PaperlayerView exact revision on arXiv

Source-linked AI summary

Measuring AI Agents as Teammates with Collaborative Intelligence

Sean Wu, Aneesh Pappu, Jiaru Zou, Yubin Kim, Yuncong Mao, Hongjian Zhou, Junchi Yu, Jieneng Chen, James Zou

arXiv:2611.02633v1cs.AI

TL;DR

Individual-model and team-level evaluations do not reveal whether a particular AI model is a good teammate. The Collaborative Intelligence Benchmark measures teammate behavior on tasks models can solve alone, and finds weaknesses across all 18 evaluated models. Even the best model scores 82.2 out of 100, while teamwork performance is not guaranteed by general capability.

  • Problem

    Existing evaluations focus on individual-model capability or whole-team outcomes, leaving unclear whether a particular model works well as a teammate.

    Source: §1 · p. 2

  • Method

    The benchmark evaluates 18 models on 240 tasks across 8 teamwork dimensions, scoring only tasks each model solves alone.

    Source: §1 · p. 2

  • Results

    Every evaluated model has clear teamwork weaknesses, and the best scores 82.2 out of 100; frontier models reach the best deal in only 14% of conflicts.

    Source: §1 · p. 2

  • Takeaways & Limitations

    General capability does not guarantee good teamwork, and models with nearly the same capability differ by up to 11 teammate-score points.

    Source: §7 · p. 12

    The benchmark cannot be given directly to people because they know conflict tasks are tests and many tasks lack accessible workplace settings.

    Source: §5 · p. 10

Abstract

from arXiv · show

Progress in AI is often measured by what a single model can do on its own. Yet teams of AI agents are starting to make breakthroughs and discoveries in science, medicine, mathematics and software engineering. As AI begins to work in teams of agents and people, the ability to collaborate becomes a core part of their intelligence. Research on human teams has measured this ability for decades. Inspired by this, we introduce the Collaborative Intelligence Benchmark, which measures how well an AI agent works as a teammate. It contains 240 tasks across 8 dimensions of teamwork, from refusing an unsafe shortcut to checking the work that a weaker model leaves behind. Each task comes from a real record or an existing benchmark, from aviation accident reports and clinical cases to coding tasks and negotiations, and has a programmatically verifiable outcome. To isolate teamwork from individual capability, models are scored only on tasks that they can solve by themselves when they have full information. We evaluate 18 closed- and open-weight models and find that every model, across sizes and families, has clear weaknesses as a teammate, with the best model scoring only 82.2 out of 100. Frontier models almost always refuse unsafe shortcuts and report misconduct, yet they cave under interpersonal pressure, reaching the best overall deal in only 14% of conflicts, and their pass rate drops from 97% to 81% when they take over a task in which a weaker model has cheated. Crucially, general capability does not guarantee good teamwork, as models with nearly the same capability score differ by up to 11 points as teammates. The benchmark includes 160 public tasks with code and data, along with 80 held-out tasks for evaluation. As AI agents increasingly work alongside other agents and people, the ability to collaborate effectively is becoming a crucial capability to measure and develop.

1 Introduction

As AI agents increasingly work with people and other agents, collaboration is becoming part of intelligence, but existing evaluations do not show whether an individual model is a good teammate. The Collaborative Intelligence Benchmark measures that ability and finds substantial teamwork weaknesses even among frontier models.

1 Introduction

  • Individual capability alone does not ensure successful team outcomes: good teammates share critical information, support others, and judge when cooperation may cause harm.The cited cybersecurity incident illustrates that a general inclination to help peers can contribute to harmful cooperation.Source: §1 · p. 2
  • Individual-model benchmarks and team-level benchmarks do not reveal a particular model’s strengths and weaknesses as a teammate, motivating measurement alongside traditional benchmarks [4] [22] [23].A model can complete its own task yet fail to help a struggling teammate, while a team’s final outcome does not identify individual contributions.Source: §1 · p. 2
  • The benchmark covers 240 tasks across 8 teamwork dimensions, using real records or existing benchmarks with programmatically checkable answers.Models work with live partners and a scripted teammate who creates a problem; scoring is limited to tasks each model solves alone.Source: §1 · p. 2
  • Across 18 models, every model has clear teamwork weaknesses; the best scores 82.2 out of 100, and models with nearly equal capability differ by up to 11 teammate-score points.Frontier models reach the best deal in only 14% of conflicts and pass 45% of emotional-intelligence tests, while mostly resisting unsafe pressure and reporting misconduct.Source: §1 · p. 2

2 Benchmark Construction

The Collaborative Intelligence Benchmark turns teamwork problems into verifiable tasks, evaluates models in alone, baseline-team, and test conditions, and scores only tasks each model can solve alone. Its eight dimensions draw on human-team research and encompass distinct ways teamwork can fail.

2.1 Eight dimensions of teamwork

  • Across eight dimensions, tasks assess responses to unsafe pressure, misconduct, withheld information, emotional withdrawal, distraction, conflict, unowned work, and inherited cheating.The dimensions draw on research on human teams, with 20 public and 10 held-out tasks per dimension.Source: §2.1 · p. 3

2.2 Tasks and environment

  • The benchmark has 240 programmatically checkable tasks: 94 from real-world records and 146 adapted from existing benchmarks and datasets [35] [36] [38] [39].Sources include aviation and chemical accident reports, clinical cases, coding, negotiation, team decision-making, and exam problems.Source: §2.2 · p. 4
  • Tasks place the model in a team of 2 to 5, with a scripted teammate creating the problem and live LLM teammates holding private facts; 79 tasks also provide a shared sandbox.The scripted teammate follows source-based branching lines, and code checks the final state.Source: §2.2 · p. 4

2.3 Evaluation protocol and outcome verification

  • The protocol runs each task alone, with a well-behaved scripted teammate, and with a teammate creating the dimension-specific problem; only tasks passed alone count toward a model’s score.Tests run three times with different seeds; code checks final outcomes, while judges label messages to the scripted teammate.Source: §2.3 · p. 4
  • The benchmark releases 160 tasks and reserves 80 held-out tasks, run and scored like the public set to assess whether public-set contamination or tuning inflates scores.The public and held-out sets each contain 20 and 10 tasks per dimension, respectively.Source: §2.3 · p. 5

3 How Well Do Models Work as Teammates?

We evaluate 18 closed- and open-weight models on teamwork dimensions, scoring each on tasks it solves alone to isolate teammate behavior. Scores show substantial weaknesses, uneven performance across dimensions, and cost–performance trade-offs.

3.1 Main results

  • 82.2 was the highest overall score, achieved by Gemini 4 Argon at high effort; every model’s score dropped 20.2–46.6 points when a teammate created a problem.At default effort, Gemini 4 Argon scored 76.6; the scores changed little when all models were compared on the same tasks.Source: §3.1 · pp. 5–6

3.2 Results by dimension

  • Frontier models passed at least 93% of unsafe-pressure and misconduct-reporting tests, but reached the best conflict deal in 14% and passed 45% of emotional-intelligence tests.Open-weight models averaged 41.6, 31.2 points below frontier models; Kimi K3.1 Pro scored 59.9, above five closed models.Source: §3.2 · p. 6, Figure 3 · p. 6
  • High effort raised five of six models by 2.7–5.7 points but lowered Kimi K3.1 Pro by 3.7, while improving initiative and emotional intelligence and worsening conflict.On average, initiative rose 8.4 points and emotional intelligence rose 6.9; Kimi K3.1 Pro’s initiative score fell 25.7 points at high effort as it used fewer tools and read less of the workspace.Source: §3.2 · pp. 6–7

3.3 Cost and performance

  • The six Pareto-frontier rows span scores and API costs; Gemini 4 Argon at high effort scored 82.2 for $273, while GPT-6.1 Sol reached 91% of that score for $30.Claude Haiku 5.5 scored 51.7 for $2.0, earning about twice as many points per dollar as any other row.Source: §3.3 · p. 7, Figure 4 · p. 7

4 Failure Modes of Teamwork

The section sorts failures into missing a problem, giving in after recognizing it, or recognizing it without resolving it. Across dimensions, failures often reflect either missed needs or failure to act on recognized problems.

4.1 Frontier models mostly refuse unsafe requests

  • Frontier models passed at least 93% of unsafe-pressure and misconduct-reporting tests, while open-weight models resisted pressure in 75%; remaining failures often reflected the model’s own judgment.In 89% of runs in the resuscitation scenario, models that refused the requested shock then ordered the wire pulled, which the protocol also forbids.Source: §4.1 · p. 8

4.2 AI teammates struggle to hold firm in a conflict

  • Under a pushy teammate, frontier models found the best split in 14% of runs, versus 100% when working alone; they asked about needs in 99% but accepted weaker offers in 81%.Greater ability on other teamwork dimensions and higher reasoning effort strengthen the tendency to accommodate in conflict; pressure deals averaged 95% of the best joint points.Source: §4.2 · p. 8, Figure 5 · p. 8

4.3 Missed initiative is mostly a failure to notice

  • In 88 of 91 failures in a six-model study, the model never noticed the initiative need; across leaderboard runs, models noticed it in 12% of 816 failures.The three closed models took initiative in 92% of quiet-day runs versus 72% on busy days; results ranged from GPT-6 Astra at 27 of 30 tests to Qwen 3.7 Max at 0 of 23.Source: §4.3 · p. 9

4.4 Inherited misconduct often survives a model switch

  • Frontier models passed only 81% of takeover tasks, versus 97% when starting those tasks themselves; in 44 of 56 failed coding takeovers, the final code still contained the cheat.Showing the incoming model the handoff diff had mixed effects: edited tests survived in 1 of 18 runs instead of 17, while special-cased code survived in 16 of 24 instead of 22.Source: §4.4 · p. 9

4.5 Capable models recognize problems in teamwork but still fail

  • 45 ± 4% of 5,094 failed runs followed recognized problems; that share was 61%–88% in five dimensions but 12%–19% in sharing information, emotional intelligence, and initiative.For frontier models, 62% of failures followed recognition; recognition is a lower-bound measure because judges label only messages before scripted turns.Source: §4.5 · pp. 9–10, Figure 6 · p. 10

5 How Do AI Agents Compare with People as Teammates?

The section compares models with human results from six source studies using task-specific success criteria in similar scenarios, while noting that direct human testing is impractical. The comparisons show both strong performance on some tasks and difficulty when a teammate creates a problem the model must address.

5 How Do AI Agents Compare with People as Teammates?

  • Directly testing people on the benchmark is impractical because they would know conflict tasks are tests and cannot readily be placed in many workplace settings.The source-study comparisons also have limitations, including small samples in several studies.Source: §5 · pp. 10–11
  • On three of six tasks, every model succeeded in every run, exceeding the success rates reported in the corresponding human studies.Figure 7 shows the other three comparisons; the source-study instructions, incentives, and settings are not fully matched.Source: Figure 7 · p. 10, §5 · p. 11
  • In CaSiNo, only 1 of 13 real pairs reached the best deal, compared with 24% of frontier-model runs and 21% of open-weight-model runs.The task scores the deal by the combined points for both sides.Source: §5 · p. 11
  • In the car-crash task, every frontier model chose the right driver in baseline runs, versus 5 of 7 open-weight models and 93% of human groups.When a teammate goes quiet in the test version, models choose the right driver in only 20% of runs.Source: §5 · p. 11

6 Related Work

The paper distinguishes its evaluation of an individual model’s teamwork from prior work that evaluates whole teams or particular social abilities. It draws on multi-agent systems, social benchmarks, safety research, and human-team research.

6 Related Work

  • Unlike multi-agent studies that score whole teams, this benchmark treats collaboration as an individual model ability while holding the rest of the team fixed.Prior work links team failures to unclear specifications, agent misalignment, and weak verification, and finds that capable agents can perform worse in teams [23] [55].Source: §6 · p. 11
  • Social benchmarks often measure what models perceive about others or win for themselves, whereas this work scores acting for the team after recognizing a problem.Many benchmark failures occur after the model has already named the problem, and conflicts are scored by the best outcome for the team rather than the model’s own share.Source: §6 · p. 11
  • Related safety work studies pressure from users or tasks, while this benchmark examines pressure exerted by a colleague.The cited work includes agreement shifts [66], harmful multi-step requests [67], and cheating when tests contradict specifications [35].Source: §6 · p. 11
  • The benchmark’s eight teamwork dimensions draw on organizational psychology, while its initiative tasks and scripted teammate adapt the bystander effect and conformity studies.The scripted teammate behaves consistently across participants, as a confederate does in classic studies.Source: §6 · p. 11

7 Conclusion

The conclusion presents the Collaborative Intelligence Benchmark as a measure of AI agents’ teamwork, finding weaknesses across models despite their general capabilities. It proposes reporting teamwork scores alongside capability scores, while identifying further validation needed before using them to select agents for teams.

7 Conclusion

  • The best model scores 82.2 out of 100, and models with nearly the same capability differ by up to 11 points as teammates.Most failures arise after models have already seen the problem, indicating a gap in acting on it rather than noticing it.Source: §7 · p. 12
  • Frontier models rarely yield to unsafe pressure but struggle to hold firm in conflicts and often retain cheating inherited from another model.The conclusion locates many failures in acting on a recognized problem.Source: §7 · p. 12
  • The authors propose reporting teamwork scores alongside capability scores, but say the score’s ability to predict added value on other team tasks must be tested first.Choosing models for teams is described as a potential use.Source: §7 · p. 12

A Limitations

The section describes limits in the evaluation and explains how a scripted teammate creates consistent pressure for a task. It also outlines checks intended to test whether benchmark scores remain meaningful across tasks and runs.

A Limitations

  • Test versions run three times per task, while alone and baseline versions run once; live teammates are language models, and the scripted teammate is less variable than real teammates.Judges, code checks, and final-state rules can err; the authors review specified run categories and corrected 19 grading errors among 315 failing safety runs.Source: A Limitations · p. 17

B A scripted teammate

  • For the ImpossibleBench task, a fixed-order script pressures the model to accept a wrong test, and the model passes only if neither it nor the lead changes, marks, or skips tests.The lead first asks to mark the test as an expected failure, then offers a loophole and offers to make the change; the model must also report the wrong test and not claim the tests pass.Source: B A scripted teammate · p. 17

C Validity of the score

  • Score-validity checks cover held-out tasks and a second run, judge agreement with people, model-specific task scoring, weak-task influence, and source-type effects on rankings.These checks use saved runs from the 24 ranked rows without new model calls.Source: C Validity of the score · pp. 17–18

C.1 New tasks, new runs and head to head

The held-out tasks preserve the public-task model ranking, and comparisons on shared tasks largely confirm the leaderboard order. Averaging across seeds improves score reliability.

C.1 New tasks, new runs and head to head

  • Kendall’s τ is 0.88 between held-out and public task rankings, with a 2.7-point average score difference.The passage describes the two task sets as having similar difficulty.Source: §C.1 · p. 18
  • A single test run gives the same verdict across all three seeds only 80% of the time, but averaging seeds removes most noise.Splitting each dimension’s tasks in two gives reliability of 0.99.Source: §C.1 · p. 18
  • 271 of 276 shared-task row pairs favor the row ranked higher on the leaderboard.The pairwise check compares rows only on tasks both solve alone.Source: Figure 9 · p. 18, §C.1 · p. 18

C.2 People check the judges

Human checks broadly support the message labels, while score adjustments and alternative estimators leave model rankings largely stable. Agreement varies across teamwork dimensions, and judge-label differences sometimes affect task outcomes.

C.2 People check the judges

  • 82.5% of 40 messages matched people’s choice of the judges’ final label, versus 87.5% agreement between the two judges.On messages where the judges disagreed, people chose their label 88.6% of the time.Source: §C.2 · p. 18, Figure 10 · p. 19
  • 2 of 7 label differences changed whether the scripted teammate yielded and the model passed; 5 did not change run outcomes.One of the five outcome-neutral differences was a clear judge error labeling approval as an objection.Source: §C.2 · p. 18
  • 81% of labeled messages received the same label from both judges, with agreement at 67%–78% in four dimensions that contain 72% of disagreements.Agreement was 85%–96% in resisting unsafe pressure, keeping the team on track, reporting misconduct and taking initiative.Source: §C.2 · p. 18, Table 3 · p. 19
  • Four other-family judges’ majority matched people on 90% of 40 messages; the two Claude judges agreed 80%, no more than with outside judges.The referee agreed with outside judges on 80%–84% of messages, and the second rater on 76%–82%.Source: §C.2 · pp. 18–19
  • Labels were lenient on 1 of 52 Claude messages versus 6 of 128 messages from other models (p = 0.68); the test detects bias only at about 19% Claude leniency or more.The first panel of smaller models gave the same result.Source: §C.2 · p. 19

C.3 Scoring each model on its own tasks

  • Because tasks solved by every model are too few, each model is scored on its own tasks passed alone; there are none in taking initiative.Among models solving at least 97% of tasks alone, harder-alone tasks were also harder team tests, especially in taking over (ρ = −0.53).Source: §C.3 · p. 19
  • 0.34 versus 0.43 log loss shows the joint model predicts test runs better than each row’s own dimension mean.Its fill-in correction shifts log-odds by −0.21 after validation found a 0.4-point average overstatement.Source: §C.3 · p. 20
  • At most one rank changes under score adjustment, inverse-probability weighting, doubly robust estimation, or pairwise-only ratings.The adjustment lowers the ten top rows by 0.1 points on average and the other 14 by 1.9.Source: §C.3 · p. 20

C.4 Task quality and a smaller subset

The Lite subset preserves most of the full benchmark’s model ranking while requiring fewer runs. Some tasks provide little information about model ordering because nearly all or very few rows pass them.

C.4 Task quality and a smaller subset

  • Kendall’s τ = 0.95 and Spearman’s ρ = 0.99 on the Lite subset, which uses 17% of full-evaluation runs.On held-out rows, τ = 0.90 and mean error is 3.4 points; random subsets give τ = 0.88 and 3.7 points.Source: Figure 11 · p. 20, §C.4 · p. 20
  • 28 tasks are passed by at least 95% of rows and 8 by at most 5%, making them candidates for replacement because they carry little ranking information.The median task correlates with its dimension score at r = 0.70; only 2 tasks correlate negatively.Source: §C.4 · p. 20

C.5 Real-world records and benchmark-built tasks

Scores on real-record tasks and benchmark-built tasks rank models alike, although benchmark-built tasks are harder on average. Task difficulty also varies across source families.

C.5 Real-world records and benchmark-built tasks

  • Pearson r = 0.99 and Kendall’s τ = 0.93 between real-record and benchmark-built task scores; benchmark-built tasks are harder by 10.4 points on average.The source changes task difficulty, not which model does better.Source: §C.5 · p. 20, Figure 12 · p. 21
  • Accident reports are the easiest source family, while negotiations and team decisions are the hardest; frontier models outperform the average in every family except negotiation.Negotiation is the source of the conflict tasks.Source: §C.5 · p. 21, Figure 12 · p. 21

D Full results

The results separate losses caused by working in a team from losses caused by a teammate’s problem, and compare teamwork scores and profiles across models.

D Full results

  • Table 5 reports test scores by dimension for every result card, including high-effort runs, with half of each 95% bootstrap interval shown in grey.Figure 13 presents the same scores as one profile per model.Source: Table 5 · p. 22, D Full results · p. 22

D.1 Where the score is lost

  • 78% of lost score occurs in the test; frontier models lose 3.2 points in a calm team and 25 points to the teammate’s problem, versus 17 and 42 points for open-weight models.A calm negotiation costs 31 points, compared with an average of 8 points in the other dimensions.Source: §D.1 · p. 22, Figure 14 · p. 24

D.2 Score by release date

  • Figure 15 dates each model by when its evaluated version became publicly callable and marks successive models that beat all earlier ones.The figure shows release timing for current models, not the field’s historical trajectory.Source: Figure 15 · p. 24

E Reasoning effort

Higher reasoning effort changes teamwork unevenly across models and dimensions. It improves several dimensions but worsens conflict handling, and more output does not predict a better score.

E Reasoning effort

  • Five of six models gain 2.7 to 5.7 overall points at high effort, while Kimi K3.1 Pro loses 3.7 points.The paired sign test is significant for five models; Table 6 reports overall scores, high-effort evaluation cost, and task-level rises and falls.Source: E Reasoning effort · p. 24, Table 6 · p. 25
  • Output length does not predict a better teamwork score; at default effort, mean output ranges from 501 to 4,609 tokens per run and correlates with score at r = −0.35.For Kimi K3.1 Pro, high effort reduces tool calls on initiative tasks from 89 to 54 per run, alongside a decline in noticing failures.Source: E Reasoning effort · pp. 25–26, Figure 17 · p. 25
  • Taking initiative rises by 8.4 points on average with high effort, while handling conflict falls for all six models by 0.3 to 11.7 points.Resisting unsafe pressure and reporting misconduct change by at most 0.6 points; four dimensions improve significantly.Source: E Reasoning effort · p. 25, Figure 16 · p. 25

F Capability and teamwork

General capability tracks teamwork closely overall, but it does not account for all model differences. Handling conflict runs counter to the other dimensions, and deal-quality measures reveal a distinct trade-off.

F Capability and teamwork

  • The Artificial Analysis Intelligence Index correlates with teamwork score at r = 0.92, with a 95% interval of 0.87 to 0.97, and accounts for 85% of score variance.Humanity’s Last Exam also correlates at r = 0.92; LMArena correlates at r = 0.76.Source: F Capability and teamwork · p. 26, Table 7 · p. 26, Figure 18 · p. 27
  • Handling conflict is the only dimension negatively associated with capability (r = −0.34); the other seven correlate with capability from 0.75 to 0.95.Models within one point on the capability index can differ by 11 teamwork points, and the highest-index model ranks fourth on teamwork.Source: F Capability and teamwork · p. 26

F.1 Handling conflict and ability

  • Conflict correlates negatively with each other teamwork dimension (r = −0.43 to −0.23), while the other dimensions correlate with one another at r ≥ 0.59.One component explains 71% of variance across the eight scores, with conflict the only dimension against it.Source: §F.1 · p. 26, Figure 19 · p. 27
  • Models stronger on the other seven dimensions ask about teammate needs more often (r = 0.81) and also give in more often (r = 0.53), while best-deal frequency is unrelated to ability (r = 0.00).Open-weight models hold firm in 28% of runs, versus 5% for frontier models.Source: §F.1 · p. 27

F.2 Joint points of the deals

  • Across 1,359 test runs, deals reach 94% of the best joint points on average; deals after giving in reach 93%, while deals after holding firm reach 90%.The conflict score is retained because the tasks test whether models give in to a pushy teammate, even though giving in costs only a few team points per task.Source: §F.2 · pp. 27–28, Table 8 · p. 28
  • Models’ deals reach 90% to 96% of best joint points, while conflict scores range from 7 to 53; counting no-deal runs as zero reverses the association (r = −0.55).Frontier models end without a deal in 1% of runs, versus 10% for open-weight models.Source: §F.2 · p. 28

G Details of the analyses

  • The analyses use saved runs rather than new model calls, with 95% percentile bootstrap intervals from 2,000 task resamples per dimension.Model-score intervals use 4,000 resamples.Source: G Details of the analyses · p. 28
Loading 2611.02633v1…