Thinking Inertia: Stanford and Oxford Reveal Why LLMs Keep Reasoning When Explicitly Told Not To

Background and the Problem

With the proliferation of reasoning-centric frontier models like o1, o3, and Claude 3.7 Sonnet, explicit 'thinking modes' have become foundational for complex problem solving. However, in low-latency production pipelines and multi-agent systems, developers frequently suppress thinking traces via system prompts (e.g., 'Return only the final answer without reasoning') to reduce costs and response times. The industry has long assumed that 'no-thinking' is a trivial setting toggled by disabling tags or requesting answer-only outputs, neglecting the empirical behavior of models instructed to halt reasoning.

Architecture and How It Works

To systematically address this evaluation gap, researchers from Stanford and Oxford established the first formal framework for quantifying no-thinking compliance:

1. Three-Tier Evaluation Metrics: Replaces fragile proxies (such as response character count) with: (i) Empty-Thinking Rate (measuring strict answer-only compliance); (ii) Instruction-aware Question-Pre-answer Relevance; and (iii) LLM-as-judge Explicit Inference Rate to reliably separate filler text from genuine reasoning;
2. Six Prompting Interventions across Three Answer Spaces: Evaluates six distinct prompt interventions across Boolean, multiple-choice, and open-ended answer spaces on six state-of-the-art LLMs;
3. Discovery of Thinking Inertia: Demonstrates that explicit negative instructions fail to suppress internal chain-of-thought tendencies. As the answer space broadens, models persistently generate unauthorized reasoning chains before emitting the final answer.

Benchmarks and Measured Results

The benchmark reveals critical empirical trade-offs:

• Pervasive Thinking Inertia: None of the tested models reliably achieved zero-inference compliance in open-ended domains; models exhibited strong latent resistance to no-think instructions;
• Compliance vs. Task Accuracy Trade-Off: While closed-domain Boolean and multiple-choice benchmarks preserve accuracy under answer-only constraints, open-ended tasks suffer dramatic performance degradations when forced to bypass thinking traces;
• Entropy Suppression via Candidate Structuring: Rewriting identical open-ended problems into multiple-choice formats increased answer-only compliance by more than 40%, demonstrating that output space constraint is essential for dampening cognitive momentum.

Getting Started for Developers

These findings provide immediate actionable guidance for engineering robust agentic architectures:

• Evaluation Toolkit: The codebase is publicly available on GitHub at https://github.com/thinking-inertia/code for benchmarking reasoning leakages across internal deployments;
• Avoid Surface-Level Prompting: Do not rely exclusively on natural language prompt negations to stop model thinking; enforce structural constraints via grammar-constrained decoding (JSON schema or CFG) or specialized stop sequences;
• Reframe Open Dispatches as Discrete Selections: Convert open-ended routing and classification steps into closed-form discrete option selection to curb unwanted inference tokens without harming upstream performance.