From Extraction to Synthesis: Entangled Heuristics for Agent-Augmented Strategic Reasoning

Giovanni Battista Raggi; Marco Pedroni; Remo Pareschi; Renato Ghisellini

REVIEW 4 major objections 4 minor 1 cited by

From Extraction to Synthesis: Entangled Heuristics for Agent-Augmented Strategic Reasoning

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Strategic heuristics should be composed, not selected: a semantic interference matrix fuses conflicting maxims into coherent narratives, and the paper reports this beats rule-ranking on coherence, novelty, and strategic depth.

desk verdict The synthesis is real, but the interference mechanism is decorative; needs major revision with ablations, matched baseline, and less circular metrics. read the letter →

arxiv 2507.13768 v1 pith:5PJS73K3 submitted 2025-07-18 cs.AI

Renato Ghisellini , Remo Pareschi , Marco Pedroni , Giovanni Battista Raggi This is my paper · ORCID

classification cs.AI

keywords heuristicextractionstrategicreasoningsemanticinterferencenarrativegenerationquantumcognitioncompositionalsynthesislargelanguagemodelssentenceembeddings

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that strategic reasoning can be automated as composition rather than selection. Its system extracts conditional if-then heuristics from classical military-theory and contemporary corporate-strategy sources, embeds them in a semantic space, and activates a subset from a scenario profile. The activated heuristics are then fused through a semantic interference matrix, built from cosine similarity between heuristic embeddings, into a weighted whole, which a large language model renders as a strategic narrative under configurable rhetorical framing. The authors report that this entangled synthesis beats a rule-ranking baseline on measured coherence, novelty, and strategic depth, and that a cross-tradition synthesis for a regulatory dispute matched how the company actually framed its defense. If correct, this turns decision support from picking the best rule into generating new, context-sensitive strategy from conflicting axioms.

What carries the argument

The carrying mechanism is the semantic interference matrix and the weighted composition it feeds. Each heuristic is a sentence-embedding vector; scenario alignment produces an activation amplitude $\alpha_i$; pairwise cosine similarity fills the interference matrix; a blend operator $\mathrm{mix}(H_i,H_j)$ merges two heuristics in proportion to their interference; and the whole weighted field is handed to a large language model as a semantic scaffold that primes narrative synthesis. The matrix is simultaneously the model of 'entanglement' and the input the language model is instructed to respect, so it is the single object that converts rule selection into composition.

What would settle it

Construct two heuristics whose prescriptions are opposites but whose surface sentences are near-identical (for example, 'always act before the rival' and 'never act before the rival'); the current cosine-based interference matrix assigns them near-maximal constructive interference. If the system then fuses them as mutually reinforcing rather than mutually contradictory, the matrix is not measuring strategic interdependence.

Watch

Extended reading notes

Core claim

The central claim is that heuristics should be modeled as semantically entangled potentials rather than discrete rules. When several heuristics are activated by a scenario, a semantic interference matrix $I_{ij}=\cos\mathrm{sim}(\vec{H_i},\vec{H_j})\cdot\kappa_{ij}$, with $\kappa_{ii}=1$ and $\kappa_{ij}=\cos\mathrm{sim}(\vec{H_i},\vec{H_j})$ for $i\neq j$, modulates a weighted composition $\Phi=\sum_i \alpha_i H_i + \sum_{i\neq j} I_{ij}\,\mathrm{mix}(H_i,H_j)$, and an LLM renders this composition as a narrative. The paper argues this produces strategy that is more coherent, more novel, and more integrative than the ranked concatenation a rule-selection engine would give, and that zero literal coverage of input heuristics is a sign of generative transformation, not failure.

Load-bearing premise

Everything hinges on treating the cosine similarity between sentence embeddings of heuristics as a faithful measure of how strategic ideas reinforce, activate, or conflict in context; if that mapping fails, both the interference mechanism and the evaluation metrics built from the same embeddings stop supporting the synthesis claim.

Editorial extensions

If this is right

Strategic decision support can shift from selecting the best rule to generating a context-sensitive synthesis of many rules.
Conflicting heuristics can be held in the same reasoning step and resolved through composition rather than elimination.
Interference-modulated synthesis should deliver outputs that are more internally consistent, more novel relative to their inputs, and denser in strategic concepts than baseline concatenation.
The framework can mix heuristics across different strategic traditions in one synthesis, not just within a single author's corpus.
A case-study alignment suggests the system can produce recommendations resembling what a real organization later does.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the semantic-interference premise holds, the same compositional machinery should transfer to other domains where contradictory principles coexist, such as legal argumentation, clinical guidelines, or public-policy trade-offs, since the framing and prompt scaffolding are domain-agnostic.
The paper's zero-coverage results imply a quantifiable trade-off: forcing the language model to cite or paraphrase input heuristics would raise literal coverage while lowering novelty, a hybrid output mode the paper mentions as an adaptation rather than an evaluated claim.
Because the implemented interference coefficients never go negative, the system cannot represent genuine opposition; implementing the signed action-constraint decomposition would let it distinguish reinforcement from contradiction, which the current evaluation cannot test.
The validation against public reporting is a single post-hoc match; a stronger test would be a blind expert comparison of machine-generated and human-generated strategies, which the paper defers.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a well-written paper that proposes a generative layer on top of an extractive heuristic-recommendation system. The idea of composing heuristics rather than ranking them is legitimate, and the qualitative examples show that GPT-4 can produce genuinely novel strategic narratives when prompted with a list of heuristics and told to synthesize. But the formal apparatus—the interference matrix and Eq. (2)—never enters the actual output path. Text is generated by GPT-4 from a prompt that includes the matrix, but there is no test of whether changing the matrix changes the text. The baseline changes the prompt format and the output template, so the +28% coherence and +209% novelty numbers could simply reflect GPT-4 following instructions ('compose a synthesis' vs. 'concatenate these rules') rather than any specific effect of the entangled-heuristic model. That's a load-bearing confound.

What the paper does well: it is clear about its own limitations (Section 11.3 openly lists coefficient arbitrariness, context independence, and evaluation gaps), it provides an OSF repository, and the qualitative outputs are interesting as examples of what structured prompting can do. The six-C scenario encoding and the SBERT activation are standard, but they are applied to a new task.

Soft spots, in order of severity. First, the causal role of the interference matrix is untested: no ablation, no shuffling, no comparison of outputs with and without Iij. Given that Iij collapses to squared cosine similarity of the same embeddings used for activation, it is not clear it carries any information beyond the activation scores. Second, the evaluation metrics are computed in the same embedding space used to build the model, so coherence and novelty are partly circular. Third, the coverage metric is zero for every output and is redefined as success; that is defensible only if the redefinition is pre-registered, which it is not. Fourth, there is a single case study, no variance, no n, and no human evaluation.

Overall: the architecture is plausible but the evidence does not support 'entanglement synthesis' as a specific mechanism. The paper is a candidate for a major revision: strip the quantum vocabulary or back it with a real interference manipulation, match the baseline prompt template, add ablations, and bring in expert judges. I would not desk-reject it, because the idea deserves referee time and the authors are honest about the gaps. But as it stands, the claims outrun the data.

Referee Report

4 major / 4 minor

Summary. The paper proposes a hybrid architecture for agent-augmented strategic reasoning. It extracts conditional heuristics from classical and contemporary strategic texts, embeds them with Sentence-BERT, computes an interference matrix I_ij = cos_sim(H_i, H_j) * kappa_ij, and uses activation scores plus this matrix to prompt GPT-4 to generate narrative syntheses. The authors contrast this 'entangled synthesis' with a rule-ranking baseline and report gains in coherence, novelty, and strategic depth, illustrated through a Meta vs. FTC case study. The stated contribution is a shift from rule selection to semantic-interference-driven composition of heuristics.

Significance. If the entanglement mechanism were causally validated, the framework would be a useful contribution to agent-augmented strategy and narrative generation, and the paper's shift from rule ranking to compositional synthesis is conceptually interesting. The paper also provides a clearly described pipeline and points to a reproducibility repository on OSF. However, the current validation does not isolate the proposed mechanism: Eq. (2) is not the component that generates text, the baseline differs in prompt content and output structure, and the evaluation metrics are computed in the same embedding space used by the mechanism. Consequently, the advertised gains cannot currently be attributed to semantic interference; the significance depends on future controlled experiments.

major comments (4)

[Section 5.3, Eq. (2) and Section 6.3] The formal synthesis operator Phi in Eq. (2), including the interference-weighted mix(H_i, H_j) term, is never executed to produce the reported outputs. Section 6.3 states that text generation is delegated to GPT-4 using a prompt containing top-N heuristics, activation scores, and a flattened interference matrix; no code path or ablation demonstrates that varying I_ij changes the generated narrative. The central claim that semantic interference drives synthesis is therefore unsupported by the current experiments; the outputs could be explained by GPT-4's generic instruction-following when asked to compose a synthesis from several heuristics.
[Section 6.5 and Section 6.7] The baseline comparison is confounded by prompt design. The rule-ranking baseline selects the top-3 heuristics and concatenates their recommendations with minimal connecting language, while the entanglement condition instructs creative composition, includes activation scores and an interference matrix, and applies a rhetorical framing. The reported +28% coherence and +209% novelty improvements therefore compare different prompt templates and output styles, not solely rule selection versus entangled synthesis. A matched-prompt baseline (e.g., the same narrative style without the interference matrix) or a direct ablation of the matrix is needed to attribute the gains to the proposed mechanism.
[Section 6.6 and Section 9] The evaluation is circular with respect to the mechanism. Coherence and novelty are measured as cosine similarities computed with the same Sentence-BERT embeddings used to derive activations and the interference matrix, and novelty is defined as divergence from the input heuristics while the entangled prompt explicitly instructs creative composition. In addition, coverage is reported as 0.00 for every output and is reinterpreted as a sign of successful generative transformation rather than as a failure; this makes the metric unfalsifiable as a quality measure. The evaluation needs independent human judgments or a metric defined in a space outside the one used to construct the synthesis.
[Table 1, Section 6.7.1, and Section 8.3.1] The quantitative evidence is based on a single scenario with no reported number of runs, seeds, or statistical tests. Table 1 gives one scalar per variant; the comparative table in Section 6.7 reports standard deviations but not the sample size or how many prompts were generated. The 'Strategic Depth' metric is human-coded but no inter-annotator agreement or rubric is provided, and the Section 8.3.1 claim of empirical validation is an anecdotal comparison to a New York Times report, not a controlled test. These gaps prevent any generalizable conclusion about the framework's advantage.

minor comments (4)

[Section 4] The sentence beginning 'This structure supports machine-readable inference and positions axioms as major premises' is repeated almost verbatim in the same subsection; the duplication should be removed or merged.
[Section 11.3] The fourth limitation bullet, 'Selection Bias: The strategist library may not reflect the full di', is truncated and does not end as a complete sentence. The authors should complete this important limitation statement.
[Section 6.2.1] The relationship between the theoretical kappa_ij in Eq. (3)-(5) and the implemented kappa_ij = cosine similarity is stated, but the value of alpha and beta in Eq. (3) is given only after the equations; it would be clearer to define them where they first appear.
[Section 8.3.1] The phrase 'Expert review identified the dominant synthesis as most clear and operationalizable' lacks detail on who the experts were, how many participated, and what criteria were used; this should be reported or removed.

Circularity Check

3 steps flagged · score 6.0 of 10

The headline novelty/coherence gains are built into the prompt and metric definitions, and the formal interference operator in Eq. (2) is never executed, so the central validation is at best partially circular.

self definitional [Section 6.5 (Baseline Implementation Details), Section 6.6 (Evaluation Methodology), Section 6.7.1 (Key Findings)]
"Direct Concatenation: Combine selected heuristic recommendations with minimal connecting language (6.5); Novelty: Semantic divergence between synthesis and inputs, signaling generativity (6.6); Baseline outputs closely paraphrased input heuristics, while entanglement synthesis generated new strategic insights (6.7.1)."

The headline +209% novelty result is measured as semantic divergence from the input heuristics. The baseline is, by design, a direct concatenation of those heuristics, while the entanglement prompt asks GPT-4 to write a synthetic narrative and includes the interference matrix and framing. A model instructed to concatenate inputs will necessarily stay close to them; a model instructed to synthesize will necessarily diverge. The metric, the baseline template, and the entanglement template are therefore aligned so that the entanglement condition wins on novelty by construction. The comparison changes both instruction and output format, so it cannot measure the effect of semantic interference; the reported advantage is entailed by the definitions rather than by the model.
renaming known result [Section 5.3 Equation (2) vs. Section 6.3 (Synthesis Generation Pipeline) and Section 7.1]
"Φ = Σ_i α_i H_i + Σ_{i≠j} I_{ij} · mix(H_i, H_j) (Eq. 2); We use the OpenAI GPT-4 API to synthesize strategic narratives. Each prompt includes: A list of top-N activated heuristics, A flattened interference matrix, Activation scores, Desired framing (Sec. 6.3)."

The formal synthesis operator that is claimed to realize entangled composition is never executed on the output path. Narrative text is generated by GPT-4 from a prompt that contains the interference matrix as an input string. There is no code path, ablation, or numerical trace showing that Φ or I_{ij} changes the generated narrative; the system architecture figure ends at 'LLM Prompting (Heuristics + Iij) → Narrative Synthesis.' Presenting standard LLM prompt-conditioned narrative writing as a validated 'entanglement synthesis engine' renames a known empirical capability of LLMs as the paper's own mechanism, while the baseline changes prompt style, so the measured advantages are not attributable to Eq. (2) or to the interference values.

1 more flagged steps

fitted input called prediction [Section 6.1 (Semantic Embedding Infrastructure), Section 6.2 (Interference Matrix Computation), Section 9 (Empirical Assessment)]
"Iij = cos sim(H_i, H_j) · κij (Eq. 6); All embeddings were computed using the all-MiniLM-L6-v2 model from Sentence-BERT. A similarity threshold of 0.4 was used to determine whether an axiom was considered 'reflected' in the synthesis (Sec. 9)."

The same embedding model defines activation (cosine similarity between scenario and heuristics), the interference matrix (cosine similarity squared), and all evaluation metrics (coherence as mean sentence similarity, novelty as semantic divergence, coverage as cosine similarity against a threshold). Therefore the 'preliminary validation' is performed inside the model's own measurement space: it cannot independently confirm the central premise that Sentence-BERT cosine similarity is a faithful proxy for strategic interdependence, because both the mechanism and the validation are the same cosine operation. The evaluation is thus a self-referential check rather than an external test of the model's core assumption.

full rationale

The paper's extraction layer, thematic clustering, and the external comparison to New York Times reporting about Meta's legal strategy are self-contained and not circular. However, the central quantitative claim that entangled synthesis outperforms rule-ranking is not an independent test. The novelty metric is defined as divergence from inputs, while the baseline is instructed to concatenate those inputs and the entanglement prompt is instructed to synthesize, so the +209% novelty advantage is built into the comparison. The coherence advantage is similarly confounded by the different prompt and output templates. Furthermore, the formal interference-weighted operator in Eq. (2) is never used to generate text; GPT-4 is the actual synthesizer, and no ablation shows that the interference matrix has any causal effect. Finally, all evaluation metrics are computed with the same Sentence-BERT embedding space that defines the model's activation and interference, so the validation shares the model's assumptions rather than testing them. These issues amount to partial circularity and confounded validation, not full definitional equivalence, so the appropriate score is 6.

Assumptions & free parameters 8 free parameters · 5 assumptions · 2 invented entities

The central claim rests on hand-set scenario weights, a manually simplified interference coefficient, and an unvalidated assumption that embedding cosine similarity equals strategic interference. The actual text generation is delegated to a proprietary LLM, so the formal synthesis equation is not a computational engine.

free parameters (8)

alpha and beta calibration = alpha=2.0, beta=1.5
Hand-set in Eq 3 (Sec 5.3) to emphasize agreement over opposition; not used in the current implementation, which falls back to raw cosine similarity.
kappa_ij approximation = cos_sim(H_i,H_j) for i != j; 1.0 for i = j
Sec 6.2.1 replaces the action-constraint decomposition of Eq 3 with plain embedding similarity, so all interference is positive and derived from the same vectors used for activation.
coverage threshold = 0.4
Sec 9 uses a cosine similarity threshold of 0.4 to decide whether an axiom is 'reflected'; no sensitivity analysis is reported.
6C scenario profile = Offensive 3.88, Defensive 4.42, Relational 4.15, Potential 4.90, Temporal 3.70, Contextual 4.55
Sec 8.1: author-assigned weights for the Meta vs. FTC scenario, which drive all activation scores.
lambda_ij mixing coefficient = unspecified
Defined only as 'derived from kappa' after Eq 2; the implemented pipeline never computes it.
theme similarity threshold = 0.6
Sec 4: themes must appear in at least three strategists with similarity > 0.6; threshold chosen arbitrarily.
top-N selected heuristics = 3 for baseline, 8 for Martin-only, 4 for cross-tradition
Sec 6.5 and Sec 8: the number of heuristics fed to the LLM differs across conditions, confounding comparisons.
LLM generation parameters = temperature=0.7, max_tokens=512
Sec 6.3: chosen by hand; these control the degree of creativity and therefore the novelty metric.

assumptions (5)

domain assumption Human strategic reasoning exhibits non-classical interference that can be approximated by cosine-similarity-based vector composition.
Sec 2 and Sec 5.3; the paper explicitly frames this as inspiration from quantum cognition, not as quantum mechanics.
domain assumption The selected thinkers (Machiavelli, Sun Tzu, Clausewitz, Liddell Hart, Martin) and the curated if-then axioms faithfully represent the strategic wisdom needed for the framework.
Sec 4; axioms are produced by close reading and LLM synthesis without expert validation; Sec 11.3 partially acknowledges selection bias.
domain assumption Sentence-BERT cosine similarity is an adequate measure of semantic activation, interference, coherence, and novelty.
Sec 6.1 and Sec 9; the same embedding space is used for scenario-to-heuristic scoring, the interference matrix, and all evaluation metrics.
domain assumption GPT-4 outputs conditioned on the structured prompt faithfully realize the entanglement synthesis model.
Sec 6.3 and Sec 7; the LLM is a black box, and the paper does not verify that the interference matrix materially changes outputs beyond the heuristic list.
domain assumption BERTopic-discovered thematic clusters correspond to genuine strategic interdependencies.
Sec 4; clustering is invoked as the 'empirical foundation' for entanglement, but no validation of the clusters is provided.

invented entities (2)

Entangled semantic field
purpose: Metaphorical space in which simultaneously activated heuristics interfere constructively or destructively before synthesis.
Sec 5.2; has no falsifiable handle beyond the embedding similarities that construct it.
Semantic interference matrix I_ij
purpose: Quantifies pairwise heuristic interaction and is fed into the LLM prompt.
Sec 5.3 and Eq 6; computed from the same embeddings used for activation and evaluation, with no behavioral or expert validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Extraction to Synthesis: Entangled Heuristics for Agent-Augmented Strategic Reasoning." pith.science (2026). https://pith.science/paper/5PJS73K3

@misc{pith2026250713768,
  author       = {Pith},
  title        = {Pith review of: From Extraction to Synthesis: Entangled Heuristics for Agent-Augmented Strategic Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PJS73K3}},
  note         = {Machine review of arXiv:2507.13768}
}

read the original abstract

We present a hybrid architecture for agent-augmented strategic reasoning, combining heuristic extraction, semantic activation, and compositional synthesis. Drawing on sources ranging from classical military theory to contemporary corporate strategy, our model activates and composes multiple heuristics through a process of semantic interdependence inspired by research in quantum cognition. Unlike traditional decision engines that select the best rule, our system fuses conflicting heuristics into coherent and context-sensitive narratives, guided by semantic interaction modeling and rhetorical framing. We demonstrate the framework via a Meta vs. FTC case study, with preliminary validation through semantic metrics. Limitations and extensions (e.g., dynamic interference tuning) are discussed.

Figures

Figures reproduced from arXiv: 2507.13768 by the authors.

**Figure 2.** Semantic interference matrix among Roger Martin axioms acti [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗

**Figure 3.** Semantic interference matrix among top axioms triggered by the [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Radar chart comparing synthesis evaluation metrics— [PITH_FULL_IMAGE:figures/full_fig_p028_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Analytic Abduction: Causal Decomposition and Governed Commitment for Human--AI Coordination
cs.AI 2026-07 conditional novelty 5.0 of 10

A κ–τ governance apparatus turns 'suspended causal decomposition' into a legible, actionable output for human-AI decision-making.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages · cited by 1 Pith paper

[1]

Quantum aspects of semantic anal- ysis and symbolic artificial intelligence

Diederik Aerts and Marek Czachor. Quantum aspects of semantic anal- ysis and symbolic artificial intelligence. Journal of Physics A: Mathe- matical and General , 37(12):L123–L132, 2004. 34

work page 2004
[2]

Concepts and their dynamics: A quantum-theoretic modeling of human thought

Diederik Aerts, Liane Gabora, and Sandro Sozzo. Concepts and their dynamics: A quantum-theoretic modeling of human thought. Topics in Cognitive Science, 5(4):737–772, 2013

work page 2013
[3]

Busemeyer and Peter D

Jerome R. Busemeyer and Peter D. Bruza. Quantum Models of Cogni- tion and Decision . Cambridge University Press, 2012

work page 2012
[4]

Busemeyer, Emmanuel M

Jerome R. Busemeyer, Emmanuel M. Pothos, Raffaella Franco, and Jennifer S. Trueblood. A quantum theoretical explanation for proba- bility judgment errors. Psychological Review, 118(2):193–218, 2011

work page 2011
[5]

The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities

Gilles Fauconnier and Mark Turner. The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities . Basic Books, 2002

work page 2002
[6]

Complaint against meta platforms, inc

Federal Trade Commission. Complaint against meta platforms, inc. https://www.ftc. gov/legal-library/browse/cases-proceedings/ 191-0134-facebook-inc-ftc-v-ftc-v-meta-platforms-inc ,

work page
[7]

Walter R. Fisher. Human Communication as Narration: Toward a Philosophy of Reason, Value, and Action . University of South Carolina Press, 1987

work page 1987
[8]

Recommending actionable strategies: A semantic approach to integrating analytical frameworks with decision heuristics

Renato Ghisellini, Remo Pareschi, Marco Pedroni, and Giovanni Bat- tista Raggi. Recommending actionable strategies: A semantic approach to integrating analytical frameworks with decision heuristics. Informa- tion, 16(3), 2025

work page 2025

Show all 34 references

[9]

Simple Heuristics That Make Us Smart

Gerd Gigerenzer, Peter M Todd, and the ABC Research Group. Simple Heuristics That Make Us Smart . Oxford University Press, 1999

1999
[10]

Topics in semantic representation

Thomas Griffiths, Mark Steyvers, and Joshua Tenenbaum. Topics in semantic representation. Psychological Review, 114:211–244, 04 2007

2007
[11]

Exploring swot analysis – where are we now? : A review of academic research from the last decade

Marilyn Helms and Judy Nixon. Exploring swot analysis – where are we now? : A review of academic research from the last decade. Journal of Strategy and Management - J Econ Manag Strat , 3:215–251, 08 2010

2010
[12]

Hogarth and Natalia Karelaia

Robin M. Hogarth and Natalia Karelaia. Simple models for multiat- tribute choice with many alternatives: When it does and does not pay to face trade-offs with binary attributes.Management Science, 51, 2005. 35

2005
[13]

Mitigating large language model hallucination with faithful fine- tuning

Minda Hu, Bowei He, Yufei Wang, Liangyou Li, Chen Ma, and Irwin King. Mitigating large language model hallucination with faithful fine- tuning. CoRR, abs/2406.11267, 2024

2024 arXiv
[14]

A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst., ...

2025
[15]

Johnson-Laird

Philip N. Johnson-Laird. How We Reason . Oxford University Press, Oxford, 2006

2006
[16]

Johnson-Laird

Philip N. Johnson-Laird. Mental models and human reasoning. Pro- ceedings of the National Academy of Sciences , 107(43):18243–18250, 2010

2010
[17]

Prospect theory: An analysis of decision under risk

Daniel Kahneman and Amos Tversky. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263–291, 1979

1979
[18]

Mark zuckerberg takes stand to defend meta against antitrust suit, 2025

Cecilia Kang, Mike Isaac, and David McCabe. Mark zuckerberg takes stand to defend meta against antitrust suit, 2025. The New York Times, accessed May 2025

2025
[19]

Lee and Wolf Vanpaemel

Michael D. Lee and Wolf Vanpaemel. Quantum models of cognition as orwellian newspeak. Behavioral and Brain Sciences , 36(3):295–296, 2013

2013
[20]

Lan- guage models as few-shot learner for task-oriented dialogue systems

Andrea Madotto, Zihan Liu, Zhaojiang Lin, and Pascale Fung. Lan- guage models as few-shot learner for task-oriented dialogue systems. CoRR, abs/2008.06239, 2020

2008 arXiv
[21]

A review of faithfulness metrics for hallucination assessment in large language models

Ben Malin, Tatiana Kalganova, and Nikoloas Boulgouris. A review of faithfulness metrics for hallucination assessment in large language models. CoRR, abs/2501.00269, 2025

2025
[22]

Roger L. Martin. Playing to Win: How Strategy Really Works . Harvard Business Review Press, 2013

2013
[23]

Efficient estimation of word representations in vector space

Tom´ as Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Yoshua Bengio and Yann LeCun, editors, 1st International Conference on Learning Representations, ICLR 2013, Scottsdale, Arizona, USA, May 2-4, 2013, Work...

2013
[24]

Strategic Nar- ratives: Communication Power and the New World Order

Alister Miskimmon, Ben O’Loughlin, and Laura Roselle. Strategic Nar- ratives: Communication Power and the New World Order . Routledge, 2013

2013
[25]

Michael E. Porter. The five competitive forces that shape strategy. Harvard Business Review , 86(1):78–93, January 2008

2008
[26]

Pothos and Jerome R

Emmanuel M. Pothos and Jerome R. Busemeyer. Quantum cognition. Annual Review of Psychology , 73(Volume 73, 2022):749–778, 2022

2022
[27]

Sentence-bert: Sentence embed- dings using siamese bert-networks

Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embed- dings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inter- natio...

2019
[28]

A systematic survey of prompt engi- neering in large language models: Techniques and applications

Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Sam- rat Mondal, and Aman Chadha. A systematic survey of prompt engi- neering in large language models: Techniques and applications. CoRR, abs/2402.07927, 2024

2024 arXiv
[29]

Teler: A general taxonomy of LLM prompts for benchmarking complex tasks

Shubhra Kanti Karmaker Santu and Dongji Feng. Teler: A general taxonomy of LLM prompts for benchmarking complex tasks. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Findings of the Asso- ciation for Computational Linguistics: EMNLP 2023, Singapore, De- cember 6-10, 20...

2023
[30]

P. Taylor. The Thirty-Six Stratagems: A modern interpretation of a strategy classic. Infinite Success. Infinite Ideas, 2013

2013
[31]

Metaphor: A Computational Perspective

Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov. Metaphor: A Computational Perspective . Synthesis Lectures on Hu- man Language Technologies. Morgan & Claypool Publishers, 2016

2016
[32]

von Senger and M.B

H. von Senger and M.B. Gubitz. The Book of Stratagems: Tactics for Triumph and Survival . Viking, 1991

1991
[33]

Software framework for topic modelling with large corpora

Radim ˇReh ˚ uˇ rek and Petr Sojka. Software framework for topic modelling with large corpora. In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks , pages 45–50, Valletta, Malta, 2010. ELRA. 37

2010
[3990]

Association for Computational Linguistics, 2019

2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Quantum aspects of semantic anal- ysis and symbolic artificial intelligence

Diederik Aerts and Marek Czachor. Quantum aspects of semantic anal- ysis and symbolic artificial intelligence. Journal of Physics A: Mathe- matical and General , 37(12):L123–L132, 2004. 34

work page 2004

[2] [2]

Concepts and their dynamics: A quantum-theoretic modeling of human thought

Diederik Aerts, Liane Gabora, and Sandro Sozzo. Concepts and their dynamics: A quantum-theoretic modeling of human thought. Topics in Cognitive Science, 5(4):737–772, 2013

work page 2013

[3] [3]

Busemeyer and Peter D

Jerome R. Busemeyer and Peter D. Bruza. Quantum Models of Cogni- tion and Decision . Cambridge University Press, 2012

work page 2012

[4] [4]

Busemeyer, Emmanuel M

Jerome R. Busemeyer, Emmanuel M. Pothos, Raffaella Franco, and Jennifer S. Trueblood. A quantum theoretical explanation for proba- bility judgment errors. Psychological Review, 118(2):193–218, 2011

work page 2011

[5] [5]

The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities

Gilles Fauconnier and Mark Turner. The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities . Basic Books, 2002

work page 2002

[6] [6]

Complaint against meta platforms, inc

Federal Trade Commission. Complaint against meta platforms, inc. https://www.ftc. gov/legal-library/browse/cases-proceedings/ 191-0134-facebook-inc-ftc-v-ftc-v-meta-platforms-inc ,

work page

[7] [7]

Walter R. Fisher. Human Communication as Narration: Toward a Philosophy of Reason, Value, and Action . University of South Carolina Press, 1987

work page 1987

[8] [8]

Recommending actionable strategies: A semantic approach to integrating analytical frameworks with decision heuristics

Renato Ghisellini, Remo Pareschi, Marco Pedroni, and Giovanni Bat- tista Raggi. Recommending actionable strategies: A semantic approach to integrating analytical frameworks with decision heuristics. Informa- tion, 16(3), 2025

work page 2025