acceptodds
Under review as a conference paper at ICLR 2027

Token as Power: How Token Resources Shape Agent Economies ?

17 views

Abstract

In evolving multi-agent economies, task performance governs economic rewards, evolutionary survival, and reproduction. However, whether intrinsic model capability guarantees competitive dominance under asymmetric resource constraints remains an open question. Motivated by debate experiments where a compute-advantaged weaker model defending an incorrect claim wins of debates at matched monetary cost, we systematically investigate token allocation within the Economy of Minds (EOM) framework across two scientific-reasoning benchmarks. We examine three regime types: persistent resource inequality, temporary budget shocks, and wealth-funded token expansion. Our investigation reveals three core dynamics. First, resource-advantaged weaker models can displace models that perform better under matched budgets, a phenomenon we term the Agentic Gresham-like Effect, even when stronger models retain a static capability advantage. Second, temporary budget shocks cause amplified losses under economic evolution compared to static populations, leaving performance gaps and changes in population and wealth after budget restoration. Third, early token investment can support subsequent lineage expansion: restricting purchases by selected agents reduces their lineages' final population shares in paired comparisons. Together, these findings establish token access as a structural driver of population dynamics rather than a passive runtime hyperparameter. Consequently, evaluating agent economies requires reporting resource distributions alongside task performance and lineage trajectories.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-Agent Systems, Agent Economies, Token Budgets

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3mJKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds on a framework called Economy of Minds and studies how token access shapes competition in evolving LLM agents, where task contributions earn wealth that determines survival and reproduction. Main contributions: larger token budgets can let weaker models displace stronger ones; temporary budget cuts can produce amplified performance losses and lasting population changes after resources are restored; and wealth-funded token purchases can reinforce lineage expansion, without necessarily improving collective performance. It also proposes "Token Productivity", which selects parents using contribution quality normalized by token consumption instead of wealth.

Critical strengths

They identified a distinction between model capability and competitive success in an evolving agent economy, which is something I found somewhat insightful. Across model families and benchmarks, increasing weaker models’ token budgets reverses population outcomes, even where stronger models retain better static performance. This makes resource allocation central to interpreting economic selection in that surviving and reproducing does not necessarily identify the most capable agents.

The experiments go beyond descriptive trajectories by controlling resource access. The fixed versus evolving population comparison isolates how economic feedback amplifies temporary token shocks, while matched purchase restrictions test whether investment access changes subsequent lineage outcomes, which strengthen the evidence that token access actively shapes population dynamics.

Critical weaknesses

First, the empirical evidence looks too limited to establish the robustness of the claimed dynamics. Held out evaluation uses only 19 FrontierScience tasks and 20 phys tasks, and evolution runs for 30 tasks. The reported comparisons do not establish variability across repeated independent runs.

The central mechanisms still remain insufficiently disentangled as icreasing token budgets might change contribution quality, reward judgments, bidding outcomes, and reproductive opportunities simultaneously. ALso under unequal budgets, weaker models already achieve comparable or better performance in several settings, so their subsequent dominance does not by itself demonstrate that economic selection favors inferior contributions.

Wealth based parent selection, bankruptcy rules, prompt mutation, and the chosen reward structure all influence which advantages become heritable population differences, but their roles are not systematically tested. Token Productivity improves performance in the reported comparisons, yet those results do not establish whether resource normalization specifically drives the improvement or whether contribution quality alone would suffice. So the conclusion seem to depend heavily on one particular economic design.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3mJK (PART I)

We sincerely thank you for taking the time to examine our work and explain your concerns. We appreciate your assessment and take the issues you raised seriously. Below, we respond to each point. We hope these responses provide a clearer basis for assessing our work.

1

First, the empirical evidence looks too limited to establish the robustness of the claimed dynamics. Held out evaluation uses only 19 FrontierScience tasks and 20 phys tasks, and evolution runs for 30 tasks. The reported comparisons do not establish variability across repeated independent runs.

For more tasks:

To expand the evaluation scale, we establish a fixed set of 100 questions on SciBench, with no overlap with either the evolution tasks or the calibration questions. We find that mean accuracy under Token Productivity parent selection is 8.10 percentage points higher than under wealth-based selection, consistent with the original experimental conclusions AppendixI1˙,Figure8,lines1512–1527Appendix I\.1, Figure 8, lines 1512–1527. Furthermore, In the SciBench flywheel experiment, descendants continue purchasing tokens after the founder exited, and the lineage ultimately occupied all five surviving Deriver positionsAppendixI2˙,Table26,lines1538–1553Appendix I\.2, Table 26, lines 1538–1553.

The two tables below both show experiments on the new scientific benchmark.

Table: SciBench | GPT parent-selection experiments

Parent selection ruleHeld-out questions per checkpointCheckpointsMean accuracy (%)Gain over Wealth (pp)
Wealth1001071.8—
Token Productivity (TP)1001079.98.1

Table: Token purchasing and lineage outcomes on SciBench (Appendix I.2, Table 26)

DatasetBackboneDominant lineageFounder purchase roundsFirst round at 100%Final count (share)Final lineage wealth
SciBenchDeepSeek-V4-FlashF03T1, T2, T7T95 of 5 (100%)4.15

For repeated independent runs:

We compile 95 completed dynamic competition trajectories seeAppendixD4˙,Tables11–14,lines921–1048see Appendix D\.4, Tables 11–14, lines 921–1048 ,covering different basemodel combinations such as gpt-6-luna vs gpt-5\.6-luna, ds-flash vs ds-pro, ds vs gpt-luna and so on\.\.\. and report four repeated runs for each of the four flywheel configurations combining two model families and two datasets, providing means, standard deviations, and individual run results seeAppendixH4˙,Tables22–24,lines1311–1399see Appendix H\.4, Tables 22–24, lines 1311–1399. Below are excerpts from the run records. Please see the appendix for details.

Table: Final dominant lineage shares across GPT-5.6-luna flywheel runs

BenchmarkRunsMean final shareSD (pp)
FrontierScience4100.00%0
Yale Physics485.84%16.76

Table: Competition outcomes across token allocations

Token budgets are listed as stronger model / weaker model. Each cell reports the number of runs in which the stronger model had more / the same number of / fewer surviving Derivers after 30 training tasks.

Stronger / weaker modelBenchmark2,048 / 2,0482,048 / 6,1442,048 / 8,1926,144 / 8,1928,192 / 8,192
DeepSeek-V4-Pro / DeepSeek-V4-FlashFrontierScience3–0–02–0–10–0–41–0–21–0–0
DeepSeek-V4-Pro / DeepSeek-V4-FlashYale Physics2–1–03–0–00–0–41–0–23–0–0
GPT-6-luna / GPT-5.6-lunaFrontierScience3–0–00–0–30–1–33–0–03–0–0
GPT-6-luna / GPT-5.6-lunaYale Physics3–0–00–0–31–0–32–0–12–1–0
GPT-6-luna / DeepSeek-V4-ProFrontierScience3–0–01–0–20–0–31–1–10–1–0
GPT-6-luna / DeepSeek-V4-ProYale Physics3–0–01–0–20–0–32–0–12–0–0
GPT-5.2 / GPT-5.1FrontierScience1–0–00–0–10–0–1——
GPT-5.2 / GPT-5.1Yale Physics1–0–00–1–00–0–1——

For evolution runs:

Regarding the evolution horizon, 30 rounds are already sufficient to demonstrate the competitive differentiation, lineage expansion, and reinvestment feedback examined in this paper and to draw the corresponding experimental conclusions. For example, combining the results across the two datasets, the eight flywheel runs for DeepSeek and the eight for GPT yield mean population proportions of the dominant lineage at round 30 of 96.25% and 92.92%, respectively seeAppendixH4˙,Tables22–24,lines1311–1399see Appendix H\.4, Tables 22–24, lines 1311–1399. This shows that, within 30 rounds, multiple runs have already established clear lineage advantages accompanied by continued reinvestment by descendants, sufficiently demonstrating and supporting the key dynamics discussed in this paper.

AuthorsOpenReview ↗

Response to Reviewer 3mJK (PART II)

2

The central mechanisms still remain insufficiently disentangled as icreasing token budgets might change contribution quality, reward judgments, bidding outcomes, and reproductive opportunities simultaneously.

We thank the reviewer for the suggestion to disentangle the mechanisms. Here, we need to clarify that contribution content, reward outcomes, bidding behavior, and reproductive opportunities are interconnected through feedback. They are not tunable hyperparameters but responses to changes in token resources. Even if the economic rules remain unchanged, they may jointly respond to token resources. Their simultaneous changes do not mean that the experiment changes multiple external settings at once.

For example, a budget change may alter the contribution generated by an agent, thereby affecting its score and reward under the same evaluation rules. Changes in wealth may then affect its subsequent participation, survival, and reproductive opportunities, ultimately changing the population participating in subsequent tasks and the competitive environment. This is one possible pathway through which resources affect evolution, and it is also the economic feedback we seek to preserve and study.

Forcing rewards, bidding outcomes, and reproductive outcomes to remain unchanged while changing the budget would interrupt some of these pathways. Therefore, in our paired intervention comparisons, we match the pre-intervention states and keep the evaluation and update rules consistent within the corresponding economic rule, while allowing agent behavior and population structure to respond to resource changes. This design examines the consequences of resource interventions within the full economic feedback process.

3

ALso under unequal budgets, weaker models already achieve comparable or better performance in several settings, so their subsequent dominance does not by itself demonstrate that economic selection favors inferior contributions.

We agree with part of your interpretation: the market can indeed favor high-performing combinations of models and token resources. However, the Gresham-like effect emphasizes the role of token resources. Beyond enabling weaker models to outperform stronger ones in static evaluations, token advantages can promote further reproduction in evolving economies, allowing models (no matter weak or strong) with larger token budgets to dominate the population. This resembles the predominance of “bad money drives out good” in circulation described by Gresham’s law. For example, our token flywheel experiments illustrate how token resources and accumulated wealth can reinforce one another, driving the rapid expansion of a single lineage.

Meanwhile, we want to claim that what performs well in a static setting (Table8) is different from what performs well during dynamic evolution. During evolution, we want the system to learn from previous high-quality responses and evolve accordingly. So the static "model+resource" performing well may not work in dynamic situation. We therefore propose Token Productivity to enable it to retain high-quality responses with high information density. As shown in our updated Figure 3, the original evolutionary mechanism harmed performance. We call for agent economies to use Token Productivity to guide the retention of high-quality models suited to dynamic evolution, or to introduce a rollback mechanism, as discussed in Section 7.

AuthorsOpenReview ↗

Response to Reviewer 3mJK (PART III)

4

Wealth based parent selection, bankruptcy rules, prompt mutation, and the chosen reward structure all influence which advantages become heritable population differences, but their roles are not systematically tested. So the conclusion seem to depend heavily on one particular economic design.

For wealth based parent selection and reward structure:

In our original experiments, we replace wealth-based parent selection with Token Productivity AppendixF3˙,Table17Appendix F\.3, Table 17 to improve the selection mechanism and increase overall accuracy. Regarding the reward structure, our token flywheel experiments combine 0.4 × individual score with 0.6 × team score, reflecting diversity in our design AppendixH3˙,Table21Appendix H\.3, Table 21.

Table: Token Productivity vs. Wealth-Based Parent Selection

Model familyBenchmarkWealth (%)Token Productivity (%)Improvement (pp)
DeepSeekFrontierScience16.3221.054.7
DeepSeekYale Physics45.7851.385.6
GPTFrontierScience5.797.892.1
GPTYale Physics37.5439.131.6

Descendant Returns under the Flywheel Reward Structure

ModelBenchmarkLineageNet return of descendants surviving at T30Net return of all descendants
DeepSeek-V4-FlashFrontierScienceF041.308−6.141
DeepSeek-V4-FlashYale PhysicsF025.2537.343
GPT-5.2FrontierScienceF030.277−1.488
GPT-5.2Yale PhysicsF023.574.607

For prompt mutation:

We change the temperature used for prompt-mutation requests from 0.9 to 0.8, and the observed results remained consistent with the Token Flywheel conclusion. The F02 founder increased its quota to 4,096, 6,144, and 8,192 tokens at T1, T2, and T3, respectively. Descendants begin purchasing tokens at T4 and completed another eight paid upgrades after the founder exits at T24. Over the full run, F02 completes 36 paid upgrades, 33 of which are made by descendants, spending a total of 0.79125 wealth units and accounting for 50 of the 56 Deriver actions. F02 occupies all Deriver positions from T18 onward, reaching 11/11 at T30, with final lineage wealth of 5.1241. See Appendix K (lines 1577–1597).

Table: Token flywheel outcomes at a requested mutation temperature of 0.8

DatasetBase modelLineageFounder’ s paid upgrade roundsFirst round occupying all Deriver positionsT30 lineage population (proportion)T30 lineage wealth
PHYSICS (Yale)DeepSeek-V4-FlashF02T1, T2, T3T1811/11 (100%)5.1241

Regarding bankruptcy rules:

We change the rent collection frequency from once every three rounds to once every four rounds, and the results remain consistent with our experimental conclusions. The F02 founder increase its quota to 4,096 and 6,144 tokens at T2 and T3, respectively. Descendants begin purchasing tokens at T5 and complete another eight paid upgrades after the founder goes bankrupt due to rent deductions at T24. Over the full run, F02 completes 31 paid upgrades, 29 of which are made by descendants, spending a total of 0.68875 wealth units and accounting for 44 of the 57 Deriver actions. See Appendix J (lines 1554–1576).

Table: Token flywheel outcomes with rent collected every four rounds

DatasetBase modelLineageFounder’ s paid upgrade roundsT30 lineage population (proportion)T30 lineage wealth
PHYSICS (Yale)DeepSeek-V4-FlashF02T2, T314/16 (87.5%)6.649
Reviewer_5p1oOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how token allocation affects performance and population dynamics in evolving LLM-agent economies. Through experiments on persistent inequality, temporary budget shocks, and token purchasing, the authors show that token-rich weaker models may displace stronger models, temporary restrictions may have lasting effects, and early token investment may create self-reinforcing lineage advantages. The paper also proposes Token Productivity, which improves average performance over wealth-based parent selection in the tested settings

Critical strengths

S1. The paper raises an important and underexplored point: token allocation can affect not only immediate performance but also wealth, reproduction, and long-term population composition. This is highly relevant to the design and evaluation of evolving agent systems.

S2. The use of fixed-population controls and paired purchase restrictions strengthens the causal interpretation of the main findings. Token Productivity also provides a practical mechanism-design contribution.

Critical weaknesses

W1. The results appear to rely on a limited number of runs, small evaluation sets, and no extensive uncertainty or significance analysis. Given the stochastic and path-dependent nature of agent economies, this substantially limits confidence in the robustness of the reported effects.

W2. The experiments are restricted to two scientific benchmarks, one economic framework, and primarily one agent role. It remains unclear whether the findings generalize to other tasks, reward systems, reproduction rules, and resource-allocation mechanisms.

Questions for authors

Q1. By Table 1, weak@8,192 matches or exceeds strong@2,048 in three of four settings. Is the “Gresham-like” displacement then just the market selecting the better-performing configuration? Is there evidence that it harms collective performance?

Q2. The debate result is consistent with LLM-judge verbosity bias [1,2]. Do token advantages earn rewards partly through output length?

References.

[1] Zheng et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS D&B.

[2] Khan et al. (2024). Debating with More Persuasive LLMs Leads to More Truthful Answers. ICML.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 5p1o (PART I)

We sincerely thank the reviewer for the positive assessment of our work and the thoughtful feedback. We appreciate your encouraging comments and constructive suggestions, and address your questions and concerns point by point below.

1

The results appear to rely on a limited number of runs, small evaluation sets, and no extensive uncertainty or significance analysis.

We have expanded our experiments and reporting of results in three ways.

For more evaluation sets:

First, we add SciBench experiments with a fixed set of 100 questions that do not overlap with either the evolution tasks or the calibration questions (see Appendix I, lines 1462–1488), expanding the original evaluation scale. On these 100 questions, mean accuracy under Token Productivity parent selection is 8.10 percentage points higher than under wealth-based selection, consistent with our experimental conclusions.

Table : SciBench | GPT parent-selection experiments

Parent selection ruleHeld-out questions per checkpointCheckpointsMean accuracy (%)Gain over Wealth (pp)
Wealth1001071.8—
Token Productivity (TP)1001079.98.1

For more basemodel:

Second, the revised manuscript provides a complete summary of 95 completed dynamic competition trajectories seeAppendixD4˙,Tables11–14,lines921–1048see Appendix D\.4, Tables 11–14, lines 921–1048, covering different basemodel combinations such as gpt-6-luna vs gpt-5\.6-luna, ds-flash vs ds-pro, ds vs gpt-luna and so on\.\.\. and token quotas. Under equal quotas, stronger models generally retain more surviving Derivers. Increasing the weaker model’s quota produces competitive reversals in some configurations, supporting the conclusion that token quotas can change competitive outcomes.

Token budgets are listed as stronger model / weaker model. Each cell reports the number of runs in which the stronger model had more / the same number of / fewer surviving Derivers after 30 training tasks.

Table: Competition outcomes across token allocations

Stronger / weaker modelBenchmark2,048 / 2,0482,048 / 6,1442,048 / 8,1926,144 / 8,1928,192 / 8,192
DeepSeek-V4-Pro / DeepSeek-V4-FlashFrontierScience3–0–02–0–10–0–41–0–21–0–0
DeepSeek-V4-Pro / DeepSeek-V4-FlashYale Physics2–1–03–0–00–0–41–0–23–0–0
GPT-6-luna / GPT-5.6-lunaFrontierScience3–0–00–0–30–1–33–0–03–0–0
GPT-6-luna / GPT-5.6-lunaYale Physics3–0–00–0–31–0–32–0–12–1–0
GPT-6-luna / DeepSeek-V4-ProFrontierScience3–0–01–0–20–0–31–1–10–1–0
GPT-6-luna / DeepSeek-V4-ProYale Physics3–0–01–0–20–0–32–0–12–0–0
GPT-5.2 / GPT-5.1FrontierScience1–0–00–0–10–0–1——
GPT-5.2 / GPT-5.1Yale Physics1–0–00–1–00–0–1——

For robustness on multi random seed:

Finally, the flywheel experiments report four repeated runs for each of the four configurations combining DeepSeek-V4-Flash and GPT-5.6-luna with the two original datasets, totaling 16 trajectories seeAppendixH4˙,Tables22–24,lines1311–1399see Appendix H\.4, Tables 22–24, lines 1311–1399. We also provide the mean and sample standard deviation of the dominant lineage’s final population proportion, together with individual run results, explicitly showing variability across runs. In all four configurations, the dominant lineage’s mean final population proportion reaches 85.84% or higher, supporting the conclusion of dominant-lineage expansion in the flywheel experiments.

Table: Token purchases and dominant lineage outcomes in flywheel runs

Model / DatasetRunDominant lineageFounder’ s purchase roundFirst round occupying all Deriver positionsFinal lineage count / total Derivers
V4-Flash / FrontierScienceOriginal runF04T1T48/8 (100%)
GPT-5.6-luna / FrontierScienceRepeat 1F01T6T122/2 (100%)
GPT-5.6-luna / PHYSICSRepeat 1F01T2Not reached10/11 (90.9%)
AuthorsOpenReview ↗

Response to Reviewer 5p1o (PART II)

2

The experiments are restricted to two scientific benchmarks, one economic framework, and primarily one agent role. It remains unclear whether the findings generalize to other tasks, reward systems, reproduction rules, and resource-allocation mechanisms.

We conduct additional validation across three dimensions: tasks, economic mechanisms, and agent roles.

For more scientific benchmark:

On SciBench, a university-level scientific computation benchmark, mean accuracy under Token Productivity–based parent selection is 79.90%, exceeding wealth-based selection by 8.10 percentage points, supporting our original conclusion on parent selection. In the flywheel experiment, descendants continued purchasing tokens after the founder exited, and the lineage ultimately occupied all five surviving Deriver positions, consistent with the original flywheel results AppendixI1˙,Figure8,lines1512–1530;AppendixI2˙,Table26,lines1538–1553Appendix I\.1, Figure 8, lines 1512–1530; Appendix I\.2, Table 26, lines 1538–1553.

The table below and the table "SciBench | GPT parent-selection experiments" for question one both show experiments on the new scientific benchmark.

Table: Token purchasing and lineage outcomes on SciBench (Appendix I.2, Table 26)

DatasetBackboneDominant lineageFounder purchase roundsFirst round at 100%Final count (share)Final lineage wealth
SciBenchDeepSeek-V4-FlashF03T1, T2, T7T95 of 5 (100%)4.15

For more Economic framework:

Few multi-agent economic frameworks currently exist, and we actively search for alternatives to test whether our findings generalize. We test resource shocks and reinvestment in Energy Society using the knowledge question answering dataset MMLU-Pro-Stratified. Participants autonomously select questions and pay for the tokens. The prize for each question is split equally among participants who answer it correctly. We find that after token quotas are restored, participants whose quotas have been reduced still earn less than the control group, supporting the conclusion that resource shocks have persistent effects AppendixL1˙,lines1613–1637Appendix L\.1, lines 1613–1637. Furthermore, allowing purchases of additional quotas increased cumulative task rewards by 18.76%, supporting the conclusion that reinvestment increases cumulative returns in the flywheel experiments AppendixL2˙,lines1639–1662Appendix L\.2, lines 1639–1662.

Table: Resource Restoration in Energy Society (Appendix L.1,Table 29)

ConditionShock accuracy (%)Recovery accuracy (%)Recovery successes / attemptsRecovery income (energy units)
A1: Frozen / uncut909014 / 206,650.00
A2: Frozen / cut859011/20.6,650.00
B1: Dynamic / uncut93.3393.7514 / 207,900.00
B2: Dynamic / cut88.3386.2514 / 204,516.67

Table: Reinvestment Outcomes in Energy Society (Appendix L.2, Table 30)

MetricPurchases allowedNo purchases
Final total energy48,441.1342,660.98
Cumulative task rewards50,000.0142,100.01
Generated tokens551,004462,686
Successful / total task attempts117 / 150112 / 150
Test accuracy under a uniform token allowance (%)9292
Reward per 1,000 generated tokens90.7490.99

For more agent roles:

We initially focus on Deriver agents because they play a central role in the system and their performance is less sensitive to output length. We extend the competition and flywheel experiments from Deriver agents to Answer agents, which are responsible for producing the final answer. Experiments show that when DeepSeek-V4-Flash receives 8,192 tokens and its opponent V4-Pro receives 2,048 tokens, it occupies all three surviving Answer positions, consistent with the original result in which the weaker model outcompeted the stronger model. In the flywheel experiment, descendants continue purchasing tokens after the founder exited, and the lineage ultimately occupies all three surviving Answer positions, consistent with the original flywheel results (Appendix E, Tables 15–16, lines 1062–1088).

Table: Answer agent competition experiment (Appendix E, Table 15)

DatasetModelToken quota per agentInitial populationSurvivors after 30 roundsFinal wealthCumulative actions
FrontierScienceDeepSeek-V4-Flash8,192232.256235
FrontierScienceDeepSeek-V4-Pro2,0482009
AuthorsOpenReview ↗

Response to Reviewer 5p1o (PART III)

Table: Answer agent flywheel experiment (Appendix E, Table 16)

DatasetModelDominant lineageFounder’ s purchase and exitDescendants’ purchase roundsFirst round occupying all surviving positionsFinal lineage count / total survivors
FrontierScienceGPT-5.6-lunaF03Purchased at T6 and exited in the same roundT28, T29, T30T93 / 3 (100%)

3

By Table 1, weak@8,192 matches or exceeds strong@2,048 in three of four settings. Is the “Gresham-like” displacement then just the market selecting the better-performing configuration?

We agree with part of your interpretation: the market can indeed favor high-performing combinations of models and token resources. However, the Gresham-like effect emphasizes the role of token resources. Beyond enabling weaker models to outperform stronger ones in static evaluations, token advantages can promote further reproduction in evolving economies, allowing models (no matter weak or strong) with larger token budgets to dominate the population. This resembles the predominance of “bad money drives out good” in circulation described by Gresham’s law. For example, our token flywheel experiments illustrate how token resources and accumulated wealth can reinforce one another, driving the rapid expansion of a single lineage.

Is there evidence that it harms collective performance?

Our concern is that the current situation of weaker models beating stronger models may narrow the range of capabilities available to the system in the future. Table 8 and 9 line811−832line 811-832 show that stronger models still achieve higher task performance under the same token budget. Therefore, if the evolved economy subsequently faces more difficult tasks, or if model budgets are increased, the capability ceiling of the entire economy will be lower. We therefore believe this leads to a long-term loss of capability.

Meanwhile, what performs well in a static setting is different from what performs well during dynamic evolution. During evolution, we want the system to learn from previous high-quality responses and evolve accordingly. We therefore propose Token Productivity to enable it to retain high-quality responses with high information density. As shown in our updated Figure 3, the original evolutionary mechanism harmed performance. We call for agent economies to use Token Productivity to guide the retention of high-quality models suited to dynamic evolution, or to introduce a rollback mechanism, as discussed in Section 7.

4

The debate result is consistent with LLM-judge verbosity bias [1,2]. Do token advantages earn rewards partly through output length?

Thank you for this interesting question. We agree that token advantages may partly influence evaluation outcomes through longer outputs, and that higher win rates or rewards cannot be attributed entirely to improvements in actual contribution quality. Actually, in our debate setting (Figure 1), the weaker model argues for an incorrect answer. Its increased win rate therefore does not indicate greater answer correctness. Instead, it shows that a resource advantage may translate into a competitive advantage under a particular evaluation mechanism. Moreover, we also use datasets with non-subjective evaluation to prove the conclusion. For example, for tasks with reference answers, such as Yale (PHYSICS), we use symbolic equivalence checking to evaluate the correctness of the final answer.

The central phenomenon in our paper is that token advantages lead weaker models to receive higher scores and thereby defeat stronger models. We identify this phenomenon, explain its implications, and call for analysis of its underlying causes. There may be many reasons for this: a weaker model may happen to hit upon the correct answer because it can produce more text; longer chains of thought may improve performance; or, of course, the judge may exhibit the verbosity bias you mentioned. We therefore find your point very reasonable and believe it provides strong analytical support for our work. Previous studies such as [1][2] have discussed these evaluation issues. Our paper shows that, when evaluation outcomes further influence wealth, resource access, and reproductive opportunities, these biases may also be amplified by economic feedback. Identifying these mechanisms helps explain why agent economies can produce outcomes in which resource advantages do not align with actual capability.

[1] Zheng et al. 20232023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS D&B.

[2] Khan et al. 20242024. Debating with More Persuasive LLMs Leads to More Truthful Answers. ICML.

Reviewer_VadHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper using multiagent systems with an economy around token budgets (or external resources, like API budgets) and how effecting that economy can effect the performance of these multiagent systems especially when they have strong and weak models. Specifically, the paper uses roles such as Literature, Planner, Deriver, Verifier, and Answer agents to solve problems like FrontierScience and PHYSICS (problems where you have a dense reward, 0 to 1, and you can iteratively improve your solution). Agents bid for opportunities to contribute to a shared task, rent is charged to each agent, and rewards are given to the agents at the end of the training task. So essentially agents must contribute and do so productively otherwise they will run out of funds and be removed from the multi-agent system.

The paper asks 3 core questions. What happens when you budget match agents (i.e. budget the total cost in running them) vs token match them (do strong models succeed or fail in this economy)? What happens when token budgets suddenly decrease and models are unable to use the same amount of tokens as prior rounds? What happens when you allow agents to buy additional tokens from their excess rewards?

The paper explores these questions on DeepSeek-V4-Flash, GPT-5.4 and 5.2. They first find that under budget constraints (i.e. API costs) compute matching models causes models that are weaker (i.e. GPT 5.2 vs GPT 5.4) to be favored in the economy and stronger models to eventually be removed. They fix this by purposing Token Productivity which is just the quality of the models answer divided by the tokens, this improves the performance of the multiagent system while keeping the budget the same.

Next, they find that if suddenly agents are restricted to using fewer tokens (i.e. in some round the max token budget goes from 8k to 1k) then the performance of the system decreases, and even if the budget goes back to what it was eventually the performance does not seem to always recover. In economies where agents can be removed and not reintroduced permanent effects can occur leading to losses compared to a control setting where the shock never happened. In fixed economies (meaning no agent is removed from the pool to account for the drought of tokens), the performance can recover or improve in all but one setting GPT-5.2 / PHYSICS. During the shock, with fixed agents, the performance drop is not as severe either.

Finally, they find that when agents can purchase token budgets from excess rewards to use later on, these agents can produce better results with these investments ultimately giving their descendants (models in future rounds) more wealth to use on buying more tokens and thus more rewards etc. However, restricting purchases (even on just 1 round) can hurt the lineage of these agents (i.e. an agent loses their ability to use their wealth as investments and thus is removed from the system).

Overall the findings show 3 unique insights into how the structure of an economy of agents can impact their performance.

Critical strengths
  • Economies are exciting and these questions are important to ask and determine. I think each of the three core findings are interesting and somewhat related to real scenarios (i.e. we will use models of varying strengths, have token shortages and want agents to be able to control their budgets to a degree).
  • Performance gain from the Token Productivity change is interesting, but a bit hard to follow and I am unsure if it generalizes. I also liked reading about how the stronger models matched under API cost budgets were slowly removed from the system and weaker models dominated.
Critical weaknesses

TL:DR; Primarily this papers weakness is in the writing. If rewritten I think it could be accepted. But as of now it looks like 3 different loosely connected experiments stapled together (each of which could be an entire paper btw) with somewhat inconclusive findings and LLM-generated prose that makes it hard to understand.

  1. The quality of writing is troubling. We need a methods section. Clearly defined terminology and consistently used terminology. The paper as of now is incredibly difficult to read. I would rewrite this paper entirely (just to fix the prose for legibility)

Explanation: I think the whole paper would need to be rewritten by a human or very delicate LM. There are overloaded terms, the experiments are not well introduced, the tables and figures are extremely complex to understand etc.

For example: "Lower prices tend to support larger Deriver populations and higher wealth shares"

This would infer that "shares" refers to the amount of reward an agent gets from a task.

"The target lineage’s final share decreases in all four comparisons (Table 5)"

table 5 however uses "share" to mean the number of agents in the population that are from that lineage?

I believe wealth/reward changes to "cumulative purchase costs in wealth units" But then in figure 5 wealth is defined as cumulative cost?

Another example,

Under equivalent token restrictions, the performance gap is larger in an active agent market than in a fixed population baseline: B1 − B2 > A1 − A2 during the reduction window

This is in the intro, we have no idea what B1, B2, A1, A2 is, nor if this is a good condition or a bad condition. The names for these settings are arbitrary, which I dislike but that's personal taste, what I cannot forgive is not introducing what these variables are before basing a result on them.

I found this paper difficult to read because new terms are being introduced or changing meaning.

  1. There is little consistency between these three questions and not a very clear outcome of them could be fixed by rewriting the paper but more likely needs a clearer result / experiment

Explanation: Sometimes science is like that, but the results feel hedged. I feel like section 7 just left me confused and without much to go on. For example, "what should economic selection reward" I think is saying that teams of agents are supposed to collaborate but the reward they collaborate for is given to a single agent, therefore, there's a conflict of incentives. But this wasn't tested?

Multi-agent system design must therefore align individual incentives with team success, rather than assuming that population growth equals better capability. This alignment does not mean banning market dominance or always favoring stronger models; weaker models with extra tokens are often more cost-effective. Instead, we should evaluate the final agent population by how well they perform tasks together using independent testing standards.

This feels unsupported. The paper did not really address this question. It's interesting that smaller models are more cost-effective than big models but this is a routing problem (it's not really the novel bit here, it's that strong models get kicked out of the population). Then finally, that last sentence doesn't really offer anything. Of course we should evaluate the population? What does that look like? How is that different from what you are doing in section 4? What are these independent testing standards?

For each experiment I was left somewhat unsure of the impact of each result. With the exception of Token Productivity (section 4).

Questions for authors

Please see my weaknesses.

  • I am confused by how agents are removed and re-entered. Maybe this is pulled over from Economy of Minds. I would restate how that works clearly.
  • Less of a question and more of a taste thing but figure 3 should be a table, figure 4 should just be bar graphs. Some of this is just hard to understand because of how the results are presented. I think table 2 could also have variables like F_{ns} and F_{s}, E_{ns} and E_{s} and that would make a lot of the future results easier to read too (especially since you compound these with deltas of means and then deltas of those deltas). Table 4 is also difficult to understand but some reworking / added columns, alignment etc. might help.
  • What's a share?
  • Did you explore the condition of all models on each task? I am curious if you always kick out strong models or if its just the setting where you have models close in performance and model family.
  • Why did GPT-5.4 / FS do better than the condition (no shock) during recovery. That seems surprising! How do fixed economies work? I was utterly perplexed by condition A.
12 comments
AuthorsOpenReview ↗

Response to Reviewer VadH (PART I)

Response

1

On the rewriting and clarification of our paper.

We are very grateful that you liked the idea and experiments in our paper. To address the writing and presentation issues, we have spared no effort and rewritten almost the entire paper to improve its clarity.

The specific changes are as follows:

  1. In Section 1, we condense the original three questions into a single core question that brings together the three subsequent experiments to answer the same central question (line 92): “How do token resources shape competitive outcomes in evolving agent economies when budgets are persistently unequal, temporarily restricted and restored, or expanded through agents’ investment of earned wealth?” We also explicitly state that we design three experiments to jointly answer this question (lines 95–102).

  2. Immediately after the core question, we rewrite the experimental overview, clarify the organization of Sections 3–6, and summarize the conclusions (lines 104–129). This establishes a logical chain across the three experiments and allows readers to understand their connections before reading the detailed experiments.

  3. In lines 114–117, we remove the previously undefined symbols A1, A2, B1, and B2 and replace them with a textual description of the experimental observations, allowing readers to understand them clearly at a glance.

  4. In Section 3.1, we explain in detail the mechanisms of EOM, one of the frameworks we use, and define variables used throughout the paper, such as wealth and budget (lines 169–184).

  5. In Section 3.3, we clearly define the performance of the overall economy, each agent’s wealth and population proportions, which are used in the subsequent sections (lines 200–226).

  6. We add our method, Token Productivity, to Section 3.4, explaining how it is calculated and the reasons behind its design.(line 232-250)

  7. In Section 4, we describe the experimental settings more clearly (lines 262–264 and 268–269) and incorporate the notation defined in Section 3 to make the descriptions clearer (lines 309–311). Meanwhile, although “share” refers to a proportion, its multiple possible interpretations made the wording unclear. We therefore have removed or replaced all ambiguous uses of “share.”

  8. We are very grateful for your positive response and are also willing to trust your taste in academic papers. We therefore replace Figure 3 with a tableline270−295line 270-295 and Figure 4 with a bar graph 324−337324-337. This has indeed made the presentation of information and the layout much clearer and more visually appealing!

  9. We rewrite almost all of Section 5, clearly explaining the procedure and design of the temporary shock (lines 345–353). We use a table to describe how the quantities, such as ΔA\Delta_A and Δamp\Delta_{\mathrm{amp}}, are calculated. Since these comparisons can be somewhat difficult to follow, we also directly provide their intuitive meanings (lines 355–366). We then rewrote each paragraph in Section 5.2 in the format of “experimental observation + conclusion + optional example,” greatly improving clarity and readability (lines 370–394).

  10. We rewrite almost all of Section 6. At the beginning, we clearly explain the reasons and logic behind the token flywheel (lines 417–422). In Section 6.1, we explain our experimental design and why we conduct counterfactual experiments.to vary only the token variable (lines 431–474). We also restate the experimental conclusions together with examples to make the whole discussion clearer (lines 475–485). We then present the experiments on token price, a hyperparameter of token investment, in Section 6.2 (lines 500–512).

  11. We rewrite Section 7 to remove ambiguity. The original paragraphs may have led you to think that the focus was on “population,” or that the central issue concerned “cost-effectiveness” or a conflict between “the individual and the whole.” That was not what we meant, so we have restated the discussion. The paper now presents two implications. First, economic evolution should assess both capability and token resources; the issue concerns how an economy evaluates agents and evolves. Second, economic frameworks should be robust to resource shocks; the issue concerns how an economy can remain robust when facing shocks or flywheel dynamics. These revisions remove the original ambiguity.

2

On the unclear definition of share, wealth, reward and cumulative purchase.

We have removed all occurrences of the word “share” because it is ambiguous. We also clearly define wealth, budget, reward, and other relevant terms in Section 3.

AuthorsOpenReview ↗

Response to Reviewer VadH (PART II)

3

On the problem of A1 A2 B1 B2 without definition in intro.

In lines 114–117, we have removed the previously undefined symbols A1, A2, B1, and B2 and replaced them with a textual description of the experimental observations, allowing readers to understand them clearly at a glance.

4

On the clarification of Section 7

We rewrite Section 7 to remove ambiguity. The original paragraphs may have led you to think that the focus was on “population,” or that the central issue concerned “cost-effectiveness” or a conflict between “the individual and the whole.” That was not what we meant, so we have restated the discussion. The paper now presents two implications. First, economic evolution should assess both capability and token resources; the issue concerns how an economy evaluates agents and evolves. Second, economic frameworks should be robust to resource shocks; the issue concerns how an economy can remain robust when facing shocks or flywheel dynamics. These revisions remove the original ambiguity.

5

On the removing and re-entering mechanism of EOM

In Section 3.1, we describe how EOM operates in detail, including its removal and re-entry mechanisms.

"Bankruptcy is checked every three rounds. Bankrupt agents are removed and regenerated after reflection, with initial wealth set to 0.20.2. Reflection revises failed prompts using task feedback and reference answers. New agents may also be created periodically. Default reproduction uses wealth-based selection to generate offspring by modifying the wealthiest agent's prompt. Descendants retain their source agent's role and lineage, and missing required roles are replenished. Task execution, wealth updates, elimination, and reproduction jointly drive economic evolution. "

6

On the connection of three main sets of experiments

In Section 1, we condense the original three questions into a single core question that brings together the three subsequent experiments to answer the same central question (line 92):

“How do token resources shape competitive outcomes in evolving agent economies when budgets are persistently unequal, temporarily restricted and restored, or expanded through agents’ investment of earned wealth?”

We also explicitly state that we designed three experiments to jointly answer this question (lines 95–102).

Meanwhile, immediately after the core question, we rewrite the experimental overview, clarify the organization of Sections 3–6, and summarize the conclusions (lines 104–129). This establishes a progression across the three experiments and allows readers to understand their connections before reading the detailed experiments.

7

On the modification of Figure 3 to Table, Figure4 to bar graph , Table 2 for clearer variable names and Table 4 for more clearer presenting.

We are very grateful for your positive response and are also willing to trust your taste in academic papers. We therefore replaced Figure 3 with a table 270−295270-295 and Figure 4 with a bar graph 324−334324-334. This has indeed made the presentation of information and the layout much clearer and more visually appealing!

In Table 2, we describe how the quantities, such as ΔA\Delta_A and Δamp\Delta_{\mathrm{amp}}, are calculated. Since these quantities can be somewhat difficult to follow, we also directly provide their intuitive meanings (lines 340–351).

We add a paragraph presenting the information in Table 4 in the format of “main conclusion + experimental example,” making it clear and intuitive (lines 400–403).

AuthorsOpenReview ↗

Response to Reviewer VadH (PART III)

8

On exploring the condition of all models on each task

Thank you for suggesting that we expand the coverage of models and benchmarks.

For BaseModels:

We have added a large number of experiments examining different token-budget allocations between DeepSeek-V4-Pro and GPT-6-luna. GPT-6-luna is the more capable model. We observe that the stronger model is more likely to be eliminated when allocated fewer tokens. As its token budget gradually increases, its position in the agent ecosystem becomes stronger than that of the weaker model. The experimental results are as follows:

Specifically, we have added records from 59 competition experiments on FrontierScience and Yale Physics to Table 1 in Section 4.2 (lines 270–288). These cover two comparisons, DeepSeek-V4-Pro versus GPT-6-luna and GPT-5.6-luna versus GPT-6-luna, across five token-budget allocations. Tables 12–13 in Appendix D.3 line957−1035line957-1035 list the results of individual runs. When both models receive 2,048 tokens, GPT-6-luna holds a majority of the final Deriver population in every recorded run across both comparisons. When GPT-6-luna receives 2,048 tokens and DeepSeek-V4-Pro receives 8,192 tokens, GPT-6-luna becomes extinct in all six recorded runs. The table below reports the number of runs in which GPT-6-luna obtains a final majority out of the total number of recorded runs under four representative token-budget allocations.

Each cell gives the number of runs in which GPT-6-luna finished with more / the same number of / fewer surviving Derivers than its opponent after 30 training tasks. Token budgets are listed as GPT-6-luna / opponent.

Table: GPT-6-luna competition outcomes across token allocations

OpponentBenchmark2,048 / 2,0482,048 / 6,1442,048 / 8,1926,144 / 8,1928,192 / 8,192
DeepSeek-V4-ProFrontierScience3–0–01–0–20–0–31–1–10–1–0
DeepSeek-V4-ProYale Physics3–0–01–0–20–0–32–0–12–0–0
GPT-5.6-lunaFrontierScience3–0–00–0–30–1–33–0–03–0–0
GPT-5.6-lunaYale Physics3–0–00–0–31–0–32–0–12–1–0

GPT-6-luna finished with a majority in all 12 runs when both models received 2,048 tokens. When GPT-6-luna received 2,048 tokens and its opponent received 8,192, it finished with fewer surviving Derivers in 12 of 14 runs, including extinction in all six runs against DeepSeek-V4-Pro.

For benchmark:

We also extend the token productivity experiments to SciBench (lines 1462–1488) using GPT-5.6-luna and GPT-6-luna. Each checkpoint is evaluated on the same set of 100 questions. Mean accuracy across ten checkpoints increases from 71.80% under wealth-based selection to 79.90% under Token Productivity selection, a difference of 8.1 percentage points. It shows that on this new dataset with a larger test set, the corrective rule based on Token Productivity remains effective. The results are presented in Figure 8 in Appendix I.1.

For robustness on multi random seed:

We also repeat the token flywheel experiments in Section 6.1 using GPT-5.6-luna. Tables 21–22 in Appendix H.4 report four runs for each benchmark (lines 1311–1321). At the end of training, the largest founder lineage accounts for an average of 100.00% of surviving Derivers on FrontierScience and 85.84% on Yale Physics, consistent with the trend of early purchasing followed by lineage expansion and dominance.

Table: Final dominant lineage shares across GPT-5.6-luna flywheel runs

BenchmarkRunsMean final shareSD (pp)
FrontierScience4100.00%0
Yale Physics485.84%16.76

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Rebuttal (PART I)

Rebuttal Summary

Dear Chairs,

We sincerely thank the reviewers for their thoughtful feedback. We are encouraged that the importance of our studying token resources in evolving agent economies, the novel insight between token resource, model capability and competitive success, the interesting result of our controlled intervention, and the value of Token Productivity mechanism.

Consensus Among Reviewers

  • Significance and Relevance: Reviewer VadH recognizes that agent economies raise important questions and finds all three core findings interesting and relevant to realistic settings involving heterogeneous models and token shortages. Reviewer 5p1o describes the effect of token allocation on wealth, reproduction, and population composition as an important and underexplored issue.

  • Conceptual Contribution: Reviewer 3mJK highlights the distinction between model capability and competitive success, recognizing that survival and reproduction do not necessarily identify the most capable agents. Reviewer 5p1o similarly emphasizes that token allocation affects both immediate performance and subsequent economic evolution.

  • Value of Controlled Interventions: Reviewer 5p1o and Reviewer 3mJK both recognize the value of the fixed-population controls and paired purchase restrictions. Their assessments highlight how these comparisons strengthen the evidence that resource access actively shapes population dynamics.

  • Promise of Token Productivity: Reviewer VadH finds the performance improvement from Token Productivity interesting, while Reviewer 5p1o recognizes it as a practical mechanism-design contribution.

Summary of Revisions and Clarifications

We have carefully addressed their feedback through substantial improvements in presentation and additional empirical validation. Although we have added supplementary texts and experiments in response to the reviewers’ feedback, these additions all clarify and strengthen the original work without changing its core research focus, contributions, or main conclusions. We summarize the main clarifications and additional experiments below.

  1. On the rewriting and clarification of our paper (VadH).
  • On the connection among the three experiments. We unified Section 1 around how persistent token inequalities, temporary shocks, and reinvestment shape competitive outcomes, and revised the overview to clarify the experiments’ connections and conclusions.

  • On mechanisms and terminology. Section 3 now explains EOM’s removal and re-entry mechanisms, defines economy performance, agent wealth, and population proportions, and introduces Token Productivity’s calculation and rationale. We also replaced ambiguous uses of “share.”

  • On the clarification of experimental descriptions. In Section 4, we describe the experimental settings more clearly and incorporate the notation defined in Section 3 to make the descriptions clearer. We then rewrote each paragraph in Section 5.2 in the format of “experimental observation + conclusion + optional example.” In Section 6.1, we explain our experimental design and why we conduct counterfactual experiments to vary only the token variable.

  • On the modification of figures and tables. We therefore replace Figure 3 with a table and Figure 4 with a bar graph. Since these quantities can be somewhat difficult to follow, we also directly provide their intuitive meanings. We add a paragraph presenting the information in Table 4 in the format of “main conclusion + experimental example,” making it clear and intuitive.

  • On the clarification of Section 7. The paper now presents two implications. First, economic evolution should assess both capability and token resources; the issue concerns how an economy evaluates agents and evolves. Second, economic frameworks should be robust to resource shocks; the issue concerns how an economy can remain robust when facing shocks or flywheel dynamics.

AuthorsOpenReview ↗

Summary of Rebuttal (PART II)

  1. On the expanded evaluation and robustness across repeated runs (VadH, 5p1o, 3mJK).

For more evaluation sets: We add SciBench experiments with a fixed set of 100 questions that do not overlap with either the evolution tasks or the calibration questions. Mean accuracy under Token Productivity parent selection is 8.10 percentage points higher than under wealth-based selection AppendixI1˙,Figure8Appendix I\.1, Figure 8.

For more basemodel: The revised manuscript provides a complete summary of 95 completed dynamic competition trajectories, covering different basemodel combinations and token quotas. Under equal quotas, stronger models generally retain more surviving Derivers. Increasing the weaker model’s quota produces competitive reversals in some configurations AppendixD4˙,Tables11–14Appendix D\.4, Tables 11–14.

For robustness on multi random seed: The flywheel experiments report four repeated runs for each of the four configurations combining DeepSeek-V4-Flash and GPT-5.6-luna with the two original datasets, totaling 16 trajectories. We provide the mean and sample standard deviation of the dominant lineage’s final population proportion, together with individual run results. In all four configurations, the dominant lineage’s mean final population proportion reaches 85.84% or higher AppendixH4˙,Tables22–24Appendix H\.4, Tables 22–24.

  1. On additional validation across tasks, economic mechanisms, and agent roles 5p1o,3mJK5p1o, 3mJK.

For more tasks: The formal 30 tasks are enough to demonstrate the conclusion. For example, in the SciBench flywheel experiment, within 30 rounds, multiple runs have already established clear lineage advantages accompanied by continued reinvestment by descendants** **AppendixI2˙,Table26Appendix I\.2, Table 26.

For more Economic framework: We test resource shocks and reinvestment in Energy Society using MMLU-Pro-Stratified. We find that after token quotas are restored, participants whose quotas have been reduced still earn less than the control group. Furthermore, allowing purchases of additional quotas increased cumulative task rewards by 18.76% AppendixL1˙–L2˙Appendix L\.1–L\.2.

For more agent roles: We extend the competition and flywheel experiments from Deriver agents to Answer agents. When DeepSeek-V4-Flash receives 8,192 tokens and its opponent V4-Pro receives 2,048 tokens, it occupies all three surviving Answer positions. In the flywheel experiment, the lineage ultimately occupies all three surviving Answer positions (Appendix E, Tables 15–16).

For prompt mutation: We change the temperature used for prompt-mutation requests from 0.9 to 0.8, and the observed results remained consistent with the Token Flywheel conclusion (Appendix K).

Regarding bankruptcy rules: We change the rent collection frequency from once every three rounds to once every four rounds, and the results remain consistent with our experimental conclusions (Appendix J).

  1. On the ablation of Token Productivity 3mJK3mJK.

To examine whether recent contribution quality alone can account for the improvement, we add a Quality-only ablation on SciBench using the same smoothed estimate of recent contribution quality, without token normalization. Across ten checkpoints evaluated on the same 100 questions, mean accuracy is 71.80% for Wealth, 72.40% for Quality-only, and 79.90% for Token Productivity. This ablation provides evidence that token normalization and contribution quality both work. See Appendix M.

  1. On the interpretation of competitive outcomes and economic feedback 5p1o,3mJK5p1o, 3mJK.

We clarify that the market can favor high-performing combinations of models and token resources. The Gresham-like effect emphasizes that token advantages can promote further reproduction and population dominance in evolving economies. We also acknowledge that token advantages may partly influence evaluation outcomes through longer outputs, and that higher win rates or rewards cannot be attributed entirely to improvements in actual contribution quality.

We further explain to reviewers that we match the pre-intervention states and keep the evaluation and update rules consistent, while allowing agent behavior and population structure to respond to resource changes. These comparisons examine the consequences of resource interventions within the full economic feedback process.

In the subsequent discussion, Reviewer 3mJK stated that the extended experiments addressed their biggest concern about limited empirical evaluation and raised their score to 3. Following their suggestion, we also revised the overview figure to improve clarity. We hope these clarifications and additional experiments provide a more complete basis for assessing our contribution.

AuthorsOpenReview ↗

Summary of Rebuttal (PART III)

  1. On concerns about terminology and presentation clarity of rebuttal modification (VadH)

We marked the revisions in blue and included labels such as “original” to make the changes and additional experiments easier to identify during the rebuttal phase. The intention behind this was to make it easier for the reviewer to see the changes in the rebuttal period, but it ended up causing misunderstandings and confusion, as Reviewer VadH points out. So we correct them now. We revised the main text, tables, and captions to integrate the rebuttal changes into a consistent presentation. We removed ambiguous labels such as “original” and “historical,” explicitly identified model comparisons and experimental settings and standardized the reporting of run counts and repeated-run labels.

We have highlighted all modifications in the revised paper in blue, and we hope the Area Chair will consider our rebuttal and the corresponding clarifications. We believe these substantial additions adequately address the reviewers' concerns and justify the positive consensus formed by reviewers.

Thank you again for your time and effort!

Best regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.