Multi-Head Attention Is a Multi-Player Game

TL;DR

This work treats transformer attention heads as strategic agents in a multi-player game and shows that standard cross-entropy training induces unpriced externalities like redundancy and correlated errors, yielding inefficient Nash equilibria. It formalizes a social objective $C^\star_{\mathrm{IB}}$ with an information-bottleneck framing and proves that the Price of Anarchy is controlled by the off-diagonal mass $\Gamma(G)$ of a head interaction matrix, linking coordination failure to hallucination and redundancy. The authors propose GAME-LoRA, a practical regularization scheme that combines Barrow Twins-inspired decorrelation and a log-determinant pressure to internalize externalities, enabling selective head coordination and coalition formation. Empirical results across hallucination and knowledge benchmarks show substantial hallucination reductions (up to $+8.1\%$) with negligible knowledge loss, validating a Pareto-improving shift on the reliability-capability frontier and highlighting the importance of modeling intra-model game dynamics in modern transformers.

Abstract

Modern transformer attention is internally multi-agent -- heads compete and coordinate -- yet we train it as if it were a monolithic optimizer. We formalize this gap: cross-entropy training induces an implicit potential game among heads, and gradient descent converges to Nash equilibria with potentially unbounded inefficiency due to unpriced externalities (redundancy, correlated errors). Our main result bounds the Price of Anarchy by $Γ(G)$, the off-diagonal mass of a head interaction matrix capturing weight and gradient coupling. Under mild smoothness assumptions, we prove that both \emph{excess hallucination probability} and \emph{excess head redundancy} scale with PoA, unifying two distinct failure modes into a single mechanism. The bound is prescriptive: regularization that reduces $Γ(G)$ provably tightens PoA. We instantiate this as GAME-LoRA, combining Barlow Twins decorrelation with log-determinant coordination pressure. Experiments validate the theory: $Γ(G)$ predicts hallucination ($p{<}0.05$), emergent coalitions exhibit selective coordination, and GAME-LoRA achieves up to 18\% hallucination reduction (8\% average) with no knowledge degradation -- a Pareto improvement inaccessible to methods ignoring the game structure.

Figures (5)

• Figure 1: Attention heads form strategic coalitions when trained with game-aware regularization. We model multi-head attention as a multi-player game where heads compete for gradient credit while imposing unpriced externalities (redundancy, correlated errors) on each other. Standard cross-entropy training finds inefficient Nash equilibria; our method, GAME-LoRA, internalizes these externalities via Pigouvian taxes (log-determinant for compression, Barlow Twins for redundancy). The figure shows the head interaction matrix $G \in \mathbb{R}^{16 \times 16}$ for layer 19 of Qwen2.5-0.5B, reordered by spectral biclustering. Left: Before GAME-LoRA: heads show diffuse, unstructured coupling. Right: After GAME-LoRA: distinct coalitions emerge (white lines), with strengthened intra-coalition coordination and inter-coalition competition. This coalitional reorganization reduces the Price of Anarchy (\ref{['thm:poa']}), achieving best-in-class hallucination performance among five methods while preserving knowledge (\ref{['tab:main']}).
• Figure 2: GAME-LoRA induces selective coordination, not uniform decorrelation. A naive interpretation of regularization losses would predict uniform weakening of all head interactions. Instead, we observe strategic differentiation: heads selectively strengthen coordination within emergent coalitions while reducing coupling across coalition boundaries. The histogram shows changes in pairwise coupling $\Delta G_{ij}$ (final $-$ initial) for head pairs identified as intra-coalition (red, $N{=}24$) versus extra-coalition (blue, $N{=}67$) by spectral biclustering. Intra-coalition pairs strengthen significantly more (mean $\Delta G = 0.010$, dashed red) than extra-coalition pairs (mean $\Delta G \approx 0$, dashed blue), with $p = 2{\times}10^{-6}$ (Mann-Whitney). This selective reorganization mirrors coalition formation in cooperative game theory: heads that benefit from joint coordination form stable alliances, while heads serving distinct functions diversify---precisely the structure predicted by our externality-aware game formulation.
• Figure 3: Instantiation of \ref{['cor:hallucination']}. Since excess hallucination $\propto \mathrm{PoA} \propto 1/(1 - c \cdot \Gamma)$, we fit $\Delta H = a - \lambda/(1 - c \cdot \Gamma)$ to training trajectories. The fit achieves $R^2 = 0.53$ ($p < 0.05$) with bootstrap 95% prediction bands (shaded). Non-vacuousness: The fitted $c \cdot \Gamma \approx 0.2 \ll 1$ confirms the bound is finite and tight---reducing $\Gamma$ yields proportional hallucination reduction.
• Figure 4: GAME-LoRA architecture. Layer 19 with 16 attention heads ($h_0$--$h_{15}$), each containing QKV LoRA adapters, plus O LoRA in the output projection. Red dashed arrows: $\mathcal{L}_{\mathrm{ABT}}$ pairwise head decorrelation. Blue dashed arrows: $\mathcal{L}_{\mathrm{LDB}}$ expansion pressure on head Gram matrix. Three losses ($\mathcal{L}_{\mathrm{CE}}$, $\mathcal{L}_{\mathrm{LDB}}$, $\mathcal{L}_{\mathrm{ABT}}$) combined via Nash-MTL arbitration.
• Figure 5: Relative change in pairwise coupling $G_{ij}$; red traces strengthen (coordinate), blue weaken (diversify).

Theorems & Definitions (23)

• Definition 2.1: Weight Coupling
• Definition 2.2: Gradient Coupling
• Definition 2.3: Head Interaction Matrix
• Definition 2.4: Information Bottleneck Social Objective
• Definition 2.5: Information Bottleneck Price of Anarchy
• Definition 2.6: MultiHeadCE
• Theorem 2.1: MultiHeadCE is a Weighted Potential Game
• proof : Proof sketch
• Definition 2.7: MultiHeadPGAC
• Theorem 2.2: Multi-Head Games Have Price of Anarchy