LRM (Large Reasoning Models) denote a class of large-scale neural architectures and associated training protocols explicitly optimized for robust, multi-step reasoning rather than only next-token prediction: they integrate mechanisms for maintaining and manipulating intermediate state (working memory), for planning multi-stage inference (explicit chain-of-thought or planner modules), and often for interfacing with symbolic components or external tools to improve compositionality, verifiability, and data efficiency. In practice LRMs extend the capabilities of contemporary LLMs by focusing model inductive biases, objective design, and curriculum selection toward tasks that require multi-hop deduction, counterfactual reasoning, algorithmic thinking, theorem proving, program synthesis, or complex decision planning; approaches include specialized architectures (e.g., modular or recurrent memory-augmented transformers), supervision strategies (chain-of-thought or intermediate supervision), reinforcement learning for multi-step policies, and neuro-symbolic hybrids that combine differentiable perception with discrete reasoning engines. For AI and ML researchers this class is significant because it targets core limitations of pattern-completion models—improving interpretability via explicit reasoning traces, enhancing out-of-distribution compositional generalization, and enabling stricter correctness guarantees in high-stakes domains—while posing new challenges in dataset design, computational cost, evaluation methodology, and alignment of stepwise reasoning with human-understandable proofs.
First appeared in research discourse around 2023–2024 and gained wider popularity in 2024–2025 as research groups and benchmarks began emphasising chain-of-thought, multi-hop reasoning, and specialised architectures/tools to push LLMs beyond surface-level correlations.