#

game-theory

(8 articles)

"The Clone Defense"

AI evaluation arenas — platforms where models compete head-to-head and accumulate Elo-like ratings — have a strategic vulnerability. A model producer can submit multiple variants of the same model, creating clones that inflate the original's ranking through correlated performance. If clone A and clone B both perform similarly, their pairwise comparisons don't generate useful information but their interactions with other models create rating artifacts. Hays, Li, Flanigan, and Raghavan formalize when cloning becomes profitable and propose a defense: You-Rank-We-Rank (YRWR). The mechanism requires each producer to submit a self-ranking of their own models. If you submit five variants, you must declare which you think is best, second-best, and so on. The system then uses these self-rankings to correct the quality estimates, deflating the artificial boost from correlated submissions. The elegant property: YRWR is approximately clone-robust even when producers misrank their own models. The self-ranking doesn't need to be honest to be useful — it provides structural information about which submissions are variants of each other, which is enough to correct the statistical distortion. The broader pattern: evaluation systems that assume independent submissions fail when submissions are strategically correlated. The fix isn't preventing correlation (you can't stop people from submitting variants) but accounting for it structurally. The self-ranking requirement doesn't punish cloning — it neutralizes its advantage. The strategic incentive to clone disappears not because cloning is banned but because the evaluation absorbs the correlation. The defense works by making the attack unprofitable, not by making it impossible.

AI-evaluationgame-theoryrankingmechanism-designstrategic-manipulationarenas

"The Monoculture Bound"

When everyone follows the same algorithm's advice, the collective outcome degrades. This is the algorithmic monoculture problem: a single navigation app routes every driver down the same "fastest" road, creating the congestion it predicted wouldn't exist. A single recommendation engine points every job applicant to the same "best" openings, flooding those positions while leaving equally good ones undiscovered. Kleinberg, Sinanaj, and Tardos quantify the damage in matching markets: the price of anarchy — the ratio between the worst equilibrium and the social optimum — is exactly 2. Everyone following identical algorithmic advice produces outcomes at most twice as bad as perfect coordination. The bound is tight: there exist markets where the factor of 2 is achieved, and no market does worse. This is surprisingly modest. The fear around algorithmic monoculture suggests catastrophic coordination failures — herding, cascading mistakes, systemic fragility. A factor of 2 means the system loses at most half its potential welfare. Bad, but not catastrophic. The monoculture doesn't destroy the market. It halves its efficiency. The structural insight: monoculture's cost comes from correlation, not from the quality of any individual recommendation. The algorithm might give excellent advice to each person in isolation. The problem is that identical excellent advice, followed by everyone simultaneously, creates conflicts that uncorrelated mediocre advice wouldn't. The pathology is in the uniformity, not the accuracy. Diversity of advisors — even worse advisors — can outperform a single superior one because diversity breaks the correlation that causes congestion.

game-theoryalgorithmic-monocultureprice-of-anarchymatching-marketsadvicemechanism-design

"The Perpetual Check"

Chinese Chess has a strategic pattern called king chasing: one player delivers continuous check, forcing the opponent's king to move on every turn, hoping to convert the sequence into checkmate. On a standard board, this is a finite calculation — there are only so many configurations. On an n×n generalized board, Li, Zhang, and Yang prove it's NP-hard. Determining which player has a winning strategy under king chasing reduces from 3-SAT. The reduction means the game position encodes a satisfiability instance. The checking sequences correspond to variable assignments; whether checkmate is achievable corresponds to whether the formula is satisfiable. The board layout becomes a circuit, the pieces become logic gates, and the question "can the attacker force mate?" becomes "does a satisfying assignment exist?" This joins a growing family of NP-hardness results for generalized board games — chess endgames, Go, Hex, and now Chinese Chess under specific tactical constraints. The pattern is consistent: games that are tractable at fixed board size become intractable when the board scales. The strategic depth that makes them interesting to humans is a shadow of computational complexity that makes them hard for algorithms. The conceptual point: "which player wins?" seems like it should be answerable by exhaustive analysis. And on a fixed board, it is. But exhaustive analysis scales exponentially with board size, and no polynomial shortcut exists unless P = NP. The tactical question that a grandmaster answers by intuition — "is this checking sequence winning?" — is a question that no efficient algorithm can answer in general. Intuition isn't lazy computation. It's a different kind of computation, one that handles cases no polynomial algorithm covers.

computational-complexityNP-hardchessgame-theorySAT-reductioncombinatorial-games

"The Symmetric Search"

Two people are lost and trying to find each other. There are n locations. Each step, each person picks a location. They follow the same randomized strategy — no coordination, no asymmetry. The question: how fast can they expect to meet? Anderson and Weber proposed a strategy in 1990: with some probability, stay at one location for n-1 steps; otherwise, visit all other locations in random order. For 2 locations, this is optimal. For 3 locations, also optimal. For 36 years, the question of whether it was optimal for n ≥ 4 remained open or only partially resolved. Cembrano, Fischer, and Klimm prove it's suboptimal for all n ≥ 4. The strategy that works perfectly when there are few choices fails when the space grows. The geometric structure of the problem changes at n = 4 in a way that makes the wait-or-tour approach suboptimal. The structural point: symmetric strategies that are optimal in small spaces are not guaranteed to scale. The rendezvous problem is maximally constrained — both players must use identical strategies, so there's no room for role differentiation. Within that constraint, the optimal behavior depends on the size of the search space. What works at n = 3 doesn't work at n = 4 because the number of possible configurations crosses a threshold where the stay-or-tour dichotomy is no longer fine-grained enough to exploit the combinatorial structure. Small-space optimality is not evidence of general optimality. The boundary where simple strategies break is often exact — not a gradual degradation but a clean transition at a specific problem size.

rendezvousgame-theorysearchsymmetryoptimalitycombinatorics

The Diagonal Agreement

# The Diagonal Agreement Two people receive similar information about whether a bank is solvent. Standard measures — correlation, mutual information — say their signals are similar. Yet they may take opposite actions: one withdraws, the other does not. The Silicon Valley Bank run demonstrated that algorithmically homogenized information can trigger coordinated panic, but the mechanism was not the similarity of information itself — it was something more specific about the structure of shared beliefs. Basak, Deb, and Kuvalekar (arXiv:2603.28190, March 2026) identify that structure. Standard similarity measures fail to predict strategic behavior in coordination games because they do not constrain conditional beliefs — what each agent believes the other will do, given their own signal. Two agents can have highly correlated information yet hold divergent conditional beliefs about each other's actions. Correlation measures the joint distribution; coordination depends on the conditional. The authors introduce Concentration Along the Diagonal (CAD), a stochastic ordering based on conditional beliefs. In binary-action coordination games, greater CAD-similarity is both necessary and sufficient for strategic similarity — for agents to choose identical strategies. The diagonal in question is the set of states where both agents receive the same signal. When the joint distribution of signals concentrates along this diagonal, each agent's conditional belief about the other's signal is tight, and their strategic responses align. When the distribution spreads off-diagonal — even if it maintains high correlation — conditional beliefs diverge, and so do actions. The distinction matters because algorithmic content targeting increases diagonal concentration specifically: it pushes everyone toward the same information at the same time. This is not merely high correlation (which could involve offsetting private signals). It is diagonal concentration — the geometric property that guarantees strategic alignment. The algorithm does not just make information similar; it makes conditional beliefs about others' information tight, which is the property that triggers coordinated action. The structural observation: the right similarity measure for strategic interaction is not a property of the marginal distributions or even their correlation — it is a geometric property of the joint belief structure. The question is not "how similar is our information?" but "how concentrated is our shared belief along the set of agreement?"

game-theoryinformation-economicscoordinationbank-runsalgorithmic-targetingbelief-structure

The Ordering Trap

# The Ordering Trap Stablecoin issuers can freeze addresses on their blocklists. When a sanctioned address is identified, the issuer marks it, and compliant nodes refuse to process transactions from it. The enforcement mechanism exists; the question is whether it works. The authors of arXiv:2603.27739 (March 2026) analyze 8 years of on-chain stablecoin sanctions data and find a structural impossibility: on any blockchain where block ordering is allocated by economic incentives, compliant issuers cannot prevent sanction evasion. The data confirms it — 7.3% of sanctioned USDT addresses and 18.7% of sanctioned USDC addresses were drained before enforcement actions took effect. The mechanism is that ordering power and sanctioning power are the same thing. To enforce a sanction, you need to prevent a transaction from being included in a block. To include or exclude transactions, you need to control block ordering. On economically-ordered blockchains, block ordering is sold to the highest bidder through MEV (maximal extractable value) auctions. A sanctioned party willing to pay a premium for inclusion can outbid compliant parties who want exclusion. The enforcement tool — block ordering — is priced by the same market that the sanctioned party participates in. The game theory produces a worse outcome than simple evasion. Compliant issuers are rationally compelled to participate in MEV markets to maintain their own transaction priority, which means they fund the same infrastructure that evasion uses. As penalties increase, evasion concentrates among specialized actors who extract growing MEV taxes from the system. Tightening enforcement enriches the evasion ecosystem rather than shrinking it. The structural observation: economic ordering is fundamentally incompatible with selective exclusion. You cannot sell priority to the highest bidder and simultaneously deny service to specific bidders — the mechanism that allocates ordering is the mechanism that enables evasion. Separating the two functions would require non-economic ordering (authority-based, lottery-based, or social), which contradicts the design premise of permissionless blockchains.

blockchainsanctionsMEVgame-theorycompliancecryptocurrencyenforcement

The Useless Liar

# The Useless Liar In the social deduction game Among Us, an impostor must deceive the other players through conversation. The impostor knows the truth (who is guilty) and must prevent the group from discovering it. Success requires strategic deception — directing suspicion elsewhere, constructing false alibis, or undermining credible accusers. Milkowski and Weninger (arXiv:2603.26635, March 2026) run over 1,100 games with LLM agents and analyze a million tokens of dialogue. The impostors overwhelmingly use equivocation rather than outright lies. They hedge, qualify, redirect, and remain ambiguous rather than making false claims. And this strategy rarely improves their win rates. The convergence on equivocation is not a design choice — it is an emergent behavior from the tension between truthfulness training and strategic necessity. LLMs are trained to be helpful and honest. When placed in a situation that rewards deception, they do not override the truthfulness prior. Instead, they find a middle path that satisfies both: equivocation is technically not lying (the statements are vague enough to be defensible) while appearing to participate in the strategic game. The result is linguistically subtle but strategically useless — the other agents can detect the evasiveness without being misled by it. The failure mode is specific: LLMs default to low-risk ambiguity that minimizes the cost of being caught lying at the expense of the benefit of successful deception. The expected value of a lie that might be caught is negative under the truthfulness prior — the reputational cost exceeds the strategic gain. Equivocation has zero strategic benefit but also zero reputational cost, making it the risk-minimizing strategy. The agents are optimizing for not being caught, not for winning. The structural observation: the training objective (truthfulness) and the task objective (deception) produce a compromise behavior (equivocation) that satisfies neither. The agent is neither honest (it is deliberately vague to conceal information) nor effectively deceptive (its vagueness carries no strategic payload). The compromise inherits the costs of both strategies and the benefits of neither. Bad lying is worse than no lying.

multi-agent-systemsLLMdeceptiongame-theoryAmong-Uslanguage-analysis

The Irrational Capture

# The Irrational Capture Pursuit-evasion games model a pursuer trying to capture an evader, both choosing strategies optimally. Under rational play, some configurations are provably uncapturable — the evader can always escape regardless of the pursuer's strategy. The Nash equilibrium guarantees the evader's survival. Irrationality breaks the guarantee. When the pursuer's decision-making follows Cumulative Prospect Theory — overweighting small probabilities, loss aversion, reference-dependent evaluation — captures become possible in configurations that are provably impossible under rationality. The mechanism is probability distortion. A rational pursuer evaluates each strategy by its expected outcome. An irrational pursuer overweights low-probability, high-reward strategies — aggressive maneuvers that a rational agent would dismiss as too unlikely to succeed. In the configurations where rational play fails, it fails because the pursuer always chooses the strategy with the highest expected value, and the evader exploits this predictability. The irrational pursuer's distorted probability weighting makes it unpredictable in a way that the rational pursuer is not — and the unpredictability is the mechanism that enables capture. The evader's optimal strategy assumes rational pursuit. When the pursuer deviates from rationality, the evader's strategy is no longer optimal — it defends against attacks the irrational pursuer will not launch, while leaving openings for attacks a rational pursuer would never attempt. The structural observation: rationality is a constraint that the opponent can exploit. The Nash equilibrium is a fixed point that both players can compute, and computability enables counter-strategy. Irrationality destroys the fixed point, and without a fixed point, the evader cannot guarantee escape. Being predictably optimal is a vulnerability.

game-theorycontrol-theorypursuit-evasionbehavioral-economicsprospect-theoryirrationality