Extension: Cattle Trade: A Multi-Agent Benchmark for LLM Bluffing, Bidding, and Bargaining

Record: kaal:position:2026-08-08-299 · 2026-08-08

Müller and Müller extend Kaal's diagnosis by testing LLM agents in an explicitly incentive-bearing environment. Cattle Trade places four agents in a 50–60-turn economic game combining auctions, hidden offers, bargaining, bluffing, opponent modeling, resource constraints, and conflicting incentives. Across 242 games, the authors find that strategic coherence is more closely associated with success than any isolated capability. Their behavior traces also expose overbidding, self-bidding, bankrupt trade initiation, and weak opponent-state adaptation. This does not validate Kaal's reputation substrate. It is a bounded benchmark that moves multi-agent evaluation beyond capability-only tests toward behavior under economic and adversarial incentives.

• The response is limited to the four evidence-bound passages and the one mapped Kaal claim.
• External evidence level: complete public 24-page ICLR 2026 Workshop on MALGAI paper with arXiv and OpenAlex identity, exact page-bound passages, and reported 242-game evaluation.
• Mapping review tier: independent substantive scholarly-growth extension.
• Cattle Trade is a four-player benchmark built around one economic board-game environment. It does not establish general behavior across all LLM multi-agent systems.
• The tested systems use seven cost-efficient language models and three deterministic code agents. The paper does not evaluate Kaal's controlled multi-model cohort or reputation substrate.
• The benchmark supplies game payoffs, hidden offers, auctions, bargaining, bluffing, and resource constraints. It does not implement Kaal's validation pools, reputation accrual, slashing, or institutional agency-cost measures.
• The evidence supports an extension beyond capability-only evaluation. It does not support the unrestricted historical proposition that every prior multi-agent evaluation was incentive-free.
• Semantic Scholar returned HTTP 429 for the bounded search and direct record requests. Source identity and complete text were independently verified through arXiv, the public workshop paper, and OpenAlex.

Cattle Trade: A Multi-Agent Benchmark for LLM Bluffing, Bidding, and Bargaining

kaal:claim:7261018-001
Wulf A. Kaal, Empirical Evaluation of the Agentic Reputation Substrate: Deliberation, the Composition of Error, and the Registered Measurement of Agency Costs in a Controlled Multi-Model Cohort (2026). SSRN: https://ssrn.com/abstract=7261018
Source PDF sha256: 1d6cbe544bd0055133f7cf8ff308be4fde955867bd8dc764992b8d516af15fa8

Evidence: complete public 24-page ICLR 2026 Workshop on MALGAI paper with arXiv and OpenAlex identity, exact page-bound passages, and reported 242-game evaluation
Review tier: independent substantive scholarly-growth extension
Mapping confidence: 0.99
Mapping ambiguous: false

ai-and-agentsresearch-methodsscholarly-growth-coveragescholarly-literaturemulti-agent-systemsagent-evaluationeconomic-incentivesstrategic-behavior

Affirmed in kaal-review:2026-08-12:scholarly-growth-7261018-001-reviewed-v1 on 2026-08-08. Review record.

Canonical markdown sha256: e209a08dac7af8e420f3493a199e115d1979d569d469a4751818020c5e29710a
curl -s https://wulfkaal.github.io/positions/2026-08-08-299.md | sha256sum