NeuroGame Transformer (NGT) Architecture

• NeuroGame Transformer (NGT) is a transformer attention mechanism that fuses cooperative game theory and statistical physics, assigning token importance via Shapley and Banzhaf indices.
• It employs mean-field inference and self-normalized importance sampling to compute Gibbs marginals from token spin configurations and pairwise interactions.
• NGT achieves competitive results on natural language inference benchmarks while providing interpretability through explicit modeling of token coalitions and dynamic weighting.

NeuroGame Transformer (NGT) most directly denotes the architecture introduced in "NeuroGame Transformer: Gibbs-Inspired Attention Driven by Game Theory and Statistical Physics" (Bouchaffra et al., 19 Mar 2026), where attention is reformulated by treating tokens simultaneously as players in a cooperative game and as interacting spins in a statistical-physics system. In that formulation, token importance is derived from Shapley values and Banzhaf indices, pairwise token relations are encoded as interaction potentials, and attention weights arise as Gibbs marginals computed by mean-field inference. The term is not uniformly used across adjacent literature, however: an explanatory overview of Deep Transformer Q-Network (DTQN) also refers to that reinforcement-learning agent as NeuroGame Transformer (Upadhyay et al., 2019), whereas "NGT" in "Neuromodulation Gated Transformer" explicitly does not mean NeuroGame Transformer (Knowles et al., 2023).

1. Nomenclature and scope

The name "NeuroGame Transformer" is official in (Bouchaffra et al., 19 Mar 2026), where it labels a transformer attention mechanism grounded in cooperative game theory and statistical physics. That paper addresses Natural Language Inference rather than game-playing environments: "game" refers to coalitional structure among tokens, not to interactive control tasks.

The same string is used differently in the overview of "Transformer Based Reinforcement Learning For Games" (Upadhyay et al., 2019). There, the model introduced in the paper is Deep Transformer Q-Network (DTQN), and the overview states that, for explanatory clarity, the transformer-based RL agent is referred to as NeuroGame Transformer (NGT). This is therefore a secondary naming usage rather than the paper’s official model name.

Two nearby names are explicitly distinct. In (Knowles et al., 2023), NGT stands for Neuromodulation Gated Transformer, and the paper states that it is not "NeuroGame Transformer." In (Liu et al., 2024), the model is NfgTransformer, a permutation-equivariant encoder for normal-form games, and the paper states that it does not mention a "NeuroGame Transformer (NGT)."

Name in source	Meaning	Relation to NeuroGame Transformer
NGT	NeuroGame Transformer	Official name in (Bouchaffra et al., 19 Mar 2026)
NGT	Neuromodulation Gated Transformer	Explicitly not NeuroGame Transformer in (Knowles et al., 2023)
DTQN / explanatory "NGT"	Deep Transformer Q-Network	Secondary usage in (Upadhyay et al., 2019)
NfgTransformer	Equivariant encoder for normal-form games	Distinct model in (Liu et al., 2024)

This suggests that the acronym must be interpreted from context rather than assumed to identify a single architecture.

2. Game-theoretic attribution in the official NGT

In the 2026 NeuroGame Transformer, tokens are the player set $\mathcal{T} = \{t_1,\dots,t_n\}$, and each coalition is a subset $C \subseteq \mathcal{T}$. A characteristic function $v:2^{\mathcal{T}} \to \mathbb{R}$ measures a coalition’s semantic value, with assumptions $v(\emptyset)=0$, monotonicity $v(C) \le v(T)$ for $C \subseteq T$, and boundedness $|v(C)| \le M$. The paper defines token embeddings $x_i = e_i + p_i \in \mathbb{R}^d$ and sets

$$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$

with learned $W_v \in \mathbb{R}^{d \times d}$ and a mild nonlinearity such as ReLU or tanh. Coalition energy is then $C \subseteq \mathcal{T}$0 (Bouchaffra et al., 19 Mar 2026).

Two attribution notions are combined. The Shapley value provides a permutation-averaged notion of global fairness:

$C \subseteq \mathcal{T}$1

The Banzhaf index measures average marginal gain across subsets of $C \subseteq \mathcal{T}$2:

$C \subseteq \mathcal{T}$3

When the sums are nonzero, both are normalized across tokens.

The model introduces a token-specific interpolation gate

$C \subseteq \mathcal{T}$4

and forms an external field

$C \subseteq \mathcal{T}$5

The paper describes this as a fairness-sensitivity trade-off: $C \subseteq \mathcal{T}$6 yields Shapley-dominant behavior, $C \subseteq \mathcal{T}$7 yields Banzhaf-dominant behavior, and intermediate values interpolate between the two. A plausible implication is that the model can assign different semantic roles to different tokens rather than enforcing a single attribution regime everywhere.

Pairwise relations are also coalition-dependent. For tokens $C \subseteq \mathcal{T}$8, the second-order interaction is

$C \subseteq \mathcal{T}$9

and the pairwise coupling is defined as an expectation over subsets $v:2^{\mathcal{T}} \to \mathbb{R}$0. By construction, $v:2^{\mathcal{T}} \to \mathbb{R}$1 and hence $v:2^{\mathcal{T}} \to \mathbb{R}$2.

3. Statistical-physics formulation and transformer integration

The statistical-physics layer treats the token set as an Ising system with spin configuration $v:2^{\mathcal{T}} \to \mathbb{R}$3, $v:2^{\mathcal{T}} \to \mathbb{R}$4. The Hamiltonian is

$v:2^{\mathcal{T}} \to \mathbb{R}$5

Positive $v:2^{\mathcal{T}} \to \mathbb{R}$6 favors active spins, positive $v:2^{\mathcal{T}} \to \mathbb{R}$7 favors same-sign configurations and is interpreted as synergy, negative $v:2^{\mathcal{T}} \to \mathbb{R}$8 favors opposite signs and is interpreted as antagonism, and $v:2^{\mathcal{T}} \to \mathbb{R}$9 indicates independence (Bouchaffra et al., 19 Mar 2026).

A Gibbs distribution with temperature $v(\emptyset)=0$0 is then defined:

$v(\emptyset)=0$1

Attention weights are the marginals

$v(\emptyset)=0$2

where $v(\emptyset)=0$3 is the magnetization. Lower $v(\emptyset)=0$4 concentrates mass on low-energy configurations; as $v(\emptyset)=0$5, $v(\emptyset)=0$6.

Because exact marginalization requires summing over $v(\emptyset)=0$7 spin configurations, the paper derives the mean-field self-consistency equations

$v(\emptyset)=0$8

with a fixed-point iteration that the paper states typically converges within 10–20 iterations and has complexity $v(\emptyset)=0$9 for $v(C) \le v(T)$0 iterations. In the reported experiments, mean-field uses $v(C) \le v(T)$1 iterations, damping $v(C) \le v(T)$2, tolerance $v(C) \le v(T)$3, and $v(C) \le v(T)$4.

The Shapley, Banzhaf, and pairwise interaction terms are estimated by self-normalized importance sampling. The target coalition distribution is Gibbs-weighted,

$v(C) \le v(T)$5

while proposals are uniform over subsets or induced prefix distributions from sampled permutations. The raw weight is

$v(C) \le v(T)$6

and normalization removes the partition function. The paper states that these estimators avoid explicit exponential factors, are consistent with sufficiently large $v(C) \le v(T)$7, are asymptotically unbiased, and have variance decreasing at the standard Monte Carlo rate $v(C) \le v(T)$8. The overall NGT complexity is summarized as $v(C) \le v(T)$9, with coupling estimation scaling in principle as $C \subseteq T$0.

Within a transformer layer, inputs are contextualized embeddings $C \subseteq T$1 and values are $C \subseteq T$2. A single head computes $C \subseteq T$3, estimates $C \subseteq T$4, $C \subseteq T$5, and $C \subseteq T$6, solves mean-field for $C \subseteq T$7, sets $C \subseteq T$8, and returns

$C \subseteq T$9

The multi-head version independently parameterizes $|v(C)| \le M$0, $|v(C)| \le M$1, $|v(C)| \le M$2, $|v(C)| \le M$3, and $|v(C)| \le M$4 per head, concatenating head outputs before projection with $|v(C)| \le M$5 (Bouchaffra et al., 19 Mar 2026).

The experimental configuration uses a BERT-Base backbone, $|v(C)| \le M$6 during training and $|v(C)| \le M$7 at evaluation, sequence length $|v(C)| \le M$8, AdamW with learning rate $|v(C)| \le M$9, linear warmup $x_i = e_i + p_i \in \mathbb{R}^d$0, MultiStepLR with a factor of $x_i = e_i + p_i \in \mathbb{R}^d$1 after epochs 3 and 4, weight decay $x_i = e_i + p_i \in \mathbb{R}^d$2, gradient clipping $x_i = e_i + p_i \in \mathbb{R}^d$3, dropout $x_i = e_i + p_i \in \mathbb{R}^d$4, label smoothing $x_i = e_i + p_i \in \mathbb{R}^d$5, mixup $x_i = e_i + p_i \in \mathbb{R}^d$6, and EMA decay $x_i = e_i + p_i \in \mathbb{R}^d$7. A notable limitation is that $x_i = e_i + p_i \in \mathbb{R}^d$8 is the expected number of active tokens rather than $x_i = e_i + p_i \in \mathbb{R}^d$9; the paper notes that downstream components expecting normalized attention may require renormalization or adaptation.

4. Empirical performance, interpretability, and limitations

The official NeuroGame Transformer is evaluated on Natural Language Inference, specifically SNLI and MNLI-matched, with labels entailment, contradiction, and neutral. The reported dataset sizes are approximately $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$0k for SNLI and $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$1k for MNLI-matched. The model uses a BERT-Base backbone and adds approximately $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$2 parameters, for a total of about $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$3M parameters (Bouchaffra et al., 19 Mar 2026).

Model	SNLI	MNLI-matched
DAM	83.30%	68.41%
ESIM	87.00%	76.63%
BERT-Base	88.86%	80.99%
BERT-Large	90.33%	84.83%
RoBERTa-Base	86.95%	85.98%
RoBERTa-Large	91.83%	90.20%
ALBERT-Base	86.49%	79.89%
ALBERT-Large	90.85%	86.76%
NGT	86.60%	79.00%

The abstract separately reports that NGT attains a test accuracy of $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$4 on SNLI with a peak validation accuracy of $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$5, outperforming some major efficient transformer baselines. The results table gives $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$6 on SNLI and $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$7 on MNLI-matched. The paper’s stated takeaway is that NGT outperforms earlier efficient baselines such as DAM, is competitive with ALBERT-Base on MNLI, and remains below larger pretrained models such as RoBERTa-Large.

Interpretability is a central claim. In the "not good movie" case study, Shapley and Banzhaf scores expose decisive tokens, while negative pairwise couplings identify antagonistic interactions such as negation. The paper describes Banzhaf as revealing "swing voters" and emphasizes that $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$8 captures antagonism relevant to NLI logic. In addition, the learnable gate $$v(C) = f\!\left(\left\lVert \sum_{t_i \in C} W_v x_i \right\rVert_2\right),$$9 is presented as adapting whether a token relies more on global fairness or local decisiveness.

The limitations are explicit. The method uses a two-stage approximation stack: self-normalized importance sampling for coalition statistics and mean-field inference for spin marginals. Estimating all $W_v \in \mathbb{R}^{d \times d}$0 scales quadratically in the number of tokens. Temperature choice and fixed-point convergence are empirically tuned rather than backed by a stated global convergence condition. The paper also notes that NGT is parameter-efficient and complementary to large pretrained backbones rather than a replacement for them (Bouchaffra et al., 19 Mar 2026).

5. Secondary usage: transformer Q-learning for games

In the overview of "Transformer Based Reinforcement Learning For Games," NeuroGame Transformer is used as an explanatory label for Deep Transformer Q-Network (DTQN), an encoder-only transformer for Q-learning in a partially observable CartPole setting (Upadhyay et al., 2019). The environment is OpenAI Gym CartPole with actions $W_v \in \mathbb{R}^{d \times d}$1, reward $W_v \in \mathbb{R}^{d \times d}$2 per time step while the pole remains upright, and a final reward of $W_v \in \mathbb{R}^{d \times d}$3 on failure. Partial observability is imposed by exposing only cart position and pole angle, omitting velocities.

The agent consumes a window of the current and three preceding time steps, so $W_v \in \mathbb{R}^{d \times d}$4. DQN concatenates these observations; DRQN and DTQN/NGT process them as sequences. The transformer component is an encoder without a decoder, using positional embeddings or encodings, scaled dot-product attention,

$W_v \in \mathbb{R}^{d \times d}$5

and multi-head composition,

$W_v \in \mathbb{R}^{d \times d}$6

The encoder’s output sequence is reduced to a single feature vector for a fully connected Q-value head over the two actions. The paper does not report the numeric choices for $W_v \in \mathbb{R}^{d \times d}$7, $W_v \in \mathbb{R}^{d \times d}$8, $W_v \in \mathbb{R}^{d \times d}$9, $C \subseteq \mathcal{T}$00, or dropout.

Training follows DQN-style Bellman regression with target networks and epsilon-greedy exploration:

$C \subseteq \mathcal{T}$01

The reported regimen is $C \subseteq \mathcal{T}$02 episodes per run and $C \subseteq \mathcal{T}$03 independent runs per algorithm. The target network is updated every $C \subseteq \mathcal{T}$04 iterations, but $C \subseteq \mathcal{T}$05 is not specified.

Empirically, DRQN achieved the best runs, including a maximum episodic score of $C \subseteq \mathcal{T}$06 in one test case. DQN and DTQN/NGT generally underperformed, and in many runs their average scores decreased over training. The paper’s interpretation is that, in this very small POMDP with only two scalar features and short context, the recurrent GRU better matches the temporal structure needed to infer hidden velocities. This suggests that the transformer-based RL usage of "NGT" belongs to a different problem class from the official 2026 NeuroGame Transformer and should not be conflated with it.

6. Distinct neighboring architectures often confused with NGT

The strongest source of acronym confusion is "Neuromodulation Gated Transformer" (Knowles et al., 2023). That model implements neuromodulation in transformers as an internal, context-dependent multiplicative gating mechanism. A gating block consumes the output activations of a transformer layer, emits a gate in $C \subseteq \mathcal{T}$07 with the same shape, and applies it element-wise:

$C \subseteq \mathcal{T}$08

In the main BERT-large experiments, a single three-layer gating block is inserted after layer 21. On SuperGLUE validation sets, the neuromodulated-gating model attains the highest mean score among the compared models, $C \subseteq \mathcal{T}$09 versus $C \subseteq \mathcal{T}$10 for no-gating-block and $C \subseteq \mathcal{T}$11 for non-neuromodulated-gating, although all reported $C \subseteq \mathcal{T}$12-values exceed $C \subseteq \mathcal{T}$13. The paper explicitly states that NGT there stands for Neuromodulation Gated Transformer and not NeuroGame Transformer.

A second neighboring system is NfgTransformer (Liu et al., 2024), a permutation-equivariant encoder for normal-form games. It represents an $C \subseteq \mathcal{T}$14-player game as action embeddings $C \subseteq \mathcal{T}$15, uses action-to-joint-action self-attention, action-to-play cross-attention, and action-to-action self-attention, and satisfies the equivariance condition $C \subseteq \mathcal{T}$16 under strong isomorphisms of players and strategies. It supports equilibrium solving, deviation-gain regression, and payoff reconstruction, with parameter count independent of the number of players and actions. The paper states directly that it does not mention a "NeuroGame Transformer (NGT)" and should not be conflated with that term.

These distinctions matter conceptually. The official NeuroGame Transformer (Bouchaffra et al., 19 Mar 2026) is a coalition-aware attention mechanism for sequence modeling; the DTQN usage (Upadhyay et al., 2019) is transformer Q-learning for partially observable control; the Neuromodulation Gated Transformer (Knowles et al., 2023) is a multiplicative gating retrofit for pretrained language encoders; and NfgTransformer (Liu et al., 2024) is an equivariant representation learner for normal-form games. The shared vocabulary of games, gating, and transformers masks substantial differences in objectives, mathematical formalism, and empirical domain.