Marginal Token Allocation Economies

• Marginal Token Allocation Economies are systems that allocate tokenized budgets based on marginal gains rather than fixed quotas.
• They bridge computational designs—such as video token pruning and agentic AI control—with non-monetary public resource mechanisms.
• These mechanisms optimize resource use by balancing marginal benefits with costs, ensuring both efficiency and fairness in allocation.

Marginal Token Allocation Economies denote a family of allocation systems in which scarce tokens, token-like budgets, or tokenized rights are distributed by marginal criteria rather than by uniform quotas or flat unit prices. Across the literature, the same allocative logic appears in efficient video token pruning, agentic AI control, non-monetary congestion pricing, divisible-resource markets with unequal budgets, coupled public-resource allocation, weighted automated market makers, and digital tokenomics: an additional token is assigned where its marginal gain, marginal social value, or marginal product exceeds its marginal compute, latency, risk, or congestion burden (Ma et al., 28 Aug 2025, Zhu, 2 May 2026, Riehl et al., 2024, Andelman et al., 2021). The term therefore spans both computational systems and socio-technical economies, but the common structure is stable: a global budget or conservation law constrains allocation, and local decisions are made at the margin.

1. Conceptual scope and token semantics

The literature uses “token” in several technically distinct but structurally related senses. In video LLMs, tokens are visual embeddings whose selection is constrained by a global token budget, and the allocative problem is to preserve salient information under redundancy (Ma et al., 28 Aug 2025). In agentic AI systems, tokens are “marginal computation units that trigger actions, occupy infrastructure, and become training data,” and they are evaluated by task value, compute cost, latency, and risk rather than by a flat unit price (Zhu, 2 May 2026). In public-good allocation, Karma tokens are a “non-monetary, fair, and efficient resource allocation mechanism” that “cannot be bought” and are “non-convertible to money,” so access is aligned with urgency and prior behavior rather than with financial power (Riehl et al., 2024). In competitive-equilibrium models, tokens are budgets with no intrinsic value, and leftover tokens are worthless (Andelman et al., 2021). In weighted AMMs, LP shares are minted or burned according to invariant scaling and thereby function as the output of a token allocation rule over multiple resources (Astarita et al., 20 Jun 2026).

This heterogeneity matters because the token object determines the surrounding constraints. Some systems preserve total token mass, as in Karma economies and thin-market token systems (Riehl et al., 2024, Ashlagi et al., 2024). Some use explicit budget constraints, as in VLLM pruning and LLM pricing (Ma et al., 28 Aug 2025, Bergemann et al., 11 Feb 2025). Some preserve an invariant, as in weighted constant-function market makers (Astarita et al., 20 Jun 2026). Some study allocation under endogenous concentration and tradability, as in governance-token fair launches and Bitcoin wealth-distribution analysis (Fernandez et al., 2022, Sadykhov et al., 19 Feb 2026). A plausible implication is that “economy” in this literature refers less to a monetary market than to a general mechanism for coordinating scarcity under marginal rules.

2. Formal structure and first-order logic

Several papers express the allocative rule as an equimarginal or marginal-gain condition. In MMG-Vid, the marginal gain of adding a visual token $t$ to a selected set $T$ is

$\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$

and the segment-level marginal gain from one more token is

$\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$

The framework then allocates budget across segments and tokens so as to maximize utility under a global budget $B$ (Ma et al., 28 Aug 2025).

In the agentic-AI formulation, all layers satisfy a common risk-adjusted master rule,

$$\text{MB}_i(\tau; x) = \text{MC}_i(\tau) + \lambda\,\Delta L_i(\tau) + \rho\,\Delta R_i(\tau; x),$$

or equivalently allocate the next token to the use

$$i^*(x) = \arg\max_{i \in I}\Big[V(x)\,\Delta Q_i(\tau; x) - \Delta C_i(\tau) - \lambda\,\Delta L_i(\tau) - \rho\,\Delta R_i(\tau; x)\Big].$$

Here the key innovation is not merely pricing computation, but making compute, latency, and risk commensurable across routing, action selection, serving, and training (Zhu, 2 May 2026).

In hardware-verification token allocation, the problem is posed directly as a constrained budget allocation across token-use categories:

$$\max C(T_1,\ldots,T_6) \quad \text{s.t.} \quad \sum_i T_i \le B,$$

with first-order conditions

$\frac{\partial C}{\partial T_i} = \lambda$

for all positively allocated categories. Category-specific marginal coverage gains are defined as $g_i = \frac{\partial C}{\partial T_i}$, and coverage efficiency is $T$0 (Patel et al., 17 Apr 2026).

In unequal-budget competitive equilibrium, each agent solves

$T$1

with equilibrium KKT conditions

$T$2

This gives a precise marginal interpretation of budgeted token allocation: consumed goods equalize marginal utility per token across uses (Andelman et al., 2021).

In Karma-based congestion pricing, the marginal token toll is set equal to marginal social cost,

$T$3

and balances evolve as

$T$4

The mechanism thereby internalizes congestion externalities without monetary transfers (Riehl et al., 2024).

These formulations differ in surface vocabulary, but they share a common architecture: a constrained objective, a marginal-value signal, and a balance law or budget law that prevents unconstrained allocation.

3. Mechanism families

The principal mechanism families can be organized by how they generate marginal signals and enforce feasibility (Ma et al., 28 Aug 2025, Elokda et al., 2024, Ashlagi et al., 2024, Pedroso et al., 18 Mar 2026, Astarita et al., 20 Jun 2026, Bergemann et al., 11 Feb 2025).

Domain	Token object	Marginal mechanism
Video LLM pruning	Visual tokens	Segment-level MV and token-level TG-DPC
Agentic verification	Inference-time token budget	Category-specific marginal coverage gain
Karma congestion pricing	Non-tradable Karma tokens	Token toll $T$5
Thin service markets	Integer token balances	Minimum-token provider selection
Divisible-resource markets	Budgets of tokens	Anonymous prices and agent-specific budgets
Multi-resource CPHS	Multiple Karma accounts	Redistribution and exchange-rate design
Weighted AMMs	LP shares	Invariant-ratio mint/burn rule

One family is greedy marginal-gain allocation. MMG-Vid first divides a video into segments by adjacent-frame cosine similarity with default threshold $T$6, then assigns a minimum per-segment allocation and distributes remaining budget by a dynamic marginal-value heuristic that rewards representativeness relative to remaining content and diversity relative to selected content. Within each segment, temporal-guided DPC selects tokens with high inter-frame uniqueness and intra-frame diversity (Ma et al., 28 Aug 2025).

A second family is price-based decentralization. Unequal-budget competitive equilibrium supports any Pareto efficient allocation with anonymous item prices and agent-specific budgets, and the optimal mechanism for LLM product design can be implemented through menus of two-part tariffs, with higher markups for more intensive users (Andelman et al., 2021, Bergemann et al., 11 Feb 2025). In the latter setting, the core distinction is whether token allocation across tasks is contractible; when it is not, the mechanism collapses to menus of token packages summarized by a representative type $T$7 (Bergemann et al., 11 Feb 2025).

A third family is token-circulation design. Karma systems specify issuance, earning, spending, expiry, redistribution, and non-convertibility rules so that token balances encode past cooperation and future priority (Riehl et al., 2024). Multi-karma economies extend this to coupled resources through redistribution rules and exchange rates among resource tokens, with Nash welfare used to evaluate heterogeneous outcomes (Elokda et al., 2024). In congestion games, integer tolls of the form

$T$8

steer the aggregate dynamics toward an optimal efficient and fair allocation from any initial condition (Pedroso et al., 18 Mar 2026).

A fourth family is balance-stabilizing matching. In thin exogenous supply markets, the minimum-token allocation rule selects the available provider with the minimum token balance. The central result is that the move from one available provider to two available providers creates a “power of two choices” effect that stabilizes the token distribution (Ashlagi et al., 2024).

A fifth family is invariant-based allocation. In weighted AMMs, arbitrary liquidity operations satisfy

$T$9

equivalently,

$\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$0

The invariant is therefore also the general allocation formula, and any non-proportional operation can be decomposed into an internal rebalancing swap combined with a proportional operation in the fee-less setting (Astarita et al., 20 Jun 2026).

4. Computational instantiations in AI and LLM systems

The most explicit computational instantiation is MMG-Vid, a training-free visual token pruning framework for video LLMs. It evaluates on MVBench, LongVideoBench, MLVU, and VideoMME; uses accuracy as the primary metric; and reports that on LLaVA-OneVision-7B it “prunes 75% of tokens (retention 25%), retains 99.5% of original performance on average, and achieves 3.9× acceleration in prefilling and 3.1× in generation.” On LLaVA-Video-7B, it achieves higher average accuracy than FastV, VisionZip, PruneVid, and FrameFusion across retention ratios 25%, 20%, 15%, and 10%, and at 10% retained tokens it maintains 93.8% accuracy versus FastV’s 77.1% (Ma et al., 28 Aug 2025). The paper also reports that replacing TG-DPC with standard DPC-KNN degrades performance markedly on MVBench, and that dynamic budget allocation yields 0.8–1.0% gains on long-video benchmarks versus uniform budgets.

A second computational instantiation appears in agentic hardware verification. The enhanced LangGraph system instruments six token usage categories—system prompt, design comprehension, stimulus generation, coverage feedback, error recovery, and agentic overhead—and uses a taxonomy of coverage holes split between methodology-bound ceilings and reasoning frontiers (Patel et al., 17 Apr 2026). Across 19 designs, the enhanced system achieves “comparable or higher cumulative coverage (95–99%) with 4–13x fewer tokens and converges 2–4x faster than the general-purpose baseline.” Large-design token allocation shifts toward design comprehension, stimulus generation, and coverage feedback, with SG reasoning share increasing from ~18% to ~28%, coverage feedback appearing at ~6%, and error recovery rising to ~8%. The paper’s derived interpretation is that $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$1 is high early on large designs, whereas $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$2 and $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$3 are highest in reasoning-frontier regimes (Patel et al., 17 Apr 2026).

A third computational instantiation concerns the economics of LLM product design rather than inference-time control. The user allocates input tokens $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$4, output tokens $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$5, and fine-tuning tokens $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$6 across tasks, with quality function $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$7 that is strictly increasing, strictly concave, and has strictly positive cross-partials (Bergemann et al., 11 Feb 2025). Under given prices, the user’s FOCs are

$\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$8

The paper shows that the optimal mechanism can be implemented through menus of two-part tariffs, with higher markups for more intensive users, thereby treating token allocation, fine-tuning, and pricing as a single screening problem (Bergemann et al., 11 Feb 2025).

Taken together, these AI papers show that marginal token allocation is not limited to pruning or billing. It governs inference-time computation, agentic orchestration, token-category budgeting, and the contract design of tokenized model access.

5. Fairness, stability, and distributional consequences

A central theme outside AI is that marginal token allocation is often proposed as a fairness-preserving alternative to money. In public-road allocation, Karma is designed to be “non-monetary, fair, and efficient,” with equal initial endowments, constant total tokens, and no debt (Riehl et al., 2024). In the NYC Lincoln Tunnel versus GW Bridge case study, the system optimum is 3,983 veh/h on the tunnel and 40.75 min average travel time, achievable either by monetary pricing of approximately $\Delta_Q(t \mid T) = Q(T \cup \{t\}) - Q(T),$925.54 or by a Karma threshold of approximately 5.36–7.71 tokens. The equity contrast is stark: monetary pricing yields approximately 20% low-income versus approximately 95% high-income tunnel access, whereas Karma yields 39.83% access for all incomes. Average urgency on the fast route is 2.32 under Karma versus 2.14 under money, and total cost per user falls to $81.20 under Karma without financial fees (Riehl et al., 2024).

Stability results sharpen the allocative interpretation. In thin markets with one requester and at most $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$0 available providers, the minimum-token rule is vacuous when $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$1, and the token distribution is unstable. When $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$2, the Markov chain is positive recurrent and admits a unique stationary distribution, with the bound

$\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$3

for all large $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$4 and all $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$5 in the general symmetric case. For two symmetric agents with $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$6,

$\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$7

The paper explicitly interprets this as a “power of two choices” effect (Ashlagi et al., 2024).

Distributional dynamics can also move in the opposite direction. In governance-token fair launches, the agent-based model finds that “regardless of the allocation, concentration persistently occurs” and that “the disease is endogenous: the cause of concentration is the tokens’ tradability” (Fernandez et al., 2022). At $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$8, the egalitarian S1 scenario requires 10.80%–11.38% of agents to control 90% of supply, compared with 2.59%–3.63% in S0 and 0.76%–2.83% in S2. Lower trading probability reduces concentration across all scenarios. This is directly relevant to token allocation economies because it shows that a marginal rule can be fair at issuance yet still generate oligarchic outcomes under unrestricted secondary trade.

Policy shocks likewise reshape token distribution. In the Bitcoin wealth-distribution study, after de-noising the wealth-bucket series, major economy-related BIPs—BIPs 32, 42, 50, 141, and 341—show significant Granger causality for poorer and mid-wealth buckets, whereas the broader economy-related set excluding major BIPs shows significant Granger causality predominantly in wealthier buckets (Sadykhov et al., 19 Feb 2026). The longest consistently significant lag in the “Full” tests is about six months. This suggests that marginal token allocation can be driven not only by internal trading rules, but also by endogenous policy changes that alter transaction semantics, throughput, or wallet standards.

In coupled-resource Karma economies, welfare analysis reinforces the same point. Redistribution design significantly affects coupled allocations and Nash welfare, whereas non-unit exchange rates play a comparatively minor role; in the studied setting, uniform redistribution to all with unit exchange rates attains maximum social welfare (Elokda et al., 2024).

6. Limitations, misconceptions, and open questions

A recurring misconception is that minimizing token count is itself the objective. The position paper on agentic AI rejects that premise directly: systems that locally minimize tokens can globally misallocate them because routing, action, serving, and training are optimizing under different or incomplete price vectors (Zhu, 2 May 2026). The relevant object is not raw token count, but risk-adjusted marginal return.

The literature also emphasizes that marginal rules may be heuristic even when they work well empirically. In MMG-Vid, the global objective is NP-hard, the method resorts to a two-stage greedy-by-marginal-gain approach, and “Q is not formally proven submodular” (Ma et al., 28 Aug 2025). In hardware verification, methodology-bound ceilings imply that $\Delta_s(n_s) = G_s(n_s + 1) - G_s(n_s).$9 for all categories beyond the reachable coverage ceiling, so further token spending increases cost without appreciable coverage gains (Patel et al., 17 Apr 2026). In multi-karma economies, existence of a Stationary Nash Equilibrium is established, but uniqueness is not; convergence of the iterative algorithms is empirical, and formal proofs that uniform redistribution with unit exchange maximizes Nash welfare beyond the studied parameter regimes remain open (Elokda et al., 2024).

Several domains expose unresolved finite-sample or implementation issues. In thin-market token systems, exact exponential tail rates for finite $B$0 remain conjectural (Ashlagi et al., 2024). In weighted AMMs, the unified fee-less formula is exact, but “the exact closed-form allocation with fees for arbitrary non-proportional baskets remains an open problem” (Astarita et al., 20 Jun 2026). In congestion-game token economies, practical bounds for the integer-toll scale parameter $B$1 and wallet cap $B$2 still require calibration, and fuller Pigouvian equivalence at the per-resource level remains open (Pedroso et al., 18 Mar 2026). In Bitcoin wealth analysis, Granger causality is predictive rather than structural, address buckets are not entities, and omitted on-chain microstructure may leave residual confounding (Sadykhov et al., 19 Feb 2026).

These limitations do not invalidate the framework. They indicate that marginal token allocation economies are best understood as a general design paradigm: a way to convert scarcity, redundancy, congestion, heterogeneity, and risk into explicit marginal signals, then enforce those signals through budgets, prices, balances, redistribution rules, matching rules, or invariants. The literature’s main contribution is to show that this paradigm is portable across video understanding, agentic inference, public-resource allocation, market equilibrium, thin-market cooperation, decentralized liquidity, and tokenized governance, while leaving open the difficult questions of dynamics, strategic behavior, and institutional robustness.