Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Super-additive Cooperation in Language Model Agents

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that combining repeated interactions with inter-group competition in an iterated prisoner's dilemma tournament raises both overall and one-shot cooperation among LLM agents, with the pattern holding for Qwen3 and Phi4 but n

desk verdict Promising pilot on super-additive cooperation in LLM agents, but a prompt-wording confound and overlapping CIs leave the causal claim unestablished; worth refereeing, not yet decisive. read the letter →

arxiv 2508.15510 v1 pith:XE6WAXHC submitted 2025-08-21 cs.AI

Filippo Tonini , Lukas Galke This is my paper
classification cs.AI
keywords super-additivecooperationlargelanguagemodelagentsiteratedprisoner'sdilemmamulti-agentinter-groupcompetitionone-shottournamentsimulationAIalignment
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 tests whether super-additive cooperation—the human pattern in which repeated interactions plus inter-group competition jointly produce more cooperation than either alone—also appears in agents built from large language models. It places identical LLM agents in iterated prisoner's dilemma tournaments under three structures: repeated interactions only, inter-group competition only, and a combined 'super-additive' condition. For Qwen3 and Phi4, the combined condition produced higher average cooperation and higher one-shot cooperation (cooperation with a stranger on the first move) than either mechanism alone; Cogito did not follow the pattern. The authors read the exception as a capability effect: models that better understood the game, as measured by post-match comprehension prompts, showed the super-additive boost. If right, the result suggests that environmental structure, not just model fine-tuning, can promote cooperative behavior in multi-agent AI systems.

What carries the argument

The load-bearing object is the super-additive tournament: a team-structured iterated prisoner's dilemma in which agents have a finite round budget, may unilaterally exit a match and take a new opponent, and refine a written plan every five rounds through a plan-then-critic loop. Across the three conditions, the structure varies the goal sentence in the prompt—personal score in RI, group score in GC, both in SA—and whether matches occur only between groups (GC) or between all players (RI, SA). The first move of every match is played against an 'unknown opponent', so first-round cooperation measures one-shot cooperation without identity or history cues. All agents in a tournament share the sam

What would settle it

Run the three conditions with a single invariant goal sentence—for example, always telling agents to maximize their own score—and vary only whether they are assigned to teams and whether matches are intra- or inter-group. If the SA advantage disappears, the reported effect is an artifact of goal phrasing rather than of the interaction structure. A complementary test: use the SA wording ('highest group and personal score') in a tournament with no actual group payoff, and compare cooperation to the real SA tournament.

Watch

Extended reading notes

Core claim

The central claim is that the combination of repeated interactions and inter-group competition—and not either element on its own—raises LLM cooperation rates, including one-shot cooperation with unknown opponents. The paper formulates this as H1 and H2 and finds support in the aggregate cooperation and OSC metrics for Qwen3 and Phi4, with the SA condition yielding higher means and non-overlapping confidence intervals versus RI and GC in those models. In the SA condition, the overall boost is driven mainly by intra-group matches: agents cooperate more with teammates than with out-group opponents, mirroring the human super-additive pattern. The paper treats Cogito as the informative exception:

Load-bearing premise

The three tournament conditions differ not only in structure but also in the exact goal wording shown to the model ('highest personal score' vs 'highest group score' vs 'highest group and personal score'), so the measured cooperation boost could come from phrasing rather than from repeated interactions or inter-group rivalry.

Editorial extensions

If this is right

  • If the super-additive effect is real, multi-agent system designers can raise baseline cooperation without fine-tuning simply by embedding agents in teams that compete against other teams while also meeting repeatedly.
  • One-shot cooperation rising with the same structure implies that team pressure can make an agent more willing to cooperate with a stranger on the first encounter, which matters for open networks where agents meet new counterparts.
  • The effect concentrates in intra-group interactions, so inter-group competition is not purely adversarial: it can be a tool for increasing within-team trust while keeping between-team competition.
  • The model-dependence result suggests a screening step: before relying on structural interventions to induce cooperation, check whether the model understands the game (the paper's meta-prompt scores); otherwise the structure may not help or may backfire.
  • Because the same structure raises cooperation in two independent open-weight models, the result is not tied to a single vendor's training run, though it has not been tested on larger proprietary models.

Reading between the lines

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

  • The paper does not control for prompt wording: the SA condition's goal sentence says 'highest group and personal score,' which may itself trigger cooperative behavior irrespective of real group competition. A condition that uses identical wording but withholds the actual team payoff would separate phrasing from structure.
  • A natural extension is to vary the visibility of the competition (public leaderboard vs private scores) and the size of teams; the theory predicts the effect should grow with the salience of the out-group threat, but the paper does not test this.
  • Because the effect appears without the words 'cooperate' or 'defect', semantic priming is unlikely to explain it; this makes the structural explanation more plausible and worth testing in other social dilemmas such as public goods games.
  • The Cogito exception suggests a capability threshold: models that cannot track group-level scoring may default to their training bias; one would predict that stronger reasoning models show larger super-additive effects, a claim the paper does not make.
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 constructs a tournament framework in which LLM agents play the Iterated Prisoner's Dilemma under three conditions: repeated interactions (RI), group competition (GC), and their combination (super-additive, SA). It tests two hypotheses: that mean cooperation (H1) and mean one-shot cooperation (H2) in the SA condition exceed the maximum of the RI and GC conditions. Experiments with Qwen3 14b, Phi4 reasoning, and Cogito 14b show that H1 and H2 hold for Qwen3 and Phi4, while Cogito is more cooperative overall and peaks under GC rather than SA. The authors attribute Cogito's exception to lower game understanding, measured by meta-prompt accuracy. The paper contributes a reusable planning/evaluation prompting scheme and open-source code, and explicitly acknowledges several limitations including scale, prompt sensitivity, and generalizability.

Significance. If the super-additive effect is real, this would be a useful result for designing multi-agent LLM systems and for testing whether human-like cooperation mechanisms transfer to AI agents. The paper is commendably transparent: it posts source code, reports raw means and confidence intervals, states hypotheses explicitly, and includes an honest limitations section. However, the central causal claim currently rests on comparisons across conditions that differ in prompt wording as well as network structure, and the one-shot cooperation result is not statistically supported by the reported intervals. The Cogito exception is interpreted through a post hoc mechanism that is not directly tested. These concerns are load-bearing because the paper's headline is a causal, model-general claim. The framework itself is a genuine contribution and the questions are well motivated, but the evidence as presented is not yet conclusive.

major comments (4)
  1. [§3.4 and §3.3] The three tournament conditions differ not only in interaction network but also in the prompt's goal wording: RI says 'highest personal score,' GC says 'highest group score,' and SA says 'highest group and personal score' and additionally provides group-score information. The paper itself states in §3.4 that 'minor changes in phrasing can lead to significantly different outcomes' and lists prompt sensitivity as a limitation in §6. Therefore the higher cooperation in SA relative to RI/GC could plausibly be driven by the wording 'group and personal' rather than by the structural combination of repeated in-group interactions and inter-group competition. This confound applies equally to H1 and H2. A factorial design that independently varies the goal wording and the interaction network—or at minimum a control condition in which SA uses the same goal text as GC with only the network changed—i
  2. [§4.2, Tables 6–8] H2 is not statistically established by the reported confidence intervals. For Qwen3, the SA one-shot cooperation rate is 0.39 [0.31, 0.47], which overlaps with the RI interval [0.20, 0.34] and the GC interval [0.16, 0.35]. For Phi4, SA is 0.43 [0.35, 0.51], overlapping with RI [0.22, 0.37] and GC [0.19, 0.38]. No significance test, effect size, or correction for multiple comparisons is reported, and the t-based intervals are applied to binary proportions without clustering by player or match. The claim that SA raises one-shot cooperation above both mechanisms is therefore not supported by the data as presented. A permutation test or a mixed-effects logistic regression on first-round choices with player and match random effects would be appropriate.
  3. [§4.4 and §5] The post hoc explanation for Cogito's exception—that its higher cooperation and GC peak reflect insufficient game understanding—is not tested quantitatively. Meta-prompt accuracy is reported only for the SA tournaments (Fig. 5), and no correlation is shown between meta-prompt accuracy and condition-specific cooperation rates. The direction of the explanation is also ambiguous: a model that cooperates uniformly across contexts might score lower on strategy-related meta-questions because it did not encounter diverse strategic situations. If this explanation is retained, the paper should measure game understanding independently of tournament condition and test the predicted relationship explicitly.
  4. [§3.1 and §3.5] The 'one-shot cooperation' measure does not match its stated definition. Section 3.5 defines OSC as cooperative behavior in the first interaction with 'no prospect of future encounters,' but in the RI and SA conditions the first round is followed by up to n−1 additional rounds of the same match (Section 3.1, 3.3). Masking the opponent's identity as 'unknown opponent' does not remove the expectation of future rounds. The measured quantity is thus first-round cooperation in an iterated game, not a true one-shot PD. This matters for H2, because super-additive theory's OSC claim concerns genuine one-off interactions. The design should use single-round matches for the OSC probe, or the terminology and interpretation should be restricted to first-round behavior in the IPD.
minor comments (5)
  1. [§4.1] The text refers to 'Phi 4 (figures 5a–5c)', but the relevant panels appear in Fig. 2 (likely 2d–2f). Please correct the cross-reference.
  2. [§3.2] Typo: 'open-wight models' should be 'open-weight models'.
  3. [§3.3] The combinatorial notation such as '(h 2)' and '(t 2)' should be typeset as binomial coefficients or explicitly described, as the current notation is ambiguous.
  4. [§4.4] It would be clearer to state explicitly that the meta-prompt accuracy results are only from the SA tournaments, since the text in §4.4 could be read as applying to all conditions.
  5. [§6] The limitation about partner choice is mentioned, but the paper does not report any analysis of how often agents exited matches or whether exit behavior differed across conditions. Reporting this would strengthen the interpretation of the cooperation rates.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: hypotheses are direct empirical comparisons, not derived from fitted inputs or self-citations.

full rationale

The paper's derivation chain is empirical rather than formal. The central hypotheses H1 and H2 (Section 1) are stated as ordinal comparisons of measured cooperation rates and are evaluated directly against the means in Tables 3–8, computed by the formulas in Section 3.5. No parameter is fitted to a subset of data and then renamed a prediction; the RI/GC/SA conditions are structurally defined in Section 3.3, and the outcomes are measured, not derived from the definition. The super-additive theory imported from Efferson et al. [9] is an external human experimental result used to motivate the hypotheses, not a self-citation, and it is not invoked as a proof of the LLM results. The post hoc explanation for Cogito's exception via meta-prompt accuracy is interpretive but not circular, since it does not feed back into the measured quantities. The prompt-wording differences across conditions are a genuine internal-validity concern acknowledged in Section 3.4 and Section 6, but a confound is a correctness risk, not a circularity. No equation reduces to another by construction, and no load-bearing step relies on the authors' own prior claims. Score 0 is appropriate because the central claim is a direct empirical comparison without fitted-input or self-citation circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new theoretical entities. The central claim rests on experimental design choices (team size, repetitions, planning interval, unreported round caps) and on assumptions that the prompt manipulations cleanly instantiate the social structures and that first-round choices measure a stable one-shot cooperation tendency.

free parameters (6)
  • K, planning-evaluation interval = 5
    Chosen as 'a good balance' (Section 3.2); affects how often agents revise strategy and therefore cooperation dynamics.
  • Tournament composition = 2 teams x 3 players
    Chosen as balance between runtime and scale (Section 3.3); group size and number of teams may affect inter-group competition pressure.
  • Number of repetitions per condition = 5
    Five repeats per model/condition (Section 3.3) determine the reported CIs; no power analysis is reported.
  • Max rounds per match n = not reported
    Referred to in constraint N < n*m (Section 3.3) but no numeric value is given; match length strongly affects repeated-interaction dynamics.
  • Total round budget N = not reported
    Used in the same constraint; not reported, so the strategic pressure to terminate matches is unspecified.
  • LLM decoding parameters (temperature, top_p, etc.) = not reported
    Model outputs depend on sampling; no decoding settings are given in the paper.
assumptions (5)
  • domain assumption The prisoner's dilemma payoff matrix and the mapping of action_a to cooperation capture meaningful social cooperation.
    The entire inference equates action_a with cooperation and treats the game as a valid model of cooperation (Section 3.1).
  • domain assumption The condition differences are caused by the tournament structures; prompt goal wording and team information are just implementations of those structures.
    Section 3.4 varies goals by condition; if phrasing itself drives behavior, the causal story fails. The authors acknowledge prompt sensitivity but do not control it.
  • domain assumption First-round actions against an 'unknown opponent' measure one-shot cooperation disposition rather than immediate response to the current group/personal objective.
    The OSC definition in Section 3.5 assumes first moves reflect stable cooperation tendency even though SA agents know their group goal.
  • ad hoc to paper Meta-prompt accuracy reflects game understanding and can explain model differences post hoc.
    Used to explain Cogito's exception (Section 5); this interpretation is introduced after seeing the results and is not a pre-registered measure.
  • domain assumption The three small open-weight LLMs are representative of LLM agents for the purposes of the conclusion.
    The limitations section acknowledges larger/stronger models may differ, but the abstract and conclusions nevertheless generalize to LLM agents.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Super-additive Cooperation in Language Model Agents." pith.science (2026). https://pith.science/paper/XE6WAXHC

@misc{pith2026250815510,
  author       = {Pith},
  title        = {Pith review of: Super-additive Cooperation in Language Model Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XE6WAXHC}},
  note         = {Machine review of arXiv:2508.15510}
}
read the original abstract

With the prospect of autonomous artificial intelligence (AI) agents, studying their tendency for cooperative behavior becomes an increasingly relevant topic. This study is inspired by the super-additive cooperation theory, where the combined effects of repeated interactions and inter-group rivalry have been argued to be the cause for cooperative tendencies found in humans. We devised a virtual tournament where language model agents, grouped into teams, face each other in a Prisoner's Dilemma game. By simulating both internal team dynamics and external competition, we discovered that this blend substantially boosts both overall and initial, one-shot cooperation levels (the tendency to cooperate in one-off interactions). This research provides a novel framework for large language models to strategize and act in complex social scenarios and offers evidence for how intergroup competition can, counter-intuitively, result in more cooperative behavior. These insights are crucial for designing future multi-agent AI systems that can effectively work together and better align with human values. Source code is available at https://github.com/pippot/Superadditive-cooperation-LLMs.

Figures

Figures reproduced from arXiv: 2508.15510 by the authors.

Figure 1
Figure 1. Workflow graph. into raw prompts and model outputs for diagnostic purposes. The complete codebase is available on GitHub.1 3.4 Prompting We use a total of 3 different prompts, one used to query the agent every round and two more used for planning and evaluating used every K rounds. All these prompts follow the same structure: 1. Game Description and Rules: A comprehensive summary of the game’s setting, objectives, a… view at source ↗
Figure 2
Figure 2. Each plot shows the cooperation rate (pc) evolution over the rounds played, for every LLM we show the results in the different tournament structures [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Each plot shows the one-shot cooperation rate ( [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: intra-group and inter-group cooperation in the super additive setting for the dif [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Meta prompt accuracy results from the SA tournaments for the different LLMs. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in 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. The Energy Society: A Simulation Environment for Studying Agent Cooperation under Survival Pressure

    cs.MA 2026-07 conditional novelty 6.0 of 10

    A new multi-agent simulation ties token-generation cost to survival, showing larger models are less efficient, shared objectives induce donations, and recommendation and memory mechanisms shape coordination.

Reference graph

Works this paper leans on

20 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abdin, M., Agarwal, S., Awadallah, A., Balachandran, V ., Behl, H., Chen, L., de Rosa, G., Gunasekar, S., Javaheripi, M., Joshi, N., Kauffmann, P., Lara, Y ., Mendes, C.C.T., Mitra, A., Nushi, B., Papailiopoulos, D., Saarikivi, O., Shah, S., Shrivastava, V ., Vi- neet, V ., Wu, Y ., Yousefi, S., Zheng, G.: Phi-4-reasoning technical report (2025), https://...

    arXiv 2025

  2. [2]

    Aher, G., Arriaga, R.I., Kalai, A.T.: Using large language models to simulate multiple hu- mans and replicate human subject studies (2023), https://arxiv.org/abs/2208.10264

    arXiv 2023

  3. [3]

    Akata, E., Schulz, L., Coda-Forno, J., Oh, S.J., Bethge, M., Schulz, E.: Playing repeated games with large language models (2023), https://arxiv.org/abs/2305.16867

    arXiv 2023

  4. [4]

    The Journal of Conflict Resolution 24(1), 3–25 (1980), http://www.jstor.org/stable/173932

    Axelrod, R.: Effective choice in the prisoner’s dilemma. The Journal of Conflict Resolution 24(1), 3–25 (1980), http://www.jstor.org/stable/173932

    work page 1980

  5. [5]

    Brookins, P., Debacker, J.: Playing games with GPT: What can we learn about a large language model from canonical strategic games? SSRN Electronic Journal (2023), https://api.semanticscholar.org/CorpusID:259714625

    work page 2023

  6. [6]

    https://doi.org/10.1257/aer.101.1.411

    Bó, Fréchette: The evolution of cooperation in infinitely repeated games: Experimental evi- dence 101 (2011). https://doi.org/10.1257/aer.101.1.411

    doi:10.1257/aer.101.1.411 2011

  7. [7]

    Christiano, P., Shlegeris, B., Amodei, D.: Supervising strong learners by amplifying weak experts (2018), https://arxiv.org/abs/1810.08575

    arXiv 2018

  8. [8]

    (Blog post) (2025), https://www.deepcogito.com/research/cogito-v1-preview

    Deep Cogito: Cogito v1 preview introducing ida as a path to general superintelligence. (Blog post) (2025), https://www.deepcogito.com/research/cogito-v1-preview

    work page 2025

Show all 20 references
  1. [9]

    Nature 626, 1–8 (02 2024)

    Efferson, C., Bernhard, H., Fischbacher, U., Fehr, E.: Super-additive cooperation. Nature 626, 1–8 (02 2024). https://doi.org/10.1038/s41586-024-07077-w

    2024 doi

  2. [10]

    Fan, C., Chen, J., Jin, Y ., He, H.: Can large language models serve as rational players in game theory? a systematic analysis (2023), https://arxiv.org/abs/2312.05488

    2023 arXiv

  3. [11]

    Fontana, N., Pierri, F., Aiello, L.M.: Nicer than humans: How do large language models behave in the prisoner’s dilemma? (2024), https://arxiv.org/abs/2406.13605

    2024 arXiv

  4. [12]

    Garcia, J.L., Hajkova, K., Marchenko, M., Patiño, C.M.: Reproducibility study of coop- eration, competition, and maliciousness: Llm-stakeholders interactive negotiation (2025), https://arxiv.org/abs/2502.16242

    2025 arXiv

  5. [13]

    PLOS Computational Biology 21(2), 1–19 (02 2025)

    Graser, C., Fujiwara-Greve, T., García, J., van Veelen, M.: Repeated games with partner choice. PLOS Computational Biology 21(2), 1–19 (02 2025). https://doi.org/10.1371/journal.pcbi.1012810, https://doi.org/10.1371/journal.pcbi.1012810

    2025 doi

  6. [14]

    Meinke, A., Schoen, B., Scheurer, J., Balesni, M., Shah, R., Hobbhahn, M.: Frontier models are capable of in-context scheming (2025), https://arxiv.org/abs/2412.04984

    2025 arXiv

  7. [15]

    Mensfelt, A., Stathis, K., Trencsenyi, V .: Logic-enhanced language model agents for trust- worthy social simulations (2024), https://arxiv.org/abs/2408.16081

    2024 arXiv

  8. [16]

    Qwen Team: Qwen3 technical report (2025), https://arxiv.org/abs/2505.09388

    2025 arXiv

  9. [17]

    Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Lan- guage agents with verbal reinforcement learning (2023), https://arxiv.org/abs/2303.11366

    2023 arXiv

  10. [18]

    Annals of Mathematics Studies, Princeton University Press (1959), https://books.google.dk/books?id=9lSVFzsTGWsC

    Tucker, A., Luce, R.: Contributions to the Theory of Games. Annals of Mathematics Studies, Princeton University Press (1959), https://books.google.dk/books?id=9lSVFzsTGWsC

    1959

  11. [19]

    Willis, R., Du, Y ., Leibo, J.Z., Luck, M.: Will systems of llm agents cooperate: An investi- gation into a social dilemma (2025), https://arxiv.org/abs/2501.16173

    2025 arXiv

  12. [20]

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., Cao, Y .: React: Synergizing reasoning and acting in language models (2023), https://arxiv.org/abs/2210.03629

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.