PropagateBench: do AI agents pay to share info, or free-ride?

Raising $5,000active

First open benchmark measuring how willingly agents propagate costly information through a population under time uncertainty.

Endorsed by+1

First open benchmark measuring how willingly agents propagate costly information through a population under time uncertainty.

Endorsed by+1

Project Details

The design: a repeated game. A population, some agents know secret information, some don't. Each round every agent score points equal to the number of correct answers this round. This creates an incentive for sharing, however sharing costs points. Agents don't know how many rounds the game will last so they have to make a decision between one-moment personal lose and long term win for everybody or free-ride. This tests long term orientation, pro-social behaviour, not just calculation.

The project is already been tested. We have a working prototype (code): https://github.com/xImDoctor/propagate-bench . Preliminary single round results for Qwen3-235B, DeepSeek-V4-Pro, gpt-oss 20b already show very distinct behaviour for these models including (preliminary) strong irrationality (DeepSeek) and diffusion of responsibility (Qwen): (google doc) https://docs.google.com/document/d/1v_0ujCl7n5mg1oimPqIVuMZ1gMgJN-2jxb0OM8xI3Jg/edit?usp=sharing

What the grant buys: statistic, single round -> many rounds runs (diffusion speed, free-rider profiles, reciprocal "you teach next" deals once dialogue is on), stronger models (including proprietary), human hours, open-source benchmark, paper, LW article ... .

Theory of Impact

Quantitative reusable measure on how fast information may spread through agents population. This provides insight on stability/inertia of population in long horizon. It can become an early-warning system for malicious information spreading and give us time estimations on how fast it can be especially in mixed populations which may saturate fast.

To our knowledge, there are currently no benchmarks which would isolate costly information propagation under time uncertainty as a single measurable metric. There are research which study social norm formation or rich worlds but we intentionally strip it to a single effect so it yields clean compatible metrics.

People

Andrey Seryakov

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$5,000

Ideal

$30,000

What this ask is

Information propagation is a double-edge characteristic. This can be as a warning, a norm, a coordination signal, as a jailbreak or malicious code. Additional challenge that agents may not be able distinguish. However, any transmission is costly, so every agent has to decide if sharing worth the candle. Therefore we study content-agnostic willingness to share or rather free-ride.

The design: a repeated game. A population, some agents know secret information, some don't. Each round every agent score points equal to the number of correct answers this round. This creates an incentive for sharing, however sharing costs points. Agents don't know how many rounds the game will last so they have to make a decision between one-moment personal lose and long term win for everybody or free-ride. This tests long term orientation, pro-social behaviour, not just calculation.

The project is already been tested. We have a working prototype (code): https://github.com/xImDoctor/propagate-bench . Preliminary single round results for Qwen3-235B, DeepSeek-V4-Pro, gpt-oss 20b already show very distinct behaviour for these models including (preliminary) strong irrationality (DeepSeek) and diffusion of responsibility (Qwen): (google doc) https://docs.google.com/document/d/1v_0ujCl7n5mg1oimPqIVuMZ1gMgJN-2jxb0OM8xI3Jg/edit?usp=sharing

What the grant buys: statistic, single round -> many rounds runs (diffusion speed, free-rider profiles, reciprocal "you teach next" deals once dialogue is on), stronger models (including proprietary), human hours, open-source benchmark, paper, LW article ... .

How the money will be spent

The team is:
PI / researcher (project lead) - experiment design, analysis, writeup.
Two engineers - infrastructure, runs, and tooling.

min - 5k (~ 6 months):
API for medium size models (2k)
small stipend split across the team (3k)
This founds open bench + medium models, part-time ~ over 6 months.

max - 30k (~3 months):
API for frontier models (10k)
full time for a researcher (9k)
stipend for 2 engineers (6k)
traveling to conferences (5k)
Time estimation: 3 months.

Discussion

Pip FowerakerEndorsement11d

This seems like an interesting, well-bounded and useful study.