Position: Cease the Humanization of Intermediate Tokens as "Reasoning" or "Thinking" Traces

---

📄 Document Metadata

The following table summarizes the archival details of this position paper:

---

👥 The Research Team

The paper was authored by a collaborative group of researchers:

• Subbarao Kambhampati (Lead)
• Karthik Valmeekam
• Siddhant Bhambri
• Vardhan Palod
• Lucas Saldyt
• Kaya Stechly
• Soumya Rani Samineni
• Durgesh Kalwar
• Upasana Biswas

---

🎯 Core Thesis

The authors address the prevalence of Intermediate Token Generation (ITG)—a technique where a Large Language Model (LLM) generates a sequence of tokens prior to delivering the final answer to boost performance on complex tasks.

The Problem: These sequences are frequently labeled as "reasoning traces" or "thinking traces". This terminology anthropomorphizes the model, falsely suggesting that the AI is following a human-like cognitive process and providing a transparent window into its "mind."

The authors argue that this is not merely a linguistic quirk but a dangerous misconception that:

1. Obscures the actual mechanical nature of the models.
2. Misleads users on how to utilize these tools effectively.
3. Encourages flawed scientific research.

---

⚙️ Technical Conceptualization

Mathematically, the process can be viewed as the model generating a sequence $S$ where: $S = \{t_{1}, t_{2}, \dots, t_{n}, A\}$ In this equation, $\{t_{1} \dots t_{n}\}$ represents the intermediate tokens and $A$ represents the final answer.

Process Flow

Example of the "Thinking" Fallacy

# How the community views it:
model.think("Step 1: Analyze the logic...") 
# Result: "The answer is 42"

# How it actually works:
model.generate_tokens("Step 1: Analyze the logic...") 
# Result: "The answer is 42"

---

🛠️ Call to Action

The authors propose a shift in how the AI community discusses these outputs.

• Stop using terms like "thinking" or "reasoning" for token sequences.
• Start describing these as intermediate computational steps or ITG.
• Avoid treating these traces as interpretable windows into human-like cognition.

---

🔄 Version History & Evolution

The paper has undergone several iterations to refine its stance:

• v1 $\rightarrow$ v4: While the background information remains largely consistent, the later versions (specifically v4) shift the focus and sharpen the specific arguments against the anthropomorphization of these traces.

---

📎 Additional Information

License:

This summary was derived from the arXiv landing page for the paper submitted to ICML 2026.