AlphaApollo: Orchestrating Foundation Models and Professional Tools into a Self-Evolving System for Deep Agentic Reasoning

2025-10-05

AI Summary

To address two fundamental bottlenecks in foundation model (FM) inference—limited inherent reasoning capability and unreliable test-time iterative reasoning—this paper proposes a self-evolving agent-based inference system. Methodologically, it introduces: (1) a shared state graph mechanism enabling multi-turn, multi-model collaborative evolution; (2) deep integration of Python’s numerical and symbolic computation libraries with external retrieval tools, supporting executable verification and grounding of reasoning decisions; and (3) a closed-loop feedback mechanism ensuring verifiability, traceability, and deliberative reasoning. Evaluated on the AIME 2024/2025 benchmark, the system achieves a +5.15-point average score improvement for Qwen2.5-14B-Instruct and a +23.34% pass-rate gain; for Llama-3.3-70B-Instruct, the pass rate increases by +26.67%. These results significantly surpass the performance ceiling of conventional single-modality FM inference.

Technology Category

• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Multiagent Systems: Other Foundations of Multi Agent Systems
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

We present AlphaApollo, a self-evolving agentic reasoning system that aims to address two bottlenecks in foundation model (FM) reasoning-limited model-intrinsic capacity and unreliable test-time iteration. AlphaApollo orchestrates multiple models with professional tools to enable deliberate, verifiable reasoning. It couples (i) a computation tool (Python with numerical and symbolic libraries) and (ii) a retrieval tool (task-relevant external information) to execute exact calculations and ground decisions. The system further supports multi-round, multi-model solution evolution via a shared state map that records candidates, executable checks, and feedback for iterative refinement. In evaluations on AIME 2024/2025 across multiple models, AlphaApollo delivers consistent gains: +5.15% Average@32 and +23.34% Pass@32 for Qwen2.5-14B-Instruct, and +8.91% Average@32 with +26.67% Pass@32 for Llama-3.3-70B-Instruct. Tool-use analysis shows that more than 80% of tool calls are successfully executed, with consistent outperformance of non-tool baselines, thereby lifting the capability ceiling of FMs. More empirical results and implementation details will be updated at https://github.com/tmlr-group/AlphaApollo.

Problem

• Addressing foundation models' limited reasoning capacity and unreliable iteration
• Integrating professional tools for verifiable calculations and grounded decision-making
• Enabling multi-round solution evolution through shared state and feedback

Innovation

• Integrates foundation models with professional computational tools
• Uses shared state map for multi-round solution evolution
• Combines retrieval tools with symbolic libraries for grounding