Pith. sign in

REVIEW 4 major objections 3 minor 31 references

Self-Evolving Multi-Agent Systems via Textual Backpropagation

T0 review · 4 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A multi-agent system trained like a neural network — with natural-language critique as its gradient — beats static multi-agent baselines on code, math, data, and writing benchmarks.

desk verdict Genuinely novel framework, but the empirical evaluation is compromised because the validation set is used inside the optimization loop and also reported as the final result, so the headline performance claims are not supported as written. read the letter →

arxiv 2506.09046 v3 pith:BAKUEWWF submitted 2025-06-10 cs.LG cs.AIcs.MA

Xiaowen Ma , Yunpu Ma , Chenyang Lin , Sikuan Yan , Jinhe Bi , Zixuan Cao , Yijun Tian , Volker Tresp
show 1 more author
Hinrich Schuetze
This is my paper · ORCID
classification cs.LGcs.AIcs.MA
keywords AgenticNeuralNetwork(ANN)multi-agentLLMsystemstextualbackpropagationself-evolvingagentsdynamicteamselectionpromptoptimizationcost-effectiveorchestration
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

This paper tries to establish that a multi-agent system of large language models can be trained the way a neural network is: a forward phase decomposes a task into subtasks and builds a cooperative team of specialized agents for each layer, and a backward phase sends natural-language critiques — textual gradients — back through the pipeline to revise agent prompts, roles, and the connections between teams. If true, this would mean multi-agent design no longer depends on hand-crafted workflows, because the system could self-evolve its own structure from data. The paper reports that this framework, the Agentic Neural Network (ANN), consistently outperforms static multi-agent baselines on code generation, mathematical reasoning, data analysis, and creative writing, with a budget model (GPT-4o-mini) reaching 93.9% on HumanEval — above the best GPT-4 baseline results it compares against. The reason to care is practical: it points toward a data-driven, lower-cost way to orchestrate LLM agents without manual prompt engineering or fixed team topologies.

What carries the argument

The load-bearing mechanism is the textual gradient: a natural-language critique, produced by a small set of reusable prompts, that plays the role of the numeric gradient in neural-network backpropagation. The forward phase performs dynamic routing, selecting for each layer an aggregation function from a candidate set conditioned on the task and the execution trajectory so far, while the backward phase mixes a global workflow-level gradient with per-layer local gradients through a coefficient $\beta$ and a momentum term, $G^t_{\text{local},\ell}' = \alpha G^t_{\text{local},\ell} + (1-\alpha)G^{t-1}_{\text{local},\ell}$. Every proposed revision is then gated by format validation and performance validation, so only changes that survive both checks enter the candidate pool. These textual gradients are what allow agent roles, prompts, and team topologies to co-evolve from data rather than from manual design.

What would settle it

Re-run the quoted baselines — Symbolic and DSPy/ToT on HumanEval and Creative Writing, Captain Agent and AutoGen on MATH and DABench — under ANN's exact training/validation splits, prompts, and scoring, and check whether the reported margins survive; for instance, whether a matched-run Symbolic still scores below ANN's 93.9% on HumanEval with GPT-4o-mini and whether Captain Agent still trails ANN's 80.0% on MATH under identical evaluation. A second falsifier targets the mechanism: retrain ANN without the backward textual-gradient phase but with the same forward team-selection budget and verify that the learning curves over 20 epochs separate as the paper's ablations show.

Watch

Extended reading notes

Core claim

The central discovery the paper claims to establish is that backpropagation has a working textual analogue for training a multi-agent LLM system end to end. The Agentic Neural Network treats each agent as a node and each cooperating team as a layer: in the forward phase a task is decomposed into subtasks and each layer selects an aggregation function, $f_\ell = \text{DynamicRoutingSelect}(F_\ell, \ell, I_\ell, I)$, then executes $O_\ell = \text{ExecuteLayer}(\ell, f_\ell, I_\ell, I)$. If the final output fails a performance threshold, the backward phase computes a global gradient $G_{\text{global}} = \text{ComputeGlobalGradient}(S, \tau)$ over the whole execution trajectory and layer-wise local gradients $G^t_{\text{local},\ell} = \beta G_{\text{global}} + (1-\beta)\, \text{ComputeLocalGradient}(\ell, f_\ell, \tau)$, applies momentum, and rewrites prompts, roles, and inter-layer connections. Under this scheme the paper reports consistent gains over the quoted baselines — for instance 93.9% on HumanEval with GPT-4o-mini against 85.8% for the best GPT-4 baseline, 82.8% on MATH, and 95.0% on DABench — and shows that the evolved teams transfer across backbones.

Load-bearing premise

The claim that ANN beats the baselines rests on the assumption that baseline scores quoted from Zhou et al. (2024) and Song et al. (2024) were obtained under evaluation conditions equivalent to the authors' own runs, since the paper states those numbers were taken from the earlier papers without re-evaluation.

Editorial extensions

If this is right

  • A budget model can carry the system: GPT-4o-mini orchestrated by ANN reaches 93.9% on HumanEval, above the 85.8% of the strongest GPT-4 baseline, so top-tier agentic performance no longer requires the largest, most expensive model.
  • Task-specific agent teams can be discovered from data rather than engineered by hand, since ANN creates and specializes new teams during optimization.
  • Optimization transfers across backbones: teams and prompts trained with GPT-3.5 or GPT-4o-mini evaluate well on GPT-4 (86.3% and 87.8% on HumanEval), so the evolved structures are not locked to the training model.
  • Each backward-phase component contributes to the reported gains, since the ablations show that removing momentum, validation-based filtering, or the backward pass itself lowers validation accuracy across all four benchmarks.

Reading between the lines

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

  • The textual-gradient loop is not obviously specific to language agents, so the same critique-as-gradient update could plausibly tune other black-box pipeline components — retrieval strategies, tool choice, or decoding settings — a direction the paper leaves to future work.
  • The paper's cost framing compares the model used rather than the total system cost; because ANN spends many agent calls per task, a fairer comparison would count tokens per solved problem against single-call baselines.
  • A natural next test, not run in the paper, is transfer across sibling benchmarks: whether a code team discovered on HumanEval helps on other code-generation sets and whether MATH-trained teams generalize to other competition-math collections.
  • The abstract cites seven benchmarks while the body reports four datasets plus one MMLU subset, so readers should count the evaluated tasks from the tables (five in total) rather than from the abstract.
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 / 3 minor

Summary. The paper introduces the Agentic Neural Network (ANN), a framework that treats multi-agent LLM collaboration as a layered neural network. In the forward phase, tasks are decomposed into subtasks and agent teams with aggregation functions are selected layer by layer; in the backward phase, textual feedback (called gradients) is used to refine prompts, roles, and inter-layer structures, optionally with momentum. The authors evaluate ANN on HumanEval, Creative Writing, MATH, DABench, and an MMLU-Machine Learning subset, reporting accuracy gains over several multi-agent baselines, and claim that a budget model (GPT-4o-mini) can outperform larger models when orchestrated by ANN.

Significance. If the empirical claims held on clean held-out data, the paper would make a useful contribution: it unifies prompt optimization, role design, and workflow topology search under one 'textual backpropagation' loop, and it provides a concrete mechanism for self-evolving multi-agent systems. The use of a small backbone during training and evaluation across larger backbones is also an interesting cost-performance angle. However, the current evaluation protocol is not clean: the same validation set is used for update acceptance and for final reporting, baselines are borrowed rather than reproduced, and several numbers are internally inconsistent. These issues place the central superiority claim on weak ground, so the significance can only be assessed after a proper held-out evaluation.

major comments (4)
  1. [§4.2.1, §4.3.3, Algorithm 2] The reported results are not clean held-out measurements. The paper describes only a train/validation split (§4.2.1), explicitly states that for MMLU 'the official validation set' is the evaluation set, and reports 'average results on the validation set at regular epoch intervals' (§4.3.3). Algorithm 2's ValidatePerformance accepts or rejects candidate prompt/structure updates based on that same validation set. Consequently, the numbers in Tables 1–4 and Figure 3 are the outcome of a search over updates that score well on the evaluation examples, which is selection bias rather than unbiased measurement. The central claim of surpassing baselines is therefore unsupported unless the authors re-run the evaluation on a separate, never-used test set and report those numbers.
  2. [§4.2.3, Tables 1–2] Baseline numbers are borrowed from (Zhou et al., 2024) and (Song et al., 2024) without reproduction or a demonstrated match of evaluation protocols. The text states these numbers are 'taken from' the prior papers. Even if the same datasets and splits are used, the baselines are presumably reported on their own held-out test sets, while ANN numbers are validation-selected (§4.3.3); comparing a validation-selected number with a test-set number is not 'under the same configurations' as claimed in the abstract. The authors should either reproduce the baselines under an identical protocol, including an identical held-out test split, or clearly restrict the comparison to numbers obtained under matching conditions.
  3. [§4.3.1, Tables 2 and 4] There are inconsistent numbers in the main results. For MATH, the text reports 82.5 for GPT-4o-mini while Table 2 lists 82.8; for DABench the text reports 88.88 for GPT-4 while Table 2 lists 92.0, and the text also reports 75.6 for GPT-3.5 while Table 4 reports 76.0 for the same configuration. These discrepancies are not explained by stochastic variation because the same table is supposed to report the same runs. The authors must reconcile the text and tables and state which set of numbers is canonical.
  4. [Tables 1–4, Figure 3] No error bars or confidence intervals are reported. The ablation section says each condition is repeated three times, but Figure 3 and the main tables give only point estimates. Given the small evaluation sets (e.g., HumanEval has 164 problems) and LLM sampling variability, the claimed 'consistent' improvements and the 93.9% HumanEval result for GPT-4o-mini need variance information or at least per-run values to be credible. Please add standard deviations or individual run results, and state the number of runs for the main tables.
minor comments (3)
  1. [Abstract and Introduction] The abstract in the manuscript header says 'Across seven benchmark datasets', while the full-text abstract and Section 4 say four datasets (and the paper actually reports five including MMLU-Machine Learning). This inconsistency should be corrected.
  2. [§4.3.1] The phrase 'even our 4o mini results93.9/% show' contains a stray slash and should be rewritten, e.g., 'our 4o-mini result of 93.9% shows'.
  3. [Appendix A, Table 5] The framework comparison in Appendix A cites a large number of references that are irrelevant to multi-agent systems (e.g., image generation, federated learning, recommendation systems). This is distracting and should be pruned to references that actually support the comparison.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported accuracies are selected on the validation set used for update acceptance, so headline gains over baselines may be optimistic and not clean held-out measurements.

  1. fitted input called prediction [§4.3.3 (Ablation Studies); §4.2.1 (Overview of Training and Validation); Algorithm 2 (Appendix B.1)]
    "we repeat each condition three times and report the average results on the validation set at regular epoch intervals (§4.3.3); "We use the official validation set as our evaluation set" (§4.2.1); Algorithm 2: "if ValidatePerformance(f'_ℓ, f_ℓ) then"; "w/o Validation Performance: Skips the validation-based filtering stage when selecting improved prompts and agent roles"."

    Algorithm 2's ValidatePerformance accepts or rejects candidate prompt/structure updates using validation-set scores, and the same validation set is the only evaluation set in §4.2.1 (no separate test set). The paper's reported headline accuracies are therefore not held-out predictions: they are the result of a search over agent configurations that maximize scores on those exact examples. Comparing such validation-selected numbers to baselines 'taken from' prior papers (presumably test-set numbers) is not 'under the same configurations'; the claimed superiority is fitted to the evaluation data rather than demonstrated on unseen data.

full rationale

The core framework—modeling multi-agent collaboration as layered networks and using textual gradients—is not circular in itself; the update equations and pseudo-code are well-defined and do not reduce to their outputs. However, the paper's central empirical claim (ANN surpasses leading multi-agent baselines) rests on numbers that are fitted to the validation set: updates are accepted via ValidatePerformance on the validation set, and the final tables report validation-set averages. This matches the fitted-input-called-prediction pattern. No load-bearing self-citation or imported uniqueness theorem is present; the many self-citations appear only in related work or appendix material and play no role in the derivation. The baselines being 'taken from' (Zhou et al., 2024) and (Song et al., 2024) without re-evaluation is a comparability risk rather than a circularity. Hence score 6: the central superiority claim partially reduces to its own evaluation signal.

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

The framework introduces no new physical or mathematical entities. It relies on the assumption that AI-generated critiques behave like gradient signals, and that borrowing baseline numbers from prior papers yields a fair comparison. The free parameters (eta, alpha, beta) are unstated hyperparameters that could affect the reported results.

free parameters (3)
  • step size eta
    Hyperparameter in the update rule f^{t+1}_l = f^t_l - eta*G^t_local,l (Section 3.2.2). No value reported; hand-set.
  • momentum coefficient alpha
    Controls mix of current and past updates in the momentum rule (Section 3.2.2). No value reported.
  • global/local weighting beta
    Balances global and local gradient contributions in G^t_local,l = beta*G_global + (1-beta)*ComputeLocalGradient (Section 3.2.2). No value reported.
assumptions (3)
  • domain assumption LLM-generated textual feedback can serve as effective gradient signals for optimizing prompts and structures
    Central to the backward pass (Section 3.2). The paper provides no evidence that these critiques reliably improve performance beyond the reported experiments.
  • domain assumption Validation performance is a valid criterion for accepting or rejecting candidate updates
    Algorithm 2 uses ValidatePerformance to filter updates; since validation set also serves as evaluation set, this biases reported results.
  • domain assumption Baseline results from (Zhou et al., 2024) and (Song et al., 2024) are directly comparable to the authors' own runs
    Section 4.2.3 states baselines are taken from these papers; no re-evaluation under identical conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Evolving Multi-Agent Systems via Textual Backpropagation." pith.science (2026). https://pith.science/paper/BAKUEWWF

@misc{pith2026250609046,
  author       = {Pith},
  title        = {Pith review of: Self-Evolving Multi-Agent Systems via Textual Backpropagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BAKUEWWF}},
  note         = {Machine review of arXiv:2506.09046}
}
read the original abstract

Leveraging multiple Large Language Models (LLMs) has proven effective for addressing complex, high-dimensional tasks, but current approaches often rely on static, manually engineered multi-agent configurations. To overcome these constraints, we present the Agentic Neural Network (ANN), a framework that conceptualizes multi-agent collaboration as a layered neural network architecture. In this design, each agent operates as a node, and each layer forms a cooperative team focused on a specific subtask. Our framework follows a two-phase optimization strategy: (1) Forward Phase - Drawing inspiration from neural network forward passes, tasks are dynamically decomposed into subtasks, and cooperative agent teams with suitable aggregation methods are constructed layer by layer. (2) Backward Phase - Mirroring backpropagation, we refine both global and local collaboration through iterative feedback, allowing agents to self-evolve their roles, prompts, and coordination. This neuro-symbolic approach enables our framework to create new or specialized agent teams post-training, delivering notable gains in accuracy and adaptability. Across seven benchmark datasets, our work surpasses leading multi-agent baselines under the same configurations, showing consistent performance improvements.

Figures

Figures reproduced from arXiv: 2506.09046 by the authors.

Figure 1
Figure 1. A conceptual comparison between classic neural networks (left) and our [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Difference between static agentic team and our framework. The left panel illustrates a static agentic team, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ablation results on HumanEval, Creative Writing, MATH, and DABench using the gpt-4o-mini model for [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Prompt-evolution trajectory for the HumanEval(Chen et al., 2021) review_code subtask. Boxes denote agent nodes, arrows indicate information flow, and shaded regions highlight components newly introduced at each iteration. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: Prompt-evolution trajectory for the DABench(Hu et al., 2024) task-analysis benchmark. Boxes denote agent nodes, arrows indicate information flow, and shaded regions highlight components newly introduced at each iteration. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Evolution of agent team structures on the [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Team structure visualizations for the HumanEval (Chen et al., 2021) and DABench (Hu et al., 2024) datasets. Each node’s color reflects its functional role within the system. The diagrams highlight how different types of agents coordinate and how task-specific configura…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [1]

    The answer is < answer extracted >

    Extract the answer from the reply in the format : 17 " The answer is < answer extracted >"

    work page

  2. [2]

    Compare the extracted answer with the ground truth

    work page

  3. [3]

    The answer is correct

    Based on your analysis , choose only one of the following outputs : ( a ) " The answer is correct ." ( b ) " The answer is approximated but should be correct ." ( c ) " The answer is incorrect . Correct Answer : < ground truth answer > </ ground truth answer > | Answer extracted : < answer extracted > </ answer extracted >." ( d ) " The reply doesn ’ t co...

    work page

  4. [4]

    - Evaluate the team collaboration structure

    Node - Level Analysis Within Blocks : For each problematic block : - Analyze the internal node _ input and node _ output . - Evaluate the team collaboration structure . - Propose improvements to intra - block agent collaboration , if necessary . - Document your suggestions in the ’ structure _ suggestion ’ section of the corresponding block . C.4 Prompt f...

    work page

  5. [5]

    Dynamic Block ID and Naming - Use { new _ block _ id } to assign a unique block _ id - Format name as { block _ name }X , where X = new _ block _ id

    work page

  6. [6]

    arXiv preprint arXiv:2308.08155

    Autogen: Enabling next-gen llm applica- tions via multi-agent conversation. arXiv preprint arXiv:2308.08155. Zikai Xiao, Zihan Chen, Liyinglan Liu, Y ANG FENG, Joey Tianyi Zhou, Jian Wu, Wanlu Liu, Howard Hao Yang, and Zuozhu Liu. 2024. Fedloge: Joint local and generic federated learning under long-tailed data. In The Twelfth International Conference on L...

    arXiv 2024

  7. [7]

    differentiation

    Large language models as optimizers. ArXiv, abs/2309.03409. Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V . Le, Denny Zhou, and Xinyun Chen. 2024. Large language models as optimizers. Preprint, arXiv:2309.03409. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and a...

    arXiv 2024

  8. [8]

    velocity

    Mitigating world biases: A multimodal multi- view debiasing framework for fake news video detec- tion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6492–6500. Gengyuan Zhang, Jinhe Bi, Jindong Gu, Yanyu Chen, and V olker Tresp. 2023. Spot! revisiting video- language models for event understanding. arXiv preprint arXiv:2311....

    arXiv 2023

Show all 31 references
  1. [12]

    Final Result Evaluation : Analyze the final result < final result > to determine if the task failed

  2. [13]

    - Document these findings in the ’ global _ analysis ’ section of the < output _ format >

    Solution Comparison : Compare < canonical solution > and < generated solution >: - Is the logic in < generated solution > aligned with < canonical solution >? - Where is the gap between the analysis and the standard answer ? - Identify specific issues in < generated solution >...

  3. [14]

    - Examine each block _ input and block _ output

    Block Input and Output Analysis : Based on the < task description > and < workflow trajectory >: - Do not compare the block outputs with the < canonical solution >. - Examine each block _ input and block _ output . - Identify which block ( s ) caused the task to fail . - Highl...

  4. [16]

    task _ data

    Evaluate Each Node - Check input _ variables for validity and consistency . - Valid sources include : * State variables : " task _ data " , " task _ prompt " , " task _ id " * Prior node outputs : e . g . , calculation _ expert 1_ output - For prompt modifications : * Include ...

  5. [17]

    Propose Structural Changes - Add or remove nodes ( max 3 additions ) - For added nodes , specify : * node _ name , agent , output format , prompt _ template * variable _ sources , constraints - Define from / to edges for new nodes - Update connected nodes ’ input _ variables i...

  6. [18]

    Impact on Other Nodes - Maintain logical consistency with the entire workflow

  7. [19]

    Use Available Agents - Refer to { available _ agents } for potential agents 19 - Check each agent ’ s constraints for fit - Modify agents as needed ( update prompt _ template , input _ variables , or define new agents )

  8. [21]

    Block Structure Description - Include : * block _ structure _ description : high - level purpose * block _ structure _ description _ details : including :

  9. [22]

    Nodes and connections

  10. [23]

    Node roles and logic

  11. [24]

    Input / output flow - Ensure clarity , accuracy , and alignment with structure

  12. [25]

    Provided Canonical Solution and Test Cases - Don ’ t over - optimize : block may not be the cause of failure - Avoid overfitting : feedback should remain generalized - Use < canonical solution > and < test cases > as reference only

  13. [26]

    Your role is to analyze recent feedback in combination with historical adjustments to guide team improvement for a specific workflow block

    Output Format - All feedback must be returned in this JSON format : { layerwise _ loss _ format } - Do not use arrows to represent edges ! C.5 Prompt for Momentum-Based Adjustment Task Description : You are an advanced strategic advisor focused on enhancing team performance . ...

  14. [27]

    - Since the current version < current team > was formed via previous adjustments , but < final result > still failed , analyze why earlier suggestions did not work

    Overlap Handling : - If < current feedback > overlaps with </ previous adjustment direction > , focus on these overlapping issues . - Since the current version < current team > was formed via previous adjustments , but < final result > still failed , analyze why earlier sugges...

  15. [28]

    - Slightly refine and consolidate all suggestions to form an updated version of feedback

    New Issues Maintenance : - If < current feedback > introduces new problems not found in < previous adjustment direction > , retain those . - Slightly refine and consolidate all suggestions to form an updated version of feedback . 20 Important Notes : - This block may not be th...

  16. [29]

    Evaluate the < task description > carefully , identifying key difficulty points and requirements

  17. [30]

    Compare block roles and structures from < list of all block ’ s structure description > to determine which best fits the task

  18. [31]

    Select the most appropriate block based on the task complexity . Output Format : - Output your selection using the exact format below : < selected _ agg _ func > X </ selected _ agg _ func > - For example , selecting Co de Fi na liz eB lo ck 3 should result in : < selected _ a...

    2024

  19. [2020]

    arXiv preprint arXiv:2009.03300

    Measuring massive multitask language under- standing. arXiv preprint arXiv:2009.03300. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. Prepr...

    2009 arXiv

  20. [2021]

    Preprint, arXiv:2107.03374

    Evaluating large language models trained on code. Preprint, arXiv:2107.03374. Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Ya-Ting Lu, Yi-Hsin Hung, Cheng Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou....

    2023 arXiv

  21. [2022]

    ArXiv, abs/2201.11903

    Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elic- its reasoning in large language ...

    2023 arXiv

  22. [2023]

    In Neural Information Pro- cessing Systems

    Reflexion: language agents with verbal re- inforcement learning. In Neural Information Pro- cessing Systems. 11 Linxin Song, Jiale Liu, Jieyu Zhang, Shaokun Zhang, Ao Luo, Shijian Wang, Qingyun Wu, and Chi Wang

  23. [2024]

    Preprint, arXiv:2405.19425

    Adaptive in-conversation team building for language model agents. Preprint, arXiv:2405.19425. Mirac Suzgun and Adam Tauman Kalai. 2024. Meta- prompting: Enhancing language models with task- agnostic scaffolding. Preprint, arXiv:2401.12954. Yijun Tian, Kaiwen Dong, Chunhui Zhan...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.