Large Reasoning Model (LRM)

Opening

A large reasoning model (LRM) is a language model trained with reinforcement learning over verifiable outcomes to produce chain-of-thought traces before emitting a final answer. The term entered widespread use around OpenAI's o1 release in September 2024 and is the canonical name for the class of models — including o1/o3, Claude with extended thinking, Gemini Thinking, DeepSeek-R1, Qwen QwQ, and Kimi K1.5 — that distinguish themselves from base LLMs by allocating inference-time compute to intermediate reasoning tokens rather than to a direct next-token prediction.

Mechanism

LRMs are trained in two stages that differentiate them from instruction-tuned LLMs. First, the model is given problems with verifiable answers (math competitions, code with test cases, formal proofs) and learns to generate extended chains of thought that include exploration, backtracking, verification, and self-correction. Second, the model's outputs are scored against the verifiable answer via reinforcement learning, often a variant of RLHF or RLAIF, with reward signals from the verified result rather than from human preference. At inference, the model produces a chain of thought whose length is itself a controllable parameter (more thinking tokens for harder problems) before producing the final answer. The mechanism assumes that the intermediate text creates additional context that helps the model arrive at the correct solution, though whether the text is faithful to the underlying computation is a separate, contested question.

Tradeoffs

The LRM framing has a sharp credibility problem: research from 2024–2026 has shown that chains of thought may not faithfully represent the model's computation (Apple's "Illusion of Thinking" paper, Anthropic's interpretability work, NYU filler-token experiments showing that literally meaningless dot-strings can substitute for human-readable reasoning). Pro-LRM researchers argue this doesn't matter because the verifiable-domain training signal produces correct answers regardless of whether the intermediate text is interpretable. Skeptics argue that taking the trace seriously as evidence of capability is wishful mnemonics in Drew McDermott's 1976 sense — labeling program components with intuitive names without evidence they implement the named function. The trade-off is between trusting the correct-answer output and understanding the process.

Open Questions

Whether LRM chains of thought can be made faithful, whether they should be, and whether interpretability methods (mechanistic interpretability, causal-tracing studies) can resolve the question. Whether LRM training generalizes beyond verifiable domains — current evidence suggests LRMs transfer poorly to open-ended reasoning (strategic planning, ethical judgment, scientific hypothesis generation) where there is no ground-truth answer to reinforce against. Whether the LRM-vs-LLM distinction will hold as base models improve (the distinction is already blurring as next-generation LLMs incorporate reasoning training natively). Whether the term "reasoning" in LRM is itself wishful mnemonics for a more specific process that has yet to be named.