pith. sign in

arxiv: 2606.17628 · v1 · pith:6TPOZUGUnew · submitted 2026-06-16 · 💻 cs.CL

OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation

Guibin Zhang , Xun Xu , Yanwei Yue , Zikun Su , Wangchunshu Zhou , Xiaobin Hu , Shuicheng Yan This is my paper

Pith reviewed 2026-06-27 01:04 UTC · model grok-4.3

classification 💻 cs.CL
keywords on-policy distillationagent evolvermemory hierarchyself-evolving agentsslow-fast co-evolutionmemory managementoutcome-calibrated attribution
0
0 comments X

The pith

OPD-Evolver trains agents to read, use, write and maintain memory through slow-fast on-policy distillation.

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

The paper argues that storing trajectories is not enough for agents to evolve; they must also learn to select, act on, write, and maintain useful experience. OPD-Evolver implements this via a fast loop that interacts with a four-level memory hierarchy for immediate test-time gains and a slow loop that applies outcome-calibrated attribution plus privileged hindsight to distill the four abilities back into the policy. The resulting system outperforms prior memory systems by up to 11.5 percent and training baselines by about 5.8 percent across multi-domain benchmarks. A sympathetic reader cares because the work shifts focus from external memory add-ons to internalized, reusable evolution skills that let a 9B model compete with far larger systems.

Core claim

OPD-Evolver cultivates holistic agent evolvers through on-policy self-distillation in a slow-fast co-evolution framework: the fast loop interacts with a four-level memory hierarchy to read, use, write, and maintain experience for rapid test-time evolution, while the slow loop uses outcome-calibrated memory attribution and privileged hindsight to internalize these abilities in the deployable policy.

What carries the argument

The slow-fast co-evolution framework with on-policy self-distillation applied to a four-level memory hierarchy that supports read, use, write, and maintain operations.

If this is right

  • OPD-Evolver surpasses memory systems such as ReasoningBank by up to 11.5 percent on multi-domain benchmarks.
  • OPD-Evolver outperforms training-based methods such as Skill0 by approximately 5.8 percent.
  • OPD-Evolver-9B challenges much larger models such as Qwen3.5-397B-A17B and Step-3.5-Flash.
  • OPD-Evolver internalizes high-value experience and memory management.

Where Pith is reading between the lines

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

  • The distillation process could be applied to other agent behaviors such as planning or tool selection beyond memory.
  • Smaller models trained this way may reduce the need for ever-larger base models in deployed agent systems.
  • The framework might support continual learning across changing environments without retraining from scratch.
  • Outcome calibration in the slow loop could serve as a general mechanism for attributing credit in multi-step agent trajectories.

Load-bearing premise

That outcome-calibrated memory attribution and privileged hindsight in the slow loop can reliably transfer the four abilities from fast-loop interactions into the deployable policy without domain-specific tuning or benchmark artifacts.

What would settle it

An experiment showing that OPD-Evolver produces no gains over baselines on a new domain or when the memory hierarchy is removed from the fast loop.

Figures

Figures reproduced from arXiv: 2606.17628 by Guibin Zhang, Shuicheng Yan, Wangchunshu Zhou, Xiaobin Hu, Xun Xu, Yanwei Yue, Zikun Su.

Figure 1
Figure 1. Figure 1: (Top) The fast loop lets the agent interact with environments and a four-level memory hierarchy; (Down) the slow loop converts outcome-calibrated hindsight into on-policy self-distillation signals for OPD-Evolver. tool-use tasks (Zhong et al., 2026). In contrast, our work uses OPD not merely to strengthen execution, but to jointly cultivate the four capabilities required by a holistic agent evolver. 3 Meth… view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of calibrated memory scores for [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of calibrated memory scores for [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Case study of experience selection and writing [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional results for memory selection on [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional results for memory writing on [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: focuses on the EXECUTOR. The exam￾ples show that the vanilla executor can violate task￾level constraints, stop before the true success con￾dition is reached, or omit required side effects. In InterCode-SQL, it issues multiple statements in one action and continues schema exploration after noisy observations. In MiniHack, it submits while merely adjacent to the goal. In LifelongAgentBench-OS, it completes g… view at source ↗
Figure 10
Figure 10. Figure 10: analyzes the WRITER. The vanilla writer tends to convert failures into broad reusable Vanilla OPD Task: sql_207 Year with most EUR gas use Outcome: failure Execution Trace USE debit_card_specializing Schema exploration attempts Wrong column Currency assumed DESCRIBE syntax error -> data rows Writer Reflect on failure? Extract broadly: 3 new_skills - schema exploration pattern - year aggregation - database… view at source ↗
read the original abstract

Memory has become a standard substrate for self-evolving agents, yet retaining experience is not the same as learning how to evolve through it. Existing memory agents can store trajectories, retrieve reflections, or accumulate skills, but often lack the holistic competence to select useful experience, act on it, write reusable knowledge, and maintain a growing repository. We introduce OPD-Evolver, a slow-fast co-evolution framework that cultivates such an agent evolver through on-policy self-distillation. In the fast loop, OPD-Evolver interacts with a four-level memory hierarchy to read, use, write, and maintain experience for rapid test-time evolution. In the slow loop, outcome-calibrated memory attribution and privileged hindsight distill these four abilities into the deployable policy. Across multi-domain benchmarks, OPD-Evolver surpasses memory systems such as ReasoningBank by up to 11.5%, and training-based methods such as Skill0 by ~5.8%. Further analysis shows that OPD-Evolver internalizes high-value experience and memory management, enabling OPD-Evolver-9B to challenge giant counterparts such as Qwen3.5-397B-A17B and Step-3.5-Flash, pointing beyond memory-augmented agents toward genuinely qualified agent evolvers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces OPD-Evolver, a slow-fast co-evolution framework that cultivates holistic agent evolvers via on-policy self-distillation. In the fast loop, the agent interacts with a four-level memory hierarchy to read, use, write, and maintain experience. In the slow loop, outcome-calibrated memory attribution and privileged hindsight distill these four abilities into the deployable policy. Across multi-domain benchmarks, OPD-Evolver claims to surpass memory systems such as ReasoningBank by up to 11.5% and training-based methods such as Skill0 by ~5.8%, with further analysis showing internalization of high-value experience that allows a 9B model to challenge much larger models.

Significance. If the central claim holds—that on-policy distillation successfully transfers the four memory abilities into the final policy without depending on privileged hindsight signals unavailable at deployment—this would advance the field by moving beyond memory-augmented agents toward genuinely self-evolving ones. The reported gains and size comparisons would be notable if shown to be robust and mechanism-driven rather than benchmark-specific.

major comments (2)
  1. [Abstract] Abstract: performance numbers (11.5% over ReasoningBank, ~5.8% over Skill0) and the claim of internalization are stated without methods, data details, verification steps, or ablations; this prevents determining whether gains arise from the claimed distillation mechanism or other factors such as benchmark artifacts or the privileged slow-loop signals.
  2. [Method] Slow-loop description: outcome-calibrated attribution and privileged hindsight are described as distilling the four abilities (read/use/write/maintain), but no equations, fitting details, or transfer analysis are provided to assess whether these abilities are internalized by the deployable fast-loop policy or remain tied to slow-loop access; this is load-bearing for the holistic evolver claim.
minor comments (1)
  1. [Abstract] Abstract: 'OPD-Evolver-9B' is referenced without specifying the base model, training configuration, or how it relates to the four-level hierarchy.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address the major comments point by point below, providing clarifications from the manuscript and committing to targeted revisions that will make the distillation mechanism and its empirical support more explicit.

read point-by-point responses

  1. Referee: [Abstract] Abstract: performance numbers (11.5% over ReasoningBank, ~5.8% over Skill0) and the claim of internalization are stated without methods, data details, verification steps, or ablations; this prevents determining whether gains arise from the claimed distillation mechanism or other factors such as benchmark artifacts or the privileged slow-loop signals.

    Authors: We agree the abstract is concise and omits setup details. The reported gains come from the multi-domain experiments and ablations described in Sections 4 and 5, which compare against ReasoningBank and Skill0 while controlling for model size and training data. The internalization claim is supported by the fast-loop-only evaluations in Section 5.3. We will revise the abstract to add one sentence referencing the benchmark domains and the ablation results that attribute gains to the on-policy distillation rather than slow-loop signals at test time. revision: yes

  2. Referee: [Method] Slow-loop description: outcome-calibrated attribution and privileged hindsight are described as distilling the four abilities (read/use/write/maintain), but no equations, fitting details, or transfer analysis are provided to assess whether these abilities are internalized by the deployable fast-loop policy or remain tied to slow-loop access; this is load-bearing for the holistic evolver claim.

    Authors: Section 3.3 describes outcome-calibrated attribution (weighting memories by outcome delta) and privileged hindsight (generating targets unavailable at deployment). The manuscript already contains transfer experiments in Section 5.3 showing the fast-loop policy retains gains without slow-loop access. To make the mechanism load-bearing claim fully verifiable, we will add explicit equations for the attribution function and distillation loss, plus expanded quantitative transfer metrics for each of the four abilities, in the revision. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The provided abstract and method description outline a slow-fast co-evolution framework using on-policy self-distillation with outcome-calibrated attribution, but contain no equations, fitted parameters presented as predictions, or self-citation chains that reduce the central claims to inputs by construction. The four abilities (read/use/write/maintain) and transfer via privileged hindsight are described as empirical outcomes across benchmarks rather than derived tautologically from the inputs. No load-bearing self-definitional steps or ansatz smuggling are identifiable from the text. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; the four-level memory hierarchy and outcome-calibrated attribution are described at high level without implementation specifics.

pith-pipeline@v0.9.1-grok · 5778 in / 1157 out tokens · 50610 ms · 2026-06-27T01:04:24.336313+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

299 extracted references · 5 canonical work pages

  1. [1]

    arXiv preprint arXiv:2505.20286 , year=

    Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution , author=. arXiv preprint arXiv:2505.20286 , year=

    arXiv

  2. [2]

    2026 , eprint=

    Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning , author=. 2026 , eprint=

    2026

  3. [3]

    2026 , eprint=

    Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning , author=. 2026 , eprint=

    2026

  4. [4]

    2026 , eprint=

    Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents , author=. 2026 , eprint=

    2026

  5. [5]

    2026 , eprint=

    Skill-R1: Agent Skill Evolution via Reinforcement Learning , author=. 2026 , eprint=

    2026

  6. [6]

    2026 , eprint=

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills , author=. 2026 , eprint=

    2026

  7. [7]

    2026 , eprint=

    OpenClaw-RL: Train Any Agent Simply by Talking , author=. 2026 , eprint=

    2026

  8. [8]

    2026 , eprint=

    SOD: Step-wise On-policy Distillation for Small Language Model Agents , author=. 2026 , eprint=

    2026

  9. [9]

    2026 , eprint=

    On-Policy Context Distillation for Language Models , author=. 2026 , eprint=

    2026

  10. [10]

    2026 , eprint=

    Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation , author=. 2026 , eprint=

    2026

  11. [11]

    2026 , eprint=

    A Survey of On-Policy Distillation for Large Language Models , author=. 2026 , eprint=

    2026

  12. [12]

    2026 , eprint=

    Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe , author=. 2026 , eprint=

    2026

  13. [13]

    2024 , eprint=

    On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes , author=. 2024 , eprint=

    2024

  14. [14]

    2015 , eprint=

    Distilling the Knowledge in a Neural Network , author=. 2015 , eprint=

    2015

  15. [15]

    2025 , eprint=

    PyVision: Agentic Vision with Dynamic Tooling , author=. 2025 , eprint=

    2025

  16. [16]

    2026 , eprint=

    EnvScaler: Scaling Tool-Interactive Environments for LLM Agent via Programmatic Synthesis , author=. 2026 , eprint=

    2026

  17. [17]

    2026 , eprint=

    On Data Engineering for Scaling LLM Terminal Capabilities , author=. 2026 , eprint=

    2026

  18. [18]

    2026 , eprint=

    Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning , author=. 2026 , eprint=

    2026

  19. [19]

    2026 , eprint=

    AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution , author=. 2026 , eprint=

    2026

  20. [20]

    2025 , eprint=

    TaskCraft: Automated Generation of Agentic Tasks , author=. 2025 , eprint=

    2025

  21. [21]

    , author =

    `DeepResearchAgent`: A Hierarchical Multi-Agent Framework for General-purpose Task Solving. , author =

  22. [22]

    arXiv preprint arXiv:2308.10848 , volume=

    Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents , author=. arXiv preprint arXiv:2308.10848 , volume=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2310.02170 , year=

    Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization , author=. arXiv preprint arXiv:2310.02170 , year=

    Pith/arXiv arXiv

  24. [24]

    2023 , eprint=

    RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text , author=. 2023 , eprint=

    2023

  25. [25]

    Embodied

    Xudong Guo and Kaixuan Huang and Jiale Liu and Wenhui Fan and Natalia V. Embodied. Language Gamification - NeurIPS 2024 Workshop , year=

    2024

  26. [26]

    , author =

    `smolagents`: a smol library to build great agentic systems. , author =

  27. [27]

    RAGBench: Explainable benchmark for retrieval-augmented generation systems.arXiv preprint arXiv:2407.11005, 2024

    RAGBench: Explainable Benchmark for Retrieval-Augmented Generation Systems , author =. 2025 , eprint =. doi:10.48550/arXiv.2407.11005 , archiveprefix =

    work page doi:10.48550/arxiv.2407.11005 2025

  28. [28]

    2024 , eprint=

    From LLM to Conversational Agent: A Memory Enhanced Architecture with Fine-Tuning of Large Language Models , author=. 2024 , eprint=

    2024

  29. [29]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

    2023

  30. [30]

    2024 , eprint=

    Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models , author=. 2024 , eprint=

    2024

  31. [31]

    arXiv preprint arXiv:1909.06146 , year=

    Pubmedqa: A dataset for biomedical research question answering , author=. arXiv preprint arXiv:1909.06146 , year=

    Pith/arXiv arXiv 1909

  32. [32]

    Langchain: Build context-aware reasoning applications , author =

  33. [33]

    Advances in Neural Information Processing Systems , volume=

    Camel: Communicative agents for" mind" exploration of large language model society , author=. Advances in Neural Information Processing Systems , volume=

  34. [34]

    Sirui Hong and Mingchen Zhuge and Jonathan Chen and Xiawu Zheng and Yuheng Cheng and Jinlin Wang and Ceyao Zhang and Zili Wang and Steven Ka Shing Yau and Zijuan Lin and Liyang Zhou and Chenyu Ran and Lingfeng Xiao and Chenglin Wu and J. Meta. The Twelfth International Conference on Learning Representations , year=

  35. [35]

    arXiv preprint arXiv:2310.10634 , year=

    Openagents: An open platform for language agents in the wild , author=. arXiv preprint arXiv:2310.10634 , year=

    arXiv

  36. [36]

    arXiv preprint arXiv:2308.08155 , year=

    Autogen: Enabling next-gen llm applications via multi-agent conversation , author=. arXiv preprint arXiv:2308.08155 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2203.11171 , year=

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2409.03733 , year=

    Planning in natural language improves llm search for code generation , author=. arXiv preprint arXiv:2409.03733 , year=

    arXiv

  39. [39]

    BioRxiv , pages=

    Cell2Sentence: teaching large language models the language of biology , author=. BioRxiv , pages=

  40. [40]

    bioRxiv , pages=

    Evaluating the utilities of foundation models in single-cell data analysis , author=. bioRxiv , pages=. 2023 , publisher=

    2023

  41. [41]

    Cell , volume=

    Mapping transcriptomic vector fields of single cells , author=. Cell , volume=. 2022 , publisher=

    2022

  42. [42]

    bioRxiv , pages=

    A Systematic Comparison of Single-Cell Perturbation Response Prediction Models , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  43. [43]

    Bioinformatics , volume=

    scPRAM accurately predicts single-cell gene expression perturbation response based on attention mechanism , author=. Bioinformatics , volume=. 2024 , publisher=

    2024

  44. [44]

    Nature methods , volume=

    Causal identification of single-cell experimental perturbation effects with CINEMA-OT , author=. Nature methods , volume=. 2023 , publisher=

    2023

  45. [45]

    Nature methods , volume=

    Learning single-cell perturbation responses using neural optimal transport , author=. Nature methods , volume=. 2023 , publisher=

    2023

  46. [46]

    Nature biotechnology , volume=

    Cell type prioritization in single-cell data , author=. Nature biotechnology , volume=. 2021 , publisher=

    2021

  47. [47]

    2024 , eprint=

    Symbolic Learning Enables Self-Evolving Agents , author=. 2024 , eprint=

    2024

  48. [48]

    2023 , eprint=

    Agents: An Open-source Framework for Autonomous Language Agents , author=. 2023 , eprint=

    2023

  49. [49]

    2021 , eprint=

    MiniHack the Planet: A Sandbox for Open-Ended Reinforcement Learning Research , author=. 2021 , eprint=

    2021

  50. [50]

    2023 , eprint=

    InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback , author=. 2023 , eprint=

    2023

  51. [51]

    2026 , eprint=

    AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications , author=. 2026 , eprint=

    2026

  52. [52]

    cell , volume=

    Perturb-Seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens , author=. cell , volume=. 2016 , publisher=

    2016

  53. [53]

    Nature , volume=

    Dissecting cell identity via network inference and in silico gene perturbation , author=. Nature , volume=. 2023 , publisher=

    2023

  54. [54]

    Bioinformatics , volume=

    AttentionPert: accurately modeling multiplexed genetic perturbations with multi-scale effects , author=. Bioinformatics , volume=. 2024 , publisher=

    2024

  55. [55]

    Nature Biotechnology , volume=

    Predicting transcriptional outcomes of novel multigene perturbations with GEARS , author=. Nature Biotechnology , volume=. 2024 , publisher=

    2024

  56. [56]

    Molecular systems biology , volume=

    Predicting cellular responses to complex perturbations in high-throughput screens , author=. Molecular systems biology , volume=

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Predicting cellular responses to novel drug perturbations at a single-cell resolution , author=. Advances in Neural Information Processing Systems , volume=

  58. [58]

    Nature methods , volume=

    scGen predicts single-cell perturbation responses , author=. Nature methods , volume=. 2019 , publisher=

    2019

  59. [59]

    Nature methods , volume=

    Large-scale foundation model on single-cell transcriptomics , author=. Nature methods , volume=. 2024 , publisher=

    2024

  60. [60]

    bioRxiv , pages=

    Perteval-scfm: Benchmarking single-cell foundation models for perturbation effect prediction , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  61. [61]

    bioRxiv , pages=

    Benchmarking AI Models for In Silico Gene Perturbation of Cells , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  62. [62]

    Nature Reviews Methods Primers , volume=

    High-content CRISPR screening , author=. Nature Reviews Methods Primers , volume=. 2022 , publisher=

    2022

  63. [63]

    Advances in Neural Information Processing Systems , volume=

    A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types , author=. Advances in Neural Information Processing Systems , volume=

  64. [64]

    2023 , howpublished =

    Daniel Burkhardt and Andrew Benz and Richard Lieberman and Scott Gigante and Ashley Chow and Ryan Holbrook and Robrecht Cannoodt and Malte Luecken , title =. 2023 , howpublished =

    2023

  65. [65]

    arXiv preprint arXiv:2204.11817 , year=

    Translation between molecules and natural language , author=. arXiv preprint arXiv:2204.11817 , year=

    arXiv

  66. [66]

    2024 , publisher=

    Peidli, Stefan and Green, Tessa D and Shen, Ciyue and Gross, Torsten and Min, Joseph and Garda, Samuele and Yuan, Bo and Schumacher, Linus J and Taylor-King, Jake P and Marks, Debora S and others , journal=. 2024 , publisher=

    2024

  67. [67]

    2024 , month =

    Benchmarking Transcriptomics Foundation Models for Perturbation Analysis: one PCA still rules them all , author =. 2024 , month =. arXiv , primaryclass =:arXiv:2410.13956 , doi =

    arXiv 2024

  68. [68]

    2024 , month =

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery , author =. 2024 , month =. arXiv , primaryclass =:arXiv:2408.06292 , doi =

    Pith/arXiv arXiv 2024

  69. [69]

    Bulaong, John E

    The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies with Experimental Validation , author =. 2024 , month =. doi:10.1101/2024.11.11.623004 , url =

    work page doi:10.1101/2024.11.11.623004 2024

  70. [70]

    scgpt: Toward building a foundation model for single-cell multi-omics using generative ai.Nature Methods, 21:1470–1480, 2024

    scGPT: toward building a foundation model for single-cell multi-omics using generative AI , author =. Nature Methods , volume =. 2024 , month =. doi:10.1038/s41592-024-02201-0 , issn =

    work page doi:10.1038/s41592-024-02201-0 2024

  71. [71]

    Theodoris, Ling Xiao, Adit Chopra, Mark D

    Transfer learning enables predictions in network biology , author =. Nature , volume =. 2023 , month =. doi:10.1038/s41586-023-06139-9 , url =

    work page doi:10.1038/s41586-023-06139-9 2023

  72. [72]

    arXiv preprint arXiv:2005.11280 , year =

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author =. arXiv preprint arXiv:2005.11280 , year =

    arXiv 2005

  73. [73]

    arXiv:1908.10084 [cs] , note =

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks , author =. arXiv:1908.10084 [cs] , note =. 2019 , month =

    Pith/arXiv arXiv 1908

  74. [74]

    2023 , month =

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author =. 2023 , month =. arXiv , primaryclass =:arXiv:2306.05685 , doi =

    Pith/arXiv arXiv 2023

  75. [75]

    Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H

    Xingyao Wang and Boxuan Li and Yufan Song and Frank F. Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H. Tran and Fuqiang Li and Ren Ma and Mingzhang Zheng and Bill Qian and Yanjun Shao and Niklas Muennighoff and Yizhe Zhang and Binyuan Hui and Junyang Lin and Robert Brennan and Hao Peng and H...

    Pith/arXiv arXiv

  76. [76]

    2024 , month =

    Large Language Model based Multi-Agents: A Survey of Progress and Challenges , author =. 2024 , month =. arXiv , primaryclass =:arXiv:2402.01680 , doi =

    Pith/arXiv arXiv 2024

  77. [77]

    Cell , volume=

    Perturb-Seq: Dissecting molecular circuits with single-cell RNA-seq of pooled genetic screens , author=. Cell , volume=. 2016 , publisher=

    2016

  78. [78]

    Genome biology , volume=

    Cell-type-specific perturbation response prediction using single-cell rna-seq , author=. Genome biology , volume=. 2021 , publisher=

    2021

  79. [79]

    Genome biology , volume=

    scGen: single-cell gene expression prediction through deep learning , author=. Genome biology , volume=. 2019 , publisher=

    2019

  80. [80]

    Briefings in bioinformatics , volume=

    Predicting single-cell perturbation responses using deep learning with application to covid-19 , author=. Briefings in bioinformatics , volume=. 2022 , publisher=

    2022

Showing first 80 references.