Controlling Reasoning Effort in LLMs

2025-07-19 · Source: Ahead of AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

OpenAI's GPT-5.6 model family, alongside DeepSeek-R1 and other flagship LLMs, now features controllable "reasoning effort" settings, allowing users to adjust response verbosity and accuracy. Reasoning models generate intermediate step-by-step traces, and their effort levels directly influence token usage and performance, as seen in GPT-5.6 Sol and Inkling benchmarks. These capabilities are typically implemented during post-training through methods like Reinforcement Learning with Verifiable Rewards (RLVR) that apply varying length penalties, or Supervised Fine-tuning (SFT) on data conditioned by effort labels. This approach enables a trade-off between computational cost and output quality, distinguishing inference scaling from training scaling.

Key takeaway

For ML engineers and AI scientists deploying LLMs, understanding and utilizing reasoning effort controls is vital for optimizing resource allocation and output quality. You should employ system prompt-based effort settings (e.g., "Reasoning effort: low/high") to dynamically adjust model verbosity and accuracy based on task requirements. Consider models trained with explicit effort conditioning, like DeepSeek V4 or Inkling, to fine-tune performance and cost without extensive retraining, ensuring efficient and effective model deployment.

Key insights

LLM reasoning effort can be dynamically controlled at inference time by training models to respond to explicit effort signals.

Principles

• Reasoning models produce intermediate thought traces.
• Higher reasoning effort correlates with increased token usage and accuracy.
• Training and inference scaling are distinct performance levers.

Method

Implement effort control by conditioning RLVR with token-cost penalties or through SFT on datasets containing varied effort-level examples, often using system prompts.

In practice

• Use system prompts to select desired reasoning effort.
• Balance model size and effort for cost-performance optimization.
• Train with alternating budgeted/unconstrained RL for efficiency.

Topics

• Reasoning Models
• LLM Inference
• Reinforcement Learning with Verifiable Rewards
• Supervised Fine-tuning
• Token Efficiency
• GPT-5.6

Best for: AI Engineer, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• A Visual Guide to Reasoning LLMs — Scaling LLM performance is shifting from training compute to inference-time reasoning, enabling models to "think longer."
• Reasoning Budgets vs. Structured CoT: Controlling LLM Thinking Tokens — Controlling LLM reasoning traces via budget or grammar can reduce token cost but risks accuracy.
• Thinking with Reasoning Skills: Fewer Tokens, More Accuracy — Distilling and retrieving reusable reasoning skills significantly improves LLM efficiency and performance.
• Pokémon Built a Robot Brain — LLMs are trained in stages, from broad knowledge acquisition to fine-tuned conversational and reasoning capabilities.
• ThinkLog: Leveraging Reasoning for Log Statement Generation — Incorporating explicit reasoning into LLM prompts significantly improves end-to-end log statement generation accuracy and cost-efficiency.
• Human Mind Reconstruction w/ AI (ToM, UserHarness) — Decoupling epistemic structure from LLM generation significantly boosts reasoning accuracy and efficiency, even for smaller models.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Ahead of AI.