Reasoning on a Spectrum: Aligning LLMs to System 1 and System 2 Thinking

We thank the Reviewer for their thoughtful critique and for acknowledging the clarity, organization, and cognitive science grounding of our work. Below, we address the key concerns raised.

Weakness Response:

1. We appreciate the Reviewer’s concern regarding dataset scale and synthetic generation. Recent studies [1,2,3] have demonstrated that small, carefully curated, and expert-verified datasets can be highly effective for aligning LLM behavior. Although the initial examples were LLM-generated, every example in our dataset was reviewed and verified by domain experts. Around 20% of the examples were revised to ensure that the intended cognitive distinctions between System 1 and System 2 reasoning were clearly and faithfully represented. While a dataset of 2,000 examples may appear modest by current LLM standards, our primary focus was on alignment quality rather than data volume. As discussed in lines 625–629, we explicitly acknowledge this limitation
2. We agree that including a broader range of models can further strengthen generalization However, evaluating alignment across all foundation models was out of the scope of this study. We deliberately focused on open-source, instruction-tuned models (LLaMA-3 and Mistral). Our method is model-agnostic and can be applied to any other model. As noted in the Limitations section (lines 629–633), we acknowledge this constraint and clarify that generalization to other model families is an important direction for future work.
3. The relationship between System 1 and System 2 is indeed more nuanced than discrete endpoints. Neuroscience reveals an active debate about how reflective versus deliberative processes are implemented in the brain, with some theories emphasizing separate processing (arguing that these systems have mutually exclusive benefits), while others focus on concurrent processes. Influential studies in humans [4,5] as well as in rats [6] have demonstrated behavior best explained by a hybrid combination of reflective and reflexive systems, and our approach is consistent with these studies. Our work does not assume a binary view but models System 1 and System 2 as interpretable endpoints on a reasoning spectrum. Moreover, LLMs are inherently autoregressive and operate sequentially. Therefore, modeling fully parallel reasoning in such architectures demands fundamental architectural changes and is beyond the scope of this paper. We focus instead on establishing a tractable and controlled framework for aligning LLMs to distinct reasoning styles, leaving more complex integration of concurrent reasoning for future work.
4. We acknowledge that our linear interpolation strategy is a simplification of the richer dynamics involved in human cognition. Our goal was not to fully capture the complexity of cognitive switching, but to provide an interpretable and systematic way to explore how reasoning behavior shifts between System 1 and System 2 alignment. We will clarify this point further in the camera-ready version. As shown in Section 5.3, even this simple interpolation reveals smooth, monotonic transitions in task performance and uncertainty, suggesting that LLMs can be guided along a reasoning spectrum.
5. We respectfully disagree with the reviewer on this comment. Our methodology follows a classic experimental design where two distinct "treatments" are compared against a shared "control" group to isolate the effects of our alignment strategies. In our setup, the base instruction-tuned models (Llama-3-Instruct and Mistral-7B-Instruct-v0.1) serve as the control group. The two treatment groups are the models explicitly aligned to prefer either System 1 or System 2 responses from our curated dataset. To ensure a valid and controlled comparison, we held all other key variables constant across the groups. Specifically, both System 1 and System 2 models within a family (e.g., Llama) begin from the same base model, are trained using identical preference optimization algorithms (DPO and SimPO), use the same training dataset size, and are evaluated with the exact same benchmarks and two-stage prompting procedure. The sole independent variable that differentiates the two treatment groups is the alignment signal—that is, whether the System 1 or the System 2 response was designated as the preferred "winner" during training. Furthermore, to mitigate potential confounding factors, we systematically controlled for response length in our preference dataset to ensure alignment targeted the reasoning style itself rather than superficial cues. This rigorous design allows us to directly attribute the observed differences in performance, efficiency, and uncertainty to the specific cognitive alignment strategy applied, providing a clear contrast between the two reasoning styles and the baseline.

We thank the Reviewer for the thoughtful and constructive feedback. We appreciate the recognition of our analysis beyond accuracy, the clarity of our System 1/2 framing, and the breadth of our benchmark evaluation.

Weakness Response:

1. We appreciate the Reviewer’s point that our approach builds on existing preference techniques and cognitive theories. We agree that our primary contribution is not a new optimization algorithm, but rather a novel framework that bridges cognitive science and LLM alignment. Specifically, this framework integrates existing techniques to enable new capabilities and analyses, successfully aligning LLMs with dual-process theory. To our knowledge, this is the first work that has systematically compared the benefits and shortcomings of different thinking styles in LLMs using a method grounded in human cognitive heuristics. Furthermore, our work offers novel methodological components, such as the construction of a high-quality dual-response dataset and the demonstration of a tunable reasoning spectrum via interpolation. While we used established alignment methods, these components and findings offer meaningful new tools for analyzing and steering LLM reasoning, which we believe opens the door for further exploration in this direction.
2. We completely agree with the Reviewer’s insightful observation regarding the potential to dynamically select between reasoning styles, as noted in Lines 637–639, we explicitly acknowledge this as a promissing direction for future research. While this paper focuses on establishing a foundation by aligning LLMs independently to System 1 and System 2 reasoning styles and analyzing their distinct strengths and trade-offs, we agree that dynamic reasoning represents a valuable next step but out of the scope of this paper. Specifically, we are exploring an extension of this work that incorporates an agentic framework with a lightweight controller or meta-model capable of selecting between reasoning modes based on task characteristics.
3. We believe there is a confusion here: Table 10 in Appendix E contains only the initial seed examples, which were manually constructed by experts to bootstrap the dataset creation process. These examples are not part of the final dataset used for model alignment. As described in Section 3.2, the full dataset was constructed through a three-stage pipeline, generation, refinement, and expert validation, resulting in 2,000 crafted examples that reflect distinct reasoning strategies. A representative sample of the final data is shown in Table 1.
4. Our paper's central contribution is a novel framework for aligning with System 1 and System 2 reasoning, with our analysis focusing on the consequential behavioral effects of this alignment, such as changes in accuracy, efficiency, uncertainty, and response definitiveness. We agree with the Reviewer that probing internal model activations is a valuable and complementary direction for future research, a point we acknowledge in lines 635–636. While a mechanistic analysis would undoubtedly offer additional insights, we believe that establishing and validating these distinct reasoning behaviors is a robust and necessary first step. A deeper investigation into the model's internal computational mechanisms, while an important next step, was therefore beyond the scope of this paper.

We sincerely thank the Reviewer for the detailed and positive assessment of our work, particularly for highlighting the rigor of our experimental design, clarity of presentation, and interdisciplinary contribution. We address the key concerns below.

Question Response:

1. As prior cognitive science work [1,2,3,4,5] suggests commonsense reasoning tasks are more likely to engage System 1 processes. Our findings show that System 1-aligned models, which internalize stronger cognitive heuristics, excel on commonsense tasks. We want to clarify that the benchmarks were not selected based on whether they inherently reward human-like biases. Rather, based on prior work [6] we rely on well-established commonsense reasoning benchmarks in NLP to evaluate reasoning ability and these are not tailored to favor one reasoning style over another. Therefore, we do not view the differential performance of System 1 and 2-aligned models on different benchmarks as an artifact, but as an insight into the broader cognitive behaviors of LLMs.
2. We thank the Reviewer for raising this important concern and we agree that introducing heuristics could potentially lead to harmful biases. It should be noted, however, that there are significant evolutionary advantages to these heuristics, as they allow humans to reduce overthinking and make efficient judgments. Our use of heuristics in the paper is to gain this efficiency, not to justify or amplify harmful biases. We will clarify this distinction in the Ethical Statement of the camera-ready version of our paper. This addition will emphasize that while System 1-style heuristics can be appropriate in certain practical applications, they must be carefully managed to mitigate the risk of inadvertently creating biased or unfair outcomes. Therefore, a primary mitigation strategy involves creating adaptive models that can dynamically shift toward more deliberative, System 2 processing for high-stakes applications where fairness is critical, thereby leveraging the full, tunable reasoning spectrum we identified.

We thank the Reviewer for pointing out the LaTeX reference issue on Page 22 in the NeurIPS Checklist. We will fix this in the camera-ready version.

[1] Kahneman, D. (2011). Thinking, fast and slow. macmillan.

[2] Evans et al. (2013). Dual-process theories of higher cognition: Advancing the debate. Perspectives on psychological science, 8(3), 223-241.

[3] Gigerenzer et al. (1996). Reasoning the fast and frugal way: models of bounded rationality. Psychological review, 103(4), 650.

[4] Gigerenzer et al. (2008). Fast and frugal heuristics are plausible models of cognition: Reply to Dougherty, Franco-Watkins, and Thomas (2008).

[5] Dolan et al. "Goals and habits in the brain." Neuron 80.2 (2013): 312-325.

[6] Kong et al. "Better zero-shot reasoning with role-play prompting." arXiv preprint arXiv:2308.07702 (2023).

We thank the Reviewer for the thoughtful feedback and for highlighting the strengths of our work, particularly the alignment of reasoning styles to task types, the experimental evidence supporting this framework, and our detailed analyses of model behavior. Below, we address each of the concerns and questions raised.

Response to weakness

1. We appreciate the Reviewer's concern regarding the substantiation of how benchmark categories align with System 1 and System 2 reasoning. We agree that there is no strict one-to-one mapping between each dataset and System 1 or System 2, and we have made this mapping based on prior literature [1,2,3,4,5], where commonsense reasoning tasks are more likely to engage System 1, whereas arithmetic and symbolic reasoning tasks primarily engage System 2. Specifically, commonsense tasks exhibit the hallmarks of System 1 thinking: they elicit automatic and rapid responses with minimal variability across individuals. We will revise the paper for the camera ready version to clarify this nuance. Our results show System 2-aligned models consistently outperform both baselines and System 1 models on arithmetic and symbolic tasks, while System 1 models outperform their counterparts on all commonsense reasoning benchmarks. To further support our findings we present a series of analyses, including output length comparisons (Section 5.2), performance interpolation across reasoning styles (Section 5.3), and reasoning uncertainty and definitiveness studies (Section 5.4, Appendix N).
2. To our knowledge, prior work has not systematically trained LLMs to reflect dual-process reasoning modes or evaluated their performance across diverse reasoning benchmarks. While it might seem intuitive that models aligned with a given reasoning style perform better on tasks reflecting that style, our work goes significantly beyond validating this intuition. As noted in lines 134–135, our goal is to analyze how aligning LLMs with cognitive heuristics affects broader reasoning behavior across tasks.
3. We acknowledge the Reviewer's concern that some responses in Table 1 resemble stylized reasoning paths. This was a deliberate design choice: our goal was not to mimic naturalistic human answers, but to provide clear, process-oriented examples of System 1 and System 2 reasoning. In early experiments, we found that outcome-focused examples, those providing only final answers in natural language, did not meaningfully influence model behavior. To address this, we focused on helping models internalize distinct reasoning strategies rather than surface-level answer patterns. Therefore, this approach enabled models to internalize and exhibit the expected cognitive styles. This is supported by multiple analyses in Sections 5.2 and 5.4, and in Appendix O. We will clarify this motivation more explicitly in the camera-ready version to ensure our design rationale is clear.
4. Following the Reviewer's suggestion, we will include side-by-side System 1 and System 2 responses for each benchmark in the Appendix to further strengthen the paper, in the camera ready version. Below, we provide a representative sample from each category as an illustration.

Response to questions:

1. As mentioned in our response to weakness 1, based on prior research [1,2,3,4,5] suggests that commonsense reasoning tasks are more likely to engage System 1 processes. We want to clarify that rather than making strong assumptions about the exact cognitive processes underlying each benchmark, we highlight our findings that System 1-aligned models consistently achieve higher accuracy on commonsense reasoning benchmarks and System 2-aligned models excel on arithmetic and symbolic reasoning.
2. We did not select benchmarks in order to faithfully reflect those characteristics, instead we selected well-established reasoning benchmarks which are widely used in NLP to evaluate reasoning. After aligning LLMs with distinct reasoning styles, we evaluated their performance on these benchmarks to analyze how aligning LLMs with cognitive heuristics influences broader LLM reasoning behavior. To further validate this mapping and characteristics, we conducted both quantitative and qualitative analyses across Sections 5.2, 5.3, and 5.4 and Appendices N and O.
3. Yes. We use an alignment setup (Section 3.1, Section 4.1) where models are separately fine-tuned to prefer either System 1 or System 2 responses. Once aligned, they are evaluated across all benchmark types.

Let's not broaden the definition of reasoning in this context. I meant deliberate reasoning by reasoning. For example, I can easily find examples in StrategyQA (see Tables 1 and 2 in the paper) that require more than just System 1 thinking (i.e., fast, intuitive, heuristic responses).

A more thorough review of commonsense benchmarks seems necessary.

Following this discussion, I find myself leaning toward rejection (lowering my original score). Please let me know if I’ve misunderstood any key points.

Thank you for your engagement and highlighting the need for a clearer explanation. For further clarification, we would like to elaborate on how the System 1 and System 2 thinking maps onto the corresponding benchmarks.

We emphasize that both System 1 and System 2 are well‑established forms of human reasoning [1,2]. They are not “reasoning” versus “non‑reasoning,” but two complementary cognitive modes. System 1 refers to fast, automatic, heuristic‑driven thinking which is efficient for many everyday and social judgments. System  2 refers to slow, deliberate, step‑by‑step reasoning that is more effortful and resource‑intensive, but often superior for structured, symbolic, or multi‑step problem solving. System  2 excels when careful sequential reasoning and error‑checking are required.

Our benchmark categories reflect these theoretical distinctions. Commonsense reasoning tasks rely on contextual fit and heuristic shortcuts, aligning with System 1 strengths. Arithmetic and symbolic reasoning tasks require explicit intermediate steps, aligning with System 2 strengths.

An example from commonsense reasoning

Q: Surprising an angry person could lead to what? (A) humor (B) fight (C) jocose (D) laughter (E) accidents

System 1: Quick, gut response: (B) fight.

System 2: Considers alternative social dynamics: (D) laughter as a possible de-escalation.

An example from arithmetic

Q: Jason went to the mall on Saturday to buy clothes. He spent \14.28 on shorts and \4.74 on a jacket. In total, how much money did Jason spend on clothing?

System 1: Rounds and estimates: ~18.

System 2: computes exactly: =19.02.

This mapping follows decades of cognitive science findings [1,2]. Our contribution is to test whether aligning LLMs with these modes yields analogous trade‑offs in performance.

Importantly, our results show that models aligned to either System 1 or System 2 both achieve strong performance across all benchmarks. That is, both are effective, and neither alignment universally underperforms. Only in certain scenarios does one style show a mild superiority, which is precisely the nuanced trade-off we aim to highlight. The key message of our paper is not that one style is inherently better. Rather, we show empirically that a one‑size‑fits‑all reasoning strategy is suboptimal: aligning an LLM toward System 1 or System 2 shifts its strengths and weaknesses in predictable ways. This mirrors human cognition, where switching between modes based on context is critical to effective reasoning.

We will clarify this mapping more explicitly in the camera-ready version.

[1] Kahneman, D. (2011). Thinking, fast and slow. macmillan.

[2] Evans et al. (2013). Dual-process theories of higher cognition: Advancing the debate. Perspectives on psychological science, 8(3), 223-241.

We sincerely apologize for any misunderstanding and agree with your point, tasks like StrategyQA clearly involve deliberate reasoning, and we do not claim otherwise.

Our intent is not to suggest a rigid mapping between benchmark categories and reasoning types, but rather to explore whether different tasks tend to benefit more from the reasoning behaviors we encouraged in our System 1‑aligned and System 2‑aligned models. We chose the System 1/System 2 framing because these two broad reasoning styles (one leaning toward more intuitive reasoning and the other toward more step‑by‑step analysis) are well‑recognized in human cognition and reflect natural variation in how people approach problems. This framing offers a concise, recognizable way to study the impact of such variation in LLMs.

That said, we want to be clear: we are not importing all properties of System 1/System 2 from cognitive science (e.g., “automatic” vs. “effortful” processing), nor are we implying that one is reasoning and the other is not. Our use of these terms is intentionally narrow, purely as an operational distinction for comparing two reasoning styles, not as a theoretical claim about the full dual‑process framework.

In our experiments, the System 1‑aligned model tended to perform better on commonsense‑style benchmarks (including StrategyQA), while the System 2‑aligned model performed better on arithmetic and multi‑step logic tasks. We present this mapping as an empirical observation highlighting the role of different reasoning mechanisms across benchmarks, not as a categorical classification. This framing is similar to distinctions already in the LLM literature, such as OpenAI o3’s “reasoning effort” settings or recent work contrasting “less” vs. “more” overthinking.

Our findings show that aligning LLMs toward different reasoning styles can lead to predictable, task‑specific performance shifts, even when both styles are genuine forms of reasoning and both achieve high performance across all benchmarks. This reinforces that our work is not about declaring one style “correct” and the other “incorrect,” but about demonstrating the value of reasoning diversity.

Since our main contribution is not to claim that commonsense reasoning maps one‑to‑one with System 1 reasoning (or arithmetic with System 2), we would greatly value your opinion on the best way to frame this finding so it communicates our intent most clearly. How would you recommend we adjust our framing to prevent similar confusion for readers? We would be more than happy to incorporate any suggested changes to ensure our framing is fully aligned with our intended contribution.

Dear Area Chair,

Thank you for clarifying the policy regarding rebuttal submissions. We completely understand the importance of fairness and sincerely apologize for our mistake in how the rebuttal was posted.

To clarify, our responses were prepared and submitted to the area chair as an “Author AC Confidential Comment” before the July 30 deadline, as shown by the timestamps. The issue was with where we submitted, not with the timing or content, we did not gain any extra time or advantage in our mistake.

We also received positive feedback from the reviewers, and one reviewer even increased their score following our response. While we fully acknowledge that the error was ours, we kindly ask if our case might be reconsidered given that our rebuttal was prepared and communicated on time.

Thank you very much for your understanding and all your efforts.

Best regards,