Park, Chanwoo, Jeong, Jaywoong, Wojtowicz, Zachary, Kim, Yubin, Lee, Taekyun, Zhang, Kaiqing, Ozdaglar, Asuman · Oct 27, 2026
TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information
Markdown · JSON · Original source
The authors introduce TeamGym, an environment for evaluating and training language model-based agents on team decision-making tasks drawn from legal, finance, medical, and software engineering corpora. They define “pivotal information” as facts that are each necessary and collectively sufficient for a correct final decision, and use it to study how agents share and retrieve information. In tests of 8 language models across 12 source corpora, distributing pivotal information among participants lowered success rates by an average of 22.9 percentage points compared with giving the same information to one model. Requiring participants to find and extract relevant information from source material lowered success rates by a further 19.9 points. The authors also report that on team decision loss improved final-decision quality in two policy-training experiments, with coordination emerging despite rewards evaluating only the final decision. The results identify information sharing and retrieval as challenges in these tested settings; the abstract does not establish how well the findings transfer to real-world teams. Do the reported training gains persist on tasks beyond the two policy-training experiments?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-05T10:16:55.677Z. First observed: 2026-10-28T05:36:43.997Z. Retrieved: 2026-10-29T23:47:09.380Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported success-rate comparisons and training findings concern the evaluated models, source corpora, and experiments; the abstract does not establish how well they transfer to real-world teams. The abstract does not give details of the two policy-training experiments or report whether their gains persist on other tasks.
Open question
Do the reported training gains persist on tasks beyond the two policy-training experiments?
Source-stated
The authors introduce TeamGym as an environment for evaluating and training language model-based agents on team decision-making problems drawn from real-world corpora.
We introduce TeamGym, an environment for evaluating and training language model-based agents on team decision-making problems constructed from real-world legal, finance, medical, and software engineering corpora.
Source-stated
The authors call a set of facts “pivotal information” when each fact is necessary and the full set is sufficient for a correct final decision.
a minimal set of facts, each element of which is necessary and all of which are collectively sufficient for a correct final decision, which we call pivotal information.
Source-stated
Across the evaluated models and corpora, distributing pivotal information across participants lowered success rates by an average of 22.9 percentage points compared with giving the information to one model.
Distributing pivotal information across participants lowers success rate by 22.9 percentage points on average relative to giving the same information to a single model
Source-stated
Requiring participants to identify and extract pivotal information from source material lowered team success rates by a further 19.9 points.
A team's success rate is lowered by a further 19.9 points when participants must identify and extract pivotal information from source material rather than receiving it directly
Source-stated
The authors report improved final-decision quality and emergent coordination after reinforcement-learning training in two policy-training experiments.
Across two policy-training experiments, we find that training improves final-decision quality and induces emergent coordination in participation, speaker selection, and communication
Source and contributor context
Park, Chanwoo, Jeong, Jaywoong, Wojtowicz, Zachary, Kim, Yubin, Lee, Taekyun, Zhang, Kaiqing, Ozdaglar, Asuman · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Reinforcement learning
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d32cd275-89f6-484d-8a46-485c13d2cedb.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-05T10:16:55.677Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 30df2238a9c2bb528c5644773388796ccf8982884b9c5491bb63a446540f4b4e. Source text hash: af598f9a8bea168ac81a4119979eb93f436482fee23ad8f3f0b67d175184f5a0.
Revision and source-check history
Version 1 · 2026-11-05T10:16:55.677Z
TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information
The authors introduce TeamGym, an environment for evaluating and training language model-based agents on team decision-making tasks drawn from legal, finance, medical, and software engineering corpora. They define “pivotal information” as facts that are each necessary and collectively sufficient for a correct final decision, and use it to study how agents share and retrieve information. In tests of 8 language models across 12 source corpora, distributing pivotal information among participants lowered success rates by an average of 22.9 percentage points compared with giving the same information to one model. Requiring participants to find and extract relevant information from source material lowered success rates by a further 19.9 points. The authors also report that on team decision loss improved final-decision quality in two policy-training experiments, with coordination emerging despite rewards evaluating only the final decision. The results identify information sharing and retrieval as challenges in these tested settings; the abstract does not establish how well the findings transfer to real-world teams. Do the reported training gains persist on tasks beyond the two policy-training experiments?
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-05T10:16:55.677Z. First observed: 2026-10-28T05:36:43.997Z. Retrieved: 2026-10-29T23:47:09.380Z.
Limits
The reported success-rate comparisons and training findings concern the evaluated models, source corpora, and experiments; the abstract does not establish how well they transfer to real-world teams. The abstract does not give details of the two policy-training experiments or report whether their gains persist on other tasks.
Open question
Do the reported training gains persist on tasks beyond the two policy-training experiments?
Source-stated
The authors introduce TeamGym as an environment for evaluating and training language model-based agents on team decision-making problems drawn from real-world corpora.
We introduce TeamGym, an environment for evaluating and training language model-based agents on team decision-making problems constructed from real-world legal, finance, medical, and software engineering corpora.
Source-stated
The authors call a set of facts “pivotal information” when each fact is necessary and the full set is sufficient for a correct final decision.
a minimal set of facts, each element of which is necessary and all of which are collectively sufficient for a correct final decision, which we call pivotal information.
Source-stated
Across the evaluated models and corpora, distributing pivotal information across participants lowered success rates by an average of 22.9 percentage points compared with giving the information to one model.
Distributing pivotal information across participants lowers success rate by 22.9 percentage points on average relative to giving the same information to a single model
Source-stated
Requiring participants to identify and extract pivotal information from source material lowered team success rates by a further 19.9 points.
A team's success rate is lowered by a further 19.9 points when participants must identify and extract pivotal information from source material rather than receiving it directly
Source-stated
The authors report improved final-decision quality and emergent coordination after reinforcement-learning training in two policy-training experiments.
Across two policy-training experiments, we find that training improves final-decision quality and induces emergent coordination in participation, speaker selection, and communication
Source and contributor context
Park, Chanwoo, Jeong, Jaywoong, Wojtowicz, Zachary, Kim, Yubin, Lee, Taekyun, Zhang, Kaiqing, Ozdaglar, Asuman · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Reinforcement learning
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d32cd275-89f6-484d-8a46-485c13d2cedb.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-05T10:16:55.677Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 30df2238a9c2bb528c5644773388796ccf8982884b9c5491bb63a446540f4b4e. Source text hash: af598f9a8bea168ac81a4119979eb93f436482fee23ad8f3f0b67d175184f5a0.
uncertain · 2026-11-05T10:16:55.677Z · anthropic/claude-opus-5.5
Publication: d32cd275-89f6-484d-8a46-485c13d2cedb
attribution — pass: The reading attributes TeamGym, the pivotal-information definition, the reported gaps and the training results to the paper's authors. It does not present them as independently verified facts.
We introduce TeamGym, an environment for evaluating and training language model-based agents
summary — uncertain: The figures are faithful, but two gaps limit understanding. First, the training result depends on unexplained jargon: reinforcement learning on team decision loss, policies and policy-training. A general reader cannot tell what was trained or how. Second, the reading omits the authors' motivation that agents are trained and benchmarked in isolation. That framing explains why the findings matter. Smallest repair: briefly explain the training in plain terms, add the isolation point, and keep the authors' labels communication gap and retrieval gap. The question is also repeated inside the summary.
Yet AI agents are still primarily trained and benchmarked in isolation.
date — pass: The originalPublished date of 2026-10-27 matches the arXiv v1 submission date. It is not the later feed date of 2026-10-29.
[Submitted on 27 Oct 2026]
support — pass: Each claim matches the abstract: the 22.9-point and 19.9-point drops, the scope of 8 models and 12 corpora, and the training results from two experiments. The limitation about transfer to real-world teams is framed as unknown rather than asserted. The question follows from the evidence without assuming an answer.
even though the reward evaluates only the final decision
media — pass: The selected media is a verified link to the arXiv abstract page for this paper, which is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTraining by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv