Best-Response Feedback: Concepts & Dynamics

• Best-response feedback is a strategic process where current system states trigger optimal or near-optimal responses, creating iterative dynamics across game-theoretic and control settings.
• It spans multiple frameworks—from finite games and online learning to sequential decision making and recommendation systems—highlighting diverse applications and equilibrium selection issues.
• Methodologies leverage oracles, opponent models, and dynamic rules to compute best responses while addressing challenges like cyclicity, noise, and computational complexity.

Searching arXiv for recent and foundational papers on best-response feedback and related best-response dynamics. Best-response feedback denotes a family of strategic, algorithmic, and control-theoretic processes in which the current state of interaction induces a best response or approximate best response, and that response is then fed back into the next state of the system. In finite games this feedback takes the form of best-response dynamics and deviator rules; in online learning it appears as a best-response oracle; in sequential decision problems it appears as planning against an explicit opponent model; and in recommendation or elicitation settings it appears as compliance or deviation relative to a recommended action. Across these settings, the central issues are equilibrium selection, convergence, robustness to noise, information revealed by feedback, and the computational cost of extracting best responses (Feldman et al., 2020, Chakrabarti et al., 2023, Hernandez et al., 2022).

1. Formal models and feedback primitives

A standard finite-game model writes a game as $G=(N,\{P_i\},\{c_i\},SC)$, where $N$ is the player set, $P_i$ is player $i$'s strategy space, $c_i$ is the individual cost, and $SC$ is a social cost. Player $i$'s best-response set at profile $p$ is

$BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$

and a pure Nash equilibrium is a profile $p$ with $N$0 for every $N$1. A best-response sequence is obtained by repeatedly selecting a suboptimal player and moving that player to a best response. The selection rule itself is part of the feedback mechanism: a deviator rule $N$2 specifies which suboptimal player moves next, and the quality of the reached equilibrium can depend strongly on that choice. The paper on best-response dynamics formalizes this dependence through the inefficiency $N$3, the worst-case ratio between the social cost of the worst equilibrium reachable under $N$4 and the best equilibrium reachable from the same initial profile, with $N$5 (Feldman et al., 2020).

A second formalization replaces direct optimization over a strategy set by an oracle. In polyhedral games each player's feasible set is

$N$6

and the only primitive available to the learner is a Best-Response Oracle (BRO), equivalently a Linear Minimization Oracle, returning

$N$7

Here best-response feedback is not a trajectory of unilateral deviations but an oracle access model for online learning and equilibrium computation (Chakrabarti et al., 2023).

In sequential games the same idea appears as a best-response objective against an opponent policy. If the opponent follows $N$8, the learner seeks

$N$9

In BRExIt this objective is operationalized by combining opponent modelling with biased Monte Carlo Tree Search, so that planning targets a best response to a known or learned opponent rather than a generic safe policy (Hernandez et al., 2022).

A fourth formalization arises in recommendation systems with unknown utilities. A moderator recommends an action profile drawn from a distribution $P_i$0, each agent receives a private recommendation, and the moderator observes the realized best response $P_i$1 rather than payoffs. In this setting, best-response feedback is the observable compliance/deviation pattern generated by recommendations (Alanqary et al., 19 Feb 2026).

2. Equilibrium selection and efficiency of best-response dynamics

In congestion games, best-response dynamics always converges to a pure Nash equilibrium, but the equilibrium reached can vary sharply with the feedback rule that selects the next deviator. This dependence is especially explicit in network formation and job scheduling. In unweighted symmetric $P_i$2 parallel-link games, the price of anarchy is $P_i$3, the local rule Max-Cost can have inefficiency $P_i$4, and the local rule Min-Path is optimal with $P_i$5. In single-source series-of-parallel networks with $P_i$6 segments, an optimal global best-response sequence can be computed in $P_i$7 by dynamic programming, while Min-Path has tight inefficiency bounds $P_i$8 and $P_i$9; with asymmetric targets on such networks, Min-Path can be exponentially bad with $i$0. In single-source extension-parallel network formation, every local deviator rule has $i$1. In weighted parallel-link games, finding an order with $i$2 is NP-hard, every local rule has $i$3, and Min-Path can be $i$4-inefficient even when all weights are nearly equal. For identical-machine scheduling with standard linear latency, no local deviator rule can guarantee inefficiency below $i$5; by contrast, in conflicting congestion scheduling with cost $i$6, a simple local rule based on the highest and lowest loaded active machines is optimal (Feldman et al., 2020).

These results make the selection rule itself a first-class design object. The paper also introduces local deviator rules and the condition of independence of irrelevant players, which requires consistent ranking of local state-vectors across profiles. The impossibility results on extension-parallel graphs and identical-machine scheduling show that this locality restriction can be intrinsically costly: in some classes no local rule improves on the price of anarchy bound (Feldman et al., 2020).

Continuous-time best-response feedback has a different structure but a related message. In finite potential games, the differential inclusion

$i$7

has, for almost every initial condition, a unique solution; in almost every regular potential game, trajectories converge to pure-strategy Nash equilibria; and convergence is exponential in the sense that $i$8 for some $i$9. The mechanism behind this result is that best responses point uphill in the potential, mixed equilibria attract only a zero-measure set of initial conditions, and once a trajectory enters a region with a unique pure best response, the dynamics reduce to a linear ODE $c_i$0 (Swenson et al., 2017).

3. Cycles, sink equilibria, and robustness to mistakes

Best-response feedback is not generically stabilizing. In random two-player normal-form games, best-reply cycles become increasingly prevalent as the number of moves grows and payoffs become more anti-correlated. For $c_i$1, the fraction of games with at least one fixed point tends to $c_i$2 as $c_i$3, while the fraction with at least one $c_i$4-cycle tends to $c_i$5. The same work reports that the fraction $c_i$6 of moves in cycles versus fixed points has strong weighted correlation with the empirical non-convergence rate of six learning rules—Bush-Mosteller reinforcement learning, fictitious play, two-population replicator dynamics, deterministic EWA, stochastic EWA, and forward-looking level-2 EWA—with $c_i$7 (Pangallo et al., 2017).

One response to cyclicity is to replace pure Nash equilibrium as the evaluative target by sink equilibrium. In two-player symmetric games, digraph-based Best-Dominating and Non-Dominated metrics rank single-population strategies via sink strongly connected components of the best-response and non-dominated digraphs. Under mild conditions, strictly best-response self-play identifies exactly the BD-preferred strategies, while weakly better-response self-play identifies the ND-preferred set when the ND digraph has a unique sink SCC, and otherwise a subset of that set. At the meta-game level, strict best-response dynamics on the profile graph leads naturally to cycle-based and memory-based metrics over sink equilibria, and perturbed SBRD can be designed so that the stochastically stable recurrent class corresponds to the sink equilibrium maximizing the chosen metric, or lies within a prescribed tolerance when the optimality gap is unknown (Yan et al., 2022, Yan et al., 2020).

Another obstacle is execution noise. In imperfect best-response mechanisms, a selected player follows a best response only with probability at least $c_i$8. A worst-case construction shows that even for an NBR-solvable game with a unique equilibrium, a $c_i$9-fair schedule and error probability $SC$0 can prevent the process from reaching equilibrium with substantial probability. The positive result is quantitative: if the game is NBR-solvable with elimination length $SC$1, the schedule is $SC$2-fair, and

$SC$3

then the $SC$4-imperfect dynamics converges to the unique equilibrium in $SC$5 steps with high probability. Under an additional payoff-gap condition, incentive compatibility is also preserved (Ferraioli et al., 2012).

4. Best-response oracles, opponent models, and strategic learning

When best-response feedback is accessed through an oracle rather than explicit enumeration, the key question becomes how much learning is possible under restricted access. In polyhedral games, approximate Reflected Online Mirror Descent can be implemented using only BRO calls by replacing Euclidean projections with approximate proximal steps solved by Away-Step Frank-Wolfe. Choosing approximation accuracy $SC$6 yields only $SC$7 BRO calls at iteration $SC$8. This leads to constant per-player regret in two-player zero-sum games and $SC$9 regret in general-sum games. In zero-sum self-play, the average iterate achieves duality gap $i$0 as a function of the total number $i$1 of BRO calls, improving on the $i$2 barrier for prior projection-free methods. Under Saddle-Point Metric Subregularity, the last iterate converges linearly; without that condition, the best iterate still converges at rate $i$3. The same paper provides lower bounds on facial distance for standard-form and integral polytopes, including sequence-form, flow, and matching polytopes, which justify the linear-rate Frank-Wolfe subroutines (Chakrabarti et al., 2023).

In sequential decision making, BRExIt shows how opponent modelling can sharpen both planning targets and learned representations. The apprentice network optimizes a combined loss

$i$4

where $i$5 is a cross-entropy loss for predicting the opponent policy. During MCTS, priors at opponent nodes are biased toward the known or learned opponent model through

$i$6

or equivalently $i$7. In Connect4, over 10 runs of 48 hours each, ground-truth BRExIt achieved a probability of improvement greater than $i$8 over vanilla ExIt, BRExIt-OMS with learned opponent models achieved greater than $i$9, ExIt-OMFS underperformed vanilla ExIt at approximately $p$0 PoI against it, and one-hot versus full-distribution opponent-modelling targets showed no statistically significant difference except in ExIt-OMFS (Hernandez et al., 2022).

Population-based multi-agent learning raises a different cost issue: PSRO typically requires one best-response training phase per agent. Joint Experience Best Response replaces these separate interactions by a single joint dataset collected under the current meta-strategy and reused offline for every agent's best-response computation. To address offline-RL distribution shift, the paper proposes Conservative JBR, Exploration-Augmented JBR, and Hybrid BR. In two-player zero-sum games, if each player learns an $p$1-best response to a $p$2-perturbed data-collection policy, the resulting strategy profile is an $p$3-Nash equilibrium of the original game. Empirically, targeted exploration yields the best accuracy-efficiency trade-off, reaching near-PSRO NashConv with about $p$4 fewer environment episodes, while hybrid variants recover full PSRO performance with approximately $p$5–$p$6 more episodes than pure JBR (Bighashdel et al., 6 Feb 2026).

5. Dynamic games and control-theoretic best-response feedback

In stochastic dynamic games, best-response feedback can be implemented directly at the policy level. SLS-BRD considers an infinite-horizon $p$7-player stochastic LQ game with linear dynamics

$p$8

and quadratic costs, and lifts the problem into System Level Synthesis coordinates $p$9. Given the other players' parameters, each player solves a robust finite-dimensional convex program for an approximate best response $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$0, then updates according to

$BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$1

If each approximate best-response map is contractive with constant $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$2, then the joint update is itself a contraction and converges geometrically to a unique fixed point, an $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$3-GFNE, at rate $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$4 (Neto et al., 2024).

A related but information-constrained setting is the zero-sum stochastic LQ dynamic game with partial and asymmetric observations. There, each best response within the class of pure linear dynamic output-feedback controllers can be written explicitly via Kalman-filter and LQR Riccati equations, but every BR iteration increases the internal state dimension by one copy of the plant state, producing higher-order belief states. The reported numerical experiments show that the game value converges after only a few iterations. The same study computes controllability and observability Gramians and Hankel singular values of the augmented belief dynamics and finds rapid decay; over 1,000 random stable scalar instances, more than $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$5 of the average Gramian eigenvalues or Hankel singular values fall below $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$6. Figure 1 in that study reports convergence of both players' costs in $BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$7–$BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$8 iterations. A Cholesky-estimate theorem then bounds the error of truncating the higher-order belief dynamics, explaining why low-order controllers can closely approximate the Nash equilibrium (Guan et al., 10 Jan 2025).

6. Recommendation, elicitation, denoising, and broader feedback systems

Best-response feedback is also a design tool for systems that do not begin with known utilities or even with stable strategic behavior. In peer-prediction, the Square Root Agreement Rule rewards an agent only when her report matches a peer's and scales the reward by the inverse square root of an empirical popularity index. In the large-$BR_i(p)=\arg\min_{p'_i\in P_i} c_i(p'_i,p_{-i}),$9 limit, truthful reporting yields expected payoff $p$0, while any misreport is bounded by a Cauchy-Schwarz argument. Under the $p$1-separation condition, truthful reporting is a strict Bayesian Nash equilibrium for sufficiently large $p$2, and in the $p$3 limit truthfulness is strongly optimal across symmetric equilibria. Synthetic experiments with mild observational bias report average lying-gain at most $p$4 and truthful-coverage approximately $p$5–$p$6 of maximal possible (Kamble et al., 2015).

In recommendation under unknown games, best-response feedback reveals only the sign structure of payoff differences under recommendation slices. The moderator observes $p$7, not payoffs, and utilities are identifiable only up to a larger indistinguishability set $p$8 characterized geometrically by equality of the normal fans of polarized utility polytopes $p$9. In that sense, the sample complexity of recovering utilities beyond $N$00 is infinite. Nevertheless, a cutting-plane recommendation algorithm achieves expected cumulative regret $N$01 under both best-response and quantal-response feedback, where $N$02 is the ambient payoff-difference dimension (Alanqary et al., 19 Feb 2026).

Distributed sensing provides a non-game-theoretic but structurally similar example. Noisy sensors form a local consensus game and update via majority best response. After one synchronous update, if the sample size satisfies the theorem's condition, every sensor at distance at least $N$03 from the true separator is correct with probability at least $N$04. When $N$05, combining this denoising step with the agnostic active-learning algorithm of Awasthi-Balcan-Long yields a separator of error at most $N$06 using only $N$07 label queries. The same work shows a sharp failure mode: under an adversarial asynchronous order, best-response updates can drive all sensors to the same wrong label (Balcan et al., 2014).

A further extension replaces strategic agents by LLMs attempting to revise their own outputs under external critique. In that setting, the relevant quantity is the gap between observed improvement after $N$08 rounds of feedback and the theoretical ceiling attainable if the model fully incorporated near-perfect feedback. The reported experiments show persistent plateaus below ceiling: for Claude 3.7 Thinking, accuracy on AIME 2024 rises from approximately $N$09 to approximately $N$10 after 10 feedback iterations, versus a ceiling of approximately $N$11; on GPQA it rises from approximately $N$12 to approximately $N$13, again below a ceiling of approximately $N$14. Progressive temperature increases produce only minor gains, and rejection sampling adds approximately $N$15–$N$16 absolute improvement without closing the gap (Jiang et al., 13 Jun 2025).

Taken together, these results portray best-response feedback not as a single algorithm but as a general architecture for turning observed strategic context into targeted updates. Its successes depend on what the feedback channel reveals, how costly best responses are to compute, whether the induced dynamics select efficient equilibria, and how robust the process is to cycles, noise, partial observability, or incomplete incorporation of corrective signals.