Thinking effort aligns between humans and reasoning models in abductive reasoning

# Thinking effort aligns between humans and reasoning models in abductive reasoning
Source: [https://arxiv.org/html/2609.01867](https://arxiv.org/html/2609.01867)
###### Abstract

A major question in cognitive modeling concerns the behavioral alignment between large language models and humans across linguistic and non\-linguistic tasks\. Unlike standard LLMs, large reasoning models \(LRMs\) are optimized with reinforcement learning from verifiable rewards, encouraging correct solutions to reasoning tasks rather than preference\-aligned responses\. Recent work[de Varda et al\. \(2025\)](https://arxiv.org/html/2609.01867#bib.bib10)investigates the cost of thinking in humans and LRMs by comparing human reaction times with model reasoning traces across a range of reasoning tasks\. We isolate this alignment by turning to abductive reasoning: unlike deductive tasks, its difficulty cannot be inferred from formal structure and offers no shortcuts a model could exploit to mimic effort without genuine search, providing firmer ground for empirical claims of shared effort\. We find further evidence of alignment between LRM and human reasoning effort, as well as evidence that models and humans tend to make similar errors\. Finally, we show that decoding methods that let models explore multiple reasoning paths increase alignment in reasoning cost between humans and LRMs across the three models tested\.

## 1Introduction

A growing area of research investigates the plausibility of LLMs as models of human language processing and cognition\([Schrimpf et al\., 2021](https://arxiv.org/html/2609.01867#bib.bib42);[Binz et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib5);[Piantadosi, 2024](https://arxiv.org/html/2609.01867#bib.bib38);[Katzir, 2023](https://arxiv.org/html/2609.01867#bib.bib23)\)\. In general, there is disagreement about the extent to which artificial neural network models are general solvers\([Lan et al\., 2023](https://arxiv.org/html/2609.01867#bib.bib28);[Brown et al\., 2020](https://arxiv.org/html/2609.01867#bib.bib6)\), especially for out\-of\-distribution problems\([Vargas Guzmán et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib51)\)\. There is an apparent tension between assessing human\-model alignment across language and general reasoning, which stems partly from the assumption that linguistic competence underlies thinking\.

fMRI evidence has shown a dissociation between the language network \(LN\) and the multiple\-demand \(MD\) network\([Fedorenko et al\., 2024a](https://arxiv.org/html/2609.01867#bib.bib15)\)\. The LN is a distinct set of brain regions in the left\-hemisphere frontal and temporal brain areas\. It is highly selective for language and causally important for linguistic decoding and encoding\. The MD network is found to be diffusely active across a range of functions, including inductive reasoning, math, and novel problem\-solving\([Fedorenko et al\., 2024b](https://arxiv.org/html/2609.01867#bib.bib16)\)\. In a study on LLM\-brain alignment over the course of training, the LN correlated primarily with formal linguistic rules and less with world knowledge and reasoning\([AlKhamissi et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib2)\)\.

Further, empirical evidence from patients with aphasia suggests that logical reasoning does not depend on natural language\([Kean et al\., 2026](https://arxiv.org/html/2609.01867#bib.bib24)\)\. Such findings may explain why LLMs learn syntactic representations faster than world knowledge\([Mahowald et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib30);[Piantadosi, 2021](https://arxiv.org/html/2609.01867#bib.bib37)\)and struggle with logic problems\([Malek et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib31)\)\.

Large reasoning models \(LRMs\) confront some of the common pitfalls in the training regime in LLMs known as reinforcement learning from human feedback \(RLHF\) by adopting a reward scheme in the fine\-tuning phase called Reinforcement learning from verifiable rewards \(RLVR\), which optimizes for correct answers on reasoning tasks\([DeepSeek\-AI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib11)\)\. LRMs produce “Chains\-of\-thought”, which are linguistic traces of the reasoning process\([Wei et al\., 2023](https://arxiv.org/html/2609.01867#bib.bib55)\), that can be used as a proxy for computational effort\. Additionally, LRMs can allocate more tokens during inference to solve more difficult problems\([Muennighoff et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib33);[Snell et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib47);[DeepSeek\-AI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib11)\), an important feature of human cognition\.

Although both LRMs and LLMs emit token sequences during CoT, it is still unclear if the underlying computation is linguistic\. Degenerate or language\-mixed traces can still confer benefits\([DeepSeek\-AI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib11)\), and meaningless filler tokens can substitute for CoT on some tasks\([Pfau et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib36)\)\. Further, reasoning can proceed entirely in latent space without emitting linguistic tokens\([Hao et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib18)\), showing the linguistic surface is separable from the computation\. As noted by[de Varda et al\. \(2025\)](https://arxiv.org/html/2609.01867#bib.bib10), CoT outputs linguistic traces that may not be faithful to the underlying computation\([Barez et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib3)\), likely because the true ’thinking’ mechanism is non\-linguistic as in humans\([Fedorenko et al\., 2024b](https://arxiv.org/html/2609.01867#bib.bib16)\)\.

### 1\.1Related Work

Recently,[de Varda et al\. \(2025\)](https://arxiv.org/html/2609.01867#bib.bib10)investigated the alignment between LRM reasoning traces measured by token count and human reaction times across seven diverse reasoning tasks\. In testing on DeepSeek\-R1, they find that models reliably track human thinking effort both within and across tasks\. They also replicate their findings to five additional open\-weight reasoning models\.

This work has garnered substantial discourse in a short period of time\. Some authors have expressed skepticism about the algorithmic correspondence between the brain and LRMs\([Dujmović, 2026](https://arxiv.org/html/2609.01867#bib.bib14);[Vankov et al\., 2026](https://arxiv.org/html/2609.01867#bib.bib50)\)\. Importantly, the investigation of such an alignment is not a mechanistic thesis but rather an empirical phenomenon that may account for similar constraints across substrates\. Other criticism has focused on the validity of CoT as a measure of thinking effort\([Hu, 2026](https://arxiv.org/html/2609.01867#bib.bib22)\), which we address in part by testing multiple decoding strategies and reasoning effort levels \(see[A\.3](https://arxiv.org/html/2609.01867#A1.SS3)\)\.

Theoretically, such observations are relevant in the contentious debate of whether symbolic processing can emerge from sub\-symbolic architectures\([Fodor and Pylyshyn, 1988](https://arxiv.org/html/2609.01867#bib.bib17);[Smolensky, 1988](https://arxiv.org/html/2609.01867#bib.bib46)\)\.

### 1\.2Abductive Reasoning

We examine the alignment between human and model reasoning effort in abductive inference\. Introduced by[Peirce \(1965\)](https://arxiv.org/html/2609.01867#bib.bib35), abduction is colloquially understood as an educated guess or best explanation\([Burch, 2024](https://arxiv.org/html/2609.01867#bib.bib7)\)\. More formally, it is the process of forming a hypothesis that serves as an explanation for underdetermined premises\. Abduction is a natural form of reasoning that pervades daily life where people operate under uncertainty\. Waking up to the sight of wet grass, one could posit several abductive explanations: my neighbor watered the grass last night, a nearby ground well began to overflow, the sprinkler system achieved sentience and rebelled, or perhaps it rained\.

It is distinct from induction and deduction and has been argued to be a core tool for generating hypotheses about scientific explananda\([Lipton, 2008](https://arxiv.org/html/2609.01867#bib.bib29);[Seddon, 2022](https://arxiv.org/html/2609.01867#bib.bib44)\)and central to reasoning between the lines in discourse\([Hobbs et al\., 1988](https://arxiv.org/html/2609.01867#bib.bib20)\)\.

In human\-model alignment, abduction serves as a uniquely fruitful type of reasoning to study\. In syllogistic and deductive tasks, a problem’s difficulty is partially recoverable from its formal structure; a model that has learned the relevant inference schemas could produce more tokens on harder\-looking items without doing real search\.111Thanks to Andrea de Varda for suggesting this framing \(personal communication\)\.Abduction lacks this property: since the evidence underdetermines the conclusion by design, no such shortcut is available, providing stronger evidence of shared effort\. As an example, consider two syllogisms of similar surface length:

- \(1\)All A are B; all B are C; therefore all A are C\. \(easy mood\)
- \(2\)Some A are not B; all C are B; therefore some A are not C\. \(harder mood\)

A model that has learned which forms are hard could allocate more tokens to \(2\) reflecting memorized structural difficulty cues rather than shared effort\. In our abductive task, there are no formal markers that would suggest which answer is correct or how difficult the discrimination is\. Difficulty arises only from searching over what each hypothesis would explain\.

### 1\.3Abduction in LLMs

The study of LLMs’ abductive ability has recently attracted interest \(for a review, see[Salimi et al\., 2026](https://arxiv.org/html/2609.01867#bib.bib41)\)\. Studies have covered abduction in syllogistic form\([Abe et al\., 2026](https://arxiv.org/html/2609.01867#bib.bib1)\), commonsense hypothesis selection/generation\([Bhagavatula et al\., 2020](https://arxiv.org/html/2609.01867#bib.bib4)\), and medical diagnosis\([Wu et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib56)\)\. Most previous work has observed that LLMs struggle with abduction and are consistently outperformed by humans\. Explicit inference to the best explanation criteria has also been shown to improve model\-based evaluation of commonsense explanations\([Dalal et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib9)\)\.

### 1\.4Abductive Reasoning Dataset

We operationalize abduction in a forced\-choice task in which two candidate hypotheses are presented as possible explanations for two observations, in the style of inference to the best explanation\. In an online experiment, we collected data from 120 participants, recording accuracy and reaction time\. The experimental items were drawn from[Bhagavatula et al\. \(2020\)](https://arxiv.org/html/2609.01867#bib.bib4), who created a commonsense dataset adapted from ROCStories\([Mostafazadeh et al\., 2016](https://arxiv.org/html/2609.01867#bib.bib32)\), a corpus of narrative short stories\.[Bhagavatula et al\. \(2020\)](https://arxiv.org/html/2609.01867#bib.bib4)used the beginnings and endings of stories as observations and recruited human participants to generate hypotheses explaining the change of state from O1 to O2\. They had a different set of participants revise the hypothesis in a minimal way that would make one hypothesis wrong\. The final step was to use BERT as an adversarial filter to select hypothesis pairs that were most difficult to discern as the correct answer\.

### 1\.5Decoding

The decoding methods in LLMs have been widely observed to affect reasoning patterns and accuracy\([Wang et al\., 2023](https://arxiv.org/html/2609.01867#bib.bib53);[Wang and Zhou, 2024](https://arxiv.org/html/2609.01867#bib.bib54);[Wang et al\., 2020](https://arxiv.org/html/2609.01867#bib.bib52)\)\. See[Renze \(2024\)](https://arxiv.org/html/2609.01867#bib.bib40)for opposing evidence\. Temperature is a hyperparameter in language models that adjusts the target distribution produced by the final softmax layer; a higher temperature yields a softer distribution over candidate tokens\([Hinton et al\., 2015](https://arxiv.org/html/2609.01867#bib.bib19)\)\. Decoding strategies that optimize for high\-probability outputs can lead to degeneration\([Holtzman et al\., 2019](https://arxiv.org/html/2609.01867#bib.bib21)\)\. Thus, running reasoning models at lower temperatures can increase the risk of thinking loops, in which models produce repetitive or near\-identical reasoning traces, whereas higher temperatures can avoid looping\([Pipis et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib39)\)\. Model providers suggest running models at sufficiently high temperatures to encourage exploration of alternative reasoning paths and to avoid loops; these temperatures usually range from 0\.6 to 1\([DeepSeek\-AI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib11);[Yang et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib57);[OpenAI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib34)\)\.

The paper is structured as follows\. We first introduce the dataset and experimental setup, including our adaptation of ’s method to abductive reasoning under an inference\-to\-the\-best\-explanation paradigm, before presenting the main results\. We then examine non\-greedy decoding, showing that temperature\-based sampling aggregated across multiple runs strengthens human–model correlations\. Finally, we situate these findings within previous work on thinking\-cost alignment and discuss why the decoding strategy may matter for measuring reasoning effort\.

Since human reaction times are strongly influenced by item length, all reported token–RTrs are partial correlations controlling for problem length\.

## 2Methods

We tested participants \(N=120N=120\) across 160 items from the Abductive Natural Language Inference Dataset from AI2 in an online forced\-choice experiment\. Participants were instructed to read two observations that described commonsense scenarios and select one of two hypotheses that best explained both observations\. Participants were instructed to answer as quickly as they could while remaining accurate\.

Observation 1:Alec works at the renaissance fair\.
Observation 2:They had to make it seem like part of the act\.
Hypothesis A:Alec jumped up in front of everyone\.
Hypothesis B:Alec’s phone started ringing during a performance\.
Correct answer: Hypothesis BFigure 1:Example item from the abductive forced\-choice task\.### 2\.1Participants

Participants \(N=127N=127\) were recruited through Prolific\. All participants were Native English speakers residing in the United States with normal to corrected\-to\-normal vision\. A total of 7 participants were removed for implausibly fast responses or disproportionately long responses, indicating they had left their computer\. The final sample includes 120 participants \(mean age = 33\.0 years, SD = 6\.6; 60% Female, 38% Male, 2% Other\)\.

### 2\.2Design

Participants were divided into 4 blocks; each block contained 40 unique abductive scenarios, covering 160 questions\. In the analysis, Block 1 \(N=28N=28\), Block 2 \(N=31N=31\), Block 3 \(N=32N=32\), Block 4 \(N=29N=29\), with no overlapping problems between blocks\.

### 2\.3Procedure

The survey was conducted in English\. Participants used the keyboard to choose between hypotheses on\-screen using the Q and P keys\. Before beginning the task, participants were tested with two multiple\-choice questions to ensure they understood the instructions\. Two practice problems were provided to help participants familiarize themselves with the layout, and they were instructed to keep their fingers on P and Q throughout the experiment\. The answer to each question was randomized between P and Q to avoid biasing participants towards one key\. Two attention trials were added to catch participants who were unfocused\. A fixation cross appeared briefly between trials \(200ms\), and time was unlimited for each problem\. The mean completion time for the study was 9\.2 minutes \(SD = 4\.0\)\.

### 2\.4Materials

The ART dataset from AllenAI consists of a multiple\-choice question\-answering task\. Each item consists of two observations \(O1, O2\) and two hypotheses \(A, B\), with the task being to select the hypothesis that better explains the observations\. The data is publicly available on[HuggingFace](https://huggingface.co/datasets/allenai/art)and consists of training and test sets\. To prevent participants from using response\-option length as a heuristic, we retained only items for which hypotheses A and B differed by at most 1 word\. Duplicate contexts \(identical O1–O2 pairs\) were removed\. From the resulting pool, 162 items were randomly sampled \(seed = 42\): 160 main items and 2 practice items\. All items came from the test set, where items underwent an adversarial filtering algorithm to retain hypotheses that are harder to distinguish, which demands deeper reasoning from participants and models\. Practice items were selected as the two shortest items \(by total word count\) in the sample to serve as clear warm\-up examples and were the same across all blocks\. The 160 main items had a mean total word count of 33\.5 \(SD = 7\.0, range = 19–70\) across both observations and both hypotheses\. Of the 160 items, 68 had hypotheses of equal length, and 92 differed by exactly one word\. The correct answer was option A on 87 items and option B on 73 items\.

In this task, you will read two observations about a situation, followed by two possible explanations \(hypotheses\)\. The observations describe the beginning and end of a short story\. Your task is to decide which hypothesis more plausibly explains what happened in between\. Read the observations and hypotheses carefully, then select the hypothesis that best connects Observation 1 to Observation 2\. Your primary goal should be to give the correct answer, but try to do so as quickly as possible\.
Observation 1:\[O1\]
Observation 2:\[O2\]
Hypothesis A:\[A\]
Hypothesis B:\[B\]
Which hypothesis \(A or B\) more plausibly explains what happened between Observation 1 and Observation 2?
Format your answer like this:‘Answer: A’or‘Answer: B’\.
\[V3 only\]:Let’s think step by step\.Figure 2:Prompt template used across conditions\. Placeholders in italics are filled with each item’s content\. The grey line was appended only in condition V3\.
### 2\.5Model evaluation

Following the method of[de Varda et al\. \(2025\)](https://arxiv.org/html/2609.01867#bib.bib10), we analyze seven large reasoning models \(LRMs\) and one non\-reasoning baseline: DeepSeek R1 \(hereafter R1\), Qwen3\-235B \(Thinking\), Qwen3\-32B \(Thinking\), GPT\-OSS\-20B, GPT\-OSS\-120B, GLM\-4\.5\-Air, Kimi\-K2 \(Thinking\), and DeepSeek V3\. Five of these were included in the original analysis; we substituted Qwen3\-32B for QwQ\-32B, which was unavailable, and added Kimi\-K2 to test robustness across reasoning models of differing size\. We use the OpenRouter API with greedy decoding \(temperature = 0\) for all models and default reasoning effort \(see Appendix[C](https://arxiv.org/html/2609.01867#A3)\)\. The prompt was designed to mirror instructions given to human participants\. For LRMs, responses include a thinking trace enclosed in<think\>\.\.\.</think\>tags, which we use to count tokens generated during reasoning; for DeepSeek V3, “let’s think step by step”\([Kojima et al\., 2023](https://arxiv.org/html/2609.01867#bib.bib25)\)is appended to the prompt to elicit chain\-of\-thought reasoning, since it is not native to the model\. Reasoning effort is measured as the total tokens generated for DeepSeek V3\.

We also expand the previous work by testing GPT\-OSS\-20B, R1, and Qwen3\-32B at temp 0\.6 and temp 1 as suggested by model\-provider guidelines\([DeepSeek\-AI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib11);[Yang et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib57);[OpenAI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib34)\)to elicit better reasoning patterns\. We run GPT\-OSS\-20B and Qwen3\-32B 25 times each, and 10 times for R1, and aggregate over item\-level responses to obtain the majority\-vote answer label and mean token cost\. To determine the minimum number of independent runs required for a stable partial correlation estimate, we ran a convergence bootstrap analysis\. For each value ofkk\(number of runs averaged, from 1 toK−1K\-1\), we drew 500 random subsets ofkkruns, computed the partial correlation between mean reasoning tokens and human RT \(controlling for prompt length\) for each subset, and recorded the mean and 95% CI across subsamples\. Convergence was declared at the firstkkfor which the 95% CI width fell below 0\.03, and the absolute change in mean partialrrfromk−1k\-1tokkfell below 0\.01\. To see whether higher temperatures adversely affect reasoning, we ran the three models at temp = 2, using the same number of runs\. Because bootstrap CI shrinks mechanically as k approaches the full pool, we cross\-validated our convergence estimates by exploiting the fact that chain\-of\-thought length converges as1/k1/\\sqrt\{k\}, solving analytically for the number of runs needed to bring the standard error of log\-transformed token counts below a fixed threshold \(Appendix[F](https://arxiv.org/html/2609.01867#A6)\)\.

## 3Results

We analyzed the variance in the forced\-choice test to determine whether there is a correlation between thinking models and human reasoning effort\. We computed Pearson correlations between the number of reasoning tokens per model and the average human RTs, both log\-transformed and per\-item\. Since item length is highly correlated with reaction time in our study \(r=0\.669r=0\.669,p<\.001p<\.001\), we use partial correlations to isolate thinking time \(residualizing log RT and log reasoning tokens on log prompt token count, as returned by each model’s native API tokenizer\)\. Regression lines show ordinary least\-squares fits\. Our results reflect the correlation between incorrect and correct items combined for models and humans\. An analysis of all responses \(model & human\), human\-correct RT only, and human\-correct and model\-correct only did not provide significantly different correlations for any model\. Results were robust to the operationalization of prompt length: rerunning partial correlations using log word count instead of log token count yielded no significant differences for all models except GPT\-OSS\-120B, for which the token\-based partial correlation was marginally higher \(Δ=\+0\.08\\Delta=\+0\.08, 95% CI\[\+0\.005,\+0\.161\]\[\+0\.005,\+0\.161\]\)\. In both cases, GPT\-OSS\-120B remained highly significant \(p<\.001p<\.001\), and the direction and ordering of results were unchanged\.

![Refer to caption](https://arxiv.org/html/2609.01867v1/r1_accuracy_alignment_t0.png)Figure 3:Human accuracy on items DeepSeek\-R1 answered correctly versus incorrectly \(T=0, greedy\)\. Items the model got wrong corresponded to substantially lower human accuracy\.Figure 4:Reasoning tokens versus mean human RT for all models tested under Greedy decoding; each dot represents one abductive item\.### 3\.1Greedy Decoding Correlations

After controlling for prompt length, all models are significantly correlated with human reasoning effort\. The strongest alignment was observed for GPT\-OSS\-120B \(r=0\.552r=0\.552,p<\.001p<\.001\) and GPT\-OSS\-20B \(r=0\.41r=0\.41,p<\.001p<\.001\), followed by Qwen3\-235B\-Thinking \(r=0\.359r=0\.359,p<\.001p<\.001\), DeepSeek\-R1 \(r=0\.327r=0\.327,p<\.001p<\.001\), Kimi\-K2\-Thinking \(r=0\.303r=0\.303,p<\.001p<\.001\), GLM\-4\.5\-Air \(r=0\.263r=0\.263,p<\.001p<\.001\), and Qwen3\-32B\-Thinking \(r=0\.229r=0\.229,p=\.002p=\.002\)\. The only non\-reasoning model, DeepSeek\-V3, was also significantly correlated \(r=0\.197r=0\.197,p=\.012p=\.012\), though it showed the weakest alignment of all models tested\. Notably, the difference between DeepSeek\-R1 and its non\-reasoning counterpart DeepSeek\-V3 did not reach significance \(Fisherz=1\.23z=1\.23,p=\.217p=\.217\), suggesting that while reasoning models tend to show numerically stronger alignment, the advantage over a capable non\-reasoning model was not reliably detected at this sample size\. As a robustness check, we recomputed all partial correlations using Spearman rank correlation \(see Appendix[D](https://arxiv.org/html/2609.01867#A4)\)\. Spearman partialρ\\rhovalues closely matched Pearson partialrrvalues across all models and conditions, confirming that the relationships are monotonic and not driven by distributional assumptions of linear correlation\.

### 3\.2Greedy Error Alignment

We were also interested in whether models and humans succeeded on similar items\. Model consensus accuracy was the simple unweighted proportion of models answering each item correctly, with all models contributing equally\. Mean human accuracy was 78\.8%\. Looking at DeepSeek\-R1, human accuracy on items R1 answered correctly was 82\.5%, compared to 47\.8% – below chance – on items R1 got wrong \(point\-biserialr=0\.53r=0\.53,p<\.001p<\.001; see Figure[3](https://arxiv.org/html/2609.01867#S3.F3)\)\. This pattern generalized across all models: the correlation between model consensus accuracy and human accuracy, computed as the Pearson partial correlation across the 160 items between per\-item model\-consensus accuracy and per\-item human accuracy, controlling for prompt length, wasr=0\.668r=0\.668\(p<\.001p<\.001\)\. When at least 75% of models failed an item, human accuracy was 48\.7%; when at least 75% of models agreed on the correct answer, human accuracy was 84\.6%\. This shared sensitivity to item difficulty argues against a simple memorization account: if these items were present in model pretraining, near\-ceiling performance would be expected\. The presence of systematic, human\-aligned errors instead suggests the models are reasoning about the problems\.

Figure 5:DeepSeek\-R1, GPT\-OSS\-20B, and Qwen3\-32B under greedy decoding \(T=0, top row\) and stochastic decoding \(averaged across K runs, bottom row\)\. Alignment increases reliably under stochastic sampling\.
### 3\.3Stochastic Decoding Correlations

To analyze the effects of decoding on reasoning alignment, we ran three models with non\-greedy decoding\. First, we used the temperatures recommended by the model providers: R1 \(temp = 0\.6\), GPT\-OSS\-20B \(temp = 1\), and Qwen3\-32B \(temp = 0\.6\)\. For each model, we ran it many times and computed the mean token response and answer label\.222The number of stochastic runs was determined iteratively: we chose an initial K a priori and increased it until estimates converged \(final values: K = 10 for R1, K = 25 for GPT\-OSS\-20B and Qwen3\-32B\)\.We assess stability of the pooled partial correlation across runs by Monte Carlo subsampling: for eachkk, we repeatedly drawkkruns at random, recompute partialrr, and summarize the bootstrap distribution’s width; we declare convergence at the smallestkk\(strictly below the full pool\) where the 95% interval is sufficiently narrow and the mean partialrrchanges little when adding another run\. Using the recommended temperatures and aggregating model responses improved partial correlations across all models\. In GPT\-OSS\-20B, the partial went fromr=0\.41r=0\.41tor=0\.55r=0\.55, \(Steiger–Williams test,p<\.001p<\.001\), and the model converged to a stable partial atk=19k=19runs\. For Qwen3\-32B, greedyr=0\.23r=0\.23and at temp = 0\.6,r=0\.32r=0\.32\(p<0\.05\)\(p<0\.05\), and converged at 18 runs\. R1 also improved from greedy decoding atr=0\.32r=0\.32andr=0\.37r=0\.37for temp = 0\.6, but the difference was not significant, and this model converged after 8 runs\. Across all three models, stochastic averaging consistently improved alignment with human RT, significantly so for GPT\-OSS\-20B \(p < \.001\) and Qwen3\-32B \(p < \.05\)\. We were also curious how the models would align at high, non\-recommended temperatures, so we tested all three models again at temp = 2 for the same number of runs\. GPT\-OSS\-20B partialrrdecreased to 0\.22 \(Δ\\Delta=−\-\.19\), and the partial did not converge after 25 runs \(Figure[6](https://arxiv.org/html/2609.01867#S4.F6)\)\. Qwen3\-32B partial increased byΔ\\Delta=\+\+\.063, but was still lower than the results with temp = 0\.6 and converged at 24 runs, 6 runs more than temp 0\.6\. R1 Temp = 2 increased from T=0 byΔ\\Delta=\+\+\.028 and converged after 6 runs\.333Token variance across stochastic runs did not predict human RT beyond what mean token count already explained \(partialrrs =−\.12\-\.12to\+\.08\+\.08, allpps\>\.12\>\.12\), confirming that a simple per\-item mean is sufficient and that results are not driven by a small number of high\-variance outlier runs\.

### 3\.4Stochastic Error Alignment

We also analyzed the alignment with human accuracy across recommended temperature settings and high temperatures\. We use a point\-biserial correlation between model responses \(correct/wrong\) and human accuracy % across all participants and items\. For the model GPT\-OSS\-20B, temp=1 had the highest correlation with human accuracy/errors \(r=0\.51r=0\.51\), followed byr=0\.5r=0\.5at temp = 2 andr=0\.42r=0\.42at temp = 0, allpps<\.001<\.001\. For R1, the correlation was stronger at temp = 0 \(r=0\.53r=0\.53\) than at temp = 0\.6 \(r=0\.51r=0\.51\), and lowest for T = 2 atr=0\.48r=0\.48, allpps<\.001<\.001\. Qwen3 showed the strongest correlation for t = 0\.6 withr=0\.63r=0\.63, temp = 0 withr=0\.6r=0\.6, and at temp = 2 withr=0\.54r=0\.54, allpps<\.001<\.001\. Even though Qwen and R1 had a higher partial at temp = 2 than temp = 0, their correlation with human accuracy was the lowest\.

## 4Discussion

Our results fall within the range expected from[de Varda et al\. \(2025\)](https://arxiv.org/html/2609.01867#bib.bib10)\. Averaged across models under greedy decoding, the ensemble partial correlation \(mean log tokens per item\) wasr=0\.42r=0\.42,p<\.001p<\.001, which closely matches their within\-task result for syllogisms \(r=0\.43r=0\.43\) and exceeds relational reasoning \(r=0\.27r=0\.27\)\. This comparison is theoretically meaningful because abductive forced\-choice reasoning, syllogisms, and relational reasoning all involve inference over structured, grounded scenarios\. However, abduction differs from deduction in that the evidence is underdetermined: the participant must select the best explanation rather than derive a conclusion that follows from sufficient premises\. For this reason, it is unsurprising that abductive reasoning falls near other scenario\-based reasoning tasks, rather than arithmetic or formal logic\.

Importantly, all LRMs outperform V3, the non\-reasoning LLM tested here, suggesting that chain\-of\-thought traces in LRMs capture thinking cost in a way that is not reducible to ordinary next\-token generation\. Human\-like sensitivity to content effects in LLMs may further explain why models and humans show similar item\-level difficulty patterns in such tasks\([Lampinen et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib27)\)\.

To assess how close model alignment comes to the maximum possible given measurement noise, we estimated a noise ceiling using bootstrapped split\-half reliability of human RTs \(Spearman\-Brown corrected,r=\.861r=\.861\)\. The strongest model, GPT\-OSS\-120B, explainsR2=\.303R^\{2\}=\.303of the residualized item\-level RT variance, or 35% of the noise\-ceiling\-explainable variance\. This constitutes a substantial degree of shared variance between two fundamentally different systems, which merits further investigation\.

We demonstrate that running models at their recommended temperatures for reasoning increases the alignment between human and model reasoning cost in forced\-choice abduction\. These results suggest that stochastic decoding can be used reproducibly if convergence is quantified across runs\. Higher temperatures introduce more variance, so more runs are required before the item\-level correlation stabilizes\. Averaging across runs is therefore analogous to pooling human reaction times across participants: it reduces noise from idiosyncratic samples \(see Figure[7](https://arxiv.org/html/2609.01867#A7.F7)\)\. This may be especially important for reasoning models, where generation\-time errors can compound over the reasoning trace\([Pipis et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib39)\), making unstable runs analogous to distracted or unusually slow participants\.

We also test temperatures that are typically considered too high for reasoning\. In particular, temp = 2 is rarely recommended outside of more creative generation settings\. We find that at temp = 2, model correlations are less stable, take longer to converge, and remain lower than those obtained with recommended settings such as temp = 0\.6\. Across both model\-human accuracy alignment and partial correlations with human reaction times, the recommended temperature produces the strongest alignment\.

Notably, model size did not predict alignment: GPT\-OSS\-20B \(20B parameters\) showed the second\-strongest correlation despite being over an order of magnitude smaller than DeepSeek\-R1 \(671B parameters\) or Kimi\-K2 \(1T parameters\), both of which showed weaker correlations\.

Figure 6:Convergence of the partial\-r estimate as a function of the number of stochastic runs \(K\) for GPT\-OSS\-20B at T=1\.0 and T=2\.0\. Shaded bands indicate 95% bootstrap confidence intervals\.An important question arises from our results: why does stochastic decoding yield better alignment with the cost of human reasoning?

Greedy decoding produced reliably more token outliers than stochastic sampling, with some items triggering extended thinking loops and very high token counts\. However, because tokens were log\-transformed, these extreme values had limited leverage on the correlations; removing the worst outliers did not meaningfully change the partialrr\. Greedy sampling also showed a higher raw token SD across all three models \(DeepSeek\-R1: 759 vs 613; GPT\-OSS\-20B: 2801 vs 224; Qwen3\-32B: 558 vs 342\), indicating greater overall spread\. Together, these results suggest that outliers alone do not account for greedy’s weaker alignment; the log transform largely neutralizes their influence\. Instead, the stochastic advantage appears to reflect attenuation\. Single\-run greedy estimates contain more noise per item, and this noise attenuates the correlation with human RT across the full difficulty range rather than at any particular point\.

## 5Conclusion

We investigated the alignment between large reasoning models and human thinking effort in a commonsense abductive reasoning task, where difficulty cannot be inferred from formal structure alone\. Across all reasoning models, there was a significant correlation, and alignment was stronger with LRMs than with the LLM DeepSeek\-V3\. Additionally, models and humans succeeded and failed on similar abductive items\.

In conjunction with results from other reasoning domains, we show that chain\-of\-thought length can serve as a proxy for cognitive effort\. We make the novel contribution of applying non\-greedy decoding methods to show that stochastic decoding, when averaged over many runs, reveals stronger alignment between human and model thinking effort\. This suggests that within\-task alignment in[de Varda et al\. \(2025\)](https://arxiv.org/html/2609.01867#bib.bib10)may be stronger than originally reported\.

While the weights of off\-the\-shelf models are fixed, hyperparameters can be treated as experimental variables to be tuned to avoid adverse effects – analogous to how we control environmental conditions in human experiments\. Alternatively, stochastic decoding may reflect a cognitive process that is more similar in humans than greedy decoding\.

## 6Future work

Future work could run stochastic sampling across a wider range of reasoning models to better understand what drives the differential increase in alignment across models\. Hyperparameters such as top\_p and top\_k were kept at default values here, but future work could examine their effects\.

It would also be useful to adapt our method to other abductive reasoning datasets, such as long\-form detective settings, which require deeper reasoning and the ability to connect clues across context\([Del and Fishel, 2023](https://arxiv.org/html/2609.01867#bib.bib13)\)\.

There is also potential to measure alignment with CoT traces beyond linguistic output, such as continuous latent space\([Hao et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib18)\), or with mechanistic properties of reasoning effort, such as token\-revision counts in deep layers\([Chen et al\., 2026](https://arxiv.org/html/2609.01867#bib.bib8)\)\.

## Limitations

We found that effort increased correctness on the hardest tercile of problems \(see Appendix[A\.3](https://arxiv.org/html/2609.01867#A1.SS3)\)\. All main analyses used each model’s default reasoning\-effort setting rather than tuning it to difficulty — a bias set a priori, much as a human told in advance to expect an easy, medium, or hard problem would shift their response speed\. This suggests alignment may be stronger with datasets calibrated to avoid ceiling effects, echoing[Lampinen \(2024\)](https://arxiv.org/html/2609.01867#bib.bib26)’s point that fair LLM\-human comparison requires matching the conditions faced by human participants\.

We tested one type of abductive reasoning, in the style of inference to the best explanation\. However, this does not represent all forms of abductive reasoning, and important differences in hypothesis selection and generation are not captured here\.

In our model runs, we used a prompt as close as possible to human directions; however, LLM outputs are known to be sensitive to prompts\([Zhao et al\., 2021](https://arxiv.org/html/2609.01867#bib.bib58);[Sclar et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib43)\), and variation in instructions could affect our results\.

Although shared error patterns with humans suggest that ground\-truth labels were not simply memorized, we cannot rule out the possibility that models encountered ART items during pretraining, which may inflate accuracy through familiarity rather than genuine reasoning\.

## Ethical Considerations

The observed alignment does not imply equivalence between models and humans\. Our focus is evidence of shared computational constraints, which matters for cognitive modeling but does not license anthropomorphizing or increased trust in model outputs\.

No personally identifiable information was collected as part of the dataset\. All participants were recruited and compensated through Prolific and gave informed consent for their anonymized behavioral data, including response accuracy and reaction times, to be used in a study of human reasoning\.

## Acknowledgments

Thanks to Andrea de Varda, Jakub Szymanik, and the anonymous reviewers for helpful discussions\.

## Data Availability

## References

- Abe et al\. \(2026\)Hirohiko Abe, Risako Ando, Takanobu Morishita Kentaro Ozeki, Koji Mineshima, and Mitsuhiro Okada\. 2026\.[Abductive Reasoning with Syllogistic Forms in Large Language Models](https://doi.org/10.48550/arXiv.2603.06428)\.*arXiv preprint*\.ArXiv:2603\.06428 \[cs\.CL\]\.
- AlKhamissi et al\. \(2025\)Badr AlKhamissi, Greta Tuckute, Yingtian Tang, Taha Osama A Binhuraib, Antoine Bosselut, and Martin Schrimpf\. 2025\.[From Language to Cognition: How LLMs Outgrow the Human Language Network](https://doi.org/10.18653/v1/2025.emnlp-main.1237)\.In*Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing*, pages 24321–24339, Suzhou, China\. Association for Computational Linguistics\.
- Barez et al\. \(2025\)Fazl Barez, Tung\-Yu Wu, Iván Arcuschin, Michael Lan, Vincent Wang, Noah Siegel, Nicolas Collignon, Clement Neo, Isabelle Lee, Alasdair Paren, Adel Bibi, Robert Trager, Damiano Fornasiere, John Yan, Yanai Elazar, and Yoshua Bengio\. 2025\.[Chain\-of\-thought is not explainability](https://aigi.ox.ac.uk/wp-content/uploads/2025/07/Cot_Is_Not_Explainability.pdf)\.
- Bhagavatula et al\. \(2020\)Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen\-tau Yih, and Yejin Choi\. 2020\.[Abductive Commonsense Reasoning](https://doi.org/10.48550/arXiv.1908.05739)\.*ICLR Conference Paper*\.ArXiv:1908\.05739 \[cs\.CL\]\.
- Binz et al\. \(2025\)Marcel Binz, Elif Akata, Matthias Bethge, Franziska Brändle, Fred Callaway, Julian Coda\-Forno, Peter Dayan, Can Demircan, Maria K\. Eckstein, Noémi Éltető, Thomas L\. Griffiths, Susanne Haridi, Akshay K\. Jagadish, Li Ji\-An, Alexander Kipnis, Sreejan Kumar, Tobias Ludwig, Marvin Mathony, Marcelo Mattar, and 21 others\. 2025\.[A foundation model to predict and capture human cognition](https://doi.org/10.1038/s41586-025-09215-4)\.*Nature*, 644\(8078\):1002–1009\.
- Brown et al\. \(2020\)Tom B\. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert\-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M\. Ziegler, Jeffrey Wu, Clemens Winter, and 12 others\. 2020\.[Language Models are Few\-Shot Learners](https://doi.org/10.48550/arXiv.2005.14165)\.*arXiv preprint*\.ArXiv:2005\.14165 \[cs\.CL\]\.
- Burch \(2024\)Robert Burch\. 2024\.[Charles Sanders Peirce](https://plato.stanford.edu/archives/sum2024/entries/peirce/)\.In Edward N\. Zalta and Uri Nodelman, editors,*The Stanford Encyclopedia of Philosophy*, summer 2024 edition\. Metaphysics Research Lab, Stanford University\.
- Chen et al\. \(2026\)Wei\-Lin Chen, Liqian Peng, Tian Tan, Chao Zhao, Blake JianHang Chen, Ziqian Lin, Alec Go, and Yu Meng\. 2026\.[Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep\-Thinking Tokens](https://doi.org/10.48550/arXiv.2602.13517)\.*arXiv preprint*\.ArXiv:2602\.13517 \[cs\.CL\]\.
- Dalal et al\. \(2024\)Dhairya Dalal, Marco Valentino, Andre Freitas, and Paul Buitelaar\. 2024\.[Inference to the best explanation in large language models](https://doi.org/10.18653/v1/2024.acl-long.14)\.In*Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pages 217–235, Bangkok, Thailand\. Association for Computational Linguistics\.
- de Varda et al\. \(2025\)Andrea Gregor de Varda, Ferdinando Pio D’Elia, Hope Kean, Andrew Lampinen, and Evelina Fedorenko\. 2025\.[The cost of thinking is similar between large reasoning models and humans](https://doi.org/10.1073/pnas.2520077122)\.*Proceedings of the National Academy of Sciences*, 122\(47\):e2520077122\.
- DeepSeek\-AI et al\. \(2025\)DeepSeek\-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z\. F\. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others\. 2025\.[DeepSeek\-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://doi.org/10.1038/s41586-025-09422-z)\.*Nature*, 645\(8081\):633–638\.ArXiv:2501\.12948 \[cs\.CL\]\.
- DeepSeek\-AI et al\. \(2024\)DeepSeek\-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, and 15 others\. 2024\.[Deepseek\-v3 technical report](https://arxiv.org/abs/2412.19437)\.*Preprint*, arXiv:2412\.19437\.
- Del and Fishel \(2023\)Maksym Del and Mark Fishel\. 2023\.[True Detective: A Deep Abductive Reasoning Benchmark Undoable for GPT\-3 and Challenging for GPT\-4](https://doi.org/10.18653/v1/2023.starsem-1.28)\.In*Proceedings of the 12th Joint Conference on Lexical and Computational Semantics \(\*SEM 2023\)*, pages 314–322, Toronto, Canada\. Association for Computational Linguistics\.
- Dujmović \(2026\)Marin Dujmović\. 2026\.[No deep insights into the alignment between human and deep learning reasoning processes: Thoughts on de Varda et al\. \(2025\)](https://doi.org/10.1073/pnas.2533685123)\.*Proceedings of the National Academy of Sciences of the United States of America*, 123\(4\):e2533685123\.
- Fedorenko et al\. \(2024a\)Evelina Fedorenko, Anna A\. Ivanova, and Tamar I\. Regev\. 2024a\.[The language network as a natural kind within the broader landscape of the human brain](https://doi.org/10.1038/s41583-024-00802-4)\.*Nature Reviews Neuroscience*, 25\(5\):289–312\.
- Fedorenko et al\. \(2024b\)Evelina Fedorenko, Steven T\. Piantadosi, and Edward A\. F\. Gibson\. 2024b\.[Language is primarily a tool for communication rather than thought](https://doi.org/10.1038/s41586-024-07522-w)\.*Nature*, 630\(8017\):575–586\.
- Fodor and Pylyshyn \(1988\)Jerry A\. Fodor and Zenon W\. Pylyshyn\. 1988\.[Connectionism and cognitive architecture: A critical analysis](https://doi.org/10.1016/0010-0277(88)90031-5)\.*Cognition*, 28\(1\):3–71\.
- Hao et al\. \(2025\)Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian\. 2025\.[Training Large Language Models to Reason in a Continuous Latent Space](https://doi.org/10.48550/arXiv.2412.06769)\.*arXiv preprint*\.ArXiv:2412\.06769 \[cs\.CL\]\.
- Hinton et al\. \(2015\)Geoffrey Hinton, Oriol Vinyals, and Jeff Dean\. 2015\.[Distilling the Knowledge in a Neural Network](https://doi.org/10.48550/arXiv.1503.02531)\.*arXiv preprint*\.ArXiv:1503\.02531 \[stat\.ML\]\.
- Hobbs et al\. \(1988\)Jerry R\. Hobbs, Mark Stickel, Paul Martin, and Douglas Edwards\. 1988\.[Interpretation as Abduction](https://doi.org/10.3115/982023.982035)\.In*26th Annual Meeting of the Association for Computational Linguistics*, pages 95–103, Buffalo, New York, USA\. Association for Computational Linguistics\.
- Holtzman et al\. \(2019\)Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi\. 2019\.[The Curious Case of Neural Text Degeneration](https://arxiv.org/abs/1904.09751v2)\.ArXiv:1904\.09751 \[cs\.CL\]\.
- Hu \(2026\)Yueqing Hu\. 2026\.[“thinking traces” in large reasoning models: Cognitive cost or performative scaffolding?](https://doi.org/10.1073/pnas.2604554123)*Proceedings of the National Academy of Sciences*, 123\(17\):e2604554123\.
- Katzir \(2023\)Roni Katzir\. 2023\.[Why Large Language Models Are Poor Theories of Human Linguistic Cognition: A Reply to Piantadosi](https://doi.org/10.5964/bioling.13153)\.*Biolinguistics*, 17:1–12\.
- Kean et al\. \(2026\)Hope Kean, Alexander Fung, Paris Jaggers, Jason Chen, Joshua S\. Rule, Yael Benn, Joshua B\. Tenenbaum, Steven T\. Piantadosi, Rosemary A\. Varley, and Evelina Fedorenko\. 2026\.[Evidence from formal logical reasoning reveals that the language of thought is not natural language](https://doi.org/10.1073/pnas.2520095123)\.*Proceedings of the National Academy of Sciences*, 123\(28\):e2520095123\.
- Kojima et al\. \(2023\)Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa\. 2023\.[Large Language Models are Zero\-Shot Reasoners](https://doi.org/10.48550/arXiv.2205.11916)\.*arXiv preprint*\.ArXiv:2205\.11916 \[cs\.CL\]\.
- Lampinen \(2024\)Andrew Lampinen\. 2024\.[Can language models handle recursively nested grammatical structures? a case study on comparing models and humans](https://doi.org/10.1162/coli_a_00525)\.*Computational Linguistics*, 50\(4\):1441–1476\.
- Lampinen et al\. \(2024\)Andrew K Lampinen, Ishita Dasgupta, Stephanie C Y Chan, Hannah R Sheahan, Antonia Creswell, Dharshan Kumaran, James L McClelland, and Felix Hill\. 2024\.[Language models, like humans, show content effects on reasoning tasks](https://doi.org/10.1093/pnasnexus/pgae233)\.*PNAS Nexus*, 3\(7\):pgae233\.
- Lan et al\. \(2023\)Nur Lan, Emmanuel Chemla, and Roni Katzir\. 2023\.[Benchmarking Neural Network Generalization for Grammar Induction](https://aclanthology.org/2023.clasp-1.15/)\.In*Proceedings of the 2023 CLASP Conference on Learning with Small Data \(LSD\)*, pages 131–140, Gothenburg, Sweden\. Association for Computational Linguistics\.
- Lipton \(2008\)Peter Lipton\. 2008\.Inference to the Best Explanation\.In W\. H\. Newton\-Smith, editor,*A Companion to the Philosophy of Science*, pages 184–193\. Wiley\-Blackwell\.
- Mahowald et al\. \(2024\)Kyle Mahowald, Anna A\. Ivanova, Idan A\. Blank, Nancy Kanwisher, Joshua B\. Tenenbaum, and Evelina Fedorenko\. 2024\.[Dissociating language and thought in large language models](https://doi.org/10.48550/arXiv.2301.06627)\.*arXiv preprint*\.ArXiv:2301\.06627 \[cs\.CL\]\.
- Malek et al\. \(2025\)Alan Malek, Jiawei Ge, Nevena Lazic, Chi Jin, András György, and Csaba Szepesvári\. 2025\.[Frontier LLMs Still Struggle with Simple Reasoning Tasks](https://doi.org/10.48550/arXiv.2507.07313)\.*arXiv preprint*\.ArXiv:2507\.07313 \[cs\.CL\] version: 1\.
- Mostafazadeh et al\. \(2016\)Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen\. 2016\.[A Corpus and Cloze Evaluation for Deeper Understanding of Commonsense Stories](https://doi.org/10.18653/v1/N16-1098)\.In*Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pages 839–849, San Diego, California\. Association for Computational Linguistics\.
- Muennighoff et al\. \(2025\)Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei\-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto\. 2025\.[s1: Simple test\-time scaling](https://doi.org/10.48550/arXiv.2501.19393)\.*arXiv preprint*\.ArXiv:2501\.19393 \[cs\.CL\]\.
- OpenAI et al\. \(2025\)OpenAI, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K\. Arora, Yu Bai, Bowen Baker, Haiming Bao, Boaz Barak, Ally Bennett, Tyler Bertao, Nivedita Brett, Eugene Brevdo, Greg Brockman, Sebastien Bubeck, Che Chang, and 107 others\. 2025\.[gpt\-oss\-120b & gpt\-oss\-20b Model Card](https://doi.org/10.48550/arXiv.2508.10925)\.*arXiv preprint*\.ArXiv:2508\.10925 \[cs\.CL\]\.
- Peirce \(1965\)Charles Sanders Peirce\. 1965\.*Collected Papers of Charles Sanders Peirce*\.Belknap Press of Harvard University Press, Cambridge, MA\.
- Pfau et al\. \(2024\)Jacob Pfau, William Merrill, and Samuel R\. Bowman\. 2024\.[Let’s think dot by dot: Hidden computation in transformer language models](https://doi.org/10.48550/arXiv.2404.15758)\.*arXiv preprint arXiv:2404\.15758*\.
- Piantadosi \(2021\)Steven T\. Piantadosi\. 2021\.[The Computational Origin of Representation](https://doi.org/10.1007/s11023-020-09540-9)\.*Minds and Machines*, 31\(1\):1–58\.
- Piantadosi \(2024\)Steven T Piantadosi\. 2024\.[Modern language models refute Chomsky’s approach to language](https://doi.org/10.5281/zenodo.12665933)\.In*From fieldwork to linguistic theory*, pages 353–414\. Language Science Press\.
- Pipis et al\. \(2025\)Charilaos Pipis, Shivam Garg, Vasilis Kontonis, Vaishnavi Shrivastava, Akshay Krishnamurthy, and Dimitris Papailiopoulos\. 2025\.[Wait, Wait, Wait… Why Do Reasoning Models Loop?](https://doi.org/10.48550/arXiv.2512.12895)*arXiv preprint*\.ArXiv:2512\.12895 \[cs\.LG\]\.
- Renze \(2024\)Matthew Renze\. 2024\.[The Effect of Sampling Temperature on Problem Solving in Large Language Models](https://doi.org/10.18653/v1/2024.findings-emnlp.432)\.In*Findings of the Association for Computational Linguistics: EMNLP 2024*, pages 7346–7356, Miami, Florida, USA\. Association for Computational Linguistics\.
- Salimi et al\. \(2026\)Moein Salimi, Shaygan Adim, Danial Parnian, Nima Alighardashi, Mahdi Jafari Siavoshani, and Mohammad Hossein Rohban\. 2026\.[Wiring the ’Why’: A Unified Taxonomy and Survey of Abductive Reasoning in LLMs](https://doi.org/10.48550/arXiv.2604.08016)\.*arXiv preprint*\.ArXiv:2604\.08016 \[cs\.AI\] version: 1\.
- Schrimpf et al\. \(2021\)Martin Schrimpf, Idan Asher Blank, Greta Tuckute, Carina Kauf, Eghbal A\. Hosseini, Nancy Kanwisher, Joshua B\. Tenenbaum, and Evelina Fedorenko\. 2021\.[The neural architecture of language: Integrative modeling converges on predictive processing](https://doi.org/10.1073/pnas.2105646118)\.*Proceedings of the National Academy of Sciences*, 118\(45\):e2105646118\.
- Sclar et al\. \(2024\)Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr\. 2024\.[Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting](https://doi.org/10.48550/arXiv.2310.11324)\.*arXiv preprint*\.ArXiv:2310\.11324 \[cs\.CL\]\.
- Seddon \(2022\)Peter B\. Seddon\. 2022\.[Nature Chose Abduction: Support from Brain Research for Lipton’s Theory of Inference to the Best Explanation](https://doi.org/10.1007/s10699-021-09811-3)\.*Foundations of Science*, 27\(4\):1489–1505\.
- Shazeer et al\. \(2017\)Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean\. 2017\.[Outrageously large neural networks: The sparsely\-gated mixture\-of\-experts layer](https://arxiv.org/abs/1701.06538)\.*Preprint*, arXiv:1701\.06538\.
- Smolensky \(1988\)Paul Smolensky\. 1988\.[On the proper treatment of connectionism](https://doi.org/10.1017/S0140525X00052432)\.*Behavioral and Brain Sciences*, 11\(1\):1–23\.
- Snell et al\. \(2024\)Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar\. 2024\.[Scaling LLM Test\-Time Compute Optimally can be More Effective than Scaling Model Parameters](https://doi.org/10.48550/arXiv.2408.03314)\.*arXiv preprint*\.ArXiv:2408\.03314 \[cs\.LG\]\.
- Team et al\. \(2025\)GLM\-4 5 Team, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, Kedong Wang, Lucen Zhong, Mingdao Liu, Rui Lu, Shulin Cao, Xiaohan Zhang, Xuancheng Huang, Yao Wei, and 152 others\. 2025\.[GLM\-4\.5: Agentic, Reasoning, and Coding \(ARC\) Foundation Models](https://doi.org/10.48550/arXiv.2508.06471)\.*arXiv preprint*\.ArXiv:2508\.06471 \[cs\.CL\]\.
- Team et al\. \(2026\)Kimi Team, Yifan Bai, Yiping Bao, Y\. Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, Zhuofu Chen, Jialei Cui, Hao Ding, Mengnan Dong, Angang Du, and 181 others\. 2026\.[Kimi K2: Open Agentic Intelligence](https://doi.org/10.48550/arXiv.2507.20534)\.*arXiv preprint*\.ArXiv:2507\.20534 \[cs\.LG\]\.
- Vankov et al\. \(2026\)Ivan I\. Vankov, Federico Adolfi, Rachel F\. Heaton, Guillermo Puebla, and Jeffrey S\. Bowers\. 2026\.[Correlations without causation do not support claims of human–LLM reasoning alignment](https://doi.org/10.1073/pnas.2536362123)\.*Proceedings of the National Academy of Sciences*, 123\(12\):e2536362123\.
- Vargas Guzmán et al\. \(2024\)Manuel Vargas Guzmán, Jakub Szymanik, and Maciej Malicki\. 2024\.[Testing the limits of logical reasoning in neural and hybrid models](https://doi.org/10.18653/v1/2024.findings-naacl.147)\.In*Findings of the Association for Computational Linguistics: NAACL 2024*, pages 2267–2279, Mexico City, Mexico\. Association for Computational Linguistics\.
- Wang et al\. \(2020\)Pei\-Hsin Wang, Sheng\-Iou Hsieh, Shih\-Chieh Chang, Yu\-Ting Chen, Jia\-Yu Pan, Wei Wei, and Da\-Chang Juan\. 2020\.[Contextual Temperature for Language Modeling](https://doi.org/10.48550/arXiv.2012.13575)\.*arXiv preprint*\.ArXiv:2012\.13575 \[cs\.CL\]\.
- Wang et al\. \(2023\)Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou\. 2023\.[Self\-Consistency Improves Chain of Thought Reasoning in Language Models](https://doi.org/10.48550/arXiv.2203.11171)\.*arXiv preprint*\.ArXiv:2203\.11171 \[cs\.CL\]\.
- Wang and Zhou \(2024\)Xuezhi Wang and Denny Zhou\. 2024\.[Chain\-of\-Thought Reasoning Without Prompting](https://doi.org/10.48550/arXiv.2402.10200)\.*arXiv preprint*\.ArXiv:2402\.10200 \[cs\.CL\]\.
- Wei et al\. \(2023\)Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou\. 2023\.[Chain\-of\-Thought Prompting Elicits Reasoning in Large Language Models](https://doi.org/10.48550/arXiv.2201.11903)\.*arXiv preprint*\.ArXiv:2201\.11903 \[cs\.CL\]\.
- Wu et al\. \(2025\)Kevin Wu, Eric Wu, Rahul Thapa, Kevin Wei, Angela Zhang, Arvind Suresh, Jacqueline J\. Tao, Min Woo Sun, Alejandro Lozano, and James Zou\. 2025\.[MedCaseReasoning: Evaluating and learning diagnostic reasoning from clinical case reports](https://arxiv.org/abs/2505.11733v2)\.ArXiv:2505\.11733 \[cs\.CL\]\.
- Yang et al\. \(2025\)An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others\. 2025\.[Qwen3 Technical Report](https://doi.org/10.48550/ARXIV.2505.09388)\.*Preprint*, arXiv:2505\.09388 \[cs\.CL\]\.
- Zhao et al\. \(2021\)Tony Z\. Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh\. 2021\.[Calibrate Before Use: Improving Few\-Shot Performance of Language Models](https://doi.org/10.48550/arXiv.2102.09690)\.*arXiv preprint*\.ArXiv:2102\.09690 \[cs\.CL\]\.

## Appendix AFrequently asked questions

### A\.1The difference between DeepSeek\-R1 and DeepSeek\-V3 is not statistically significant\. Does this undermine the claim that LRMs show stronger alignment with human reasoning effort?

Not necessarily\. Commonsense abductive reasoning relies heavily on plausible, linguistically\-grounded scenarios well\-represented in pretraining corpora, meaning a capable LLM like V3 can approximate human performance through world knowledge alone without explicit reasoning chains\. Notably, V3 still showed the weakest alignment of all models tested\. The gap between LRMs and standard LLMs is likely to become more pronounced on harder abductive problems that rely less on surface linguistic patterns, such as long\-form detective reasoning, where extended chain\-of\-thought would provide a clearer advantage over exclusive next\-token prediction\.

### A\.2Why does alignment not scale with model size?

All models tested \(with the exception of Qwen3\-32B\) use a Mixture\-of\-Experts architecture[Shazeer et al\. \(2017\)](https://arxiv.org/html/2609.01867#bib.bib45), where only a fraction of parameters are activated per token\. Across models, activated parameter counts are actually fairly similar, so raw size comparisons may not impact this task meaningfully\. Differences in alignment more likely reflect training dynamics particular to each provider, such as differences in RLVR reward schemes and fine\-tuning curricula\. It is also worth noting that commonsense reasoning tasks may not be the best setting to detect scale effects, since all models can draw on world knowledge from pretraining\. Tasks that rely less on linguistic and commonsense patterns may reveal a clearer relationship between model characteristics and human alignment\.

### A\.3Do the results change under different reasoning\-effort parameters?

We tested GPT\-OSS\-20B at low and high reasoning effort across all 160 items\. Token–RT alignment held at both settings \(lowr=\+\.34r=\+\.34, highr=\+\.30r=\+\.30, bothp<\.001p<\.001; Steiger’s testp=\.60p=\.60\)\. Aggregate accuracy was similar \(82\.5% vs\. 86\.3%\)\. Stratifying by human difficulty, high effort outperformed low effort on the hardest tercile \(n=52n=52: 69\.2% vs\. 53\.8%; McNemar exactp=\.022p=\.022\) but not the easiest, with the advantage growing with item difficulty \(effort×\\timeshuman\-accuracy interaction, item\-clustered SEs:β=−3\.71\\beta=\-3\.71,p=\.016p=\.016\)\. These results further suggest that reasoning alignment is clearer on hard problems and that hyperparameters should be chosen faithfully a priori\.

## Appendix BModel Specifications

#### DeepSeek\-R1

is built on DeepSeek\-V3 and uses reinforcement learning to optimize for correct answers relative to ground\-truth labels on verifiable tasks such as math\([DeepSeek\-AI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib11)\)\.

#### DeepSeek\-V3

is a Mixture\-of\-Experts model with 671B total parameters, of which 37B are activated\([DeepSeek\-AI et al\., 2024](https://arxiv.org/html/2609.01867#bib.bib12)\)\.

#### GPT\-OSS\-20B / 120B

are both Mixture\-of\-Experts models; the larger model has 128 experts, and the smaller has 32\([OpenAI et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib34)\)\.

#### Qwen3\-235B / 32B

Qwen3\-235B is a Mixture\-of\-Experts model, and Qwen3\-32B uses a dense causal architecture with a dedicated thinking mode that can be turned on/off\([Yang et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib57)\)\.

#### GLM 4\.5 Air

is a Mixture\-of\-Experts model with 355 billion total parameters and 32 billion activated parameters\([Team et al\., 2025](https://arxiv.org/html/2609.01867#bib.bib48)\)\.

#### Kimi\-K2

is a Mixture\-of\-Experts model with 1 trillion total parameters and 32 billion activated parameters\([Team et al\., 2026](https://arxiv.org/html/2609.01867#bib.bib49)\)\.

## Appendix CHyperparameters

Table 1:Hyperparameters per model\. Italics = provider default, not set in code\. n\.s\. = not specified in code; n\.s\.p\. = not specified in code or by provider; n/a = not applicable\.
## Appendix DPearson and Spearman Correlations

Table 2:Partial correlations \(Pearson and Spearman\) for models under greedy and stochastic decoding conditions\.
## Appendix EHuman vs Model Accuracy

Table 3:Accuracy on the abductive reasoning task under greedy decoding \(T=0T=0\), sorted by model\. Model CIs are Wilson score 95% intervals \(n=160n=160\); human CI is a bootstrap 95% interval \(2,000 resamples\)\.
## Appendix FChain\-of\-Thought Convergence

We estimated per\-item token variabilityσ¯\\bar\{\\sigma\}from our existing runs and solved analytically for the number of runsk∗k^\{\*\}needed to bring the standard error below a thresholdθ\\thetain log\-token space \(SE=σ¯/k\\text\{SE\}=\\bar\{\\sigma\}/\\sqrt\{k\}\), i\.e\.k∗=\(σ¯/θ\)2k^\{\*\}=\(\\bar\{\\sigma\}/\\theta\)^\{2\}\. For lower\-temperature models,k∗k^\{\*\}ranged from 5 to 18 runs; the two criteria agree on the ordering of models and temperatures, and, critically, the partialrratk=Kk=Kand atk=k∗k=k^\{\*\}did not differ significantly\. For GPT\-OSS\-20B and Qwen3\-32B at temp = 2\.0, the token variability is so large thatk∗k^\{\*\}far exceeds our pool of available runs; for GPT\-OSS\-20B this corroborates our original finding of non\-convergence, while for Qwen3\-32B it reflects a stricter criterion than the bootstrap method, which found convergence at 24 runs\.

Table 4:σ¯\\bar\{\\sigma\}: median per\-item SD of log\(thinking tokens\);k∗k^\{\*\}computed from unroundedσ¯\\bar\{\\sigma\}using the SE threshold above\.†\\daggerPool exhausted
## Appendix GStochastic Attenuation

Figure 7:GPT\-OSS\-20B thinking tokens per item \(log scale\), with items sorted by stochastic mean token count \(low to high difficulty\)\.