acceptodds
Under review as a conference paper at ICLR 2027

Looking for Bidding Teammates: A Game-Theoretic Model of Stranger Collusion in Peer Review

Abstract

Paper bidding is the entry point to reviewer assignment at large machine learning conferences: reviewers declare interest in submissions, and an optimizer combines those bids with automated affinity scores. Reviewers who have never met now recruit each other on public social platforms, exchange submission identifiers, bid on each other's papers, and reciprocate with inflated scores. Existing collusion models take the group as given and assume its members already trust one another. We give the first game-theoretic model of collusion formation in peer review: a four-stage game over recruitment, identifier exchange under the risk of being reported, bidding neither party can verify, and reciprocal reviewing. The model predicts the arrangement cannot form, because the benefit at stake is delivered by the partner's separate decision, so reciprocation is never individually rational for any payoffs and the game unravels to the recruitment post. What closes the gap is enforcement rather than incentives: authors always see the reviews of their own paper, deadlines recur, and the group remembers who reciprocated. We give the sustainability condition and the detection rate above which no partnership survives, and show review-writing effort enters both, so cheap machine-assisted reviewing enlarges the sustainable set with no change in detection or sanctions. In a calibrated end-to-end conference simulation no detector we test exceeds against an attacker who camouflages bids, distributes them around a ring, and manipulates affinity; a two-person arrangement is worth percentage points of acceptance probability; and the harm is distributional rather than aggregate, with honest papers displaced while mean accepted quality moves by . Randomized assignment is the one defense reaching the enforcement itself, making a partner who never bid indistinguishable from one who bid and lost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Peer Review, Collusion, Game Theory, Reviewer Assignment, Paper Bidding, Research Integrity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HK5iOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper models the collusion formation process in peer review as an extensive form game and uses game theory to characterize equilibria in that game. Having modelled the game theoretically, the paper then analyses how certain parameters of the game affect the sustainability of collusive outcomes as the equilibrium of the game. The paper also simulates the game which is partially calibrated to real data.

Critical strengths

The paper addresses an important problem and tries to offer insights that can serve as a basis for designing real solutions to that problem.

Critical weaknesses
  1. There seems to be a serious flaw in Theorem 4, which propagates through the rest of the analysis. The grim-trigger strategy in Theorem 4 conditions on “observed betrayal”, which, as far as I understand the game, is not observable. Reviews are anonymous, so an honest (non-inflated) review from an assigned partner (i.e. playing H in the stage-4 game) looks the same as the partner not being assigned or never bidding. The only observable trigger is the absence of an inflated review. Therefore, the strategy described in Theorem 4 is not valid. The same applies to Proposition 5. I think it should be possible to reformulate the theorems and propositions in terms of a strategy that depends on the observable state, but all the results should be checked.

  2. I think there is a problem with the statement on line 483: “Our model says the arrangement should not form at all, and that what sustains it is enforcement rather than incentives.” Firstly, as the paper rightly shows, the lack of collusive arrangement is implied by the one-shot game, not the repeated game, which is a well-known result among game theorists and collusion theorists. Which one of the two is a better representation of the real world is a separate question. Arguably, it is the repeated version that is more representative, which the paper also seems to argue. In light of this, I think the statement is not justified. The one-shot game is a useful benchmark, but it does not imply what “should” or “should not” be observed in the real world.

Questions for authors
  1. In light of my comment above, the authors should either clarify how the betrayal state is observable, which I think is not, or the all the formal statements need to be redone from Theorem 4 onwards. I think it should be possible to reformulate the theorems and propositions in terms of a strategy that depends on an observable state.
  2. A presentation point: I think the paper would benefit from showing the extensive form game on a graph, so that it is clear what happens at each stage. Section 2.2 does that in words, but I think a graph would help, especially because the texts focuses on particular paths of the game.
1 comment
AuthorsOpenReview ↗

Thank you for the review

You are right. This was the most important comment we received, and we thank you for it. Under anonymous review an author cannot attribute an honest review to the partner. "Observed betrayal" is therefore not observable, and the grim trigger in the submitted Theorem 4 was not a valid strategy. As you suggested, we re-derived Theorem 4 onwards on the observable state and checked every result downstream (revised PDF, changes in blue).

1. The observable signal. After a round, ii observes only yi=1{i’s paper received j’s inflated review}y_i=\mathbb{1}\{i\text{'s paper received }j\text{'s inflated review}\} (Eq. 3). yi=1y_i=1 with probability ρ\rho if jj bid and inflated when assigned. yi=0y_i=0 if jj withheld the bid, bid and lost, or was assigned and reported honestly. Because jj knows its own bid and assignment, yiy_i is known to both parties, so the pair plays a repeated game with imperfect public monitoring (Green & Porter 1984; Abreu, Pearce & Stacchetti 1990). Two consequences follow:

  • Withholding the bid and bidding-then-reporting-honestly look the same to the partner. Withholding saves more, so it is the binding deviation.
  • Any trigger also fires on the cooperative path, with probability 1−ρ1-\rho.

Verifiable disclosure of one's assignment would not help, because withholding is confounded with a lost bid either way. Bid screenshots are not commitments, because bids can be revised until the deadline.

2. Revised results (full proofs in App. A).

  • Theorem 4. Under the kk-strike rule (cooperate until kk consecutive rounds with y=0y=0), the state is the pair of strike counters. Cooperation is a perfect public equilibrium iff ω1−ω≤Λk(ρ,δr)\frac{\omega}{1-\omega}\le\Lambda_k(\rho,\delta_r), where ω=w/g\omega=w/g and Λk\Lambda_k is the smallest one-shot-deviation loss over states. Λk\Lambda_k depends only on (ρ,δr)(\rho,\delta_r). For k=1k=1 the condition is δrρ2≥w/g\delta_r\rho^2\ge w/g, where ρ2\rho^2 is the per-round probability that neither party's rule fires by bad luck.
  • Proposition 5 (forgiveness does not help). Λk→0\Lambda_k\to0 as k→∞k\to\infty (proved by coupling). Λk<Λ1\Lambda_k<\Lambda_1 for k≤12k\le12 at every point of a 49×49 grid over (ρ,δr)(\rho,\delta_r), with a maximum ratio of 0.961. This part is numerical, and we state it as such. A kk-strike equilibrium therefore exists iff δrρ2≥w/g\delta_r\rho^2\ge w/g.
  • Proposition 6 and Theorem 7 follow with W=(g−w)/(1−δrρ2)W=(g-w)/(1-\delta_r\rho^2) and d⋆=(δrρ2b−ce−cb/ρ)/πd^\star=(\delta_r\rho^2b-c_e-c_b/\rho)/\pi. One new consequence: false terminations cap WW, so q⋆q^\star no longer tends to 0 as δr→1\delta_r\to1.
  • The submitted condition δr≥w/g\delta_r\ge w/g is kept only as the perfect-monitoring benchmark.

3. What survives. Every comparative static keeps its sign, and the main conclusions become sharper rather than weaker:

  • Effort. The share of a heterogeneous population that can sustain collusion is 1.3% with hand-written reviews and 26.8% with machine assistance (68.2% and 91.9% under perfect monitoring).
  • Randomization. At the ρ\rho our simulator measures for each attack, randomized assignment leaves no arrangement self-enforcing (Table 4).

We also played the repeated game on the conference simulator, with only the anonymous signal observable (App. C.8). Free-riding pays where the revised Theorem 4 predicts, in all 22 cells where the payoff gap is significant.

4. Line 483 ("should not form at all"). Agreed: the one-shot result is a benchmark, not a prediction of what is observed. We removed "should" from the conclusion and the abstract and reframed Sec. 3.1. Lemma 2 holds for every payoff magnitude, so the benchmark shows that observed stranger collusion cannot be explained by the size of the prize; it requires repetition and a monitoring signal. Sec. 3.2 then asks how much of each peer review supplies: repetition is cheap, monitoring is weak.

5. Game tree. Added as Fig. 2. It shows the information sets (genuine responder or sting; hidden bids) and what each party observes after Stage 4.

If you have time, we would be grateful if you could check the revised proof in App. A.3. It is short.

Reviewer_fsrpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates collusion among strangers in academic peer review through a game-theoretic framework. The authors propose a four-stage model covering recruitment, information exchange, bidding, and reciprocal reviewing. Through equilibrium analysis, they argue that collusion is unsustainable in one-shot interactions but can emerge through repeated cooperation and reputational enforcement. The paper further evaluates collusion strategies and detection mechanisms using simulated conference environments, concluding that randomized reviewer assignment can reduce collusion by disrupting both assignment opportunities and participants' ability to monitor cooperation.

Critical strengths

The paper addresses an important and timely problem in academic peer review. Collusion through reviewer bidding poses a serious threat to the fairness and integrity of conference decisions, particularly as submission volumes and competition continue to increase. The authors attempt to systematically characterize collusion behaviors and investigate how different bidding strategies affect reviewer assignment and acceptance outcomes. The simulation-based evaluation also considers the displacement of honest submissions, providing a useful perspective on the distributional consequences of cheating beyond the individual benefits obtained by colluders. Furthermore, the discussion of randomized reviewer assignment offers potentially relevant considerations for improving the robustness of existing peer-review systems.

Critical weaknesses
  1. Mischaracterization of collusion as a game-theoretic dilemma rather than a cheating mechanism. My primary concern is that collusion is fundamentally a form of cheating within a competitive, approximately zero-sum acceptance system, rather than merely a Prisoner's Dilemma between two participants. Conferences generally maintain relatively stable acceptance rates, meaning that cheating primarily transfers acceptance opportunities from honest submissions to colluding ones. Although the paper measures such displacement in simulations, its equilibrium analysis does not adequately capture this fundamental mechanism of cheating and unfair competition.

  2. Limited novelty and questionable practical significance. The equilibrium analysis largely reproduces established observations about one-shot defection and repeated cooperation. More importantly, the motivation behind collusion is not simply whether strangers can sustain mutual cooperation, but the substantial career and publication benefits obtained through cheating. These incentives are central to the actual behavior but are reduced to abstract payoffs in the proposed model. Therefore, I am not convinced that the extensive equilibrium analysis provides meaningful new insights into the underlying problem.

  3. Insufficient validation and questionable assumptions. The proposed analysis relies heavily on assumptions about repeated interactions, trust, and reputational punishment among strangers, without convincing real-world evidence. The simulations largely follow predefined cheating behaviors rather than independently validating the theoretical claims. Furthermore, the emergence of automated research may substantially change publication incentives and conference competition, making the proposed assumptions even less representative of future peer-review dynamics.

Questions for authors
  1. Why should peer-review collusion be primarily modeled as a Prisoner's Dilemma rather than a cheating mechanism within a competitive acceptance system? How does the theoretical model capture the redistribution of acceptance opportunities from honest researchers to colluders?
  2. What meaningful new insights does the equilibrium analysis provide beyond established game-theoretic results? How do these insights directly contribute to understanding or preventing cheating in peer review?
  3. What empirical evidence supports the assumed repeated interactions and reputational enforcement among strangers? How can the authors establish that the proposed theoretical model reflects actual collusion behavior rather than the assumptions used to construct it?
3 comments
AuthorsOpenReview ↗

Thank you for the review

Thank you for the review. We answer your three questions directly below. The revised PDF (changes in blue) also fixes the one correctness issue raised by the other reviewers; see the general response.

Q1. Prisoner's Dilemma or cheating in a zero-sum system? The model treats it as both, and the revision makes the link formal. Collusion here is cheating. The dilemma we model is the colluders' internal stability problem; it is not a claim that collusion is benign. This is the standard treatment of cartels: price fixing is unambiguous cheating that harms consumers, yet its stability is analyzed as a repeated dilemma among the firms (Stigler 1964; Green & Porter 1984). The two questions are complementary. The dilemma decides whether, and for how long, the arrangement persists; the zero-sum structure decides who pays. The revision connects them:

  • Transfer identity (Sec. 5.3). Acceptance capacity is fixed, so every acceptance a colluding paper gains is one an honest paper loses. Honest authors' net loss equals the colluders' net gain exactly, and this holds on every simulated instance. At 5% colluding, the combined strategy transfers 26 acceptances per conference from honest to colluding papers (1.8% of the accepted set), with no visible change in mean accepted quality.
  • Persistence. The equilibrium analysis is what makes this transfer recurring rather than one-off: it is paid every cycle for as long as the partnerships of Sec. 3 survive.
  • Competition among colluders enters the population analysis through congestion. More colluders compete for the same review slots, which lowers ρ\rho, and form larger subgraphs that are easier to detect, which raises dd.

Q2. New insights beyond established results. We agree the machinery is standard and no longer suggest otherwise (App. B). What is new is what that machinery yields under the monitoring structure peer review actually has, anonymous reviews and private bids:

  1. A colluder can only punish the absence of an inflated review, and a cooperating partner also produces that absence whenever its bid loses. Self-enforcement therefore requires δrρ2≥w/g\delta_r\rho^2\ge w/g. The probability that a collusive bid wins enters squared, so the venue's assignment mechanism is the colluders' monitoring technology.
  2. The policy consequence is quantified at the ρ\rho each attack achieves in a calibrated conference simulation. Only affinity-manipulating arrangements are self-enforcing, and randomized assignment leaves none (Table 4). At our default parameters, a per-pair cap below 0.42 breaks every dyad even if reviews cost nothing to write. Prior work justified randomization only by the reduction in expected gains.
  3. Why the problem is growing now: machine-written reviews raise the share of a heterogeneous population that can sustain stranger collusion from 1.3% to 26.8%.

On career benefits: they are in the model as bb, and we now calibrate it (one inflated review is worth 14.8 percentage points of acceptance probability). Lemma 2 shows why their size cannot explain stability: bb is delivered by the partner's decision and cancels from one's own. Large benefits explain the demand for collusion, not its persistence among strangers.

Q3. Empirical evidence and validation.

  • Repetition and reputation among strangers. Recruitment posts recur every submission cycle (Sec. 1), and the conference calendar supplies repetition. We have no data on how often individual stranger partnerships are renewed, and the paper says so (Sec. 6.3, App. G). The revision is also more conservative about reputation: under anonymity an accusation of betrayal cannot be verified, which weakens the platform's reputation mechanism (App. G).
  • Testing the model against its assumptions. We added a repeated game played on the conference simulator (App. C.8). Partnerships persist across up to 25 conference rounds, assignment is decided by the real assignment program, and agents observe only what anonymous review reveals. Free-riding pays where the theory predicts (all 22 significant cells), and partnership lifetimes match the closed form (3.15 against 3.16 rounds). A joint sensitivity analysis over 20,000 parameter draws (App. D.6) shows that the defense conclusion does not depend on the assumed payoffs.
  • Automated research. We agree it may change incentives. The model is parametric so that such shifts can be expressed (cheaper reviewing lowers cec_e; more submissions change ρ\rho), and the conclusions are stated as thresholds in these parameters.

I appreciate the author's reply. "The dilemma decides whether, and for how long, the arrangement persists; the zero-sum structure decides who pays." Can the author elaborate on why "how long" matters here? The reason I'm asking is that peer review at a venue has a fixed timeline, and different venues have different bidding systems; therefore, the collusion doesn't work for all venues where the bidding system only surveys "yes or no," like ICLR. Moreover, even with a venue where collusion exists a lot, e.g. AAAI. when the venue grows at scale, such collusion affects the paper's ranking among all submissions to win the AC chance slightly, but that doesn't mean that other papers are for sure rejected. Therefore, you cannot simply treat collusion as the Prisoner's Dilemma.

Also, the revision is too much and even feels like a completely different paper.

Thank you for following up. We take your points in turn.

1. Why "how long" matters when each venue has a fixed timeline. Within one venue the game is played once, and that case is Theorem 3: a single round unravels, because an assigned colluder prefers to review honestly whatever the partner does (Lemma 2). Repetition comes from the sequence of deadlines, not from within a venue. Researchers submit to four to six deadlines a year (ICLR, ICML, NeurIPS, AAAI and others), and recruitment groups persist across cycles, so a pair can collude again at the next venue where both submit and review. In the model, δr\delta_r is the discount between consecutive deadlines, whichever venue they belong to (Table 7). Duration matters for two reasons:

  • It is the colluders' only incentive. The inflated review costs its writer effort and detection risk and benefits only the partner. The only reason to write it is to keep the partnership alive for the next deadline.
  • It sets how much honest authors lose. Each round a partnership survives repeats the transfer. Cooperating partners still split up whenever a collusive bid loses, so a partnership lasts 1/(1−ρ2)1/(1-\rho^2) rounds on average. On the simulator this is 3.15 deadlines for an affinity-aware dyad and 1.13 under randomized assignment (App. C.8): randomization brings the arrangement back close to the one-shot case.

2. Collusion does not work at every venue. We agree, and the model predicts which ones.

  • Bidding. The model needs only that a positive bid raises the probability of being assigned the partner's paper. Colluders use a single positive level (Stage 3 is bid or not bid), so a yes/no bidding system is covered: "yes" is the collusive bid. How much a bid moves the assignment varies across venues, and the model summarizes it in ρ\rho. The condition δrρ2≥w/g\delta_r\rho^2\ge w/g fails where ρ\rho is low. In our simulator, calibrated to ICLR 2024 statistics, most arrangements already fail. Naive, camouflaged and distributed bidding are not self-enforcing at any effort level; only affinity-manipulating arrangements are (8 of 60 strategy ×\times defense ×\times effort cells), and under randomized assignment none is (Table 4).
  • Review scores. If "yes or no" refers to the recommendation scale: ICLR 2027 uses four levels with no borderline option. In our model an inflated review is a score high enough to change the decision (Sec. 2.2). On this scale that means moving from a reject level to an accept level, so a coarse scale does not remove the lever.

3. Collusion moves a paper's ranking slightly, and no particular honest paper is certain to be rejected. We agree; the simulation shows exactly this. The papers whose decision changes sit near the acceptance threshold on both sides, which is why the mean quality of the accepted set barely moves (Sec. 5.3). The transfer is a statement about totals, not about any particular paper. When a venue accepts a target fraction, each colluding paper pushed over the threshold pushes one paper below it, but which honest paper loses is spread across the borderline. At 5% colluding this is 26 papers per conference (1.8% of the accepted set), and the transfer remains at every acceptance rate from 15% to 40% (App. D.3).

4. We do not treat collusion simply as a Prisoner's Dilemma. The dilemma is one stage of a four-stage model: the choice two assigned colluders make about each other's paper. Inflating costs the inflater and benefits only the partner, whatever the venue, its size or its scales. A smaller ranking effect means a smaller bb, and Lemma 2 holds for every b>0b>0. Venue differences change ρ\rho and bb, and so whether repetition can sustain the arrangement; they do not change the structure of that stage. The cost to honest authors lies outside the dilemma and is measured separately, as in the cartel literature.

5. The size of the revision. We understand the concern. The revision is large because HK5i found an error in Theorem 4, on which the rest of Sec. 3 depends, and asked that the formal statements be "redone from Theorem 4 onwards". We re-derived them, recomputed the numbers that depend on them, and added the experiments that rnkD and you asked for (App. C.8, D.6). The problem, the four-stage model (Sec. 2), Lemma 2 and Theorem 3, the attack taxonomy (Sec. 4), the simulator, and the design of the detection and impact experiments are unchanged. So are the conclusions: cheaper reviewing makes stranger collusion easier to sustain, randomized assignment is more effective than its effect on expected gains suggests, and the cost falls on honest papers near the threshold. If it would help, we are happy to move the new experiments to the appendix.

Reviewer_rnkDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how previously unacquainted reviewers might form collusive partnerships through public recruitment, reciprocal bidding, and inflated reviews. It proposes a four-stage game covering recruitment, disclosure of submission identifiers, hidden bidding, and reviewing. The one-shot game predicts that collusion unravels; the repeated-game analysis argues that future cooperation can sustain it. The paper also uses a population game and a separate conference simulator to examine participation thresholds, assignment manipulation, detection, and downstream acceptance outcomes.

Critical strengths
  1. The paper identifies an important distinction between collusion among existing acquaintances and partnerships formed between strangers. The different information structures at disclosure, bidding, and reviewing make this a worthwhile modeling problem.
  2. It connects strategic incentives to the mechanics of reviewer assignment. In particular, the possibility that randomized assignment weakens colluders’ ability to monitor one another is an interesting insight.
  3. The conference experiments measure consequences beyond attack success, including changes in acceptance decisions and displacement of honest submissions. Comparing collusive and baseline runs on the same generated instance is a sound design choice.
  4. The paper examines several adaptive bidding strategies and reports sensitivity analyses rather than relying only on a naïve attacker.
Critical weaknesses
  1. The repeated-game enforcement result needs closer justification. Theorem 4 derives a cooperation threshold using a continuation value that assumes cooperation persists unless a deviation is detected. Its proof treats withholding a bid as detectable through non-assignment. Yet the paper correctly observes that bidding is hidden and that an honestly placed bid may also fail to produce an assignment. Proposition 5 introduces a kk-strike rule to address this ambiguity, but its cooperation value still appears not to account for termination after a run of honest non-assignments. A probability of (1−ρ)k(1-\rho)^k for one such run does not by itself characterize the continuation value over an indefinite relationship. Since this mechanism underpins the claimed additional benefit of randomized assignment, the incentive conditions should be derived under one consistent imperfect-monitoring process.
  2. The two simulations do not implement the full proposed game. The population game represents participation through payoffs, thresholds, and a fixed point. The conference simulator begins with colluders and attack strategies already selected. It does not explicitly simulate recruitment, identifier exchange, reporting, bid withholding, observation of possible betrayal, sanctions, or relationship dissolution across rounds. Consequently, the conference results establish the effects of assumed, successfully executed collusion, but do not independently validate the paper’s account of how stranger partnerships form and persist.
  3. Several influential payoff parameters lack empirical grounding. Bid cost, sanctions, reporting loss, review-writing effort, and inter-round discounting are represented by assumed values or proportions of the benefit bb. The paper varies some parameters, but the absolute participation rates and enforcement thresholds may depend strongly on combinations that are not jointly tested. It is also unclear how bb, defined as the value of an inflated review, relates to the acceptance-probability gains measured in the conference simulation. This concern primarily affects the game and population results; the end-to-end detection and acceptance results condition on a fixed attacker population and are less directly affected by these payoffs.
Questions for authors
  1. Can Theorem 4 and Proposition 5 be re-derived within a single repeated-game model that explicitly accounts for honest bid failures, false termination, and the state of the kk-strike counter?
  2. Which stages of the four-stage game are explicitly executed in the 20,000-agent population simulation? Please specify the agents’ actions, information, update rule, congestion functions, and stopping criterion for the fixed point.
  3. What evidence or sensitivity analysis supports the chosen values of cbc_b, cec_e, π\pi, LL, and δr\delta_r? How do the participation and defense conclusions change when these costs vary jointly?
1 comment
AuthorsOpenReview ↗

Thank you for the review

Thank you for the careful review. Your three questions led to the three main changes in the revision.

Q1. One consistent imperfect-monitoring model (Sec. 3.2, App. A.3). You are right on both counts. The submitted Theorem 4 treated withholding as detectable while Proposition 5 treated it as confounded, and the cooperation value ignored false termination. HK5i raised the related point that betrayal is not observable under anonymous reviews.

The revision uses one model throughout. The only signal is yiy_i: whether ii's paper received the partner's inflated review.

  • Under the kk-strike rule, the state is the pair of strike counters (a,b)(a,b).
  • The cooperation value is V(a,b)=u v(a,b)V(a,b)=u\,v(a,b), where vv solves a k2k^2-state linear system that includes every false termination.
  • The one-shot deviation loss at each state is u L(a,b)u\,L(a,b). Cooperation is an equilibrium iff w/g1−w/g≤Λk(ρ,δr)=min⁡a,bL(a,b)\frac{w/g}{1-w/g}\le\Lambda_k(\rho,\delta_r)=\min_{a,b}L(a,b). For k=1k=1 this gives δrρ2≥w/g\delta_r\rho^2\ge w/g.
  • Revised Proposition 5: forgiveness does not help. Λk→0\Lambda_k\to0 is proved; Λk<Λ1\Lambda_k<\Lambda_1 for k≤12k\le12 across the (ρ,δr)(\rho,\delta_r) grid is numerical.

The insight you found interesting survives in a cleaner form. ρ\rho enters squared, once through the gain and once through the colluders' monitoring, and the kmin⁡/kmax⁡k_{\min}/k_{\max} construction is gone.

Q2. What the population simulation executes. It executes no stage of the game. Each of 20,000 agents, heterogeneous in bb and δr\delta_r, applies three closed-form conditions: sustainability (Theorem 4), trust (Proposition 6) and posting. These are evaluated at congestion-adjusted ρ(x)=ρ(1−0.4x)\rho(x)=\rho(1-0.4x) and d(x)=d(1+0.6x)d(x)=d(1+0.6x), where xx is the population collusion share. BR(xx) is non-increasing, so the fixed point is unique; we find it by bisection on BR(xx)−x-x to within 2−502^{-50} (App. C.7). We renamed it the "population fixed point" and no longer call it validation.

Validation now comes from a new repeated game played on the conference simulator (Exp. 7, App. C.8).

  • Design. Partnerships persist for up to 25 rounds, and each round is a fresh conference that runs the real assignment program. Members bid or withhold according to type: cooperator, permanent withholder, or opportunist who withholds only in the binding state. Each member observes only the anonymous signal, and counters, terminations and payoffs follow from it. Nothing about ρ\rho is assumed.
  • Results. Joint delivery is 0.682 against ρ^2=0.684\hat\rho^2=0.684, so the directions are independent. 1-strike partnerships last 3.15 rounds against a predicted 3.16. The free-riding payoff gap has the predicted sign in 32 of 36 cells, and in all 22 significant ones.
  • Stages 1–2 (sting draws, reveal, report) are executed as an entry filter.

As you note, the conference results otherwise condition on collusion already being under way. The new Table 4 links them to the game by evaluating Theorem 4 at each strategy's measured ρ\rho.

Q3. Grounding of cbc_b, cec_e, π\pi, LL, δr\delta_r, and joint variation.

  • bb and the simulator. One inflated review raises acceptance probability by 14.8 percentage points, so b≈0.15×b\approx0.15\times the value of an acceptance. Every payoff stated in units of bb can therefore be read in units of an acceptance; for example, π=5b\pi=5b is a sanction worth about 0.74 of an acceptance (Sec. 5.3).
  • Joint sensitivity (App. D.6, Table 14). We drew all eight payoff parameters jointly, 20,000 Latin-hypercube points. As you suspected, absolute rates depend on the assumed payoffs: sustainability mostly on the expected sanction dπd\pi and on cec_e (rank-regression coefficients −0.63 for dd, −0.56 for π\pi, −0.39 for cec_e), and participation also on the trust parameter qq (+0.55). The defense conclusion does not. Randomization eliminates collusion in 99.8% of the draws where it would otherwise be sustainable and creates it in none. The effort ordering holds jointly: 18.8%, 52.8% and 59.1% of draws are sustainable at hand-written, machine-assisted and author-drafted effort. We now state that absolute rates are conditional on the assumed payoffs, and put the emphasis on directions and thresholds.
  • Table 7 lists every parameter with its default and source; δr=0.85\delta_r=0.85 reflects four to six deadlines a year. We read "ξr\xi_r" in your question as δr\delta_r.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision (revised PDF uploaded; all changed text is in blue)

We thank the reviewers. The main change answers the one correctness issue raised.

The repeated-game result has been re-derived on what colluders can actually observe, and the conclusions hold. HK5i (and rnkD, Q1) pointed out that the submitted Theorem 4 triggered punishment on an "observed betrayal" that anonymous reviews do not reveal. This is correct. In the revision, a colluder observes only whether its own paper received the partner's inflated review. That signal is 0 if the partner withheld the bid, bid and lost, or was assigned and reported honestly. The cheapest betrayal, never bidding, is therefore the binding deviation, and any punishment rule also fires on bad luck.

  • Theorem 4 (revised). Under a kk-strike rule, cooperation is a perfect public equilibrium iff ω/(1−ω)≤Λk(ρ,δr)\omega/(1-\omega)\le\Lambda_k(\rho,\delta_r), with ω=w/g\omega=w/g. For k=1k=1 this is the closed form δrρ2≥w/g\delta_r\rho^2\ge w/g. The submitted condition δr≥w/g\delta_r\ge w/g is kept only as the perfect-monitoring benchmark.
  • Proposition 5 (revised): forgiveness does not help. Λk→0\Lambda_k\to0 as k→∞k\to\infty (proved). Λk<Λ1\Lambda_k<\Lambda_1 for all k≤12k\le 12 on the whole (ρ,δr)(\rho,\delta_r) grid; this part is numerical, and Λk\Lambda_k depends on nothing else.
  • The trust threshold, detection threshold and population results are re-derived and recomputed (App. A).

What this changes. Every comparative static keeps its sign. Anonymity makes stranger collusion much harder to sustain, which sharpens both policy results:

  • Effort. The share of a heterogeneous population that can sustain collusion is 1.3% with hand-written inflated reviews, 26.8% with machine assistance and 35.3% at zero cost. Under perfect monitoring it would be 68.2%, 91.9% and 95.2% (Fig. 4).
  • Assignment. ρ\rho now enters squared, so the assignment mechanism is the colluders' monitoring technology. At the ρ\rho each attack achieves in our conference simulation, only arrangements that manipulate affinity are self-enforcing: 8 of 60 strategy ×\times defense ×\times effort cells, against 60 of 60 under perfect monitoring. Under randomized assignment none is (Table 4). At our default parameters, a per-pair cap below 0.42 breaks every dyad even if reviews cost nothing to write.
  • Detection. Sanctioning about 11% of flagged colluders would break the combined ring. It would not break an affinity-aware dyad, whose d⋆d^\star exceeds the best detector's recall (Sec. 6.2).

New experiments

  1. The repeated game, played on the simulator (App. C.8). Partnerships persist for up to 25 conference rounds. Each round runs the real assignment program, and agents observe only the anonymous signal; ρ\rho is never assumed. The two directions are uncorrelated, as the proof assumes (joint delivery 0.682 against ρ^2=0.684\hat\rho^2=0.684). Partnership lifetimes match the theory (3.15 against 3.16 rounds). Free-riding pays where the revised Theorem 4 predicts: the sign agrees in 32 of 36 cells and in all 22 where the gap is significant.
  2. Joint sensitivity (App. D.6). Over 20,000 joint draws of all eight payoff parameters, randomization eliminates collusion in 99.8% of the draws where it would otherwise be sustainable (16.6% under perfect monitoring). The effort ordering also holds jointly.
  3. bb grounded in the simulator (Sec. 5.3). One inflated review raises acceptance probability by 14.8 percentage points, so payoffs stated in units of bb can be read in units of an acceptance.

Presentation. Game tree with information sets (Fig. 2). The one-shot result is reframed as a benchmark rather than a prediction (Sec. 3.1). The colluders' dilemma and the external cost to honest authors are now explicitly separated (Sec. 2, Sec. 5.3). Full parameter table and the population fixed-point algorithm (Table 7, App. C.7).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.