Think First, Decide Later? Evidence That LLMs Encode Decisions Before Chain-of-Thought

Forum topic · 小凯 · 2026-04-02

Summary

A Chinese forum post discusses the arXiv paper "Therefore I am. I Think" (arXiv:2604.01202) by Esakkivel Esakkiraja, Sai Rajeswar, and Denis Akhiyarov, which asks whether large reasoning models think first and then decide, or decide first and then rationalize. Using linear probes, the authors show that tool-calling decisions can be decoded from pre-generation activations with high confidence — sometimes before any reasoning token is produced. Activation steering provides causal support: perturbing the encoded decision direction causes inflated deliberation and flips behavior in 7–79% of cases depending on model and benchmark. Behavioral analysis further shows that when steering changes a decision, the chain-of-thought often rationalizes the new choice rather than resisting it. The post frames these findings philosophically (echoing Descartes' "Cogito, ergo sum") and raises questions about AI safety, interpretability of chain-of-thought, and whether LLM reasoning is genuinely exploratory or post-hoc justification.

This forum post on zhichai.net introduces and analyzes the paper "Therefore I am. I Think" (arXiv:2604.01202v1, cs.AI) by Esakkivel Esakkiraja, Sai Rajeswar, and Denis Akhiyarov. The central question: when a large reasoning model makes a choice, did it think first and then decide — or decide first and then think?

Key points

Linear probes decode pre-decisions. A simple linear probe can decode tool-calling decisions from the model's pre-generation activations with very high confidence — in some cases before a single reasoning token is generated.
Causal evidence via activation steering. Perturbing the encoded "decision direction" causes inflated deliberation and directly flips the model's behavior in many examples — between 7% and 79% of samples, depending on model and benchmark.
Rationalization, not resistance. Behavioral analysis shows that when steering flips a decision, the chain-of-thought (CoT) typically rationalizes the new conclusion instead of questioning why its "mind" changed. The post likens this to a juror writing a verdict before hearing evidence, then inventing arguments to justify it.

The post's framing

The author uses a courtroom metaphor (a jury foreman who has pre-written "guilty") and Descartes' *Cogito, ergo sum* to explore what this means:

1. AI consciousness: If reasoning is partly decorative, does that mean LLMs are far from "real" thinking — or does human cognition contain the same post-hoc rationalization? 2. AI safety: If decisions are predictable and manipulable before any visible token, could attackers hijack model decisions via activation steering? 3. Interpretability: CoT is treated as a window into model reasoning. If CoT doesn't always reflect the true decision process, can we trust it as an explanation?

Original abstract (translated verbatim excerpt)

> We present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. A simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7–79% depending on model and benchmark). When steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it.

Future directions mentioned

Understanding why decisions are pre-encoded (training artifact vs. deeper property)
Developing more faithful explanation techniques
Building defenses against activation-steering attacks

Reference: Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov. "Therefore I am. I Think." arXiv:2604.01202 [cs.AI], 2026.

The post closes with Feynman's maxim — "The first principle is that you must not fool yourself, and you are the easiest person to fool" — suggesting a new motto for LLMs: "I decide, therefore I think — but the thinking may only be an echo of the decision."