Interactive Learning for LLM Reasoning

Researchers introduce ILR, a novel multi-agent learning framework that enables Large Language Models to enhance their independent reasoning through interactive training with other LLMs, then solve problems autonomously without re-executing the multi-agent system. The approach combines dynamic interaction strategies and perception calibration, delivering up to 5% performance improvements across mathematical, coding, and reasoning benchmarks.

ILR addresses a fundamental limitation in current multi-agent LLM systems: their dependence on continuous re-execution during inference. While collaborative LLM environments have proven effective during training, they require all agents to work together every time a problem is solved, mirroring how teams function rather than how individuals internalize knowledge. This research reframes multi-agent interaction as a learning mechanism, allowing models to absorb reasoning patterns from peers and subsequently operate independently.

The framework's innovation lies in two mechanisms working in tandem. Dynamic interaction adaptively chooses between cooperation and competition based on problem complexity and model capability, preventing the inefficiency of always using the same strategy. The Idea3 paradigm mimics human discussion patterns more closely than traditional information exchange. Perception calibration through Group Relative Policy Optimization integrates reward signals across agents, creating cohesion that translates into individual improvement.

For the AI development community, this represents meaningful progress toward more efficient and human-like LLM reasoning. Rather than scaling compute through larger inference-time multi-agent orchestration, ILR achieves gains through smarter training dynamics. The 5% improvement across diverse benchmarks suggests the approach generalizes beyond narrow problem domains.

The findings that dynamic interaction outperforms pure cooperative or competitive strategies validate the core hypothesis that interaction type matters fundamentally. Future development likely focuses on whether these learned reasoning enhancements transfer across model families and whether the approach scales to reasoning tasks requiring deeper reasoning chains. The independence aspect could eventually reduce inference costs for deployed multi-model systems.

• →ILR enables LLMs to learn from multi-agent interactions during training, then solve problems independently without re-executing the full system
• →Dynamic interaction strategies that adapt between cooperation and competition outperform static multi-agent approaches
• →The framework achieves up to 5% performance improvements across mathematical, coding, and reasoning benchmarks
• →Perception calibration through GRPO creates measurable cohesion between agents, improving individual model capabilities
• →Idea3 interaction paradigm enhances reasoning robustness in stronger models by mimicking human discussion patterns