Together AI Unveils RARO at ICML 2026: 'Escaping the Verifier' Paradigm Enables LLM Reasoning Without Task-Specific Rewards

In a paradigm-shifting presentation at the 43rd International Conference on Machine Learning (ICML 2026) in Seoul, researchers from Together AI and MIT have introduced a groundbreaking methodology that fundamentally rethinks how Large Language Models (LLMs) acquire reasoning capabilities www.together.ai . Unveiled on July 9, 2026, during Day 3 of the main conference, the paper titled "Escaping the Verifier: Learning to Reason via Demonstrations" proposes a novel framework dubbed RARO, which effectively liberates model training from the verifier bottleneck that has long constrained reinforcement learning pipelines www.researchgate.net . The heuristic Bottleneck of Modern Reasoning For the past several years, the dominant ubiquitous approach to teaching LLMs complex reasoning—such as mathematical proofs or multi-step coding—has relied heavily on Reinforcement Learning from Human or AI Feedback (RLHF/RLAIF) paired with task-specific verifiers quantumzeitgeist.com . These verifiers act as rigid judges, assigning reward signals based on whether a final output matches a predefined correct answer. However, this architecture is inherently brittle; if a verifier cannot be easily constructed for a nuanced or stochastic task, the model struggles to learn effectively arxiv.org . Enter RARO: Reasoning via Demonstrations Together AI researchers Locke Cai and Ivan Provilkov, in collaboration with MIT, address this challenge with RARO, a methodology that enables strong reasoning capabilities without requiring a task-specific verifier ribbitribbit.co . Instead of relying on a binary reward signal from an external judge, RARO leverages high-quality, step-by-step reasoning demonstrations to guide the model's amelioration process quantumzeitgeist.com . By training the model to recognize and replicate the structural integrity of logical thought processes rather than just the final destination, RARO achieves superior generalization across diverse cognitive tasks. The team utilized Together AI's managed fine-tuning infrastructure to train both Supervised Fine-Tuning (SFT) and Rationalization baselines, monitoring validation loss to ensure the model was internalizing the reasoning pathways rather than merely memorizing outputs arxiv.org . The empirical results presented at ICML 2026 demonstrate that RARO not only matches but frequently exceeds the performance of traditional verifier-dependent models, particularly in domains where ground-truth verification is computationally expensive or ambiguous. Broader Implications for the AI Ecosystem The introduction of RARO represents a critical step toward more autonomous and adaptable AI systems. By "escaping the verifier," developers can now deploy reasoning models in highly subjective or complex domains—such as legal analysis, strategic planning, and creative synthesis—where traditional programmatic verifiers fail entirely. As the machine learning community continues to gather in Seoul for the remainder of ICML 2026, Together AI's contribution stands out as a seminal milestone in the pursuit of Artificial General Intelligence (AGI), proving that the path to advanced cognition may rely more on the quality of the journey than the rigidity of the destination www.together.ai .