ERIM Teaches Models to Induce Rules, Not Guess

🔥 What happened Researchers from Tongji and Fudan dropped ERIM, a framework that forces multimodal models to explicitly induce the latent rules behind demonstrations instead of treating in-context learning as blind input-output mapping. Training is two-stage: cold-start supervised fine-tuning, then reinforcement learning. 💡 Why it matters Prior MICL methods skip rule induction entirely and just pattern-match. ERIM shows that correctly inducing the underlying rule yields consistently stronger and more stable performance across diverse MICL settings. That's the gap between mimicking and actually reasoning. ⚡ Our take Finally, someone stops selling in-context learning as a magic trick. Explicit rule induction is the honest path — and probably the only one that stays interpretable.