zettlemoyer, L., and Yih, W.-t. REPLUG: Retrieval-Augmented Black-Box Language Models . arXiv preprint arXiv:2301.12652, 2023

    work page arXiv 2023

  54. [59]

    Direct retrieval-augmented optimization: Synergizing knowledge selection and language models

    Shi, Z., Yan, L., Sun, W., Feng, Y., Ren, P., Ma, X., Wang, S., Yin, D., de Rijke, M., and Ren, Z. Direct retrieval-augmented optimization: Synergizing knowledge selection and language models. arXiv preprint arXiv:2505.03075, 2025

    work page arXiv 2025

  55. [60]

    and Ghahramani, Z

    Snelson, E. and Ghahramani, Z. Sparse gaussian processes using pseudo-inputs. Advances in neural information processing systems, 18, 2005

    2005

  56. [61]

    Snoek, J., Larochelle, H., and Adams, R. P. Practical bayesian optimization of machine learning algorithms. Advances in neural information processing systems, 25, 2012

    2012

  57. [62]

    X., and Wen, J.-R

    Tang, T., Li, J., Zhao, W. X., and Wen, J.-R. Context-tuning: Learning contextualized prompts for natural language generation. In Proceedings of the 29th International Conference on Computational Linguistics (COLING), pp.\ 6340--6354, 2022

    2022

  58. [63]

    Reinforced Informativeness Optimization for Long-Form Retrieval-Augmented Generation

    Wang, Y., Ren, R., Wang, Y., Zhao, W. X., Liu, J., Wu, H., and Wang, H. Reinforced informativeness optimization for long-form retrieval-augmented generation. arXiv preprint arXiv:2505.20825, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  59. [64]

    and Luo, H

    Wei, C.-Y. and Luo, H. Non-stationary reinforcement learning without prior knowledge: An optimal black-box approach. In Proceedings of the 34th Conference on Learning Theory (COLT), pp.\ 4300--4354. PMLR, 2021

    2021

  60. [65]

    G., Hu, Z., Salakhutdinov, R., and Xing, E

    Wilson, A. G., Hu, Z., Salakhutdinov, R., and Xing, E. P. Deep kernel learning. Artificial intelligence and statistics, pp.\ 370--378, 2016

    2016

  61. [66]

    Online convex programming and generalized infinitesimal gradient ascent

    Zinkevich, M. Online convex programming and generalized infinitesimal gradient ascent. In Proceedings of the Twentieth International Conference on Machine Learning (ICML'03), pp.\ 928--936, 2003

    2003