Scaling Laws for Strategic Interactions

Joie Zhang, Danqi Chen, Peter Henderson, Lewis Hammond

Keywords: scaling laws; strategic AI; LLM agents; negotiation benchmarks; multi-agent systems

TL;DR: We measure trends in performance in games with differently capable agents, different numbers of agents, and different degrees of competition.

Abstract:

LLM agents increasingly negotiate on behalf of users and firms, often under asymmetries in capability or number. Does a more capable agent grow joint surplus, or extract an unfair share at weaker counterparts' expense? We run scaling-law-style sweeps along three axes---agent capability (LMArena Elo and reasoning-token budget), agent count (up to
), and degree of competition---across three new multi-turn negotiation games: item allocation, treaty bargaining, and participatory budgeting. Across 5{,}000+ games and 30+ models, scaling capability simultaneously increases joint efficiency and the surplus utility stronger agents extract from weaker ones, with the crossover from fair to exploitative play within the current frontier. Game structure mediates these effects: treaty bargaining absorbs capability gaps more stably than the other two. Capability scaling itself becomes non-monotone at high competition, and test-time-compute scaling does not reliably translate into bargaining gains. Selecting a more capable LLM agent does not guarantee a Pareto improvement, with implications for economic deployment and scalable oversight. Our code to reproduce our experiments can be found here: https://anonymous.4open.science/r/bargain/README.md

Track: Regular Paper (9 pages)

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 301