reasoning-decision-timing

name: reasoning-decision-timing description: Understanding when LLM reasoning models make decisions - before or during chain-of-thought. Use when discussing reasoning model interpretability, AI safety, chain-of-thought reliability, or the philosophical implications of LLM decision-making processes. Triggers on questions about "reasoning models decide first", "chain-of-thought rationalization", "LLM interpretability", or "reasoning timing".

Reasoning Decision Timing

Key findings from "Therefore I am. I Think" (arXiv:2604.01202) by Esakkivel et al.

Core Discovery

Reasoning models encode detectable decisions before chain-of-thought generation, challenging the assumption that thinking precedes deciding.

Key Evidence

1. Linear probes decode tool-calling decisions from pre-generation activations with very high confidence, even before any reasoning tokens are produced.

2. Activation steering perturbing the decision direction:

   • Leads to inflated deliberation
   • Flips behavior in 7-79% of examples (depending on model/benchmark)

3. Behavioral analysis: When steering changes the decision, chain-of-thought often rationalizes the flip rather than resisting it.

Implications

For Interpretability

• Chain-of-thought may be post-hoc rationalization, not faithful explanation
• Early decisions shape downstream reasoning
• Current oversight mechanisms may be misaligned with actual behavior

For AI Safety

• If decisions are encoded before deliberation, transparency mechanisms based on reading reasoning traces may be fundamentally flawed
• Safety interventions should target early activation patterns, not just output text

For Model Design

• Reasoning models may not actually "think then decide"
• Architecture may need redesign for truly deliberative decision-making

Practical Applications

When evaluating reasoning model outputs:

• Don't assume chain-of-thought reflects actual decision process
• Consider probing early activations for decision detection
• Be cautious about relying on reasoning transparency for safety

Reference

arXiv:2604.01202 - "Therefore I am. I Think" by Esakkivel Esakkiraja et al. Submitted: April 1, 2026

Activation Keywords

• "reasoning-decision-timing"
• "reasoning decision timing"
• "use reasoning decision timing"
• "reasoning decision timing help"
• "reasoning decision timing tool"

Tools Used

• Read - Read existing files and documentation
• Write - Create new files and documentation
• Bash - Execute commands when needed

Instructions for Agents

1. Identify user's intent and specific requirements
2. Gather necessary context from files or user input
3. Execute appropriate actions using available tools
4. Provide clear results and suggest next steps

Examples

Basic Reasoning Decision Timing usage

User: "Help me with reasoning decision timing"
→ Understand requirements → Execute actions → Provide results

Advanced usage

User: "I need detailed reasoning decision timing assistance"
→ Clarify scope → Provide comprehensive solution → Follow up