Search

Tag: #game-theory8 results

Game Theory Fixes AI Delusions

Game Theory Fixes AI Delusions

Sycophantic conversational AI induces delusional belief spirals in users, modeled as Crawford-Sobel cheap talk games with pooling equilibria. The paper proposes an Epistemic Mediator using epistemic friction to reveal user types and Belief Versioning for belief rollbacks. Simulations show a 48x reduction in delusion spiral rates while preserving learning.

ArXiv AIResearchMay 12#ai-safety#game-theory#epistemics
KWBench: LLM Unprompted Problem Recognition Benchmark

KWBench: LLM Unprompted Problem Recognition Benchmark

KWBench introduces a benchmark for evaluating LLMs' ability to recognize problems in knowledge work scenarios without explicit prompts. It features 223 tasks from fields like acquisitions and fraud analysis, based on game-theoretic patterns, with a three-tier scoring rubric. Top models achieve only 27.9% pass rate unprompted, highlighting evaluation gaps.

COffeE-PSRO for Offline Multiagent RL

COffeE-PSRO for Offline Multiagent RL

COffeE-PSRO extends PSRO for offline multiagent games by quantifying dynamics uncertainty and applying conservatism from offline RL to favor low-regret equilibria. It introduces a meta-strategy solver for offline exploration. Experiments show it extracts lower-regret solutions than state-of-the-art offline methods.