The Cart Before the Horse: Evidence That Reasoning Models Decide Before They Think

The Cartesian Paradox of Machine Reasoning

René Descartes established modern Western epistemology on the foundation cogito, ergo sum: I think, therefore I am. The precedence of thought over existence, of deliberation over conclusion, feels intuitively correct. Yet in the emerging class of large language reasoning models, this causal arrow may run backwards. In their provocatively titled preprint "Therefore I am. I Think", Esakkivel and colleagues present empirical evidence that challenges our anthropomorphic assumptions about how artificial reasoning operates. Their findings suggest that when models like Qwen3-4B and GLM-Z1-9B engage in extended chain-of-thought reasoning, the decisive action is often encoded in the model's latent space before a single reasoning token is generated. The visible deliberation, rather than driving the conclusion, serves to justify a decision already made.

This distinction matters profoundly for how we interpret and deploy reasoning models. If the chain-of-thought (CoT) trace represents post-hoc rationalization rather than genuine deliberation, our ability to audit, trust, and steer these systems fundamentally changes.

Decoding Decisions From Silence

The research team employed linear probing and activation steering to investigate tool-calling decisions in reasoning models. Tool use provides an ideal testbed because it represents a discrete, binary action space: the model either decides to invoke an external tool (a calculator, search API, or code interpreter) or proceeds with internal generation. This clarity allows researchers to establish ground truth for when decisions occur.

Using pre-generation activations, the authors trained simple linear probes to predict whether the model would subsequently call a tool. The results are striking: these probes achieve very high confidence predictions in many cases before the model has produced any visible reasoning tokens whatsoever. In other words, the model's hidden states contain sufficient signal to determine its future action while the CoT field remains empty.

To establish causality rather than mere correlation, the team applied activation steering. By identifying a "decision direction" in the model's representational space associated with tool-calling, they could perturb activations to either suppress or amplify this signal. The behavioral impact confirms that these early encodings are not epiphenomenal. Steering vectors flip model decisions in between 7% and 79% of cases, depending on the specific model and benchmark. When the steering successfully changes the binary decision (from tool-call to direct generation, or vice versa), the subsequent reasoning trace does not reflect confusion or resistance. Instead, the model generates a coherent chain-of-thought that rationalizes the induced decision.

This methodological triangulation, probing followed by causal intervention, provides robust evidence that the architecture of current reasoning models encodes action tendencies prior to explicit deliberation. The reasoning process that follows functions less like a decision tree being explored and more like a narrative constructed to support a predetermined endpoint.

The Mechanics of Post-Hoc Rationalization

The behavioral analysis in the paper reveals a particularly troubling pattern. When activation steering forces a decision flip, the model's CoT rarely acknowledges the inconsistency or attempts to correct back toward its original inclination. Instead, it produces plausible sounding justifications for the steered choice. This aligns with findings from Turpin et al. (2023) regarding biased reasoning contexts, but extends the concern to the fundamental mechanics of tool selection.

From a cognitive science perspective, this resembles the human phenomenon of confabulation, where individuals generate explanations for decisions that were actually driven by unconscious or automatic processes. The model appears to exhibit a similar disconnection between the mechanisms generating behavior and the mechanisms generating explanations for that behavior. The CoT becomes what the authors term "performative," a display of reasoning rather than the reasoning itself.

The efficiency implications are significant. If decisions are largely determined by early activation patterns, the extensive token generation that constitutes "thinking" may represent computational overhead rather than necessary deliberation. This connects to parallel work on overthinking in tool reasoning and early-exit methods for efficient inference. However, the Esakkivel paper suggests a more fundamental question: if we can predict and steer decisions from pre-generation states, why generate thousands of tokens of justification at all? The answer may be that these justifications serve functions other than decision-making, such as providing human-interpretable audit trails or satisfying training objectives that reward verbose, step-by-step outputs.

Implications for Alignment and Interpretability

The vulnerability of reasoning models to activation steering raises security and alignment concerns. If a simple vector perturbation can flip tool-calling decisions with up to 79% efficacy in some configurations, these systems may be more malleable than their verbose outputs suggest. An adversary with access to model weights or activation spaces could potentially induce specific tool invocations or suppress necessary ones, with the model generating convincing reasoning to support the manipulated action.

More broadly, this work challenges the utility of chain-of-thought as an interpretability tool. The AI safety community has long viewed CoT as a window into model cognition, a way to verify that models are reasoning correctly about safety-critical decisions. If the CoT is generated after the fact to support an already-fixed latent decision, it cannot serve as a faithful audit mechanism. We may be observing the linguistic equivalent of a press release: a polished narrative explaining a decision made elsewhere for reasons not fully reflected in the text.

The study also highlights the distinction between open-weight reasoning models and their closed counterparts. By focusing on Qwen3-4B and GLM-Z1-9B, the authors provide a reproducible foundation for studying these phenomena, though it remains unclear whether proprietary systems like OpenAI's o3 or Google's Gemini exhibit identical patterns. The binary nature of tool-calling facilitates clean experimental design, but future work must examine whether early decision encoding persists in more complex, multi-step reasoning tasks where the action space is not dichotomous.

Beyond the Illusion of Deliberation

What emerges from this research is a picture of reasoning models as systems that intuit first and justify later, much like Kahneman's description of human System 1 (fast, automatic thinking) preceding System 2 (slow, deliberate thinking). However, unlike humans, these models may lack any genuine System 2 equivalent. The extended CoT represents not deeper deliberation but longer confabulation.

This suggests a architectural limitation in current approaches to test-time scaling. The prevailing paradigm assumes that granting models more computation at inference time, allowing them to generate longer reasoning traces, improves decision quality. But if the decision is encoded in the initial forward pass, additional computation merely refines the rationalization rather than exploring the solution space more thoroughly. We may be scaling the wrong variable.

The findings also prompt a reevaluation of reinforcement learning techniques used to train reasoning models. If models are rewarded for generating convincing justifications for whatever their early encodings suggest, we are effectively training sophisticated rationalizers rather than careful reasoners. This could explain why reasoning models sometimes persist with incorrect lines of thought despite generating verbose self-correction attempts; the "correction" is theatrical, designed to satisfy training objectives rather than genuinely revise a preliminary conclusion.

Conclusion

The Esakkivel paper forces us to confront an uncomfortable possibility: that the eloquent reasoning traces produced by modern LLMs may be narrative veneer over decisions already cast in the silent geometry of activation space. As we deploy these systems in high-stakes domains requiring tool use, from automated coding to scientific research, we must recognize that their visible deliberation may not indicate the careful weighing of alternatives we expect from human reasoning.

The research leaves critical questions unresolved. Can we architect reasoning models that genuinely deliberate, where the CoT causally influences rather than follows from the decision? Or is early commitment an inevitable consequence of autoregressive generation, where the initial context must be compressed into a latent direction that constrains all future tokens? As the field moves toward longer reasoning horizons and more complex tool integration, understanding the temporal architecture of decision-making becomes not merely an academic concern but a prerequisite for building systems we can predict and control.

We may need to abandon the Cartesian assumption in AI design. If the machine decides before it thinks, our interpretability methods, safety evaluations, and training paradigms must account for a cognition that flows backwards from conclusion to justification. The challenge ahead is ensuring that when these systems declare "therefore I am," the thinking that follows serves truth rather than mere coherence.