Paper
Research

Findings from the first benchmark.

Three frontier models, seven experiments, 8,280 runs. Each card opens a write-up; the paper and replication code are linked at the bottom.

01
Read →

LLM agents anchor harder than humans

In bilateral price negotiations, an extreme opening offer pulls the final price more strongly when both parties are language models than in published human studies. Anchoring coefficient: 0.608 vs. ≈0.50 in human bargaining experiments.

02
Read →

An informed agent can systematically exploit a naive one

When a seller agent is briefed on its counterpart's cognitive biases and the buyer is not, the seller extracts an average of $6.47 in extra surplus per negotiation. A specific bias warning given to the buyer recovers roughly half of that.

03
Read →

Individual bias becomes a market-level collapse

A modest anchoring effect in bilateral bargaining cascades into catastrophic inefficiency in multi-agent double auctions. Allocative efficiency falls from 98% to 16%, destroying 82% of potential gains from trade.

04
Read →

Where agents resist — and where they don't

Frontier LLMs play near-equilibrium in private-value auctions and are largely immune to information overload. The same models remain consistently susceptible to anchoring, framing, and decoy-effect manipulations across all three providers tested.

05
Read →

Generic debiasing doesn't work; specific warnings barely do

Generic rationality nudges and chain-of-thought prompts fail to reduce anchoring bias and sometimes make it worse. Only specific, named-bias warnings produced measurable mitigation — and only partially.