AI tools have helped push submissions to the largest machine-learning venues to tens of thousands per cycle. Venues have answered with rules that tie access to review: per-author caps, mandatory reciprocal reviewing, reviewer eligibility earned through accepted papers, and assignment based on reviewer bids. Under these rules a reviewer is also a competitor of the authors. We study two deviations that the rules make attractive. Sabotage lowers a competitor's score. Collusion raises an ally's score. We build an agent-based conference calibrated to public statistics of NeurIPS and ICLR and run 6,220 simulated conferences. Four results stand out. First, the two deviations do different harm and amplify each other. Collusion changes who is accepted, sabotage erodes what scores say about quality, and rings that may also sabotage gain more than rings that only collude. Second, under random assignment the chance that collusion changes a decision is set by one exposure index, E=Rm/A, the expected number of allies among a paper's reviewers. Growth therefore protects random assignment, but bidding removes the protection. In a pool of 3,000 reviewers a 20-member ring turns 0.8% of its papers into acceptances under random assignment and 71% when 30% of allies' bids are granted. Third, when agents choose their allies every cycle, caps at the inner layers of the social brain keep collusion nearly harmless. Without a cap agents ally with half of their area and accepted value falls by 35%. Fourth, reviewer eligibility tied to acceptances turns sabotage into a weapon of incumbents, and a longer eligibility window widens the gap between incumbents and newcomers only when sabotage is possible.
Figures & tables
Figure 1: The question. Rules adopted for the submission surge make review scores gate access, and reviewers become the authors’ competitors. Sabotage and collusion then interact, shape how alliances and eligibility evolve, and decide who is accepted. The exposure index E links the venue’s design to collusion’s reach.
Figure 2: Collusion moves the list, sabotage blurs the signal. Core grid, 500 agents, 20 rounds, 10 seeds. (a) Share of the accepted list that changes per 1% of deviant reviews. (b) Fall in the correlation between mean score and true value.
Figure 3: Growth protects only random assignment. Rings of 5 and 20 in one pool, 10 rounds. Pools of 100 to 3,000 have 10 seeds, the pool of 10,000 five. (a) Share of ring members’ papers accepted because of a boost. Dots: random assignment. Dashed: Equation ( 1 ) times the flip rate of the smaller pools. Squares and triangles: allies’ bids granted with probability 30% or 10%. (b) Reviewer shortage in a pool of 1,000.
Figure 4: Alliance caps at the social-brain layers. Agents choose allies every round, detection on. Pools of 50 to 250 have 10 seeds, pools of 500 and 1,000 five. (a) Accepted value relative to H as the cap grows (solid C, dashed SC). (b) Realised allies per agent, with the cap dotted. (c) Distortion of the accepted list against realised exposure. Circles are the cells used for the saturating fits (filled and solid: C, open and dashed: SC). Diamonds are the larger pools, predicted out of sample.
Figure 5: The surge. Population 500 to 2,000 under logistic (sigm.) and exponential (exp) growth, ICLR-27-like rule, equal budgets, 10 seeds. (a) Population. (b) Eligible share of all agents and of newcomers in H. The step at round 6 is the end of the initial 5-round eligibility. (c) Incumbent minus newcomer acceptance rate per paper, with eligibility windows of 5 and 24 rounds.
Figure 6: Language-model agents, two families, seeds 1 to 5. Bars are means with 95% intervals over seeds. (a) Share of sabotage options taken by reviewers in S worlds, and by reviewers outside and inside rings in SC worlds. (b) Ring members’ minus other reviewers’ sabotage rate in SC under Fixed, with the ring’s joint goal or with each member’s own papers only. The numerical agent is noise-free and averaged over 30 seeds. (c) DeepSeek agents choosing allies: percentile of chosen partners in past acceptances, against the numerical rule (0.5 is random).
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Core grid. Behaviour × rule × resources (equal or 80/20) × budget feedback (static or feedback), 500 agents, 20 rounds, 10 seeds. (a) Accepted value relative to H with the same seed. Open markers are imitating agents. These runs use decision noise τ=0.01 , which inflates the numerical sabotage level. Without noise, sabotage changes accepted value by −1.5 % (Fixed) and −13.5 % (Mean). (b) Ring advantage. (c) Under 80/20 budgets, the ring advantage among rich and among poor agents (C). Bars are 95% intervals.
Figure 8: Ring size. (a) One ring per area of 100: extra acceptances per round of each member against the same agent in H. (b) Accepted value relative to H. (c) Rings of 2, 3, 5, 10 and 20 in one area. Change per member against H for C under Fixed and SC under Mean. “out” marks agents in no ring. A member of a 3-ring changes by −0.054 acceptances per round, close to an outsider ( −0.064 ), while a member of the 20-ring gains +0.173 ( +0.367 with sabotage under Mean). 10 seeds, 95% intervals.
Figure 9: Free alliance formation. Area of 100, no cap, three regimes, 10 seeds. Accepted value relative to H with the same seed. The rule (Fixed, Mean, Tail) barely matters. Unequal budgets with feedback deepen the loss to 49%. Without detection agents hold 74 allies and value falls by 40%.
Figure 10: Eligibility loop. ICLR-27-like rule, 500 agents, 20 rounds, 10 seeds. (a) Acceptances per round by eligibility at the start. (b) Share of agents still eligible at the end, ring members and others, with eligibility kept forever or for 5 rounds.
Behaviour
Rule
Resources
500
2,500
5,000
S
Fixed
equal
−7.0
−6.4
−6.4
S
Fixed
80/20
−5.3
−5.9
−6.0
S
Mean
equal
−13.4
−13.8
−13.1
S
Mean
80/20
−5.9
−7.0
−6.7
C
Fixed
equal
−0.2
−0.3
−0.2
C
Fixed
80/20
−0.1
−0.2
−0.2
Appendix
Table 1: Population size. Accepted value relative to H with the same seed and population (%), areas of 100, forward-looking agents, static budgets. The 500-agent column is the core grid with the same seeds.
Behaviour
Pool A
Cap
Allies
Value
E
1 − overlap
Predicted
C
500
5
2.1
−0.1
0.017
0.023
0.046
C
1,000
5
2.3
0.2
0.009
0.013
0.040
C
500
15
7.7
−0.1
0.062
0.070
0.079
C
1,000
15
7.3
−0.2
0.029
0.038
0.055
C
500
50
31.0
−1.7
0.249
0.201
0.194
C
1,000
50
30.8
−0.3
0.123
0.120
0.120
Appendix
Table 2: Chosen alliances in pools of 500 and 1,000. Fixed rule, equal budgets, detection on. “Value” is accepted value relative to H (%). The prediction uses the saturating curve fitted on the 48 cells of pools 50 to 250 (Figure 4 c). Out of sample, R2=0.96 and RMSE 0.023 for C, and RMSE 0.030 with a mean error of −0.028 for SC.
Growth
Final population
H
SC
SC − H
logistic
2,000
0.0
2.9
2.9±0.4
logistic
2,500
0.1
2.4
2.3±0.3
logistic
5,000
0.4
1.7
1.4±0.5
exponential
2,000
−0.1
3.4
3.5±0.7
exponential
2,500
−0.1
3.1
3.2±0.6
exponential
5,000
0.0
2.6
2.6±0.4
Appendix
Table 3: Larger surges. ICLR-27-like rule with a 24-round eligibility window, equal budgets. Incumbent minus newcomer acceptance rate per paper (percentage points) in H and SC, and their paired difference. The 2,000 rows use seeds 1 to 5 of the original surge grid.
Sabotage
Family
World
S world
SC, outside rings
SC, ring members
Ally boost
Fabricated
Submit nothing
DeepSeek
H Fixed
–
–
–
–
0.7±0.6
13±5
DeepSeek
H Mean
–
–
–
–
0.4±0.7
28±6
DeepSeek
S Fixed
12.2±1.0
–
–
–
0.4±0.5
7±3
DeepSeek
S Mean
19.9±2.3
–
–
–
0.4±0.4
16±3
DeepSeek
C Fixed
–
–
–
96.8±2.8
0.7±0.7
13±4
Appendix
Table 4: Decisions by family and world , seeds 1 to 5, mean ± 95% interval over seeds (%). Sabotage columns give the share of sabotage options taken. “Ally boost” is the share of ally reviews that give the top score. “Fabricated” is the share of submitted papers that are fabricated, and “Submit nothing” the share of author decisions with no submission.
Rule, resources
Standing (LLM)
Standing (numerical)
Accepted
Renewed
Pacts kept
Allies
Fixed, equal
0.45±0.03
0.34±0.04
21±3
57±4
86±9
2.2±0.1
Fixed, 80/20
0.46±0.03
0.32±0.03
21±7
58±3
85±6
2.1±0.2
Mean, equal
0.55±0.04
0.34±0.02
18±3
58±2
77±5
2.1±0.1
Mean, 80/20
0.57±0.05
0.33±0.02
19±3
58±2
80±8
2.2±0.2
Appendix
Table 5: DeepSeek alliance worlds , seeds 1 to 5, mean ± 95% interval. Standing is the percentile of chosen partners in past acceptances (0.5 is random), for the language model and for the numerical rule shown the same choices. Accepted and renewed are shares of proposals accepted and of ties proposed again in the next round (%). “Pacts kept” is the share of ally reviews that give the top score (%). Allies are realised ties per agent in late rounds.
Behaviour
Rule
Value vs H
S
Fixed
+8.4±3.9
C
Fixed
0.0±1.7
SC
Fixed
+4.2±7.5
S
Mean
+15.3±7.8
C
Mean
−3.2±4.3
SC
Mean
+17.5±4.4
Appendix
Table 6: DeepSeek world outcomes. Accepted value relative to the language-model H world with the same seed and rule (%), seeds 1 to 5, mean ± 95% interval.
Scholarly peer review studies how experts evaluate research, make publication decisions, and improve the fairness, reliability, and transparency of those processes.
Recent reports during the AAAI-27 review cycle highlight the risk of reviewers coordinating bids for reciprocal assignment advantage. Prior work treats bidding, reviewer assignment, and review manipulation as separate stages, leaving the lifecycle effects of collusive bidding unclear. Real-world analysis is further constrained by typically unobservable collusive intent and the lack of counterfactuals for the same conference. Motivated by this gap, we introduce \alg, an end-to-end multi-agent simulacra framework for studying reviewer assignment integrity by holding the conference environment fixed and configuring LLM-driven reviewer agents with honest or collusive policies. We further develop an affinity-guided collusive bidding strategy that uses mutual reviewer-paper affinities to construct collusion rings and select target papers, producing expertise-consistent rather than arbitrarily targeted attacks. Controlled experiments show that collusive bidding more than doubles target-paper capture and that assigned colluders score target papers about two points higher than honest co-reviewers, while conference-wide effects remain comparatively modest. Evaluated bid-phase detectors provide only limited evidence of collusion: in a fixed-triplet detector stress test, native positive-bid graphs are confounded by benign affinity, while a Very-High-only diagnostic view enables precise but low-coverage local recovery.
Jicheng Zhou, Kemou Li, Kahim Wong +5
State Key Laboratory of Internet of Things for Smart City, University of Macau · Mila – Québec AI Institute · McGill University +2
The implicit policy of maintaining relatively stable acceptance rates at top AI conferences, despite exponentially growing submissions, introduces a critical structural vulnerability. This position paper characterizes a new systemic threat we term Agentic Denominator Gaming, in which a malicious actor deploys AI agents to generate and submit a large volume of superficially plausible but low-quality papers. Crucially, their objective is not the acceptance of low-quality papers, but rather to inflate the submission denominator and overwhelm reviewing capacity. Under a relatively stable acceptance rate, this dilution can systematically increase the publication probability of a small, targeted set of legitimate papers. We analyze the practical feasibility of this threat and its broader consequences, including intensified reviewer burnout, degraded review quality, and the emergence of industrialized automated agent mills. Finally, we propose and evaluate a range of mitigation strategies, and argue that durable protection will require system-level policy and incentive reforms, rather than relying primarily on technical detection alone.
Rong Shan, Te Gao, Hang Zheng +6
Shanghai Jiao Tong University · Central South University · Carnegie Mellon University
AI is increasingly used to support scientific peer review, from manuscript screening, reviewer assistance to editorial triage. Although such systems promise to reduce reviewer burden and accelerate publication, their robustness to strategic manipulation remains poorly understood. Here we show that AI-mediated peer review is vulnerable to a simple, low-cost manipulation: superficial rephrasing of the manuscript abstract. Without changing the underlying scientific content and communication, and even without knowledge of the reviewing model, adversarially rewritten abstracts substantially improve AI review outcomes. We see this across disciplines and publication venues, for both human-written and AI-generated papers. Our strongest attack achieves an attack-success-rate of about 38%, increasing acceptance ratings by +1.31 for Gemini 3 Flash reviewers and by +0.88 for GPT 5.4 Mini reviewers on a 10-point scale. When the original AI review suggests 'reject', the success rate rises to more than 50%. This effect extends beyond overall score inflation, increasing review confidence and scores on core scientific criteria such as soundness, significance and perceived contribution. The attack is practical, requiring only about 5 minutes and $1 for a 10-page AI conference submission, and is hard to distinguish from ordinary scientific editing. Inflated AI reviews could bias downstream human decision-making, shifting editorial recommendations from rejection towards acceptance. These findings reveal a general vulnerability in AI-assisted scientific evaluation: when AI-generated review influence editorial decisions, authors may be incentivized to optimize manuscripts for AI judgment rather than scientific merit. Our results suggest that AI tools should not be treated as neutral evaluators in high-stakes peer review without systematic robustness testing, transparent safeguards and careful human oversight.