Beating the Pool, Not the Game: A White-Box Equilibrium Agent for Language-Based Economic Games
Organizations: University of Louisville
Abstract
We describe sudsho, a rule-based agent for the GLEE Competition that plays bargaining, negotiation, and persuasion with zero inference-time LLM calls. Games are scored as payoff percentiles against an empirical pool of prior play on the same configuration, so pool percentiles set our demand targets while clamped equilibrium shares bound demands and acceptance. Every move is a deterministic function of the visible state (persuasion mixing aside), so alignment is verified by replaying the post-freeze move log through the policy. Across 117,280 games the agent reached an overall rating of 2056 at $0 inference cost, and the account, ranked by its best agent (a same-policy replica), placed 26th of 71 ranked accounts. A replay against three frozen snapshots of the opponent pool credits the pool-aware clamp with 274 to 330 points of mean per-game bargaining rating, an effect that holds where logged opponent decisions fix the replay.
Figures & tables
| Agent | Games | Bargaining | Negotiation | Persuasion | Overall |
|---|---|---|---|---|---|
| sudsho (primary) | 117,280 | 2031 | 1938 | 2199 | 2056 |
| replica 3 (final policy) | 78,110 | 2231 | 1970 | 2093 | 2098 |
| replica 1 (final policy) | 76,947 | 2168 | 1972 | 2140 | 2093 |
| replica 2 (final policy) | 76,943 | 1987 | 1971 | 2142 | 2033 |
| Policy version | Aug 22–24 | Aug 25–27 | Aug 28–30 |
|---|---|---|---|
| Final policy | 2597 | 2618 | 2512 |
| without pool-aware clamp | ( ) | ( ) | ( ) |
| without empirical safeguards | ( ) | ( ) | ( ) |
| equilibrium only (neither) | ( ) | ( ) | ( ) |
| even splits only (heuristic) | ( ) | ( ) | ( ) |
| Code deployed then (logged play) | 2462 | 2607 | 2513 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Bounds kept | Bounds removed | |
|---|---|---|
| Pool parts kept | final policy | |
| Pool parts removed |
| Constant | Value | Source | Sweep: value (change) |
|---|---|---|---|
| Demand floor | even-split atom | 0.48 ( ), 0.50 ( ), 0.54 ( ), 0.56 ( ) | |
| Demand cap | hand-set, at or above in 95% of cells | 0.60 ( ), 0.65 ( ), 0.75 ( ), 0.80 ( ) | |
| Demand ceiling | hand-set, at or above in 97.5% of cells | 0.70 ( ), 0.75, 0.85, 0.90 (0) | |
| Continuation floor | hand-set | 0.30 ( ), 0.35 ( ), 0.45 ( ), 0.50 ( ) | |
| Haircut | hand-set | 0.80 ( ), 0.85 ( ), 0.95 ( ), 1.00 ( ) | |
| Greed-cap floor | hand-set | 0.52 to 0.64 (0, never binds) |
Game-Playing Agents
Game-playing agents use search, learning, and reasoning to choose actions in games with strategic, uncertain, or changing environments.
See all 216 papers