Pith. sign in

REVIEW 4 major objections 5 minor 65 references

Stabilizing Black-Box Prompt Optimization with Textual Regularization and Signal Aggregation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TRAS stabilizes black-box prompt optimization by adding a success-derived textual regularizer and Monte Carlo signal aggregation to textual-gradient search, achieving relative accuracy gains of 4.9% to 21.5% over PromptAgent.

desk verdict TRAS is a plausible plug-in with useful gains, but the central attribution mechanism is untested and the evaluation is too thin to take the headline numbers at face value. read the letter →

arxiv 2507.09839 v2 pith:JNOHV5J2 submitted 2025-07-14 cs.LG

MohammadReza Davari , Utkarsh Garg , Weixin Cai , Eugene Belilovsky This is my paper
classification cs.LG
keywords black-boxpromptoptimizationtextualgradientsregularizationMonteCarlosignalaggregationmigrationinstructionlosslargelanguagemodelsengineering
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

Automatic prompt optimization typically revises prompts using only critiques of the model's mistakes, which the paper argues makes the process unstable: fixes for local errors can erase globally useful instructions and force the optimizer into costly remove-and-rediscover cycles. TRAS (Textual Regularization with Aggregated Signals) is a plug-and-play framework that adds a success-derived textual regularizer, telling the optimizer which prompt components to preserve, and Monte Carlo Signal Aggregation, which compresses several noisy gradient or regularizer samples into a single directive. The paper also formalizes Automatic Prompt Migration (APM), the task of adapting an expert prompt to a new model version without losing load-bearing instructions. Across five reasoning tasks, TRAS reports consistently higher accuracy, lower variance across seeds, faster convergence, and fewer API calls than the PromptAgent baseline, in both the standard optimization setting and the migration setting. If the attribution behind the regularizer is sound, this gives practitioners a concrete recipe for making black-box prompt search cheaper, more stable, and portable across model generations.

What carries the argument

The load-bearing mechanism is the pairing of a success-conditioned textual regularizer with Monte Carlo Signal Aggregation (MCSA). The regularizer is produced by an attribution policy $\pi_r$ that reads the current prompt and the batch of correct predictions, then emits preservation directives — do-not-remove, keep-unchanged, strengthen — attached to specific prompt components; a warm-up schedule ($\tau_{\mathrm{warmup}} = 3$ or $4$ in APO, $0$ in APM) delays activation so the optimizer explores before committing. MCSA samples $K$ independent signals from the gradient policy $\pi_g$ or the regularizer policy $\pi_r$ and uses an LLM summarizer to consolidate them into one directive, emphasizing consistent advice and filtering outliers, with a characterized variance–dilution trade-off that peaks near $K \approx 6$.

What would settle it

Run TRAS on a task where the initial prompt contains a plausible-sounding but irrelevant instruction that happens to be present in most successful runs; if the regularizer preserves that instruction and final accuracy ends at or below an error-only baseline, the attribution premise fails. A complementary control replaces the attribution policy with a random preserve-something signal: if the random version matches TRAS's accuracy, the regularizer contributes no causal information.

Watch

Extended reading notes

Core claim

The paper's central claim is that the update signal, not just the search strategy, limits black-box prompt optimization: error-only textual gradients are stochastic and one-sided, saying what to change but never what to keep, so optimization suffers semantic drift. TRAS keeps the standard textual gradient for error correction and adds a complementary textual regularizer, in which an LLM attribution policy inspects correct predictions and issues do-not-remove, keep-unchanged, or strengthen constraints on specific prompt components, with a warm-up schedule so early exploration is not anchored to a weak prompt. It then aggregates $K$ independent samples of either signal, via an LLM summarizer, into one actionable directive, trading variance reduction against dilution; the sweet spot is $K = 6$ in cold-start optimization and $K = 2$ in the warm-started migration regime. On five reasoning tasks, TRAS beats the PromptAgent backbone in every configuration, with relative accuracy gains of 4.9% to 21.5% on GPT-3.5-turbo and 3.5% to 16.0% when migrating expert prompts to GPT-4o, alongside narrower variance and lower total API calls.

Load-bearing premise

The method's load-bearing premise is that the success-derived textual regularizer is informative: the LLM attribution policy must correctly identify which prompt components actually caused the correct predictions, because if it misattributes success to irrelevant or harmful components, preserving them will anchor a mediocre prompt and cancel the benefit of corrective gradients.

Editorial extensions

If this is right

  • Existing textual-gradient search backbones can be upgraded with the regularizer and MCSA without changing the search strategy itself, since TRAS is orthogonal to how the next prompt is selected.
  • Expert prompts optimized on one model can be migrated to a newer model with far less instruction loss: TRAS preserves transferable structure while correcting target-specific mismatches.
  • Despite extra sampling calls, total API usage falls by 0.5% to 3.3% in the standard setting and 4.2% to 6.2% under migration, because avoiding remove-and-rediscover cycles shortens optimization.
  • Aggregation has a measured sweet spot: accuracy rises with the sample count $K$ up to roughly $K = 6$, then declines as over-compression dilutes the directive into generic summaries.

Reading between the lines

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

  • The paper never directly verifies that the components the attribution policy nominates are causally responsible for successes; a clean check would ablate regularizer-preserved components from the final prompt and confirm accuracy drops.
  • A stated cost constraint — pilot runs rather than a sweep chose $K = 2$ for APM because GPT-4o calls scale with $K$ — leaves the migration-regime aggregation optimum uncharacterized, and adaptive per-iteration $K$ tuning is a natural untested extension.
  • The explore-then-commit warm-up schedule generalizes beyond prompts: any iterative LLM-refinement loop that mixes corrective and preservative feedback should delay the preservative signal until enough evidence accumulates.
  • The same success-derived preservation signal could apply to other discrete black-box search spaces, such as tool-use configurations or retrieval pipelines, wherever noisy feedback can erase components that already work.
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

4 major / 5 minor

Summary. The paper proposes TRAS, a framework that augments existing black-box prompt optimization (APO) methods with a textual regularizer derived from successful predictions and with Monte Carlo Signal Aggregation (MCSA) to reduce the variance of LLM-generated update signals. It also formalizes Automatic Prompt Migration (APM) as the problem of adapting a source-model expert prompt to a target model while preventing instruction loss. The method is evaluated on five reasoning tasks using GPT-3.5-turbo for standard APO and GPT-4o for APM, with PromptAgent as the only search backbone. The reported results show consistent accuracy improvements over PromptAgent (4.9%–21.5% in APO, 3.5%–16.0% in APM), reduced variance, and lower API call counts.

Significance. If the mechanism is validated, the paper makes a useful contribution: it identifies a practical failure mode (semantic drift / instruction loss) and offers a plug-and-play module that can be attached to existing textual-gradient search methods. The MCSA aggregation idea is simple and likely transferable. The formalization of APM could also be of practical value given rapid model churn. However, the central causal claim—that the success-derived regularizer preserves load-bearing prompt components—is not yet established, and the limited experimental baselines weaken the generality of the empirical claims. The paper ships code, which is a strength for reproducibility.

major comments (4)
  1. [Section 3.3, Tables 1 and 3] The central claim that the success-derived textual regularizer preserves load-bearing prompt components is not causally tested. The +TR ablation adds π_r's preservation directives to PromptAgent, but any generic preservation bias that discourages destructive edits would register as a gain, so the ablation cannot separate informative attribution from an anchoring effect. Please add (i) a control whose regularizer is a fixed 'keep all current instructions' directive (or a random subset), (ii) a component-level test where the components identified by π_r are deleted from the prompt and target accuracy is measured, and (iii) a qualitative analysis of regularizer outputs showing that they track components actually responsible for correctness, especially after migration to GPT-4o.
  2. [Section 4.3, Tables 1 and 3] The hyperparameters τ_warmup and K are selected on the same five tasks used for the main results (Figure 3), and then Table 1 uses those values while reporting paired t-test p-values without correction for multiple comparisons. This inflates the significance; for example, the Penguins row in Table 1 gives p=0.007 for TRAS but the +MCSA-only effect is p=0.083, and with five tasks and multiple variants the family-wise error is not controlled. Please either evaluate on held-out tasks (or tasks not used for hyperparameter selection), apply a multiple-comparison correction, or explicitly state that the p-values are uncorrected and exploratory.
  3. [Section 3.6 versus Figure 3a] The APM setting sets τ_warmup=0 to preserve the transferred expert prompt, yet Figure 3a in the cold-start setting shows τ_warmup∈{0,1} underperforms because early regularization 'anchors weak prompts.' The paper justifies the difference by stating the expert prompt is already strong, but no APM-specific warm-up sweep or transfer analysis is provided. Because the APM gains in Table 3 are attributed to immediate preservation, the reader cannot tell whether the regularizer preserves transferable structure or anchors to source-specific components that are irrelevant or harmful on GPT-4o. Please present an APM ablation over τ_warmup and a component-level transfer test.
  4. [Sections 4.1 and 4.2] The empirical evaluation uses a single search backbone (PromptAgent) and no comparison against other textual-gradient APO methods such as ProTeGi or OPRO. Since the paper claims TRAS is plug-and-play with 'existing APO search backbones' (Section 1, Figure 2), at least one experiment instantiating TRAS with a different backbone (or a comparison to a second baseline family) is needed to support that claim and to establish that the gains are not specific to PromptAgent's MCTS strategy.
minor comments (5)
  1. [Section 3.4] The aggregation operator Φ is described as an 'LLM summarizer' but no details are given about its prompt, temperature, or whether the same LLM (GPT-3.5-turbo vs. GPT-4o) is used for aggregation as for generation. Specifying these details is important for reproducibility.
  2. [Equation (3.7)] The constraint 'p ∈ N(p⋆_S)' is written as if N is a neighborhood, but the method implements preservation only as a soft regularizer via r_t. Please clarify whether N is a hard constraint or a soft regularizer.
  3. [Figure 3] The normalized improvement curves in Figure 3 have no error bars or multi-seed variance. Given the stochasticity of LLM-generated signals, showing at least the standard deviation across seeds would make the sweet-spot conclusions (τ_warmup≈3–4, K≈6) more convincing.
  4. [Appendix A] There is a typo in the Penguins paragraph: 'examines' should be 'examines' or 'examines' (the text reads 'This dataset examinsines'). Also, the CB example prompt in Appendix B contains a repeated 'Hypothesis:' line that should be cleaned up.
  5. [Table 2] The p-values for 'Initial' rows compare DP and EP direct transfer; please clarify whether these are paired across the same five seeds, since the DP and EP runs appear to be independent optimizations with different initializers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; empirical claims are benchmarked externally against PromptAgent.

full rationale

TRAS's update rules (Eqs. 3.2-3.6) are not reductions of the reported outcome: the accuracy gains in Tables 1 and 3 are measured on held-out test sets relative to the external PromptAgent baseline, so they cannot be forced by the definition of the regularizer or the aggregation operator. The regularizer r_t is an LLM-generated attribution, a modeling assumption with causal content that is untested, but an assumption is not a circularity. The only self-citation is [42], the first author's unpublished thesis, cited for the APM label; however, the paper itself defines APM in Eq. 3.7, so the thesis is not needed to derive any result. Hyperparameters K and tau_warmup are selected from ablations and then reused in the main runs; this is transparent tuning on validation-style data, not a fitted parameter renamed as a prediction. The tension that APM sets tau_warmup=0 while Figure 3a shows early warm-up can hurt is a correctness/robustness concern, not a circular step. No step of the claimed derivation chain reduces to its own input by construction.

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

The central empirical claims rely on several hand-set hyperparameters chosen on the same evaluation tasks, and on the assumption that LLM-generated textual signals are reliable. No new physical or conceptual entities are postulated; textual regularization and MCSA are procedural additions, not entities.

free parameters (4)
  • K (MCSA sample count, APO) = 6
    Chosen by ablation (Fig. 3b) on the same five tasks to maximize relative improvement; used in all APO runs.
  • K (MCSA sample count, APM) = 2
    Chosen based on pilot runs and API cost; no accuracy gain observed for K >= 3.
  • tau_warmup (regularization activation iteration) = 3 (Causal Judgment, Geometric Shapes, Penguins), 4 (Biosses, CB)
    Per-dataset value selected from ablation (Fig. 3a) that gives peak accuracy on the same tasks.
  • Aggregation and update prompt templates = Not specified in paper; reside in code
    The behavior of MCSA and the prompt update depends on hand-written LLM prompts (Phi, pi_upd) that are not included in the manuscript.
assumptions (4)
  • domain assumption The risk R(p) over the validation set is a reliable proxy for test performance (Eq. 3.1).
    Standard in prompt optimization; small validation sets (30-95 examples) make this proxy noisy but the paper uses it without discussion.
  • domain assumption LLM-generated textual critiques and regularizers are informative enough to guide optimization.
    The whole method assumes pi_g and pi_r produce useful signals; no verification of signal quality beyond final accuracy.
  • domain assumption Aggregating multiple sampled signals via an LLM summarizer reduces variance without systematic bias (Eq. 3.5).
    MCSA is justified empirically, not theoretically; the dilution effect is acknowledged but not modeled.
  • ad hoc to paper A warm-up period prevents premature anchoring to a weak initial prompt.
    The schedule (Eq. 3.2) and the per-dataset tau are introduced specifically to make the method work; no general principle given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stabilizing Black-Box Prompt Optimization with Textual Regularization and Signal Aggregation." pith.science (2026). https://pith.science/paper/JNOHV5J2

@misc{pith2026250709839,
  author       = {Pith},
  title        = {Pith review of: Stabilizing Black-Box Prompt Optimization with Textual Regularization and Signal Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JNOHV5J2}},
  note         = {Machine review of arXiv:2507.09839}
}
read the original abstract

An increasing number of NLP applications interact with large language models (LLMs) through black-box APIs, making prompt engineering critical for controlling model behavior. Recent Automatic Prompt Optimization (APO) methods iteratively refine prompts using model-generated critiques (often called textual gradients), but they predominantly optimize from failures and underutilize information contained in correct predictions, leading to instability and semantic drift. We propose TRAS (Textual Regularization with Aggregated Signals), a feedback-centric framework that is plug-and-play with existing APO search backbones. It retains the standard textual gradient signal from prior work for error correction and introduces a complementary textual regularizer derived from successful predictions to preserve beneficial prompt components. Because both signals are stochastic and can be noisy, we further introduce Monte Carlo Signal Aggregation (MCSA), which samples multiple gradients or regularizers and aggregates them into a single actionable directive, emphasizing consistent, actionable advice while filtering out outliers. Motivated by rapid model churn, we also formalize Automatic Prompt Migration (APM), the practical problem of adapting an expert prompt across model versions or API providers without losing critical instructions. Across standard APO and APM scenarios, our approach consistently outperforms strong baselines, yielding higher accuracy, faster convergence, and lower query cost, while substantially reducing the degradation observed under naive prompt migration.

Figures

Figures reproduced from arXiv: 2507.09839 by the authors.

Figure 1
Figure 1. TRAS preserves prompt structure across optimization. Error-only APO (or￾ange) updates prompts from failures alone; the trajectory oscillates with a wide variance band and working prompt components get erased (instruction loss). TRAS (blue) adds a textual regularizer from successes and aggregates multiple samples per signal (MCSA), producing a smoother trajectory with a narrower variance band that better approaches t… view at source ↗
Figure 2
Figure 2. Overview of our proposed frame￾work for automatic prompt optimization (APO). The framework consists of five pri￾mary modules. The Search Module is ab￾stracted to allow for the integration of var￾ious search and planning methods. APO methods can be broadly categorized by their level of access to model internals. Methods with full or partial access to parameters, gradients, or output probabilities, applicable to open-… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 35 canonical work pages

  1. [1]

    On the opportunities and risks of foundation models

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, et al. “On the opportunities and risks of foundation models”. In:arXiv preprint arXiv:2108.07258(2021)

    arXiv 2021

  2. [2]

    Bert: Pre-training of deep bidirec- tional transformers for language understanding

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. “Bert: Pre-training of deep bidirec- tional transformers for language understanding”. In:Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). 2019, pp. 4171–4186

    work page 2019

  3. [3]

    BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension

    M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V. Stoyanov, and L. Zettlemoyer. “BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension”. In:Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, July ...

    work page 2020

  4. [4]

    Improving language under- standing by generative pre-training

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al. “Improving language under- standing by generative pre-training”. In: (2018)

    work page 2018

  5. [5]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J. Liu. “Exploring the limits of transfer learning with a unified text-to-text transformer”. In: Journal of machine learning research21.140 (2020), pp. 1–67

    work page 2020

  6. [6]

    Toponym identification in epidemiology articles– a deep learning approach

    M. Davari, L. Kosseim, and T. D. Bui. “Toponym identification in epidemiology articles– a deep learning approach”. In:International Conference on Computational Linguistics and Intelligent Text Processing. Springer. 2019, pp. 26–37

    work page 2019

  7. [7]

    Semantic Similarity Matching Using Contextualized Representations

    F. Farahnak, E. Mohammadi, M. Davari, and L. Kosseim. “Semantic Similarity Matching Using Contextualized Representations.” In:Canadian AI. 2021

    work page 2021

  8. [8]

    TIMBERT: toponym identifier for the medical domain based on BERT

    M. Davari, L. Kosseim, and T. Bui. “TIMBERT: toponym identifier for the medical domain based on BERT”. In:Proceedings of the 28th International Conference on Computational Linguistics. 2020, pp. 662–668

    work page 2020

Show all 65 references
  1. [9]

    Toxbuster: In-game chat toxicity buster with BERT

    Z. Yang, Y. Maricar, M. Davari, N. Grenon-Godbout, and R. Rabbany. “Toxbuster: In-game chat toxicity buster with BERT”. In:arXiv preprint arXiv:2305.12542(2023)

    2023 arXiv

  2. [10]

    Clac at semeval-2024 task 2: Faithful clinical trial infer- ence

    J. Marks, M. Davari, and L. Kosseim. “Clac at semeval-2024 task 2: Faithful clinical trial infer- ence”. In:Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval- 2024). 2024, pp. 1673–1677

    2024

  3. [11]

    Neural network approaches to medical toponym recognition

    M. Davari. “Neural network approaches to medical toponym recognition”. PhD thesis. Con- cordia University, 2020

    2020

  4. [12]

    A primer in BERTology: What we know about how BERT works

    A. Rogers, O. Kovaleva, and A. Rumshisky. “A primer in BERTology: What we know about how BERT works”. In:Transactions of the association for computational linguistics8 (2021), pp. 842–866

    2021

  5. [13]

    Probing representation for- getting in supervised and unsupervised continual learning

    M. Davari, N. Asadi, S. Mudur, R. Aljundi, and E. Belilovsky. “Probing representation for- getting in supervised and unsupervised continual learning”. In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 16712–16721

    2022

  6. [14]

    Similarityofneuralnetworkrepresentations revisited

    S.Kornblith,M.Norouzi,H.Lee,andG.Hinton.“Similarityofneuralnetworkrepresentations revisited”. In:International conference on machine learning. PMLR. 2019, pp. 3519–3529

    2019

  7. [15]

    Do vision transform- ers see like convolutional neural networks?

    M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. “Do vision transform- ers see like convolutional neural networks?” In:Advances in neural information processing systems34 (2021), pp. 12116–12128. 10

    2021

  8. [16]

    Reliability of CKA as a Similarity Measure in Deep Learning

    M. Davari, S. Horoi, A. Natik, G. Lajoie, G. Wolf, and E. Belilovsky. “Reliability of CKA as a Similarity Measure in Deep Learning”. In:The Eleventh International Conference on Learning Representations. 2023

    2023

  9. [17]

    On the inadequacy of CKA as a measure of similarity in deep learning

    M. Davari, S. Horoi, A. Natik, G. Lajoie, G. Wolf, and E. Belilovsky. “On the inadequacy of CKA as a measure of similarity in deep learning”. In:ICLR 2022 Workshop on Geometrical and Topological Representation Learning. 2022

    2022

  10. [18]

    Prefix-tuning: Optimizing continuous prompts for generation

    X. L. Li and P. Liang. “Prefix-tuning: Optimizing continuous prompts for generation”. In: arXiv preprint arXiv:2101.00190(2021)

    2021 arXiv

  11. [19]

    The power of scale for parameter-efficient prompt tuning

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

    2021 arXiv

  12. [20]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al. “Lora: Low-rank adaptation of large language models.” In:ICLR1.2 (2022), p. 3

    2022

  13. [21]

    Model breadcrumbs: scalable upcycling of finetuned foundation models via sparse task vectors merging

    M. Davari and E. Belilovsky. “Model breadcrumbs: scalable upcycling of finetuned foundation models via sparse task vectors merging”. In:ICML 2024 Workshop on Foundation Models in the Wild. 2024

    2024

  14. [22]

    Ties-merging: Resolving inter- ference when merging models

    P. Yadav, D. Tam, L. Choshen, C. A. Raffel, and M. Bansal. “Ties-merging: Resolving inter- ference when merging models”. In:Advances in Neural Information Processing Systems36 (2023), pp. 7093–7115

    2023

  15. [23]

    Model breadcrumbs: Scaling multi-task model merging with sparse masks

    M. Davari and E. Belilovsky. “Model breadcrumbs: Scaling multi-task model merging with sparse masks”. In:European Conference on Computer Vision. Springer. 2024, pp. 270–287

    2024

  16. [24]

    Language models are super mario: Absorbing abilities from homologous models as a free lunch

    L. Yu, B. Yu, H. Yu, F. Huang, and Y. Li. “Language models are super mario: Absorbing abilities from homologous models as a free lunch”. In:Forty-first International Conference on Machine Learning. 2024

    2024

  17. [25]

    Gpt-4 technical report

    OpenAI. “Gpt-4 technical report”. In:arXiv preprint arXiv:2303.08774(2023)

    2023 arXiv

  18. [26]

    Bubeck, V

    S. Bubeck, V. Chadrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y. T. Lee, Y. Li, S. Lundberg, et al.Sparks of artificial general intelligence: Early experiments with gpt-4. 2023

    2023

  19. [27]

    BioGPT: generative pre-trained transformer for biomedical text generation and mining

    R. Luo, L. Sun, Y. Xia, T. Qin, S. Zhang, H. Poon, and T.-Y. Liu. “BioGPT: generative pre-trained transformer for biomedical text generation and mining”. In:Briefings in bioin- formatics23.6 (2022), bbac409

    2022

  20. [28]

    Large language models are zero- shot reasoners

    T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa. “Large language models are zero- shot reasoners”. In:Advances in neural information processing systems35 (2022), pp. 22199– 22213

    2022

  21. [29]

    Self-consistency improves chain of thought reasoning in language models

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou. “Self-consistency improves chain of thought reasoning in language models”. In:arXiv preprint arXiv:2203.11171(2022)

    2022 arXiv

  22. [30]

    Least-to-most prompting enables complex reasoning in large language models

    D. Zhou, N. Schärli, L. Hou, J. Wei, N. Scales, X. Wang, D. Schuurmans, C. Cui, O. Bousquet, Q. Le, et al. “Least-to-most prompting enables complex reasoning in large language models”. In:arXiv preprint arXiv:2205.10625(2022)

    2022 arXiv

  23. [31]

    Self-refine: Iterative refinement with self-feedback

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, et al. “Self-refine: Iterative refinement with self-feedback”. In:Ad- vances in Neural Information Processing Systems36 (2023), pp. 46534–46594

    2023

  24. [32]

    Constitutional ai: Harmlessness from ai feedback

    Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, et al. “Constitutional ai: Harmlessness from ai feedback”. In:arXiv preprint arXiv:2212.08073(2022)

    2022 arXiv

  25. [33]

    Teaching large language models to self-debug

    X. Chen, M. Lin, N. Schärli, and D. Zhou. “Teaching large language models to self-debug”. In:arXiv preprint arXiv:2304.05128(2023)

    2023 arXiv

  26. [34]

    Structgpt: A general framework for large language model to reason over structured data

    J. Jiang, K. Zhou, Z. Dong, K. Ye, W. X. Zhao, and J.-R. Wen. “Structgpt: A general framework for large language model to reason over structured data”. In:arXiv preprint arXiv:2305.09645(2023)

    2023 arXiv

  27. [35]

    Chain-of-thought prompting elicits reasoning in large language models

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. “Chain-of-thought prompting elicits reasoning in large language models”. In:Advances in neural information processing systems35 (2022), pp. 24824–24837

    2022

  28. [36]

    Better zero-shot reasoning with role-play prompting

    A. Kong, S. Zhao, H. Chen, Q. Li, Y. Qin, R. Sun, X. Zhou, E. Wang, and X. Dong. “Better zero-shot reasoning with role-play prompting”. In:arXiv preprint arXiv:2308.07702(2023). 11

    2023 arXiv

  29. [37]

    Promp- tagent: Strategic planning with language models enables expert-level prompt optimization

    X. Wang, C. Li, Z. Wang, F. Bai, H. Luo, J. Zhang, N. Jojic, E. P. Xing, and Z. Hu. “Promp- tagent: Strategic planning with language models enables expert-level prompt optimization”. In:arXiv preprint arXiv:2310.16427(2023)

    2023 arXiv

  30. [38]

    Large language models as optimizers

    C. Yang, X. Wang, Y. Lu, H. Liu, Q. V. Le, D. Zhou, and X. Chen. “Large language models as optimizers”. In:arXiv preprint arXiv:2309.03409(2023)

    2023 arXiv

  31. [39]

    Large language models are human-level prompt engineers

    Y. Zhou, A. I. Muresanu, Z. Han, K. Paster, S. Pitis, H. Chan, and J. Ba. “Large language models are human-level prompt engineers”. In:arXiv preprint arXiv:2211.01910(2022)

    2022 arXiv

  32. [40]

    Automatic prompt optimization with

    R. Pryzant, D. Iter, J. Li, Y. T. Lee, C. Zhu, and M. Zeng. “Automatic prompt optimization with" gradient descent" and beam search”. In:arXiv preprint arXiv:2305.03495(2023)

    2023 arXiv

  33. [41]

    Symbolic prompt program search: A structure-aware approach to efficient compile-time prompt optimization

    T. Schnabel and J. Neville. “Symbolic prompt program search: A structure-aware approach to efficient compile-time prompt optimization”. In:arXiv preprint arXiv:2404.02319(2024)

    2024 arXiv

  34. [42]

    Continual Learning in Constrained Scenarios: Bridging Real-World Needs and Practical Constraints

    M. Davari. “Continual Learning in Constrained Scenarios: Bridging Real-World Needs and Practical Constraints”. Unpublished. PhD thesis. Concordia University, 2025.url:https: //spectrum.library.concordia.ca/id/eprint/995684/

    2025

  35. [43]

    Tempera: Test-time prompting via reinforcement learning

    T. Zhang, X. Wang, D. Zhou, D. Schuurmans, and J. E. Gonzalez. “Tempera: Test-time prompting via reinforcement learning”. In:arXiv preprint arXiv:2211.11890(2022)

    2022 arXiv

  36. [44]

    Let’s Do a Thought Experiment: Using Counterfactuals to Improve Moral Reasoning

    X. Ma, S. Mishra, A. Beirami, A. Beutel, and J. Chen. “Let’s Do a Thought Experiment: Using Counterfactuals to Improve Moral Reasoning”. In:arXiv preprint arXiv:2306.14308 (2023)

    2023 arXiv

  37. [45]

    When do you need Chain-of-Thought Prompting for ChatGPT?

    J. Chen, L. Chen, H. Huang, and T. Zhou. “When do you need Chain-of-Thought Prompting for ChatGPT?” In:arXiv preprint arXiv:2304.03262(2023)

    2023 arXiv

  38. [46]

    Prototype-Sample Rela- tion Distillation: Towards Replay-Free Continual Learning

    N. Asadi, M. Davari, S. Mudur, R. Aljundi, and E. Belilovsky. “Prototype-Sample Rela- tion Distillation: Towards Replay-Free Continual Learning”. In:International Conference on Machine Learning. PMLR. 2023, pp. 1093–1106

    2023

  39. [47]

    Llama: Open and efficient foundation language models

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, et al. “Llama: Open and efficient foundation language models”. In:arXiv preprint arXiv:2302.13971(2023)

    2023 arXiv

  40. [48]

    Llama 2: Open foundation and fine-tuned chat models

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al. “Llama 2: Open foundation and fine-tuned chat models”. In: arXiv preprint arXiv:2307.09288(2023)

    2023 arXiv

  41. [49]

    The llama 3 herd of models

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al. “The llama 3 herd of models”. In:arXiv preprint arXiv:2407.21783(2024)

    2024 arXiv

  42. [50]

    Mistral 7b. arxiv

    A. Q. Jiang, A Sablayrolles, A Mensch, C Bamford, D. S. Chaplot, D. Casas, F Bressand, G Lengyel, G Lample, L Saulnier, et al. “Mistral 7b. arxiv”. In:arXiv preprint arXiv:2310.06825 10 (2023)

    2023 arXiv

  43. [51]

    M. A. team.Mistral nemo. Accessed: 2025. 2024.url:https://mistral.ai/news/mistral- nemo

    2025

  44. [52]

    Multitask prompt tuning enables parameter-efficient transfer learning

    Z. Wang, R. Panda, L. Karlinsky, R. Feris, H. Sun, and Y. Kim. “Multitask prompt tuning enables parameter-efficient transfer learning”. In:arXiv preprint arXiv:2303.02861(2023)

    2023 arXiv

  45. [53]

    Learning how to ask: Querying LMs with mixtures of soft prompts

    G. Qin and J. Eisner. “Learning how to ask: Querying LMs with mixtures of soft prompts”. In:arXiv preprint arXiv:2104.06599(2021)

    2021 arXiv

  46. [54]

    Autoprompt: Eliciting knowledge from language models with automatically generated prompts

    T. Shin, Y. Razeghi, R. L. Logan IV, E. Wallace, and S. Singh. “Autoprompt: Eliciting knowledge from language models with automatically generated prompts”. In:arXiv preprint arXiv:2010.15980(2020)

    2020 arXiv

  47. [55]

    Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery

    Y. Wen, N. Jain, J. Kirchenbauer, M. Goldblum, J. Geiping, and T. Goldstein. “Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery”. In:Ad- vances in Neural Information Processing Systems36 (2023), pp. 51008–51025

    2023

  48. [56]

    Making pre-trained language models better few-shot learn- ers

    T. Gao, A. Fisch, and D. Chen. “Making pre-trained language models better few-shot learn- ers”. In:arXiv preprint arXiv:2012.15723(2020)

    2020 arXiv

  49. [57]

    Instructzero: Efficient instruc- tion optimization for black-box large language models

    L. Chen, J. Chen, T. Goldstein, H. Huang, and T. Zhou. “Instructzero: Efficient instruc- tion optimization for black-box large language models”. In:arXiv preprint arXiv:2306.03082 (2023)

    2023 arXiv

  50. [58]

    Optimizing prompts for text-to-image generation

    Y. Hao, Z. Chi, L. Dong, and F. Wei. “Optimizing prompts for text-to-image generation”. In: Advances in Neural Information Processing Systems36 (2023), pp. 66923–66939. 12

    2023

  51. [59]

    Efficient selectivity and backup operators in Monte-Carlo tree search

    R. Coulom. “Efficient selectivity and backup operators in Monte-Carlo tree search”. In:In- ternational conference on computers and games. Springer. 2006, pp. 72–83

    2006

  52. [60]

    MARS: A Multi-Agent Framework Incorporating Socratic Guidance for Automated Prompt Optimization

    J. Zhang, Z. Wang, H. Zhu, J. Liu, Q. Lin, and E. Cambria. “MARS: A Multi-Agent Framework Incorporating Socratic Guidance for Automated Prompt Optimization”. In:arXiv preprint arXiv:2503.16874(2025)

    2025

  53. [61]

    A Survey of Automatic Prompt Engineering: An Opti- mization Perspective

    W. Li, X. Wang, W. Li, and B. Jin. “A Survey of Automatic Prompt Engineering: An Opti- mization Perspective”. In:arXiv preprint arXiv:2502.11560(2025)

    2025 arXiv

  54. [62]

    Challenging big-bench tasks and whether chain-of-thought can solve them

    M. Suzgun, N. Scales, N. Schärli, S. Gehrmann, Y. Tay, H. W. Chung, A. Chowdhery, Q. V. Le, E. H. Chi, D. Zhou, et al. “Challenging big-bench tasks and whether chain-of-thought can solve them”. In:arXiv preprint arXiv:2210.09261(2022)

    2022 arXiv

  55. [63]

    The commitmentbank: Investigating projection in naturally occurring discourse

    M.-C. De Marneffe, M. Simons, and J. Tonhauser. “The commitmentbank: Investigating projection in naturally occurring discourse”. In:proceedings of Sinn und Bedeutung. Vol. 23. 2019, pp. 107–124

    2019

  56. [64]

    BIOSSES: a semantic sentence similarity estima- tion system for the biomedical domain

    G. Soğancıoğlu, H. Öztürk, and A. Özgür. “BIOSSES: a semantic sentence similarity estima- tion system for the biomedical domain”. In:Bioinformatics33.14 (2017), pp. i49–i58

    2017

  57. [65]

    Superglue: A stickier benchmark for general-purpose language understanding systems

    A. Wang, Y. Pruksachatkun, N. Nangia, A. Singh, J. Michael, F. Hill, O. Levy, and S. Bow- man. “Superglue: A stickier benchmark for general-purpose language understanding systems”. In:Advances in neural information processing systems32 (2019). 13 Dataset T rain V alidation T e...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.