Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

KIMAs: A Configurable Knowledge Integrated Multi-Agent System

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read KIMAs is a configurable multi-agent pipeline that enriches queries with conversation and knowledge context, routes each query to the best-suited knowledge sources using embedding-cluster centroids, and generates cited answers with…

desk verdict A coherent configurable RAG multi-agent system design with a genuinely useful look-back citation trick, but no quantitative evaluation and a real routing/context parallelization tension that the paper leaves unaddressed. read the letter →

arxiv 2502.09596 v1 pith:25YLTEQV submitted 2025-02-13 cs.AI cs.MA

Zitao Li , Fei Wei , Yuexiang Xie , Dawei Gao , Weirui Kuang , Zhijian Ma , Bingchen Qian , Yaliang Li
show 1 more author
Bolin Ding
This is my paper · ORCID
classification cs.AIcs.MA
keywords retrieval-augmentedgenerationmulti-agentsystemsqueryrewriteknowledgeroutingcitationconfigurablepipelineknowledge-intensiveQAlow-latencyRAG
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

The paper describes KIMAs, a configurable multi-agent system for building knowledge-intensive question-answering applications on top of large language models (LLMs). Its central claim is that the combination of conversation-aware query rewriting, per-agent knowledge routing by embedding-cluster centroids, reranking-based filtering, and a two-stage look-back citation mechanism lets developers handle heterogeneous knowledge sources without building heavy data pipelines. The system is designed so that each piece can be switched on or off by configuration, and the optimized pipeline runs query routing and context analysis in parallel to keep latency close to simple retrieval-augmented generation (RAG). The authors support the claim by showing how the same framework is configured for three operational applications, from a small repository Q&A bot to a high-throughput social-media bot with sub-ten-second latency.

What carries the argument

The load-bearing mechanism is embedding-cluster routing. For each retrieval agent, the embeddings of its knowledge chunks are clustered and the centroids serve as compact synopses of what that agent knows; at query time, the query embedding is compared with all centroids and only the top-K agents are activated for retrieval. The other pillars are the context manager's two rewrites (one for retrieval, one for the summarizer), the retrieval agents' pluggable rewrite strategies (prompt, retrieval, keyword, HyDE, and translation), and the rerank-then-look-back citation pipeline. The parallel execution is what reconciles these sequential-looking stages with low latency.

What would settle it

Take a set of knowledge sources with known topics, plant queries whose answers lie in exactly one source, and check whether the embedding-centroid routing places that source in the top-K activated agents. A second decisive test: use a follow-up question with pronouns whose antecedent is in the conversation history and run the optimized parallel pipeline; if routing activates the wrong source because it used the un-enriched query, the claim that context management improves retrieval in multi-turn conversations fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is a system design: KIMAs decomposes a RAG application into three agent types — a context manager that rewrites a user query with conversation context and distills history for the summarizer, retrieval agents that rewrite queries to fit their own knowledge sources, and a summarizer that reranks and filters retrieved chunks and generates the final answer. Knowledge sources are represented by embeddings of their chunks, and routing compares the query embedding to each agent's cluster centroids, with optional developer-supplied mix-in text and score scaling to impose preference. Citation generation uses a look-back strategy: the answer is streamed first, then the model is asked which retrieved chunks support it. The authors report that three production configurations run with reliable performance, including an endpoint that responds in under ten seconds by disabling the context manager and using keyword rewrite.

Load-bearing premise

The system works only if the routing step reliably picks the knowledge sources that contain the answer, and in the optimized pipeline it makes that choice from the raw user query before conversation context has been added.

Editorial extensions

If this is right

  • A developer can build an application over local vector databases, search-engine APIs, and domain HTTP APIs by editing a configuration file rather than writing a new retrieval stack.
  • Because routing activates only the top-K agents, cost and latency scale with the number of relevant sources rather than the total number of configured sources.
  • Disabling the context manager is a usable latency lever; the Olympic use case reports under ten seconds end-to-end with only keyword rewrite and direct summarization.
  • The look-back citation generation lets the answer stream to the user immediately, with references appearing after a short pause, avoiding the latency and formatting risks of one-shot structured-output citation.
  • The same agent types and pipeline can be reconfigured for small, large, and turbo-scale applications, which is the paper's evidence of generality.

Reading between the lines

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

  • The routing mechanism is not limited to knowledge-source selection: the same centroid-comparison idea could route queries to tools, functions, or specialized agents whose descriptions are embedded, making it a general component-routing primitive.
  • The paper does not report routing accuracy; an implicit testable claim is that centroid similarity survives the heterogeneity of chunks (code versus prose) after score scaling. A direct measurement of recall@K of correct sources would settle this.
  • The score-scaling knob is hand-tuned per source; a natural extension is to learn the scaling factors from implicit user feedback (for example, which cited references users click) rather than setting them manually.
  • The parallelization structure implies that end-to-end latency is governed by the slowest parallel branch, typically retrieval or reranking, so further latency gains would come from accelerating reranking rather than from the LLM calls.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents KIMAs, a configurable multi-agent framework for knowledge-intensive question answering. The system is organized around three agent types: a context manager that rewrites conversational queries and digests conversation history, retrieval agents that access heterogeneous knowledge sources (local vector databases, online search engines, domain-specific HTTP APIs), and a summarizer that generates final answers with references. Routing is performed by embedding-based clustering over each agent's knowledge chunks, with optional developer-supplied mix-in text and score scaling. The pipeline is parallelized at the query-ingest and retrieval stages. The paper reports three deployment use cases: an AgentScope Q&A chatbot, a ModelScope Q&A chatbot, and an Olympic-commentary bot on Weibo. No quantitative evaluation is provided; the use cases are described through screenshots and qualitative configuration summaries.

Significance. If validated, KIMAs would be a practically useful open-source framework for building RAG-based conversational applications over heterogeneous knowledge sources. The paper's design choices are reasonable and clearly motivated: the separation of conversation-context and knowledge-context rewriting, the centroid-based routing with manual intervention, and the two-stage look-back citation generation are all sensible engineering contributions. The paper also gives concrete configuration recipes for three deployed systems, which is valuable for practitioners. However, the central claims about improved retrieval accuracy, multi-turn coherency, efficient routing, and low latency are currently unsupported by any quantitative measurement. There are no baselines, ablations, benchmarks, or error analyses, and the primary deployments are the authors' own products, so the effectiveness claims rest on self-referential demonstration. The architectural ideas are promising, but the evidence needed to support the paper's stated contributions is not yet present.

major comments (4)
  1. [Section 4 and Section 5] The paper makes quantitative claims in the abstract (improved retrieval accuracy, multi-turn conversational coherency, efficient routing, low latency), but Section 4 provides only screenshots and qualitative descriptions, and Section 5 concludes on the basis of these demonstrations. There are no retrieval accuracy numbers, routing precision/recall figures, end-to-end answer quality metrics, latency distributions, or comparisons against a single-agent RAG baseline or an existing framework such as LlamaIndex or LangChain. The only quantitative statement is the claim in Section 4.3 that end-to-end latency is reduced to less than 10 seconds per post, but no measurement methodology or data is given. Without this evidence, the core claims of the paper are not testable. I would request a systematic evaluation with at least retrieval recall, routing accuracy, answer faithfulness or correctness, and latency, on both the deployed use cases and a public benchmark.
  2. [Section 3.5, Parallelization 1] The optimized pipeline runs query routing on the raw user query in parallel with the context manager's conversation-context rewrite. As the motivating example in Section 3.1 shows, a follow-up such as "Where can I find the code for it?" contains an unresolved pronoun that only the conversation history can resolve. Under the parallel execution described in Section 3.5, the embedding used for centroid routing is computed from the unresolved fragment, so routing may activate the wrong retrieval agents. The later context-aware rewrite is then applied only to the wrong sources, and retrieval cannot recover the omitted information. The paper does not report any routing accuracy measurements for multi-turn queries, so there is no evidence that this ordering is safe. This design tension directly affects the central claim of combining multi-turn coherency with efficient routing; I would expect either a sequential routing-after-rewrite option, or an evaluation showing that routing on raw queries is sufficient in conversational settings.
  3. [Section 3.3.2] The routing mechanism has several free parameters that are not analyzed: the weight of the manual mix-in relative to local knowledge embeddings, the score scaling factor, and the number of activated retrieval agents (top-K). These parameters are described as developer preferences, but no guidance or sensitivity analysis is given, and no measurement shows how routing quality or end-to-end answer quality changes with them. Since the paper claims "efficient knowledge routing and retrieval" as a key property, the evaluation should include routing accuracy against a labeled set of queries, the overhead of the centroid-based routing compared with LLM-based routing, and the effect of the mix-in and scaling parameters on routing decisions.
  4. [Section 3.4] The summarization section claims that the reranking model is "more reliable and general" than reciprocal rank fusion, and that the look-back citation strategy is robust and low-latency, but neither claim is measured. In particular, there is no evaluation of citation accuracy, no comparison of the look-back approach against the one-step approach, and no measurement of the latency added by the second-stage citation generation. The paper also notes that "we tested several approaches" but reports only the one-step approach and the final look-back design, without data from those tests. These mechanisms are load-bearing for the claims of "simple but effective filter and reference generation," so they need direct experimental support.
minor comments (5)
  1. [References] Reference [13] is titled "Kimi.ai" but the URL points to perplexity.ai; the reference should be corrected to the actual source.
  2. [References] References [29] and [30] are duplicate entries for the same Chain-of-Thought paper; one should be removed and the citations merged.
  3. [Section 4.2] In the "Knowledge sources configuration" paragraph, the sentence "these knowledge sources can be hosted and retrieved locally because the retrieval standard is more" is incomplete and should be finished.
  4. [Section 4.2] In the "Pipeline configuration" paragraph, "routine mechanism" should be "routing mechanism."
  5. [Figure 3 caption] The caption says "Agent A is roused to conduct knowledge retrieval"; "roused" should be "routed" or "activated."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KIMAs is a system/technical report whose components are described transparently; no prediction or derivation reduces to its inputs by construction.

full rationale

KIMAs does not present a mathematical derivation or quantitative prediction that could be forced by construction. The routing mechanism is explicitly adopted from the authors' prior work [32] ('We adopt an algorithm similar to [32]'), and the system is built on AgentScope [5], but these are disclosed imports rather than hidden premises; the paper's contributions are the configurable integration, query-rewrite strategies, filtering, citation look-back, and parallelization, which are described algorithmically. The manual mix-in and score scaling in Section 3.3.2 are presented as developer configuration knobs, not as parameters fitted to a dataset and then re-reported as predictions. The use cases in Section 4 are illustrative deployments with no benchmark numbers (the only metric is '<10 seconds' latency for the Olympic bot), so the claim of 'reliable performance' is unquantified, but that is an evaluation gap rather than circularity. No equation, definition, or uniqueness theorem is used to equate an output to an input, and no self-citation is invoked to forbid alternatives. The absence of external benchmarks raises correctness/validation risk but does not make the derivation circular.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new theoretical entities are introduced. The configuration knobs (mix-in weight, score scaling, top-K) are user-set parameters, not fitted constants. The paper's claims rest on unverified assumptions about LLM behavior, embedding routing, and reranking.

free parameters (3)
  • Routing mix-in weight = not specified
    Section 3.3.2 lets developers set a weight balancing similarity between the query and local knowledge centroids versus manually provided description embeddings; no default or tuning protocol is reported.
  • Score scaling factor = not specified
    Section 3.3.2 allows scaling similarity scores for specific knowledge types to compensate for different score ranges; no values or calibration procedure are given.
  • Top-K activated retrieval agents = not specified
    Routing activates agents whose centroid similarity is in the top-K; K is a configurable hyperparameter with no reported value.
assumptions (5)
  • domain assumption LLMs prompted with conversation history and task instructions produce query rewrites and context analyses that are accurate enough to improve retrieval.
    Sections 3.2 and 3.5 rely on the context manager's LLM-generated analysis and indices_of_related_messages without evaluating their correctness.
  • domain assumption Embedding-space similarity is a reliable signal for routing queries to the correct knowledge sources.
    Section 3.3.2 builds routing on clustering query and chunk embeddings with centroid matching; no routing accuracy measurements are provided.
  • domain assumption The routing algorithm of reference [32] works as stated.
    KIMAs adopts a similar routing method from reference [32], which overlaps in authorship with this paper, and does not test the algorithm independently.
  • domain assumption A reranking model can order and filter heterogeneous retrieved chunks from multiple sources more reliably than raw similarity scores.
    Section 3.4 rejects reciprocal rank fusion in favor of a reranking model but gives no comparison or model specification.
  • domain assumption AgentScope's message passing and agent lifecycle correctly support the parallelized pipeline.
    Section 3.5 assumes parallel execution of context rewriting, routing, retrieval, and context analysis is safe and correct, without latency or correctness measurements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KIMAs: A Configurable Knowledge Integrated Multi-Agent System." pith.science (2026). https://pith.science/paper/25YLTEQV

@misc{pith2026250209596,
  author       = {Pith},
  title        = {Pith review of: KIMAs: A Configurable Knowledge Integrated Multi-Agent System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/25YLTEQV}},
  note         = {Machine review of arXiv:2502.09596}
}
read the original abstract

Knowledge-intensive conversations supported by large language models (LLMs) have become one of the most popular and helpful applications that can assist people in different aspects. Many current knowledge-intensive applications are centered on retrieval-augmented generation (RAG) techniques. While many open-source RAG frameworks facilitate the development of RAG-based applications, they often fall short in handling practical scenarios complicated by heterogeneous data in topics and formats, conversational context management, and the requirement of low-latency response times. This technical report presents a configurable knowledge integrated multi-agent system, KIMAs, to address these challenges. KIMAs features a flexible and configurable system for integrating diverse knowledge sources with 1) context management and query rewrite mechanisms to improve retrieval accuracy and multi-turn conversational coherency, 2) efficient knowledge routing and retrieval, 3) simple but effective filter and reference generation mechanisms, and 4) optimized parallelizable multi-agent pipeline execution. Our work provides a scalable framework for advancing the deployment of LLMs in real-world settings. To show how KIMAs can help developers build knowledge-intensive applications with different scales and emphases, we demonstrate how we configure the system to three applications already running in practice with reliable performance.

Figures

Figures reproduced from arXiv: 2502.09596 by the authors.

Figure 1
Figure 1. KIMAs system with agentive modularization and configurable pipeline. repository. An initial response can be “Yes. For example, there is a simulation for the game Werewolf that...” After that, the user may ask a follow-up question, “Where can I find the code for it?” In this case, the context manager must rely on context to determine that “it” refers to the Werewolf game in the AgentScope repository and enrich the us… view at source ↗
Figure 2
Figure 2. Query rewrite mechanisms in KIMAs. • Whether routing is required (with human intervention). We also prove some flexibility for the routing mechanism. Besides the built-in mechanism routing, we allow developers to cast their human preferences for the routing by providing additional manually written mix-in text or scaling the importance of some weights. More details are deferred to Section 3.3. With these key agents a… view at source ↗
Figure 3
Figure 3. A simple visualization of the routing mechanism. Because the query embedding is closer to centroids [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Use case in Q&A group of AgentScope Goals. In this use case, we adapt KIMAs to help answer questions about AgentScope’s GitHub repository. The expectation for this application is to serve as a chatbot in a Q&A group for developers building their multi-agent application…
Figure 5
Figure 5. Figure 5: Three demonstration QA pairs using different knowledge resources in Modelscope QA. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications

    cs.AI 2025-08 unverdicted novelty 4.0 of 10

    AgentScope 1.0 packages the components needed to build, evaluate, and deploy LLM agent applications into one developer framework.

Reference graph

Works this paper leans on

35 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Improving language models by retrieving from trillions of tokens

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. InInternational conference on machine learning, pages 2206–2240. PMLR, 2022

    2022

  2. [2]

    Chateval: Towards better llm-based evaluators through multi-agent debate.arXiv preprint arXiv:2308.07201, 2023

    Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. Chateval: Towards better llm-based evaluators through multi-agent debate.arXiv preprint arXiv:2308.07201, 2023

    arXiv 2023

  3. [3]

    Reciprocal rank fusion outperforms condorcet and individual rank learning methods

    Gordon V Cormack, Charles LA Clarke, and Stefan Buettcher. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. InProceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 758–759, 2009

    2009

  4. [4]

    Elasticsearch.software], version, 6(1), 2018

    BV Elasticsearch. Elasticsearch.software], version, 6(1), 2018

    work page 2018

  5. [5]

    Agentscope: A flexible yet robust multi-agent platform.arXiv preprint arXiv:2402.14034, 2024

    Dawei Gao, Zitao Li, Xuchen Pan, Weirui Kuang, Zhijian Ma, Bingchen Qian, Fei Wei, Wenhao Zhang, Yuexiang Xie, Daoyuan Chen, et al. Agentscope: A flexible yet robust multi-agent platform.arXiv preprint arXiv:2402.14034, 2024

    arXiv 2024

  6. [6]

    Precise zero-shot dense retrieval without relevance labels

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762–1777, 2023

    work page 2023

  7. [7]

    Enabling large language models to generate text with citations

    Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. Enabling large language models to generate text with citations. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6465–6488, 2023

    2023

  8. [8]

    Memory matters: The need to improve long-term memory in llm-agents

    Kostas Hatalis, Despina Christou, Joshua Myers, Steven Jones, Keith Lambert, Adam Amos-Binks, Zohreh Dannenhauer, and Dustin Dannenhauer. Memory matters: The need to improve long-term memory in llm-agents. InProceedings of the AAAI Symposium Series, volume 2, pages 277–280, 2023

    work page 2023

Show all 35 references
  1. [9]

    Metagpt: Meta programming for a multi-agent collaborative framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, et al. Metagpt: Meta programming for a multi-agent collaborative framework. InThe Twelfth International Conference on Learning Representations

  2. [10]

    Leveraging passage retrieval with generative models for open domain question answering.arXiv preprint arXiv:2007.01282, 2020

    Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering.arXiv preprint arXiv:2007.01282, 2020

    2007 arXiv

  3. [11]

    Swe-bench: Can language models resolve real-world github issues? In The Twelfth International Conference on Learning Representations

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R Narasimhan. Swe-bench: Can language models resolve real-world github issues? In The Twelfth International Conference on Learning Representations

  4. [12]

    Dense passage retrieval for open-domain question answering.arXiv preprint arXiv:2004.04906, 2020

    Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering.arXiv preprint arXiv:2004.04906, 2020

    2004 arXiv

  5. [13]

    Kimi.ai. Kimi.ai. https://www.perplexity.ai/, 2023. Accessed: 2025-01-09

    2023

  6. [14]

    Langchain

    LangChain. Langchain. https://www.langchain.com/, 2023. Accessed: 2025-01-09

    2023

  7. [15]

    Camel: Commu- nicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems, 36:51991–52008, 2023

    Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Commu- nicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems, 36:51991–52008, 2023

    2023

  8. [16]

    Arxiv copilot: A self-evolving and efficient LLM system for personalized academic assistance

    Guanyu Lin, Tao Feng, Pengrui Han, Ge Liu, and Jiaxuan You. Arxiv copilot: A self-evolving and efficient LLM system for personalized academic assistance. In Delia Irazu Hernandez Farias, Tom Hope, and Manling Li, editors,Proceedings of the 2024 Conference on Empirical Methods ...

    2024

  9. [17]

    Llm+ p: Empowering large language models with optimal planning proficiency.arXiv preprint arXiv:2304.11477, 2023

    Bo Liu, Yuqian Jiang, Xiaohan Zhang, Qiang Liu, Shiqi Zhang, Joydeep Biswas, and Peter Stone. Llm+ p: Empowering large language models with optimal planning proficiency.arXiv preprint arXiv:2304.11477, 2023

    2023 arXiv

  10. [18]

    LlamaIndex: Build ai knowledge assistants over your enterprise data

    LlamaIndex. LlamaIndex: Build ai knowledge assistants over your enterprise data. https://www. llamaindex.ai/, 2023. Accessed: 2025-01-09

    2023

  11. [19]

    Chameleon: Plug-and-play compositional reasoning with large language models.Advances in Neural Information Processing Systems, 36, 2024

    Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models.Advances in Neural Information Processing Systems, 36, 2024

    2024

  12. [20]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinkingtheroleofdemonstrations: Whatmakesin-contextlearningwork? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048–...

    2022

  13. [21]

    Talm: Tool augmented language models.arXiv preprint arXiv:2205.12255, 2022

    Aaron Parisi, Yao Zhao, and Noah Fiedel. Talm: Tool augmented language models.arXiv preprint arXiv:2205.12255, 2022

    2022 arXiv

  14. [22]

    Perplexity

    Perplexity. Perplexity. https://www.perplexity.ai/, 2023. Accessed: 2025-01-09

    2023

  15. [23]

    In-context retrieval-augmented language models.Transactions of the Association for Computational Linguistics, 11:1316–1331, 2023

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models.Transactions of the Association for Computational Linguistics, 11:1316–1331, 2023

    2023

  16. [24]

    Toolformer: Language models can teach themselves to use tools.Advances in Neural Information Processing Systems, 36:68539–68551, 2023

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools.Advances in Neural Information Processing Systems, 36:68539–68551, 2023

    2023

  17. [25]

    Bing search.https://www.bing.com/

    Bing Search. Bing search.https://www.bing.com/. Accessed: 2025-01-09

    2025

  18. [26]

    Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36, 2024

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36, 2024

    2024

  19. [27]

    Medagents: Large language models as collaborators for zero-shot medical reasoning

    Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, and Mark Gerstein. Medagents: Large language models as collaborators for zero-shot medical reasoning. arXiv preprint arXiv:2311.10537, 2023

    2023 arXiv

  20. [28]

    Emergent abilities of large language models.arXiv preprint arXiv:2206.07682, 2022

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.arXiv preprint arXiv:2206.07682, 2022

    2022 arXiv

  21. [30]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  22. [31]

    LLM powered autonomous agents

    Lilian Weng. LLM powered autonomous agents. https://lilianweng.github.io/posts/ 2023-06-23-agent/, 2023. Accessed: 2025-01-09

    2023

  23. [32]

    Talk to right specialists: Routing and planning in multi-agent system for question answering, 2025

    Feijie Wu, Zitao Li, Fei Wei, Yaliang Li, Bolin Ding, and Jing Gao. Talk to right specialists: Routing and planning in multi-agent system for question answering, 2025

    2025

  24. [33]

    Autogen: Enabling next-gen llm applications via multi-agent conversation framework

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 15

    2023 arXiv

  25. [34]

    Swe-agent: Agent-computer interfaces enable automated software engineering.arXiv preprint arXiv:2405.15793, 2024

    John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering.arXiv preprint arXiv:2405.15793, 2024

    2024 arXiv

  26. [35]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in Neural Information Processing Systems, 36, 2024

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in Neural Information Processing Systems, 36, 2024

    2024

  27. [36]

    React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022. 16

    2022 arXiv

Pith tools

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