Xolver: Generalist Reasoning and Problem Solving through Federated Multi-Agent Dynamics and Holistic Experience Learning

LLMs solve every problem in two senses of isolation: across problems they discard prior reasoning traces, and within a single problem they cannot compare or reuse parallel attempts. We close this gap with Xolver, a training-free framework with three contributions. (i) Holistic Experience, a typed formalism mapping four human experience sources—mentorship, prior-self, tool mastery, peer collaboration—onto retrieval, self-retrieval, tool use, and multi-agent solving. (ii) Federated Agent Learning: agents independently solve the full problem and a judge curates a fixed-size top-m ones; subsequent rounds condition on this buffer rather than on free-form peer messages. (iii) A dual-memory inference algorithm} coupling episodic and intermediate, with quality-gated insertion and a relevance/recency/novelty eviction rule. Extensive experiments across 14 reasoning, agentic, and coding benchmarks demonstrate that Xolver consistently outperforms not only collaborative baseline methods such as debate, panel discussion, and cascading with identical setups (e.g., agents, tools, and iterations) but also specialized evolving/iterative, memory-augmented, and tool-using agents, including OctoTools, Search-o1, AWorld, OpenHands, OAgents, and Agent S2.5. Even with lightweight backbones (e.g., QWQ-32B), Xolver surpasses state-of-the-art proprietary models, and with stronger backbones (e.g., o3-mini-high), it achieves new state-of-the-art performance across a broad range of benchmarks. We will release all code/data.