How to Control LLM Reasoning Effort
Ahead of AI (Sebastian Raschka) · rss · 2026-07-18
This article explores how to control an LLM's reasoning effort, reviewing the basic concepts of reasoning models before discussing scaling approaches at both the training and inference levels.
Core Concepts
- A reasoning model doesn't equal human reasoning; it's a model that outputs intermediate reasoning traces and solves problems step-by-step.
- There are two main paths to improving reasoning: training scaling and inference-time scaling.
Training Level
- The article highlights RLVR (reinforcement learning with verifiable rewards): using outcome correctness as a reward for verifiable tasks.
- Verifiable domains are primarily math and coding, using tools like symbolic calculators, compilers, and unit tests to verify correctness.
- Training doesn't truly depend on the intermediate reasoning traces themselves; rewards are based mainly on the final answer and format.
- Driven solely by output rewards, models can learn to self-correct, backtrack, and revise—phenomena dubbed "Aha" moments.
Inference Level
- Extra compute can be applied during inference to improve answer quality, known as inference scaling.
- An example is self-consistency: sampling the same prompt multiple times and using majority voting to select the final answer.
- This method works for standard LLMs and can be layered on top of reasoning models.
Further Discussion
- The article notes that many modern LLMs have, in a sense, already become reasoning models.
- It introduces the concept of "different effort levels," allowing models to trade off speed, cost, and performance by controlling reasoning intensity.
More from Research
- Linear Digressions returns with a new season of audio essays on AI agents — ChrisGPotts · 2026-07-21
- Async OPD distillation doubles throughput while matching synchronous math accuracy — _lewtun · 2026-07-21
- A forecasting lesson on why R-squared alone led to overfitting and worse predictions — mdancho84 · 2026-07-21
- Google DeepMind’s Project Genie talk shows how creatives feed into model research — alexanderchen · 2026-07-21
- Nat Lambert says RL distillation does not use the strongest models as teachers — natolambert · 2026-07-21
- Thread claims GPT-5.6 Sol helped build a new counterexample factory for the Jacobian conjecture — LucaAmb · 2026-07-21