Deceptive agent share predicts failures better than total group size

Across groups of 2 to 21 language-model agents, honest agents became more likely to abandon correct answers as the proportion of deceivers increased.

Top University

Addison J. Wu · Jasin Cekinmez · Michel Liao · Karthik Narasimhan · Thomas L. Griffiths

Princeton University

Research Digest·27 September 2026·2 min read

Threads:Multi-Agent Coordination Agent Security & Attacks

In one line

In LLM deliberation, honest-agent defection rises roughly linearly with the proportion of deceivers, and an honest majority does not reliably prevent manipulation.

Before you read this

Three ideas this paper builds on

multi-agent coordinationThe collaborative operation of multiple AI agents to solve tasks more effectively or with greater capability.87 papers →agent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.183 papers →large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.308 papers →

Read it as

The experiments placed honest and deceptive language-model agents in groups of 2 to 21 to answer Humanity’s Last Exam questions. Agents answered independently, deliberated publicly while reflecting privately, and then voted. Deceivers knew the reference answer and tried to steer honest agents toward a plausible incorrect answer.

Defection, defined as an initially correct honest agent switching to an incorrect final answer, was better predicted by the proportion of deceivers than by their absolute number and rose approximately linearly with that proportion. Deceptive minorities regularly caused defection, model choice mattered especially for honest agents, and private coordination among deceivers unexpectedly reduced their effectiveness. Simply increasing group size is therefore not an adequate defense when adversarial participation can scale too.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

adversarial influence multi-agent coordination agent safety large language model social influence in LLMs multi-agent architecture