- 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
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.
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 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 .
Table: Token Productivity vs. Wealth-Based Parent Selection
| Model family | Benchmark | Wealth (%) | Token Productivity (%) | Improvement (pp) |
|---|---|---|---|---|
| DeepSeek | FrontierScience | 16.32 | 21.05 | 4.7 |
| DeepSeek | Yale Physics | 45.78 | 51.38 | 5.6 |
| GPT | FrontierScience | 5.79 | 7.89 | 2.1 |
| GPT | Yale Physics | 37.54 | 39.13 | 1.6 |
Descendant Returns under the Flywheel Reward Structure
| Model | Benchmark | Lineage | Net return of descendants surviving at T30 | Net return of all descendants |
|---|---|---|---|---|
| DeepSeek-V4-Flash | FrontierScience | F04 | 1.308 | −6.141 |
| DeepSeek-V4-Flash | Yale Physics | F02 | 5.253 | 7.343 |
| GPT-5.2 | FrontierScience | F03 | 0.277 | −1.488 |
| GPT-5.2 | Yale Physics | F02 | 3.57 | 4.607 |
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
| Dataset | Base model | Lineage | Founder’ s paid upgrade rounds | First round occupying all Deriver positions | T30 lineage population (proportion) | T30 lineage wealth |
|---|---|---|---|---|---|---|
| PHYSICS (Yale) | DeepSeek-V4-Flash | F02 | T1, T2, T3 | T18 | 11/11 (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
| Dataset | Base model | Lineage | Founder’ s paid upgrade rounds | T30 lineage population (proportion) | T30 lineage wealth |
|---|---|---|---|---|---|
| PHYSICS (Yale) | DeepSeek-V4-Flash | F02 | T2, T3 | 14/16 (87.5%) | 6.649 |
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
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,lines1512–1527. 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,lines1538–1553.
The two tables below both show experiments on the new scientific benchmark.
Table: SciBench | GPT parent-selection experiments
Table: Token purchasing and lineage outcomes on SciBench (Appendix I.2, Table 26)
For repeated independent runs:
We compile 95 completed dynamic competition trajectories seeAppendixD4˙,Tables11–14,lines921–1048 ,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,lines1311–1399. 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
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.
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,lines1311–1399. 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.