Playing with Strangers: A New Benchmark for Ad-Hoc Human-AI Teamwork

TL;DR for operators

Teamwork is the awkward part of agentic AI. It is easy to show a model completing a task when the environment is clean, the instructions are explicit, and the other “teammates” behave exactly as expected. Real deployments are less polite. Humans omit context, follow local conventions, adapt unevenly, and occasionally do something that looks wrong only because the system has misunderstood the room.

The paper behind AH2AC2, the Ad-Hoc Human-AI Coordination Challenge, is useful because it moves evaluation closer to that problem.¹ It does not ask whether an agent can win Hanabi with a clone of itself. It asks whether an agent can coordinate with human-like partners in a game built around hidden information, constrained communication, and convention-heavy interpretation.

The evidence is more interesting than a leaderboard victory lap. Pure behavioural cloning looks human-like on prediction metrics but collapses in play. Self-play can learn conventions that do not transfer. Population methods do not automatically rescue coordination. Small human datasets help less than one might hope. A prompted reasoning model shows some capability, especially in three-player games, but the result is explicitly preliminary and not a declaration that LLMs have mastered teamwork. Naturally, the robots still have homework.

For business use, the lesson is not “use Hanabi to evaluate your sales assistant”. The lesson is: if your AI product must work with people, then task accuracy is only one slice of evaluation. You also need tests for convention alignment, partner robustness, partial observability, and behaviour under limited human data. AH2AC2 offers a research version of that evaluation discipline.

The uncomfortable result: clever agents still make poor teammates

The headline result is not that one method wins. The headline result is that almost every method fails in a recognisably different way.

The AH2AC2 leaderboard compares several families of agents against hosted human-proxy partners. In the two-player setting, Off-Belief Learning, or OBL, reaches a mean score of 21.04 and median score of 22 without using the open human data. Best response to behavioural cloning, or BR-BC, follows at 19.41 mean and 20 median. HDR-IPPO, despite using human-data regularisation, reaches only 12.76 mean and 15 median in the data-limited challenge setting. Behavioural cloning itself lands at 2.12 mean and median 0. A model can imitate human moves and still be a terrible collaborator. That sentence should be printed on procurement decks, preferably in red.

In the three-player setting, the picture shifts. DeepSeek-R1 with H-Group conventions reaches 14.62 mean and 18 median, narrowly above HDR-IPPO at 14.03 mean and 16 median. But the authors are careful: DeepSeek-R1 was evaluated on only 100 games because of resource and time constraints, while the other methods use the standard 1,000-game evaluation. That makes the LLM result an exploratory signal, not a crown.

This is why the paper deserves an evidence-first reading. A standard summary would say: new benchmark, new dataset, new baselines. Accurate, but anaemic. The useful story is that human-compatible coordination remains a measurement problem before it is an optimisation problem.

What AH2AC2 actually evaluates

Hanabi is a cooperative card game where players can see everyone else’s cards but not their own. Players must give hints, infer intent, preserve information tokens, avoid wasting limited lives, and coordinate around conventions. The maximum score is 25. If the team loses all lives, the score is zero.

That makes Hanabi unusually punishing for brittle teamwork. An agent cannot simply “know the answer”. It must interpret why a teammate gave a clue, whether a card should be saved, when silence is information, and when a move is legal but socially nonsensical. In other words, Hanabi turns coordination into a laboratory without pretending the laboratory is the world.

AH2AC2 has two main components.

First, there is the coordination challenge. Candidate agents are evaluated by playing with hosted human-proxy agents. These proxies are trained from a large hidden dataset of human Hanabi games from the hanab.live community: 101,096 two-player games and 46,525 three-player games, for 147,621 games in total. Participants do not receive this full dataset. They receive a deliberately limited public dataset of 3,079 games: 1,858 two-player games and 1,221 three-player games. The public games are high-quality human play, with average scores above 23 in both settings.

Second, there is an optional human-action prediction task. Here, agents must predict human moves on held-out games using teacher-forced cross-entropy loss. This matters because coordination and prediction are related but not identical. A model may predict plausible human actions yet fail when its own actions alter the future state. In business terms, watching humans in the workflow is not the same as being a good teammate inside the workflow. The latter is where the expensive mistakes usually live.

The human proxies are not released directly. They are hosted behind an API, with registered evaluation access and a leaderboard. The reason is not mysterious: if researchers could download the test partners, they could overfit to them. The benchmark would become a contest in memorising the examiner’s habits. Academia has enough hobbies.

The proxy agents are the benchmark’s real product

The most important engineering move in the paper is the construction of human-proxy agents. These are not just imitation models. The authors first train behavioural cloning policies on human games, then refine them using Human-Data-Regularised Independent Proximal Policy Optimisation, or HDR-IPPO.

The mechanism is simple in outline. Behavioural cloning gives the agent a human-like starting point. Reinforcement learning improves robustness in gameplay. A KL regularisation term keeps the learned policy close to the behavioural cloning policy, preventing the agent from drifting into alien conventions that score well with itself but confuse human-like partners.

That last part is the whole game. Without the regularisation, RL may discover strategies that look effective in isolation but stop being interpretable to partners. With too much regularisation, the model may remain too close to brittle imitation. The useful range is a compromise: improve play without shedding the social grammar.

The paper validates the human proxies through several tests, each serving a different purpose.

The self-play gains are large. In the paper’s proxy validation, two-player human proxies score mean self-play results around 22.55 and 22.97, while their corresponding BC policies are 3.0 and 4.0 points lower. In three-player self-play, the improvement is much more dramatic: the human proxies score 20.88 and 21.21, improving over BC by 15.7 and 13.9 points. The authors note that three-player BC agents frequently lose all lives, producing zero-score games. Regularised RL cuts those collapses sharply.

This is not merely “RL makes the score go up”. The more precise interpretation is: imitation captures conventions, but it cannot reliably recover from unfamiliar states. Reinforcement learning improves robustness, but unregularised optimisation can drift away from convention. Human-data regularisation is the bridge between “looks human in supervised data” and “plays robustly with human-like partners”.

That distinction matters for enterprise agents. A support agent trained on historical conversations may sound like a good employee. Put it into a live escalation workflow, and it may encounter states that never appeared in the archive: half-resolved tickets, contradictory customer claims, legacy rules known only by one team, or a human colleague using shorthand. Behavioural imitation is useful until reality starts improvising.

The ablation is not decorative; it explains the failure mode

The appendix’s regularisation ablation is worth treating as mechanism evidence, not academic upholstery. The authors vary the KL regularisation weight in a two-player setting from 0.00 to 0.70. The goal is to see how strongly the RL policy must be anchored to the behavioural cloning policy to retain human-compatible conventions.

The pattern is not “more regularisation is always better”. The agent with a regularisation weight of 0.20 records the strongest self-play performance in the ablation, with a reported mean score of 23.42. In 5,000 self-play games, that same weight produces 1,821 perfect games, zero zero-score games, and a minimum score of 15. By contrast, no regularisation produces only 143 perfect games, although it also avoids zero-score games in that specific table. Higher regularisation weights can also degrade robustness: at 0.70, the agent records 1,166 perfect games but 113 zero-score games.

The held-out prediction results show the other side of the trade-off. The non-regularised policy performs badly against human data, with test loss 7.385 and test accuracy 0.327. Higher regularisation brings the model back towards human-like action distributions. At 0.70, test loss is 0.469 and test accuracy 0.668, close to the behavioural cloning policy’s 0.468 loss and 0.674 accuracy.

So the ablation’s message is not a neat slider labelled “human-likeness”. It is a coordination frontier. Too little anchoring and the agent wanders into private conventions. Too much anchoring and it may inherit imitation’s fragility. The valuable agent is neither a parrot nor a solo genius. Tedious, yes. Also true.

The leaderboard is a taxonomy of coordination failures

The baseline section is useful because it refuses to make the problem look easier than it is. Each method represents a plausible philosophy of coordination.

Self-play says: let agents learn together. The failure is over-coordination with familiar partners. Behavioural cloning says: imitate humans. The failure is brittleness when the model’s own actions push the game into unfamiliar states. BR-BC says: train a best response to a human-like model. The failure is that performance may improve while action-prediction alignment becomes poor. Population methods say: train against diversity. The failure is that diversity is not automatically the right kind of human compatibility. Prompted LLMs say: use language and reasoning. The failure is that language-level competence does not yet translate into reliable interactive coordination under benchmark pressure.

This is the paper’s most useful business translation: a strong agent architecture does not guarantee a strong teammate. The coordination failure can sit in the training regime, the evaluation partner, the data scale, the convention set, or the interface between prediction and action. “Our model passed the task benchmark” is therefore not enough. Passed with whom? Under what hidden information? With what conventions? After whose unexpected move?

Those questions sound fussy until a deployed agent sends the wrong instruction to a nurse, routes a finance approval around the wrong stakeholder, or “helps” an operator by overriding a local exception. Then everyone rediscovers evaluation design, usually while wearing a fireproof vest.

The LLM result is interesting because it is not triumphant

The DeepSeek-R1 experiment is a useful inclusion precisely because it is modest. The authors prompt the model with Hanabi state information and test two variants: one with only the current game state, and another with H-Group conventions included. There is no Hanabi-specific fine-tuning.

In two-player games, the H-Group prompt improves performance from 5.43 mean to 9.91 mean, but both have median 0. That suggests the model sometimes coordinates but often collapses. In three-player games, the model performs better: 14.38 mean without conventions and 14.62 mean with conventions, both with median 18. This is competitive with the strongest standard 3P baseline in the table, HDR-IPPO at 14.03 mean and 16 median.

But the paper places the necessary boundary around the result. DeepSeek-R1 was evaluated on only 100 games, unlike the 1,000-game standard used for other methods, and the action-prediction challenge is left for future work. That makes the result a research lead, not a product conclusion.

For business readers, the implication is pleasantly inconvenient. LLMs may bring useful priors for reasoning about conventions and hidden information. They may also be unstable, expensive to evaluate, and sensitive to prompt framing. The operational question is not whether an LLM can explain teamwork. Most can. At length. The question is whether it can coordinate reliably across many interaction trajectories where its own previous moves change the state space. That is a different exam.

What this paper directly shows

The paper directly shows three things.

First, a standardised human-AI coordination benchmark can be built without releasing the full evaluation machinery. AH2AC2 uses a small public dataset for participants, hosted proxy partners for evaluation, and a leaderboard with controlled access. This is not just a technical detail. It is benchmark governance.

Second, human-proxy agents can be made stronger by combining behavioural cloning with regularised RL. The proxies improve substantially over BC in gameplay while retaining evidence of human-like behaviour through cross-play, action prediction, and behavioural metrics such as information per play and communicativeness.

Third, existing baselines leave a large gap. OBL is strong in two-player games without using human data. Limited-data human adaptation methods underperform in important ways. Prompted LLMs show some capability but not enough evidence for confident claims. Coordination remains unsolved, which is rude but helpful.

What Cognaptus infers for enterprise AI systems

The enterprise interpretation is not that Hanabi should become the new Salesforce demo. The practical lesson is that organisations need coordination benchmarks for their own workflows.

A useful internal benchmark would copy AH2AC2’s evaluation philosophy rather than its game environment. It would include human-like proxy partners, hidden evaluation cases, limited public training examples, and both action-quality and action-prediction metrics. It would test how agents behave when partners follow different but plausible conventions. It would separate “does the agent know the policy?” from “does the agent coordinate under ambiguity?”.

For example, consider an AI operations assistant that helps triage procurement approvals. A normal benchmark might test whether it classifies invoices correctly. A coordination benchmark would test whether it routes ambiguous cases to the right human, preserves implicit approval conventions, adapts when a manager changes the sequence, and avoids confidently completing a task after misunderstanding a partial instruction.

The same pattern applies to clinical decision support, robotics, call-centre escalation, logistics control rooms, and finance operations. In each case, the agent’s value depends not only on individual correctness but on compatibility with human partners.

The ROI is not “agents get smarter”. That phrase has done enough damage. The ROI is cheaper diagnosis before deployment: finding out which agent fails with which partner type, which conventions it has mislearned, and whether adding human demonstrations actually improves coordination or only makes the confusion more fluent.

Where the evidence stops

The boundaries are important because they affect how this paper should be used.

Hanabi is structured. Workplace coordination is messier, slower, noisier, and entangled with incentives, permissions, liability, and organisational politics. The benchmark tests convention-heavy partial-information coordination, but it does not prove that a method will work in hospitals, vehicles, warehouses, or enterprise software.

The human proxies are trained from hanab.live players using H-Group conventions. The authors argue that H-Group conventions are not a single rigid strategy but a collection of techniques players mix and adapt. Still, this is a specific community and convention culture. Generalising beyond it requires care.

The proxies are validated indirectly. Cross-play, action prediction, IPP, communicativeness, and qualitative convention analysis are meaningful signals, but the authors themselves identify direct human-AI play with the proxies as future work. That is the correct next validation step.

The LLM evaluation is preliminary. The 100-game DeepSeek-R1 result is enough to justify further study, not enough to guide procurement. Anyone turning that table into “LLMs beat RL at teamwork” should be gently escorted away from the dashboard.

Finally, the benchmark’s integrity depends on controlled access. Hosted proxies reduce overfitting risk, but any leaderboard can become a game if submissions are frequent, incentives are high, and evaluation partners become predictable. Benchmark governance is not a one-time design choice. It is maintenance.

The real contribution is evaluation discipline

AH2AC2 is not exciting because it makes Hanabi agents score higher. It is exciting because it makes coordination failure easier to see.

That is a different kind of progress. In agentic AI, the industry is still over-invested in demos where the agent works alone, or with a perfectly obedient toolchain, or with a human reduced to a prompt box. But many valuable applications require agents to work with people who have partial knowledge, tacit conventions, changing priorities, and the occasional talent for ambiguity. So, humans.

The paper’s best lesson is therefore methodological. Do not evaluate only the agent. Evaluate the relationship between the agent, the partner, the data limit, the convention set, and the hidden state. Measure task outcome and behavioural alignment. Keep the test partners protected. Treat prediction and coordination as related but separate competencies.

That is less glamorous than announcing a universal teammate. It is also more useful. Before AI agents can become reliable colleagues, they need to pass tests built for strangers, not mirrors.

Cognaptus: Automate the Present, Incubate the Future.