Researchers have built a training framework that lets language models absorb reasoning skills from other models — then use those skills solo.
The system, called ILR (Interactive Learning for LLM Reasoning), pairs models during training using two mechanisms. The first, Dynamic Interaction, picks cooperative or competitive strategies based on question difficulty and model capability. Models then exchange ideas through a structured discussion protocol called Idea3 before each model produces its own answer. The second mechanism, Perception Calibration, uses a reinforcement learning technique — Group Relative Policy Optimization — to fold one model's reward signals into another's training objective, tightening how well the models learn from each other. ILR was tested across three models from two families on eight benchmarks covering math, coding, general question answering, and scientific reasoning, outperforming single-agent baselines by up to 5%.
The practical appeal is the inference-time story: once trained, a model does not need to re-run the full multi-agent pipeline to get an answer. That matters because multi-agent inference is expensive — every additional model call adds latency and cost. The catch is that the source findings are not entirely clean on this point: Idea3 was also shown to improve robustness when stronger models do run in multi-agent mode at inference, so the independence gains are real but not a full substitute for the pipeline.
The result sits in a crowded lane — frameworks like RLVR and self-play training have pushed similar ideas — but ILR's explicit split between training-time collaboration and inference-time independence is a more honest accounting of where the compute bill actually lands.