Cognitive Foundations for Reasoning and Their Manifestation in LLMs

Shuyue Stella Li; Priyanka Kargupta; Haocheng Wang; Jinu Lee; Dean Light; Shan Chen; Orevaoghene Ahia; Thomas L. Griffiths

cs.AINov 20, 2025

Cognitive Foundations for Reasoning and Their Manifestation in LLMs

Authors: Shuyue Stella Li, Priyanka Kargupta, Haocheng Wang, Jinu Lee, Dean Light, Shan Chen, Orevaoghene Ahia, Thomas L. Griffiths,

Organizations: University of Washington · University of Illinois Urbana-Champaign · Princeton University · Harvard University

Abstract

Large language models (LLMs) solve complex problems yet fail on simpler variants, suggesting they achieve correct outputs through mechanisms fundamentally different from human reasoning. To understand this gap, we synthesize cognitive science research into a taxonomy of 28 cognitive elements spanning reasoning invariants, meta-cognitive controls, representations for organizing reasoning and knowledge, and transformation operations. We introduce a fine-grained evaluation framework and apply it to 192K traces from 18 models across text, vision, and audio, alongside 1,346 human think-aloud traces on a new, domain-agnostic dataset for human-model comparison. We find that models are able to deploy cognitive elements seen in human reasoning but differ in how they use them. Leveraging these patterns, we develop test-time reasoning guidance that automatically scaffolds successful reasoning topologies, improving the accuracy of the strongest models by 9.7 percentage points on ill-structured problems. Meta-analysis of 1.6K LLM reasoning papers reveals the research community concentrates on easily quantifiable elements but rarely targets self-awareness, whose use differs between human and model reasoning. By establishing a shared vocabulary between cognitive science and LLM research, our framework enables diagnosis of reasoning failures and principled development of models that reason through robust cognitive mechanisms rather than spurious shortcuts, while providing tools to test theories of human cognition at scale.

Figures & tables

Figure 1: An example of cognitive elements present in a reasoning trace for building a LEGO spaceship. We characterize the different elements along the four dimensions of our taxonomy, as shown in Table 1 .

A. Reasoning Invariants: Properties that valid reasoning maintains across steps, problems, and domains.
height Logical coherence	Maintain consistency across reasoning steps and contexts ( Fodor and Pylyshyn, 1988 ) .
height Compositionality	Build complex ideas from simpler components ( Fodor, 1975 ) .
height Productivity	Formulate an indefinite number of thoughts or solutions using a finite set of elements ( Halford, 1989 ) .
height Conceptual processing	Operate over abstract concepts and relations, beyond surface and linguistic form ( Halford, 1989 ) .
B. Meta-Cognitive Controls: Higher-order abilities that select, monitor, and adapt processes.
height Self-awareness	Assess own knowledge state, capabilities, and task solvability ( Wicklund, 1979 ) .

Table 1: Taxonomy of cognitive elements, organized along four main dimensions: Reasoning Invariants, Meta-cognitive Controls, Reasoning Representations, and Reasoning Operations.

Modality	Dataset	#	Model	#	# Traces
Text	GeneralThought ( Taylor, 2024 )	10,307	Qwen3 (8B, 14B, 32B) ( Yang et al., 2025 )	3	31,791
	ClaimSpect ( Kargupta et al., 2025 )	290	DeepSeek-R1-Distill-Qwen2.5 (1.5B, 7B, 14B, 32B) ( DeepSeek-AI et al., 2025 )	4	42,388
			DeepSeek-R1-Distill-Llama3 (8B, 70B) ( DeepSeek-AI et al., 2025 )	2	21,194
			DeepScaleR-1.5B-Preview ( Luo et al., 2025 )	1	10,597
			s1.1-32B ( Muennighoff et al., 2025 )	1	10,597
			OpenThinker-32B ( Guha et al., 2025 )	1	10,597

Table 2: Overview of Data Sources and Models by Modality

Panel of Figure 2 — Figure 2: Dataset composition and difficulty across problem types, ordered along Jonassen’s structuredness continuum. Accuracy is lower and varies more across models on less structured problems. Details in Appendix A.7 .

Panel of Figure 3 — Figure 3: (Left) Length-normalized presence rate of each cognitive element in the traces of each model. (Right) Average reasoning trace length (# characters, log colour scale) for each model and problem type; cells marked $\emptyset$ have no traces. All models share a similar element profile and differ mainly in how many elements they exhibit.

Figure 4: Length normalized mean presence score. Most elements appear more often in model traces. Correct traces contain denser elements than incorrect ones in both models and humans ( $p<10^{-17}$ ).

Problem summary	You feel disconnected from a close friend: they take 24–48 hours to answer your texts, yet are engaged in person, active on social media, and say they are “bad at texting.” Your other friends reply within hours, and you have started to feel hurt. Diagnose what’s happening and propose how to address this.
Reasoner	Reasoning Trace Excerpt
Human (285 words)	height so what i think is happening is that there’s a mismatch in communication styles not the lack of care height <selective attention> my friend is really present and engaged when we are together in person … height at the same time height they are slow texting especially when i see them acting on social media makes it feel personal even if height they say it isn’t <context awareness> height I think the reason it hurts is because timely texting is important height to me. It’s how I stay connected and reassured. ... Even though logically, I know that might not be height true. <self-awareness> The way I address it is by being honest but calm. … height If after communication, clearly, height I still feel hurt over time, then I’ll have to reflect on whether this friendship is meeting my emotional height needs. <evaluation> …
Olmo-3-7B (744 words)	height Okay, so I need to figure out what’s going on with this friend and how to address the disconnection. height <goal management> height Let’s start by breaking down the situation the user described. First, the friend height takes 24-48 hours to respond to texts, sometimes even longer. <decomposition> … height The key here is the height discrepancy between in-person interaction and text communication. <selective attention> … height Possible height diagnoses: The issue is likely a mismatch in communication styles between the two friends. height <pattern recognition> … height Possible steps: 1. Have a calm, non-confrontational talk with the friend. height 2. Express feelings honestly but kindly. 3. ... <sequential organization> …

Table 3: Example diagnosis-solution problem. Both human and model reach the same diagnosis with shared elements, but the model packs in more elements and proceeds step by step as an outside observer, while the human reasons from their own feelings, with the context awareness and self-awareness that are more frequent in human traces.

Figure 5: Difference in allocation between correct and incorrect traces is larger in humans than in models.

Panel of Figure 6 — Figure 6: Problem type analysis. (Left) Length-normalized presence rate of each element for each problem type, ordered from well-structured to ill-structured. (Right) Positive pointwise mutual information (PPMI) between the presence of each element and trace success. Elements predict success more on less structured problems.

Figure 7: (a) Length-normalized element presence per trace. Both humans and models use fewer elements on less structured problems. (b–e) Length-normalized mean presence score of height verification , height evaluation , height goal management , and height self-awareness . Models monitor their own reasoning less on less structured problems, while humans consistently self-monitor. Per-element details in App. A.6 .

Figure 8: Successful ( $G^{*}$ ) and typical ( $\hat{G}$ ) reasoning topologies on algorithmic and diagnosis-solution problems, with each element’s association with success (NPMI) and presence rate (Prob.). Successful topologies begin with height selective attention , and the typical topologies begin with height forward chaining .

		Summary			Well-structured			Intermediate			Ill-structured			Other
	Model	Avg.	Well	Ill	Algo	Story	Rule	Decision	Troub	Diag	Case	Design	Dilem	Log	Fact
Low	R1-Distill-Qwen-1.5B	-9.5	+2.7	-10.7	+4.0	+12.0	-8.0	-32.0	-16.0	-4.0	+2.0	-14.0	-20.0	-10.0	-18.0
	DeepScaleR-1.5B	-13.1	-7.3	-15.3	-4.0	-4.0	-14.0	-20.0	-16.0	-2.0	-8.0	-16.0	-22.0	-20.0	-18.0
	Hermes-3-Llama-3-8B	+1.8	-22.7	+21.3	-20.0	-36.0	-12.0	+8.0	+8.0	+12.0	+26.0	+8.0	+30.0	-6.0	+2.0
	Olmo-3-7B-Think	+6.7	+17.3	+10.7	+16.0	+14.0	+22.0	-6.0	-6.0	-8.0	+10.0	0.0	+22.0	+8.0	+2.0
	R1-Distill-Llama-8B	-4.0	-3.3	-2.0	-14.0	+8.0	-4.0	-4.0	-10.0	-6.0	-2.0	-10.0	+6.0	-8.0	0.0

Table 4: Change in accuracy under test-time reasoning guidance (percentage points with bootstrapped standard errors). Guidance helps high-capability models most on ill-structured problems.

Figure 9: % of LLM reasoning papers in which each cognitive element is present or partially present in the research design. Research concentrates on context awareness, verification, and decomposition, but rarely targets self-awareness, causal/spatial reasoning, or backward chaining.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

	Presence		Allocation (pp)
Element	Human	Model	Human	Model
Reasoning invariants
Logical coherence	$\mathbf{+0.79}$	$+0.17$	$\mathbf{+2.40}$	$\mathbf{-1.06}$
Compositionality	$\mathbf{+0.63}$	$\mathbf{+0.50}$	$\mathbf{+2.63}$	$\mathbf{+0.63}$
Productivity	$-0.05$	$+0.11$	$\mathbf{-1.49}$	$\mathbf{-0.31}$
Conceptual processing	$\mathbf{+0.14}$	$\mathbf{+0.14}$	$-0.20$	$\mathbf{-0.53}$

Appendix

Table 5: Difference between correct and incorrect traces (correct minus incorrect) in each element’s length-normalized presence score and allocation share (percentage points) on the matched corpus (bold: BH-corrected

q<0.05

); in both groups, correct traces contain more forward chaining, compositionality, selective attention, and decomposition.

Element	Human change	Model change	Interaction $q$
Reasoning invariants
Logical coherence	$\mathbf{-0.37}$	$\mathbf{-0.23}$	$\mathbf{0.034}$
Compositionality	$-0.26$	$-0.32$	$0.391$
Productivity	$\mathbf{+0.63}$	$\mathbf{+0.84}$	$\mathbf{0.001}$
Conceptual processing	$\mathbf{+0.50}$	$\mathbf{+0.64}$	$\mathbf{0.021}$
Meta-cognitive controls

Appendix

Table 6: Change from story problems to case analysis in length-normalized presence score of each element on the matched corpus, with the BH-corrected

q

of the group

\times

problem type interaction (bold:

q<0.05

); models drop self-monitoring elements on case analysis far more than humans do.

Problem Type	Qwen3-32B	Qwen3-14B	Qwen3-8B	R1-Qwen-32B	R1-Qwen-14B	R1-Qwen-7B	R1-Qwen-1.5B	R1-Llama-70B
Algorithmic	78.2% (6059)	77.2% (6060)	74.4% (6059)	69.9% (6061)	65.6% (6061)	61.7% (6061)	46.2% (6061)	67.9% (6061)
Story Problem	84.5% (304)	87.2% (304)	85.2% (304)	77.6% (304)	78.0% (304)	67.4% (304)	57.9% (304)	77.6% (304)
Rule-Using	80.4% (491)	62.1% (491)	61.5% (491)	59.7% (491)	54.4% (491)	49.9% (491)	28.5% (491)	58.5% (491)
Decision-Making	71.9% (89)	78.7% (89)	76.4% (89)	57.3% (89)	52.8% (89)	37.1% (89)	23.6% (89)	65.2% (89)
Troubleshooting	81.7% (115)	65.2% (115)	65.2% (115)	65.2% (115)	54.8% (115)	19.1% (115)	10.4% (115)	66.1% (115)
Diagnosis-Solution	75.0% (84)	66.7% (84)	56.0% (84)	46.4% (84)	51.2% (84)	7.1% (84)	4.8% (84)	60.7% (84)

Appendix

Table 7: Accuracy by problem type and model for text reasoning problems (Part 1). Numbers in parentheses give the number of problems; Mean over types weights types equally and Overall pools all problems.

Problem Type	R1-Llama-8B	Olmo-3-32B	Olmo-3-7B	OpenThinker	s1.1-32B	Hermes-8B	DeepScaleR	R1-671B	Avg.
Algorithmic	50.5% (6061)	57.8% (6061)	42.6% (6061)	72.2% (6060)	61.6% (6061)	30.5% (6061)	49.6% (6061)	81.3% (6061)	61.7%
Story Problem	62.8% (304)	68.8% (304)	64.1% (304)	79.9% (304)	68.4% (304)	48.4% (304)	65.1% (304)	85.5% (304)	72.4%
Rule-Using	36.7% (491)	59.3% (491)	25.7% (491)	75.8% (491)	49.5% (491)	33.0% (491)	30.8% (491)	85.9% (491)	53.2%
Decision-Making	47.2% (89)	66.3% (89)	58.4% (89)	74.2% (89)	55.1% (89)	37.1% (89)	14.6% (89)	80.9% (89)	56.0%
Troubleshooting	36.5% (115)	67.8% (115)	38.3% (115)	73.9% (115)	66.1% (115)	44.3% (115)	10.4% (115)	84.3% (115)	53.1%
Diagnosis-Solution	27.4% (84)	65.5% (84)	29.8% (84)	69.0% (84)	56.0% (84)	25.0% (84)	6.0% (84)	86.9% (84)	45.8%

Appendix

Table 8: Accuracy by problem type and model (Part 2), and average across all 16 models. DeepSeek-R1’s ClaimSpect problems are judged by comparison with a reference answer.

Topology	Elements
Case analysis problems (ill-structured)
Human $G^{*}$	Selective attention (0.39, 0.63) $\rightarrow$ Logical coherence (0.54, 0.48) $\parallel$ Compositionality (0.29, 0.49) $\parallel$ Sequential organization (0.25, 0.50) $\parallel$ Context awareness (0.13, 0.40) $\rightarrow$ Strategy selection (0.20, 0.40) $\rightarrow$ Conceptual processing (0.17, 0.40)
Model $G^{*}$	Compositionality (0.74, 0.42) $\parallel$ Selective attention (0.65, 0.36) $\rightarrow$ Logical coherence (0.99, 0.04) $\parallel$ Conceptual processing (0.37, 0.29) $\parallel$ Sequential organization (0.81, 0.20) $\supset$ Productivity (0.34, 0.24) $\rightarrow$ Goal management (0.19, 0.18)
Human $\hat{G}$	Logical coherence (0.54, 0.48) $\parallel$ Selective attention (0.39, 0.63) $\parallel$ Context awareness (0.13, 0.40) $\rightarrow$ Strategy selection (0.20, 0.40) $\parallel$ Compositionality (0.29, 0.49) $\parallel$ Sequential organization (0.25, 0.50) $\rightarrow$ Goal management (0.11, 0.32)
Model $\hat{G}$	Logical coherence (0.99, 0.04) $\parallel$ Sequential organization (0.81, 0.20) $\parallel$ Selective attention (0.65, 0.36) $\rightarrow$ Compositionality (0.74, 0.42) $\parallel$ Conceptual processing (0.37, 0.29) $\rightarrow$ Strategy selection (0.29, 0.20) $\parallel$ Productivity (0.34, 0.24)

Appendix

Table 9: Successful (

G^{*}

) and typical (

\hat{G}

) seven-element topologies of human and model traces on case analysis problems in the matched corpus, with elements in order of extraction, each followed by its presence rate and NPMI (

\rightarrow

NEXT ,

\parallel

PARALLEL ,

\supset

CONTAINS ; bold : selective attention and logical coherence). Successful topologies of both groups begin with selective attention within their first two elements and do not include forward chaining.

Structured Reasoning in Language Models

See all 65 papers

The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modes
Jun 9, 2026Avinash Anand, Mahisha Ramesh, Avni Mittal +7CoT Reasoning RL for Language Model Reasoning
SIT × Nvidia AI Center (SNAIC) · MIDAS Lab, IIIT Delhi · MIDAS Lab, IIT Mandi +4
BrainBench: Exposing the Commonsense Reasoning Gap in Large Language Models
Mar 16, 2026Yuzhe TangCommonsense Reasoning Language Model Reasoning Evaluation
Georgia Institute of Technology
Think $^{2}$ : Grounded Metacognitive Reasoning in Large Language Models
Feb 21, 2026Abraham Paul Elenjical, Vivek Hruday Kavuri, Vasudeva VarmaLanguage Model Self-Assessment Metacognition in Language Models
IIIT Hyderabad

Jun 9, 2026J/K move · Enter open · S save

The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modes

Avinash Anand, Mahisha Ramesh, Avni Mittal, Ashutosh Kumar+6

SIT × Nvidia AI Center (SNAIC) · MIDAS Lab, IIIT Delhi · MIDAS Lab, IIT Mandi+4

CoT Reasoning RL for Language Model Reasoning Mathematical Reasoning LLM Reasoning

Open paper AskPDF arXiv

Large Language Models (LLMs) have achieved strong performance across natural language processing tasks, yet reliable reasoning remains an open challenge. Although modern LLMs show progress in structured inference, multi-step problem solving, and contextual understanding, their reasoning behavior is often inconsistent and sensitive to prompting strategies, task design, and model scale. This survey provides a systematic analysis of more than 300 recent papers from arXiv, Semantic Scholar, Google Scholar, Papers with Code, and the ACL Anthology to examine how reasoning capabilities emerge in LLMs and where they fail. We make three main contributions. First, we introduce a structured taxonomy of LLM reasoning research, covering Chain-of-Thought reasoning, multi-hop reasoning, mathematical reasoning, common sense reasoning, visual and temporal reasoning, code and algorithmic reasoning, retrieval-augmented reasoning, tool-augmented and agentic reasoning, and reinforcement learning-based reasoning. Second, we analyze methodological trends across these paradigms, including prompting methods, model architectures, training objectives, reward modeling, and evaluation benchmarks. Third, we synthesize recurring limitations and failure modes, such as reasoning hallucinations, brittle multi-step inference, weak causal abstraction, and poor cross-domain generalization. By organizing a rapidly expanding literature, this survey offers a unified view of the current capabilities and limitations of reasoning in LLMs. We also identify emerging research directions, including meta-reasoning, self-evolving reasoning frameworks, multimodal reasoning, and socially grounded reasoning. Overall, this work aims to serve as a reference for developing more robust, interpretable, and generalizable reasoning systems in future language models.

Abstract

Structured Reasoning in Language Models

Explore similar work

The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modes

BrainBench: Exposing the Commonsense Reasoning Gap in Large Language Models

Think2^{2}2: Grounded Metacognitive Reasoning in Large Language Models

Think $^{2}$ : Grounded Metacognitive Reasoning in Large Language Models