Game State Impact Rate Analysis

• Game State Impact Rate is a measure of how changes in game states affect strategic outcomes, defined variably across domains such as AI, market dynamics, evolutionary games, and esports.
• Its applications span human-AI interaction equilibria, market impact in Minority Games, and dynamic evaluations in hidden-role games, employing diverse methodologies.
• Frameworks range from adaptation rates and Markov transitions to state-dependent reinforcement learning and win probability deltas, emphasizing practical evaluation.

In the cited literature, Game State Impact Rate refers to how a game state, or the rate at which the state changes, affects strategic outcomes. The expression is used for the effect of the AI’s adaptation rate $\alpha$ on the joint action $(h_t,m_t)$ in a human-AI continuous game (Isa et al., 2024), for the susceptibility-controlled permanent impact of a meta-order in the Minority Game (Barato et al., 2011), for transition rates $\lambda_i,\mu_i$ between game states in an evolutionary Markov model (Feng et al., 9 Feb 2025), for state-dependent reinforcement in a SARSA-based spatial Prisoner’s Dilemma (Yang et al., 2024), and for action-level changes in a game-state evaluation or win probability in Secret Hitler and Counter-Strike: Global Offensive (Bauer, 9 Apr 2026, Xenopoulos et al., 2020). This suggests that the term functions as a cross-domain label rather than as a single standardized invariant.

1. Major meanings in the literature

The available works use closely related but non-identical formalizations. In some settings, the quantity of interest is a control parameter that changes the trajectory of play; in others it is a transition rate, a state-dependent response coefficient, or an action-level delta computed from a state evaluator.

Setting	Game state	Impact-rate notion
Human-AI continuous game	joint action vector $(h_t,m_t)$	AI adaptation rate $\alpha$
Minority Game	stationary regime with $H$ and $\chi$	permanent impact $\Delta^*$
Evolutionary Markov game	$X_j(t)\in\{G_0,G_1,\cdots,G_n\}$	transition rates $\lambda_i,\mu_i$
Spatial Prisoner’s Dilemma with SARSA	local cooperation state	state-dependent reinforcement effect
Bayesian repeated zero-sum game	random state $(h_t,m_t)$0	communication rate $(h_t,m_t)$1 with strategic randomization
Secret Hitler / CSGO	evaluated state before and after an action	mean state delta or win-probability delta

A common misconception is to treat all of these quantities as interchangeable. The cited papers instead distinguish between equilibrium-selection dynamics, market response, population redistribution across game states, reinforcement-learning feedback, adversarial state communication, and action valuation.

2. Adaptation rate as a game-state driver in human-AI continuous games

In "Effect of Adaptation Rate and Cost Display in a Human-AI Interaction Game" (Isa et al., 2024), the game state is the joint action vector $(h_t,m_t)$2, where the human chooses $(h_t,m_t)$3 and the AI chooses $(h_t,m_t)$4. The AI updates its action by gradient descent on its own cost:

$(h_t,m_t)$5

For the quadratic AI cost

$(h_t,m_t)$6

this becomes

$(h_t,m_t)$7

The paper interprets these cases directly: $(h_t,m_t)$8 means the AI is fixed at its Nash-equilibrium action, small $(h_t,m_t)$9 means the AI changes slowly, and $\lambda_i,\mu_i$0 means the AI instantaneously reaches its best response to the human.

The equilibrium concepts are the Nash equilibrium and the human-led Stackelberg equilibrium. The Nash equilibrium satisfies

$\lambda_i,\mu_i$1

with

$\lambda_i,\mu_i$2

Letting

$\lambda_i,\mu_i$3

the human-led Stackelberg equilibrium satisfies

$\lambda_i,\mu_i$4

The experiments used two feedback conditions. In Experiment 1, the human saw only the current cost $\lambda_i,\mu_i$5, displayed as an expanding or shrinking circle, with the instruction to keep the circle as small as possible. In Experiment 2, the human saw a $\lambda_i,\mu_i$6 local heat map of cost values around the current cursor position,

$\lambda_i,\mu_i$7

with lighter color meaning lower cost and the instruction to keep the color inside the black circle cursor as light as possible.

Across both experiments, the distribution of human and AI actions moved from near Nash equilibrium at low $\lambda_i,\mu_i$8 to near Stackelberg equilibrium at high $\lambda_i,\mu_i$9. Slow AI adaptation shifted outcomes toward Nash equilibrium, whereas fast adaptation shifted outcomes toward the human-led Stackelberg equilibrium. Human cost generally decreased as $(h_t,m_t)$0 increased, and AI cost showed a U-shaped trend across rates. The local landscape display shifted outcomes toward Nash relative to the current-cost display. The paper’s own summary identifies this as the “game state impact rate”: $(h_t,m_t)$1 changes the trajectory of $(h_t,m_t)$2 over time and thus the eventual equilibrium neighborhood.

3. Stationary-state susceptibility and market impact in the Minority Game

"Impact of meta-order in the Minority Game" (Barato et al., 2011) studies market impact by adding a trader who buys or sells a fixed amount $(h_t,m_t)$3 for a finite time $(h_t,m_t)$4. In the grand-canonical Minority Game, speculators update scores according to

$(h_t,m_t)$5

with excess demand

$(h_t,m_t)$6

The meta-order modifies this to

$(h_t,m_t)$7

so the total executed size is

$(h_t,m_t)$8

Price formation is assumed linear:

$(h_t,m_t)$9

Impact is then defined as

$\alpha$0

with permanent impact

$\alpha$1

The central state variable is predictability

$\alpha$2

The paper distinguishes a predictable phase with $\alpha$3 and finite susceptibility $\alpha$4, and an unpredictable / information-efficient phase with $\alpha$5 and divergent $\alpha$6. The exact stationary response formula is

$\alpha$7

and this yields

$\alpha$8

The resulting interpretation is explicit. In the unpredictable state, $\alpha$9, so $H$0: the market absorbs the meta-order without leaving a permanent trace. In the predictable state, $H$1, so the permanent impact is non-zero and linear in $H$2. The paper therefore defines state dependence through the stationary-state susceptibility $H$3: the more efficient and unpredictable the state, the smaller the lasting impact per unit traded volume; the more predictable the state, the larger the response.

The same framework also connects execution cost to the impact trajectory and notes that the impact law is more linear than empirical concave laws because of the assumed linear relation between excess demand and price. This is a precise example in which “game state impact rate” is neither a behavioral heuristic nor a generic metric, but an exact state-dependent response coefficient.

4. Markov transition rates between game states in evolutionary dynamics

In "An Evolutionary Game With the Game Transitions Based on the Markov Process" (Feng et al., 9 Feb 2025), the impact rate is the rate at which an individual’s current game state changes over time. Each individual $H$4 has a game-state process

$H$5

with state space

$H$6

The model assumes adjacent-state transitions: $H$7 at exponential rate $H$8, and $H$9 at exponential rate $\chi$0. The authors explicitly interpret this as a continuous-time homogeneous Markov chain.

Transition probabilities over a short interval $\chi$1 are

$\chi$2

and the transition rate is defined by

$\chi$3

The chain is described as homogeneous, irreducible, and continuous.

For the limiting probability $\chi$4, the stationary distribution satisfies the global balance equation

$\chi$5

where $\chi$6 is the outflow rate from $\chi$7 and $\chi$8 is the inflow rate. The paper emphasizes that occupancy of each game state is determined by the chain rates.

The simulations specialize to three game states: PDG, SDG, and SHG, with transitions PDG $\chi$9 SDG at rate $\Delta^*$0, SDG $\Delta^*$1 SHG at rate $\Delta^*$2, SDG $\Delta^*$3 PDG at rate $\Delta^*$4, and SHG $\Delta^*$5 SDG at rate $\Delta^*$6. The stationary expected numbers are

$\Delta^*$7

$\Delta^*$8

$\Delta^*$9

Increasing $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$0 decreases the number of individuals in PDG and increases SDG/SHG; increasing $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$1 decreases PDG and SDG and increases SHG; increasing $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$2 pushes mass back toward PDG; increasing $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$3 reduces SHG and increases PDG/SDG. The reported relative errors between theory and simulation are all below about $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$4.

The qualitative conclusion is equally direct: $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$5 and $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$6 promote cooperation, whereas $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$7 and $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$8 inhibit cooperation. The mechanism is that each game state $X_j(t)\in\{G_0,G_1,\cdots,G_n\}$9 has its own payoff matrix

$\lambda_i,\mu_i$0

so moving probability mass toward SDG and SHG changes the local incentive structure in a more cooperative direction.

The paper further includes a reputation mechanism,

$\lambda_i,\mu_i$1

neighbor selection probability

$\lambda_i,\mu_i$2

and Fermi strategy adoption

$\lambda_i,\mu_i$3

Game transitions alone help cooperation somewhat, but adding reputation enhances the cooperative outcome more strongly. The model also distinguishes the time scale of game transitions from the time scale of strategy updates, which may be fixed, exponential, or power-law distributed.

5. State-dependent reinforcement in the spatial Prisoner’s Dilemma

"The State-Action-Reward-State-Action Algorithm in Spatial Prisoner’s Dilemma Game" (Yang et al., 2024) states explicitly that “game state impact rate” is not given as a single explicit scalar formula. Instead, it is reflected through how the state definition, action choice, and reward feedback shape the evolution of cooperation and defection over time.

Agents are placed on an $\lambda_i,\mu_i$4 square lattice with periodic boundaries, with $\lambda_i,\mu_i$5, and each agent interacts with its four neighbors. The Prisoner’s Dilemma payoff matrix is

$\lambda_i,\mu_i$6

with

$\lambda_i,\mu_i$7

The paper later sets the single dilemma-strength measure to

$\lambda_i,\mu_i$8

SARSA is used in two ways. In one version, the state $\lambda_i,\mu_i$9 is determined from the cooperation rate between the agent and its neighbors, the action $(h_t,m_t)$00 is selecting one neighboring agent as the learning target, and the reward $(h_t,m_t)$01 is the cumulative payoff obtained after interacting with neighbors using the chosen strategy. In the second version, the state is again based on the cooperation rate between the agent and its neighbors, but the action is to cooperate or defect directly.

The SARSA update is

$(h_t,m_t)$02

For comparison, the paper gives the Q-learning update

$(h_t,m_t)$03

The interaction process is described as a Markov decision process

$(h_t,m_t)$04

and action choice is $(h_t,m_t)$05-greedy: with probability $(h_t,m_t)$06, the agent chooses $(h_t,m_t)$07; with probability $(h_t,m_t)$08, it chooses a random action. The reported parameters are $(h_t,m_t)$09, $(h_t,m_t)$10, $(h_t,m_t)$11, and noise $(h_t,m_t)$12.

The paper’s central claim is that the current local neighborhood composition determines the state, the SARSA policy chooses an action, and the resulting payoff becomes the reward signal that drives future behavior. Cooperative clusters, mixed boundaries, and defecting regions therefore produce different Q-values and different action probabilities. Reported spatial dynamics include tight cooperator clusters, difficulty for defectors in penetrating stable cooperative regions, expansion of cooperative clusters over time, and a shown case in which defectors are ultimately eliminated.

Heatmaps show that as $(h_t,m_t)$13 and $(h_t,m_t)$14 increase, cooperation becomes harder, but SARSA consistently improves cooperation relative to the traditional baseline, especially under harsher dilemma conditions. When the proportion $(h_t,m_t)$15 of SARSA agents is varied, cooperation first increases and then decreases, with maximum cooperation occurring when about $(h_t,m_t)$16–$(h_t,m_t)$17 of agents are SARSA agents. Average reward under SARSA rises over time and becomes higher than the traditional case, while cooperation and reward curves both tend to fall initially and then rise later as SARSA experience accumulates.

6. State information, communication rate, and adversarial unpredictability

"State Information in Bayesian Games" (0911.0874) studies a two-player repeated zero-sum Bayesian game with a random state $(h_t,m_t)$18 drawn i.i.d. across stages from $(h_t,m_t)$19. At each stage, Player A chooses action $(h_t,m_t)$20, Player B chooses action $(h_t,m_t)$21, and the stage payoff to Player A is $(h_t,m_t)$22. A helper observes the state sequence $(h_t,m_t)$23 non-causally and can communicate to Player A over a noiseless link of rate $(h_t,m_t)$24 bits per game iteration.

The paper considers several information structures: neither player knows $(h_t,m_t)$25; only Player A knows $(h_t,m_t)$26; only Player B knows $(h_t,m_t)$27; both know $(h_t,m_t)$28; and the helper knows $(h_t,m_t)$29 and communicates to Player A at rate $(h_t,m_t)$30. Its central claim is that communication for adversarial games is not the same as rate-distortion coding, because mixed-strategy randomization is strategically essential.

The erasure game provides the canonical example. Here $(h_t,m_t)$31 is equally likely; Player A has actions $(h_t,m_t)$32; Player B has actions $(h_t,m_t)$33; and if Player A chooses the wrong symbol relative to the state, the payoff is $(h_t,m_t)$34 or a very large negative number in a finite version. If Player A knows $(h_t,m_t)$35, the optimal mixed strategy is

$(h_t,m_t)$36

The reported values under different information structures are: neither knows $(h_t,m_t)$37, value $(h_t,m_t)$38; both know $(h_t,m_t)$39, value $(h_t,m_t)$40; A knows $(h_t,m_t)$41, value $(h_t,m_t)$42; B knows $(h_t,m_t)$43, value $(h_t,m_t)$44.

The helper’s scheme uses an auxiliary variable $(h_t,m_t)$45 with

$(h_t,m_t)$46

so that $(h_t,m_t)$47 forms a Markov chain. The paper frames this in terms of strong coordination: the aim is to synthesize a joint law on $(h_t,m_t)$48 close in total variation to i.i.d. samples from a target $(h_t,m_t)$49. The required rate for full synthesis is tied to Wyner common information $(h_t,m_t)$50 rather than only $(h_t,m_t)$51.

The main theorem uses the quantities $(h_t,m_t)$52, $(h_t,m_t)$53, and $(h_t,m_t)$54 in the rate-performance tradeoff. The paper’s interpretation is that communication rate has two roles: to convey the state information needed to correlate behavior with $(h_t,m_t)$55, and to support the randomness needed to keep the action sequence hard for the adversary to predict. For the erasure game, when Player B knows the state and the optimal strategy is

$(h_t,m_t)$56

the minimum description rate needed to achieve the optimal payoff is

$(h_t,m_t)$57

The paper also describes non-stationary effects over a block: early actions may appear random and well correlated with the state, while later actions may become predictable as the opponent learns the message. This motivates layered coding with multiple auxiliaries. A plausible implication is that, in adversarial repeated games, a “state impact rate” can be understood not only as a dynamical parameter but also as a communication budget governing how much state correlation and strategic unpredictability can be maintained simultaneously.

7. Action-level state deltas in hidden-role games and esports analytics

"Evaluating LLMs in a Complex Hidden Role Game" (Bauer, 9 Apr 2026) introduces Game State Impact Rate (GSIR) as an explicit action-level metric in Secret Hitler. It is designed to measure how much a player’s individual action changes the strategic position of the game. The game state is mapped to a continuous evaluation function, gamestate, ranging from $(h_t,m_t)$58 to $(h_t,m_t)$59, where negative values mean a fascist advantage and positive values mean a liberal advantage. The score combines five components: policy progress score, deck composition score, president score, role identification accuracy, and Hitler danger score. The normalized game-state score is

$(h_t,m_t)$60

where $(h_t,m_t)$61 is the raw weighted sum and $(h_t,m_t)$62 is the round number.

GSIR is computed from the change in game-state score caused by an action:

$(h_t,m_t)$63

The paper then defines GSIR as the mean delta in gamestate attributable to actions by an agent in role $(h_t,m_t)$64. Fascist-affiliation scores are inverted so that higher GSIR always means better play for the player’s side. GSIR is presented alongside Role Identification Accuracy (RIA), which measures whether a model correctly identifies others, and Deception Retention Rate (DRR), which measures whether others are fooled about the model’s identity.

The empirical findings are role-sensitive. R1 Distill 70B achieved the best overall GSIR, with an average impact of $(h_t,m_t)$65. Qwen 3 32B had the strongest liberal GSIR at $(h_t,m_t)$66. R1 Distill 70B was the only model with consistently positive GSIR across all roles, including Hitler, with Hitler GSIR reported as $(h_t,m_t)$67. Most models showed negative average impact when playing Fascist or Hitler. In the prompting ablation, a Role Prompt as Hitler produced a delta of $(h_t,m_t)$68 on average; Memory achieved $(h_t,m_t)$69 overall; and CoT + Memory had the strongest liberal GSIR at $(h_t,m_t)$70. The broader benchmark used five-player games against reputation-based agents, and the abstract reports that models playing as Fascists consistently yielded negative impact scores and failed to sustain deception, resulting in roughly $(h_t,m_t)$71 shorter games compared to humans.

A closely related construct appears in "Valuing Player Actions in Counter-Strike: Global Offensive" (Xenopoulos et al., 2020), although the paper does not use the label GSIR. Its action-value framework is Win Probability Added (WPA). A game state $(h_t,m_t)$72 contains round context at tick $(h_t,m_t)$73, and the predictor estimates

$(h_t,m_t)$74

where $(h_t,m_t)$75 if CT wins round $(h_t,m_t)$76. The value of action $(h_t,m_t)$77 is

$(h_t,m_t)$78

This is the change in predicted win probability caused by the action. Player WPA is the sum of action values, reported as WPA per round.

The state features used for prediction are Map, Ticks Since Start, Equipment Value, Players Remaining, HP Remaining, Bomb Planted, Bomb Plant Site, and Team Bombsite Distance. The paper evaluates logistic regression, CatBoost, and XGBoost, with XGBoost performing best: log loss $(h_t,m_t)$79, Brier score $(h_t,m_t)$80, and AUC $(h_t,m_t)$81. It also reports that WPA has the highest month-to-month correlation among the compared metrics at $(h_t,m_t)$82, ahead of KDR at $(h_t,m_t)$83, ADR at $(h_t,m_t)$84, KAST\% at $(h_t,m_t)$85, and Rating 2.0 at $(h_t,m_t)$86. An illustrative use case is a ZywOo clutch in which a headshot in a 1-vs-2 situation with 13 HP increases win probability by $(h_t,m_t)$87.

Taken together, these works suggest that “game state impact rate” is best understood as a state-relative measure of consequence. Depending on the domain, it may be the rate at which an adaptive opponent updates, the susceptibility of a stationary regime, the transition intensity between game states, the reinforcement effect of local social state, the communication rate required for strategically useful state information, or the change in a continuous game-state evaluation induced by a single action. The common structure is that impact is defined relative to a state representation and a dynamical or evaluative rule, not as an isolated property of actions alone.