Pith. sign in

REVIEW 3 major objections 5 minor 5 cited by

AgentDistill: Training-Free Agent Distillation with Generalizable MCP Boxes

T0 review · 3 major / 5 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 AgentDistill claims that a small frozen student agent can inherit a strong teacher's tool-use competence by mounting a curated MCP-Box, with no training at all.

desk verdict A promising training-free distillation idea undercut by an in-sample evaluation that tests the MCP-Box on the same 100 examples used to build it. read the letter →

arxiv 2506.14728 v1 pith:4M7O643Y submitted 2025-06-17 cs.AI

Jiahao Qiu , Xinzhe Juan , Yimin Wang , Ling Yang , Xuan Qi , Tongcheng Zhang , Jiacheng Guo , Yifu Lu
show 6 more authors
Zixin Yao Hongru Wang Shilong Liu Xun Jiang Liu Leqi Mengdi Wang
This is my paper
classification cs.AI
keywords agentdistillationtraining-freeModelContextProtocol(MCP)MCP-BoxtoolreusesmalllanguagemodelsvisualquestionansweringGameof24
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

AgentDistill claims that the hardest part of agent knowledge—knowing which tools to build and how to use them—can be transferred from a large teacher model to a small student model without any gradient updates. The teacher generates self-contained Model-Context-Protocol (MCP) tools while solving tasks; these tools are abstracted, clustered, and consolidated into a compact MCP-Box that is mounted on the student at inference time. On three benchmarks, frozen 8B-scale students equipped with the box match or beat a strong GPT-4o-based tool agent, despite never being trained. If the claim holds, agent distillation becomes a library-construction problem rather than a weight-training problem.

What carries the argument

The load-bearing object is the MCP-Box, a repository of consolidated tools written in FastMCP-compatible Python, each with a functional label and a parameterized interface. It is built by a three-stage pipeline: abstraction rewrites raw teacher MCPs into configurable, task-agnostic forms; clustering groups them by function; consolidation merges each cluster into one documented, production-ready tool. At inference the entire box is exposed to the student as a callable tool set, with no retrieval or selection step. The student still does the planning and argument-filling, but the space of possible tool calls is constrained to functional, verified options, which is what lets a frozen policy absorb teacher competence without gradient updates.

What would settle it

Take a held-out split from PathVQA, SLAKE, and Game of 24 that shares no questions with the 100 examples used to construct the MCP-Box, run AgentDistill, and compare the post-distillation gains. If the reported improvements, especially the large Game-of-24 jumps, disappear or shrink sharply, the central generalization claim fails; if they persist, the transfer claim is supported.

Watch

Extended reading notes

Core claim

The paper's central discovery is that distillation for agents can be done by reusing executable tool protocols rather than by imitating reasoning traces or fine-tuning a student. The teacher agent solves a sample of tasks and emits structured MCPs; successful, syntactically correct MCPs are rewritten into parameterized forms, clustered by function, and merged into one consolidated tool per cluster. The resulting MCP-Box is mounted wholesale into a frozen student agent built on a small language model. The authors report that this procedure lifts student accuracy on every tested model and dataset, with gains up to +48.4 points on Game of 24 for GPT-3.5-turbo, and that average student performance after distillation surpasses OctoTools with GPT-4o on PathVQA and SLAKE, while remaining below the teacher on Game of 24.

Load-bearing premise

The generalization claim depends on evaluating the student on problems that were not used to build the MCP-Box; the paper draws both the generation sample and the evaluation set from the same 100 validation examples, so if those sets overlap, the gains could come from tools tuned to the test questions rather than transferable skill.

Editorial extensions

If this is right

  • The MCP-Box can be mounted onto any frozen small-agent policy, so the distillation cost is paid once per domain and then reused across different student models.
  • Because the box is external to the model, student capability can be upgraded by swapping in a better box without retraining the underlying policy.
  • On Game of 24, a single distilled MCP is invoked on 100 percent of test cases, showing that a very small tool library can encode an entire task skill.
  • Student agents with the box beat a retrieval-based competitor that uses a much larger base model, suggesting tool-library quality can matter more than backbone size for these tasks.
  • Since no gradients are applied, the same pipeline can be repeated for a new domain using only teacher rollouts and one curation pass.

Reading between the lines

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

  • Editorial extension: the untested boundary is whether the box generalizes to task families not represented in the 100 construction examples; a cluster-level holdout, building the box on one question type and testing on another, would separate memorized tools from transferable ones.
  • Editorial extension: if the evaluation set overlaps the construction set, the method still offers a cheap way to specialize agents to a known evaluation distribution, but it would be better described as test-time tool customization than as general distillation.
  • Editorial extension: an MCP-Box is a reusable artifact, so the pipeline naturally supports sharing, versioning, and safety review of distilled tool libraries, a property that trajectory-replay distillation lacks.
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

3 major / 5 minor

Summary. AgentDistill is a training-free framework for transferring tool-using competence from a large teacher agent to small frozen student agents. The teacher generates executable Model-Context-Protocols (MCPs) from successful task trajectories; these MCPs are then abstracted, clustered, and consolidated into a compact MCP-Box, which is mounted into student agents at inference time without any gradient updates. Experiments on PathVQA, SLAKE, and Game of 24 report that MCP-Box-equipped students improve over their unassisted baselines and approach or exceed the performance of OctoTools (GPT-4o). The central claim is that this constitutes generalizable, scalable agent distillation that transfers problem-solving skill to small models while avoiding training. The formal objective in Eq. (2) selects an MCP-Box B to maximize accuracy on a dataset D, and Section 4.1.1 states that the same 100 validation examples are used both for MCP-box generation and for evaluation.

Significance. If the reported gains survive evaluation on data fully disjoint from MCP construction, AgentDistill would be a simple and cost-effective distillation alternative: no training, no trajectory replay, and a clear mechanism for compressing teacher tool-use behavior into executable, parameterized tools. The pipeline is described in enough detail to be reproducible in principle, and the training-free property (Eq. (1)) is exact rather than approximate. The proposed abstraction-clustering-consolidation procedure is a sensible way to reduce a noisy pool of teacher-generated MCPs into a manageable tool library. However, the current experimental protocol does not establish the central generalization claim, and no code or data are released, so the ambiguity cannot be resolved from the artifact.

major comments (3)
  1. [§4.1.1, Eq. (2), Tables 2–3] The evaluation is circular as reported. The text states that 100 examples are sampled from the validation set for MCP-box generation and that the student agent is then evaluated on the same dataset. Under Eq. (2), B is selected to maximize accuracy on D, so the reported gains may simply reflect tools tailored to the exact prompts used for evaluation. This does not support the abstract's claim of 'generalizable MCP boxes' or the contribution bullet about 'unseen tasks.' Please re-run the evaluation on a held-out split disjoint from the MCP-construction examples and report the exact split sizes; without this, the comparison to OctoTools (GPT-4o) in Table 3 is also in-sample.
  2. [§3.2, Table 3] Only successful teacher trajectories are used for MCP creation, and the teacher itself achieves only 52% on PathVQA and 66% on SLAKE. This means that for a large fraction of evaluation examples no teacher-generated MCP should exist, raising the question of what exactly is being transferred on those examples. The paper should report the number of successful trajectories per dataset, the coverage of the MCP-Box over the evaluation set, and an ablation separating cases with and without teacher success; otherwise the mechanism behind the reported improvements is unclear.
  3. [§4.1.1, §4.2, abstract] The claimed cross-domain generalization is not tested. Each MCP-Box is constructed per dataset and the student is evaluated on that same dataset; no experiment applies a box built on one domain to another domain or to a genuinely unseen task. The abstract's statement that students 'generalize their capabilities across domains and solve new problems' needs direct evidence, such as a held-out task or a train/test domain split.
minor comments (5)
  1. [§2 heading] The heading 'Releated Works' contains a typo and should read 'Related Works.'
  2. [§4.1.2] The text says 'We compare four settings' but then lists five items; please correct the enumeration.
  3. [§4.2, paragraph on Game of 24] The sentence 'while the teacher asignificantly outperforming Octotools with GPT-4o' is grammatically incomplete and should be rewritten.
  4. [Table 1] The 'MCP-Box Calling Rate' measures only whether at least one tool was invoked, not whether the invocation was correct or necessary; the interpretation in the text should be softened accordingly, especially for Game of 24 where a single MCP is called in 100% of cases.
  5. [Eq. (2)] The use of an expectation over D is informal; please state explicitly that D is the empirical distribution of the sampled examples and clarify that the optimization is over a finite pool L.

Circularity Check

2 steps flagged · score 7.0 of 10

Evaluation constructs the MCP-Box from the same 100 examples used to report after-distillation accuracy, so the reported gains and 'generalization' claims are in-sample fits rather than predictions on unseen tasks.

  1. fitted input called prediction [Section 3.1, Eq. (2) and Section 4.1.1; results in Tables 2 and 3]
    "For each dataset, we sample 100 examples from validation set for MCP-box generation, same as benchmark dataset construction introduced in Octotools[52], and evaluate the student agent before distillation (without MCP box integration), after distillation (with MCP box integration), Student Agent with pre-defined tools (Octotools Framework), and the teacher agent on the same dataset."

    Eq. (2) optimizes B to maximize accuracy over D. Section 4.1.1 builds the MCP-Box from teacher trajectories on the same 100 examples used for evaluation. The 'After Distillation' column in Table 2 is therefore the in-sample value of the objective being optimized, not a held-out prediction; calling this distillation and claiming 'generalizable MCP boxes' for 'unseen tasks' presents a fitted construction output as a prediction.

  2. fitted input called prediction [Section 4.2, 'Generalizability and Usage Frequency of Distilled MCPs' (Table 1 discussion)]
    "A high MCP-box calling rate indicates that distilled MCPs are broadly applicable across diverse inputs and consistently reused by student agents. These results confirm that our framework produces reusable and transferable MCPs that generalize well without requiring any additional training."

    The calling rate is measured on the same 100 examples from which the MCPs were generated via successful teacher trajectories. A tool library tailored to those inputs will naturally be invoked on them; the high rate is evidence of in-sample coverage, not of transfer to diverse or unseen inputs. This uses the construction set as evidence for the generalization claim it is supposed to test.

full rationale

The paper's own Eq. (2) defines the MCP-Box B as the maximizer of expected accuracy over supervision set D, and Section 4.1.1 states that the 100 validation examples used for MCP-box generation are also the dataset on which the student is evaluated before and after distillation. Consequently, the after-distillation numbers in Tables 2 and 3 are values of the optimized objective on the construction set, not out-of-sample predictions. The abstract's claim of 'generalizable MCP boxes' and the contribution bullet about 'unseen tasks' are therefore not supported by the reported experiments; at best the experiments show that a repository selected to maximize accuracy on D scores well on D. The 'MCP-Box Calling Rate' argument in Section 4.2 has the same problem: high calling rates on the construction set are expected for tools distilled from successful trajectories on those very inputs. I do not find independent load-bearing self-citation here: Alita [20] and the other self-citations are background and do not supply the derivation, and no uniqueness theorem is imported. The manuscript does not release code or data, and the wording leaves a possible disjoint-split reading, but as written the reduction is explicit and the central generalization claim is not established.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

The method relies on the assumption that teacher-generated tools generalize, and the evaluation setup may conflate tool construction and test evaluation.

free parameters (1)
  • Number of validation examples used for MCP generation = 100
    Hand-chosen sample size; affects the content and coverage of the MCP-Box and is not swept or justified.
assumptions (3)
  • domain assumption Successful teacher trajectories are sufficient to generate reusable, generalizable MCPs
    Section 3.2 only keeps trajectories where the teacher solved the example; the paper assumes tools extracted from these are applicable beyond those examples.
  • domain assumption An LLM can abstract, cluster, and consolidate raw MCPs into correct, general-purpose FastMCP tools
    Section 3.3 uses Claude-Sonnet-4 for abstraction, clustering, and consolidation; no verification is provided that the consolidated tools are correct or generalizable beyond the prompts.
  • domain assumption Mounting all MCPs into the student agent at once, without retrieval, does not degrade planning
    Section 3.4 states no retrieval or reranking is used; the paper assumes a small model can select relevant tools from a large library.
invented entities (1)
  • MCP-Box
    purpose: A curated library of distilled, reusable Model Context Protocol tools mounted into student agents
    The MCP-Box is introduced by this paper; its effectiveness is only evaluated on the same benchmarks used to construct it, and no external validation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AgentDistill: Training-Free Agent Distillation with Generalizable MCP Boxes." pith.science (2026). https://pith.science/paper/4M7O643Y

@misc{pith2026250614728,
  author       = {Pith},
  title        = {Pith review of: AgentDistill: Training-Free Agent Distillation with Generalizable MCP Boxes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4M7O643Y}},
  note         = {Machine review of arXiv:2506.14728}
}
read the original abstract

While knowledge distillation has become a mature field for compressing large language models (LLMs) into smaller ones by aligning their outputs or internal representations, the distillation of LLM-based agents, which involve planning, memory, and tool use, remains relatively underexplored. Existing agent distillation methods typically replay full teacher trajectories or imitate step-by-step teacher tool usage, but they often struggle to train student agents to dynamically plan and act in novel environments. We propose AgentDistill, a novel, training-free agent distillation framework that enables efficient and scalable knowledge transfer via direct reuse of Model-Context-Protocols (MCPs), which are structured and reusable task-solving modules autonomously generated by teacher agents. The reuse of these distilled MCPs enables student agents to generalize their capabilities across domains and solve new problems with minimal supervision or human intervention. Experiments on biomedical and mathematical benchmarks demonstrate that our distilled student agents, built on small language models, can achieve performance comparable to advanced systems using large LLMs such as OctoTools (GPT-4o), highlighting the effectiveness of our framework in building scalable and cost-efficient intelligent agents.

Figures

Figures reproduced from arXiv: 2506.14728 by the authors.

Figure 1
Figure 1. Comparison between traditional LLM distillation (top) and our proposed training-free agent distillation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Performance comparison across three benchmarks. After AgentDistill, student agents with small language model backbone achieve performance comparable to agents using pre￾defined tools (e.g., OctoTools with GPT-4o), demonstrating the effectiveness of our distilla￾tion framework. ∗ These authors contributed equally to this work. † Correspondence to: leqiliu@utexas.edu, mengdiw@princeton.edu. arXiv:2506.14728v1 [cs.AI] … view at source ↗
Figure 3
Figure 3. Overview of AgentDistill, the training-free agent distillation framework via Model–Context–Protocols (MCPs). The teacher agent with large language model solves tasks by decomposing them through a Manager Agent and generating task-specific MCPs via open-source search, script generation, and virtual execution. Valid MCPs are abstracted, clustered, and consolidated into a reusable MCP-Box. At inference, the student age… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustrative example of the MCP-Box construction process. Starting from two raw MCP drafts (green and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: AgentDistill constructs a generalizable MCP from teacher-generated subtasks. Green and blue MCPs target [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. AgentBrew: Lifelong Knowledge Brewing from Strong Teachers to Weak LLM Agents

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A training-free method where a strong teacher turns a weak agent's failures into environment-validated notes that significantly boost the weak agent's test-time performance.

  2. From Trajectories to Prefixes: Reusing Teacher Trajectories via Replayed Prefixes and Online Continuation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Replaying teacher trajectory prefixes and explicitly optimizing the historical prefix tokens improves small-model agent success rates over distillation and response-only GRPO baselines in TextCraft, BabyAI, and ALFWorld.

  3. SOD: Step-wise On-policy Distillation for Small Language Model Agents

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    SOD reweights on-policy distillation strength step-by-step using divergence to stabilize tool use in small language model agents, yielding up to 20.86% gains and 26.13% on AIME 2025 for a 0.6B model.

  4. MENTOR: Reinforcement Learning via Flexible Teacher-Optimized Rewards for Tool-Use Distillation

    cs.CL 2025-10 conditional novelty 6.0 of 10

    A teacher-guided dense reward for GRPO distillation improves small-model tool-use accuracy and out-of-domain generalization on executable-tool benchmarks.

  5. Limits of Spatial Imagery Reasoning in Frontier LLM Models

    cs.CV 2026-03 unverdicted novelty 5.0 of 10

    An external 3D imagery tool does not fix frontier MLLM mental-rotation performance (≤62.5%), revealing missing visual-spatial primitives rather than a pure working-memory limit.

Reference graph

Works this paper leans on

52 extracted references · 11 canonical work pages · cited by 5 Pith papers

  1. [1]

    Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    arXiv 2015

  2. [2]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108, 2019

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108, 2019

    arXiv 1910

  3. [3]

    Patient knowledge distillation for bert model compression.arXiv preprint arXiv:1908.09355, 2019

    Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model compression.arXiv preprint arXiv:1908.09355, 2019

    arXiv 1908

  4. [4]

    Tinybert: Distilling bert for natural language understanding.arXiv preprint arXiv:1909.10351, 2019

    Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding.arXiv preprint arXiv:1909.10351, 2019

    arXiv 1909

  5. [5]

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

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

    2020

  6. [6]

    Mobilebert: a compact task-agnostic bert for resource-limited devices.arXiv preprint arXiv:2004.02984, 2020

    Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. Mobilebert: a compact task-agnostic bert for resource-limited devices.arXiv preprint arXiv:2004.02984, 2020

    arXiv 2004

  7. [7]

    Gkd: A general knowledge distillation framework for large-scale pre-trained language model.arXiv preprint arXiv:2306.06629, 2023

    Shicheng Tan, Weng Lam Tam, Yuanchun Wang, Wenwen Gong, Yang Yang, Hongyin Tang, Keqing He, Jiahao Liu, Jingang Wang, Shu Zhao, et al. Gkd: A general knowledge distillation framework for large-scale pre-trained language model.arXiv preprint arXiv:2306.06629, 2023

    arXiv 2023

  8. [8]

    Large language models are reasoning teachers.arXiv preprint arXiv:2212.10071, 2022

    Namgyu Ho, Laura Schmid, and Se-Young Yun. Large language models are reasoning teachers.arXiv preprint arXiv:2212.10071, 2022

    arXiv 2022

Show all 52 references
  1. [9]

    Distilling reasoning capabilities into smaller language models.Findings of the Association for Computational Linguistics: ACL 2023, pages 7059–7073, 2023

    Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. Distilling reasoning capabilities into smaller language models.Findings of the Association for Computational Linguistics: ACL 2023, pages 7059–7073, 2023

    2023

  2. [10]

    Symbolic chain-of- thought distillation: Small models can also" think" step-by-step.arXiv preprint arXiv:2306.14050, 2023

    Liunian Harold Li, Jack Hessel, Youngjae Yu, Xiang Ren, Kai-Wei Chang, and Yejin Choi. Symbolic chain-of- thought distillation: Small models can also" think" step-by-step.arXiv preprint arXiv:2306.14050, 2023

    2023 arXiv

  3. [11]

    Explanations from large language models make small reasoners better.arXiv preprint arXiv:2210.06726, 2022

    Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, et al. Explanations from large language models make small reasoners better.arXiv preprint arXiv:2210.06726, 2022

    arXiv

  4. [12]

    Keypoint-based progressive chain-of-thought distillation for llms.arXiv preprint arXiv:2405.16064, 2024

    Kaituo Feng, Changsheng Li, Xiaolu Zhang, Jun Zhou, Ye Yuan, and Guoren Wang. Keypoint-based progressive chain-of-thought distillation for llms.arXiv preprint arXiv:2405.16064, 2024

    2024 arXiv

  5. [13]

    Structured agent distillation for large language model.arXiv preprint arXiv:2505.13820, 2025

    Jun Liu, Zhenglun Kong, Peiyan Dong, Changdi Yang, Tianqi Li, Hao Tang, Geng Yuan, Wei Niu, Wenbin Zhang, Pu Zhao, et al. Structured agent distillation for large language model.arXiv preprint arXiv:2505.13820, 2025

    2025 arXiv

  6. [14]

    Distilling llm agent into small models with retrieval and code tools.arXiv preprint arXiv:2505.17612, 2025

    Minki Kang, Jongwon Jeong, Seanie Lee, Jaewoong Cho, and Sung Ju Hwang. Distilling llm agent into small models with retrieval and code tools.arXiv preprint arXiv:2505.17612, 2025

    2025

  7. [15]

    Magdi: Structured distillation of multi-agent interaction graphs improves reasoning in smaller language models.arXiv preprint arXiv:2402.01620, 2024

    Justin Chih-Yao Chen, Swarnadeep Saha, Elias Stengel-Eskin, and Mohit Bansal. Magdi: Structured distillation of multi-agent interaction graphs improves reasoning in smaller language models.arXiv preprint arXiv:2402.01620, 2024

    2024 arXiv

  8. [16]

    Sub-goal distillation: A method to improve small language agents.arXiv preprint arXiv:2405.02749, 2024

    Maryam Hashemzadeh, Elias Stengel-Eskin, Sarath Chandar, and Marc-Alexandre Cote. Sub-goal distillation: A method to improve small language agents.arXiv preprint arXiv:2405.02749, 2024

    2024 arXiv

  9. [17]

    Introducing the model context protocol

    Anthropic. Introducing the model context protocol. https://www.anthropic.com/news/ model-context-protocol, November 2024. Accessed on 2025-06-09

    2024

  10. [18]

    Model context protocol (mcp): Landscape, security threats, and future research directions.arXiv preprint arXiv:2503.23278, 2025

    Xinyi Hou, Yanjie Zhao, Shenao Wang, and Haoyu Wang. Model context protocol (mcp): Landscape, security threats, and future research directions.arXiv preprint arXiv:2503.23278, 2025

    2025 arXiv

  11. [19]

    Mcip: Protecting mcp safety via model contextual integrity protocol.arXiv preprint arXiv:2505.14590, 2025

    Huihao Jing, Haoran Li, Wenbin Hu, Qi Hu, Heli Xu, Tianshu Chu, Peizhao Hu, and Yangqiu Song. Mcip: Protecting mcp safety via model contextual integrity protocol.arXiv preprint arXiv:2505.14590, 2025

    2025

  12. [20]

    Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution.arXiv preprint arXiv:2505.20286, 2025

    Jiahao Qiu, Xuan Qi, Tongcheng Zhang, Xinzhe Juan, Jiacheng Guo, Yifu Lu, Yimin Wang, Zixin Yao, Qihan Ren, Xun Jiang, et al. Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution.arXiv preprint arXiv:2505.20286, 2025

    2025 arXiv

  13. [21]

    Alpaca: A strong, replicable instruction-following model.Stanford Center for Research on F oundation Models

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpaca: A strong, replicable instruction-following model.Stanford Center for Research on F oundation Models. https://crfm. stanford. edu/2023/03/13/al...

    2023

  14. [22]

    Orca: Progressive learning from complex explanation traces of gpt-4.arXiv preprint arXiv:2306.02707, 2023

    Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Awadallah. Orca: Progressive learning from complex explanation traces of gpt-4.arXiv preprint arXiv:2306.02707, 2023. 11

    2023 arXiv

  15. [23]

    Super- correct: Advancing small llm reasoning with thought template distillation and self-correction

    Ling Yang, Zhaochen Yu, Tianjun Zhang, Minkai Xu, Joseph E Gonzalez, Bin Cui, and Shuicheng Yan. Super- correct: Advancing small llm reasoning with thought template distillation and self-correction. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  16. [24]

    Learning to maximize mutual information for chain-of-thought distillation.arXiv preprint arXiv:2403.03348, 2024

    Xin Chen, Hanxian Huang, Yanjun Gao, Yi Wang, Jishen Zhao, and Ke Ding. Learning to maximize mutual information for chain-of-thought distillation.arXiv preprint arXiv:2403.03348, 2024

    2024 arXiv

  17. [25]

    Scott: Self-consistent chain-of-thought distillation.arXiv preprint arXiv:2305.01879, 2023

    Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. Scott: Self-consistent chain-of-thought distillation.arXiv preprint arXiv:2305.01879, 2023

    2023 arXiv

  18. [26]

    Skip-thinking: Chunk-wise chain-of-thought distillation enable smaller language models to reason better and faster.arXiv preprint arXiv:2505.18642, 2025

    Xiao Chen, Sihang Zhou, Ke Liang, Xiaoyu Sun, and Xinwang Liu. Skip-thinking: Chunk-wise chain-of-thought distillation enable smaller language models to reason better and faster.arXiv preprint arXiv:2505.18642, 2025

    2025 arXiv

  19. [27]

    Reasonflux: Hierarchical llm reasoning via scaling thought templates.arXiv preprint arXiv:2502.06772, 2025

    Ling Yang, Zhaochen Yu, Bin Cui, and Mengdi Wang. Reasonflux: Hierarchical llm reasoning via scaling thought templates.arXiv preprint arXiv:2502.06772, 2025

    2025 arXiv

  20. [28]

    Unicott: A unified framework for structural chain-of-thought distillation

    Xianwei Zhuang, Zhihong Zhu, Zhichang Wang, Xuxin Cheng, and Yuexian Zou. Unicott: A unified framework for structural chain-of-thought distillation. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  21. [29]

    Disco: Distilling counterfactuals with large language models.arXiv preprint arXiv:2212.10534, 2022

    Zeming Chen, Qiyue Gao, Antoine Bosselut, Ashish Sabharwal, and Kyle Richardson. Disco: Distilling counterfactuals with large language models.arXiv preprint arXiv:2212.10534, 2022

    2022 arXiv

  22. [30]

    Specializing smaller language models towards multi-step reasoning

    Yao Fu, Hao Peng, Litu Ou, Ashish Sabharwal, and Tushar Khot. Specializing smaller language models towards multi-step reasoning. InInternational Conference on Machine Learning, pages 10421–10430. PMLR, 2023

    2023

  23. [31]

    Learning by distilling context.arXiv preprint arXiv:2209.15189, 2022

    Charlie Snell, Dan Klein, and Ruiqi Zhong. Learning by distilling context.arXiv preprint arXiv:2209.15189, 2022

    2022 arXiv

  24. [32]

    Efficient llm context distillation

    Rajesh Upadhayayaya, Zachary Smith, Chritopher Kottmyer, and Manish Raj Osti. Efficient llm context distillation. arXiv preprint arXiv:2409.01930, 2024

    2024

  25. [33]

    In-context learning distillation: Transferring few-shot learning ability of pre-trained language models.arXiv preprint arXiv:2212.10670, 2022

    Yukun Huang, Yanda Chen, Zhou Yu, and Kathleen McKeown. In-context learning distillation: Transferring few-shot learning ability of pre-trained language models.arXiv preprint arXiv:2212.10670, 2022

    2022 arXiv

  26. [34]

    In-context learning distillation for efficient few-shot fine-tuning

    Yifei Duan, Liu Li, Zirui Zhai, and Jinxia Yao. In-context learning distillation for efficient few-shot fine-tuning. arXiv preprint arXiv:2412.13243, 2024

    2024 arXiv

  27. [35]

    Knowledge distilla- tion from language-oriented to emergent communication for multi-agent remote control

    Yongjun Kim, Sejin Seo, Jihong Park, Mehdi Bennis, Seong-Lyun Kim, and Junil Choi. Knowledge distilla- tion from language-oriented to emergent communication for multi-agent remote control. InICC 2024-IEEE International Conference on Communications, pages 2962–2967. IEEE, 2024

    2024

  28. [36]

    Embodied cot distillation from llm to off-the-shelf agents.arXiv preprint arXiv:2412.11499, 2024

    Wonje Choi, Woo Kyung Kim, Minjong Yoo, and Honguk Woo. Embodied cot distillation from llm to off-the-shelf agents.arXiv preprint arXiv:2412.11499, 2024

    2024 arXiv

  29. [37]

    Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation.arXiv preprint arXiv:2505.23885, 2025

    Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, et al. Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation.arXiv preprint arXiv:2505.23885, 2025

    2025 arXiv

  30. [38]

    Autoagents: A framework for automatic agent generation.arXiv preprint arXiv:2309.17288, 2023

    Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, Börje F Karlsson, Jie Fu, and Yemin Shi. Autoagents: A framework for automatic agent generation.arXiv preprint arXiv:2309.17288, 2023

    2023 arXiv

  31. [39]

    Long term memory: The foundation of ai self-evolution.arXiv preprint arXiv:2410.15665, 2024

    Xun Jiang, Feng Li, Han Zhao, Jiaying Wang, Jun Shao, Shihao Xu, Shu Zhang, Weiling Chen, Xavier Tang, Yize Chen, et al. Long term memory: The foundation of ai self-evolution.arXiv preprint arXiv:2410.15665, 2024

    2024 arXiv

  32. [40]

    Finrobot: an open-source ai agent platform for financial applications using large language models.arXiv preprint arXiv:2405.14767, 2024

    Hongyang Yang, Boyu Zhang, Neng Wang, Cheng Guo, Xiaoli Zhang, Likun Lin, Junlin Wang, Tianyu Zhou, Mao Guan, Runjia Zhang, et al. Finrobot: an open-source ai agent platform for financial applications using large language models.arXiv preprint arXiv:2405.14767, 2024

    2024 arXiv

  33. [41]

    Chemagent: Self-updating library in large language models improves chemical reasoning.arXiv preprint arXiv:2501.06590, 2025

    Xiangru Tang, Tianyu Hu, Muyang Ye, Yanjun Shao, Xunjian Yin, Siru Ouyang, Wangchunshu Zhou, Pan Lu, Zhuosheng Zhang, Yilun Zhao, et al. Chemagent: Self-updating library in large language models improves chemical reasoning.arXiv preprint arXiv:2501.06590, 2025

    2025 arXiv

  34. [42]

    Clinicalagent: Clinical trial multi-agent system with large language model-based reasoning

    Ling Yue, Sixue Xing, Jintai Chen, and Tianfan Fu. Clinicalagent: Clinical trial multi-agent system with large language model-based reasoning. InProceedings of the 15th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics, pages 1–10, 2024

    2024

  35. [43]

    Agentcoder: Multi-agent- based code generation with iterative testing and optimisation.arXiv preprint arXiv:2312.13010, 2023

    Dong Huang, Jie M Zhang, Michael Luck, Qingwen Bu, Yuhao Qing, and Heming Cui. Agentcoder: Multi-agent- based code generation with iterative testing and optimisation.arXiv preprint arXiv:2312.13010, 2023

    2023 arXiv

  36. [44]

    Atomagents: Alloy design and discovery through physics-aware multi-modal multi-agent artificial intelligence.arXiv preprint arXiv:2407.10022, 2024

    Alireza Ghafarollahi and Markus J Buehler. Atomagents: Alloy design and discovery through physics-aware multi-modal multi-agent artificial intelligence.arXiv preprint arXiv:2407.10022, 2024. 12

    2024 arXiv

  37. [45]

    Protagents: protein discovery via large language model multi-agent collaborations combining physics and machine learning.Digital Discovery, 3(7):1389–1409, 2024

    Alireza Ghafarollahi and Markus J Buehler. Protagents: protein discovery via large language model multi-agent collaborations combining physics and machine learning.Digital Discovery, 3(7):1389–1409, 2024

    2024

  38. [46]

    On path to multimodal historical reasoning: Histbench and histagent.arXiv preprint arXiv:2505.20246, 2025

    Jiahao Qiu, Fulian Xiao, Yimin Wang, Yuchen Mao, Yijia Chen, Xinzhe Juan, Siran Wang, Xuan Qi, Tongcheng Zhang, Zixin Yao, et al. On path to multimodal historical reasoning: Histbench and histagent.arXiv preprint arXiv:2505.20246, 2025

    2025 arXiv

  39. [47]

    Emoagent: Assessing and safeguarding human-ai interaction for mental health safety.arXiv preprint arXiv:2504.09689, 2025

    Jiahao Qiu, Yinghui He, Xinzhe Juan, Yimin Wang, Yuhan Liu, Zixin Yao, Yue Wu, Xun Jiang, Ling Yang, and Mengdi Wang. Emoagent: Assessing and safeguarding human-ai interaction for mental health safety.arXiv preprint arXiv:2504.09689, 2025

    2025 arXiv

  40. [48]

    ‘smo- lagents‘: a smol library to build great agentic systems

    Aymeric Roucher, Albert Villanova del Moral, Thomas Wolf, Leandro von Werra, and Erik Kaunismäki. ‘smo- lagents‘: a smol library to build great agentic systems. https://github.com/huggingface/smolagents, 2025

    2025

  41. [49]

    Math twenty four (24 -game) dataset

    Nathan Lile. Math twenty four (24 -game) dataset. https://huggingface.co/datasets/nlile/24-game, March 2025. Accessed on 2025-06-10

    2025

  42. [50]

    Pathvqa: 30000+ questions for medical visual question answering.arXiv preprint arXiv:2003.10286, 2020

    Xuehai He, Yichen Zhang, Luntian Mou, Eric Xing, and Pengtao Xie. Pathvqa: 30000+ questions for medical visual question answering.arXiv preprint arXiv:2003.10286, 2020

    2003 arXiv

  43. [51]

    Slake: A semantically-labeled knowledge- enhanced dataset for medical visual question answering

    Bo Liu, Li-Ming Zhan, Li Xu, Lin Ma, Yan Yang, and Xiao-Ming Wu. Slake: A semantically-labeled knowledge- enhanced dataset for medical visual question answering. In2021 IEEE 18th international symposium on biomedical imaging (ISBI), pages 1650–1654. IEEE, 2021

    2021

  44. [52]

    Octotools: An agentic framework with extensible tools for complex reasoning.arXiv preprint arXiv:2502.11271, 2025

    Pan Lu, Bowen Chen, Sheng Liu, Rahul Thapa, Joseph Boen, and James Zou. Octotools: An agentic framework with extensible tools for complex reasoning.arXiv preprint arXiv:2502.11271, 2025. 13

    2025 arXiv

Pith tools

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