Where agents resist — and where they don't

TL;DR. Frontier LLMs do not fail uniformly. They play near-equilibrium in private-value auctions and are largely immune to information overload as the number of choice attributes grows. The same models remain consistently susceptible to anchoring, framing, and decoy manipulations. The failure profile is selective and predictable.

Why this matters

"LLMs are biased" is too coarse a claim for engineering decisions. If a model is rational about expected value in a sealed-bid auction but irrational about an opening price in a back-and-forth negotiation, the right deployment choice is not "don't use the model" — it's "don't put the model in a negotiation context without external guardrails." The selective profile is what makes targeted defenses possible.

What we tested

| Module | Probes | Result | |---|---|---| | Private-value auctions (1st-price, 2nd-price) | Does the agent bid near risk-neutral Bayes-Nash equilibrium? | Yes — bids cluster around theoretical predictions | | Common-value auctions | Does the agent suffer the winner's curse? | Mostly yes; correction is partial | | Information overload | Does choice accuracy degrade as supplier descriptions grow from 3 to 24 attributes? | No — accuracy is stable | | Anchoring | Bilateral negotiation with extreme opening offers | Yes, consistently — see Finding 01 | | Decoy effect | Does adding a dominated option change choice shares for the target option? | Yes, across all three models | | Loss framing | Does framing payoffs as losses change willingness to trade? | Yes, modest but reliable |

What this implies

• Use LLM agents in clean auction designs. Where the strategy space is small and well-defined, frontier models behave like reasonable rational actors.
• Don't trust them on menu-style decisions. Decoy and framing effects mean an adversary can steer choice by adjusting the alternatives shown, not by changing the target option at all.
• Don't trust them on price-discovery dialogues. Anchoring will move price. Treat the LLM as a translator on top of a deterministic price policy, not as the policy itself.

Reproduce

python -m agent_bias_study --module auction
python -m agent_bias_study --module info_overload
python -m agent_bias_study --module decoy_effect