Artificial entrepreneurial cognition: Locating and causally steering an opportunity recognition dial inside large language models (LLMs)
Christian Fisch, Angela Altmeier, Martin Obschonka, Michal Kosinski, Pin Ni
cs.CL · cs.LG
Quick verdict
First causal steering of an entrepreneurship construct inside LLMs: rigorous, honest, cross-model work; but selectivity is partial and replications uneven.
Summary
This paper introduces 'artificial entrepreneurial cognition' — the functional organisation of entrepreneurship-relevant representations inside AI systems — and tests it for opportunity recognition (OR). Using Claude Opus 4.7, the authors generate 636 matched OR-present/OR-absent scenario pairs, filtered through vocabulary, pair-level, and batch-level rules plus an AFLite-style adversarial filter that cuts bag-of-words leakage accuracy from 0.912 to 0.528. In Llama 3.1 8B-Instruct, a difference-in-means direction at layer 12 classifies held-out pairs at 0.748 (logistic probe 0.839). Steering along this 'OR dial' at a KL-gated coefficient shifts forced-choice recognition log-odds (+0.401 up, −0.716 down), directional ablation returns the retrained probe to chance, and an LLM judge prefers OR-up free-text explanations in 0.656 of pairs (0.708 order-doubled). The direction is geometrically distinct from feasibility, desirability, and exploitation directions (cosines ≤0.328 versus 0.925 split-half reliability), yet steering it shifts all three downstream judgments. Recovery, signed steering, and geometric separation replicate in four additional models (DiM 0.705–0.744), though selectivity controls were omitted or weakened in the replications — null-tier separation reverses in Qwen3-8B. The authors candidly note that OR-up steering fails a strengthened null test, no causal double dissociation from positive valence was found, and the battery lacks independent human coding.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An OR-related linear direction is recoverable in Llama 3.1 8B-Instruct at layer 12 (held-out DiM accuracy 0.748; LR probe 0.839) and survives lexical, topical, and estimation-artefact controls.
- The direction is causally operative: signed steering shifts recognition log-odds (+0.401 up / −0.716 down at α=3.5) and directional ablation removes the signal relative to matched random directions.
- The OR direction is geometrically distinct from feasibility, desirability, and exploitation directions (max cosine 0.328 vs 0.925 split-half reliability; joint span R²=0.116), yet steering it shifts all three downstream judgments.
- Recovery, signed steering, and geometric separation replicate across four additional LLMs spanning families and scales (DiM accuracy 0.705–0.744).
- This constitutes the first causal intervention on an internal representation of an entrepreneurship construct inside an LLM.
Strengths
Disciplined contrastive battery construction
Section 3.1.1 + Appendix A: 5,000 Claude Opus 4.7 pairs pass vocabulary-, pair-, and batch-level rules, deterministic validation, and an AFLite-style TF-IDF filter (τ=0.750), leaving 636 pairs; bag-of-words separability falls 0.912→0.528 while a size-matched random subsample stays at 0.867, showing the filter — not sample size — removes lexical shortcuts.
Multi-pronged causal validation with calibrated interventions
Section 3.2: steering coefficient α=3.5 chosen on 35 held-out calibration scenarios under a 0.100 KL coherence gate (Fig. 4); signed steering (+0.401/−0.716 log-odds, Table 3), directional ablation, label permutation (p=0.002), and shuffled-label control (0.501) jointly support causal operativity beyond generic perturbation.
Theory-driven discriminant testing
Unusually transparent boundary reporting
Cross-generator and cross-model robustness
Weaknesses
Construct specificity is incomplete at the causal level
Replication depth is asymmetric relative to the headline claim
Label validity rests entirely on the generator
Modest practical effect sizes at the coherence-gated dose
Readout confounds in both behavioural channels
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper introduces *artificial entrepreneurial cognition* — the functional organisation of entrepreneurship-relevant representations inside AI systems — and operationalises it for opportunity recognition (OR). Three propositions structure the theory: representational existence (P1), causal operativity (P2), and representational integration (P3). The empirical claims are that an OR direction is recoverable in Llama 3.1 8B-Instruct (§3.1.3: held-out DiM 0.748, LR 0.839, at layer 12 / read position −5), is causally operative under steering and ablation (§3.2), and is geometrically distinct from feasibility, desirability, and exploitation while causally coupled to them (§3.3). The framing is explicitly bounded: third-person, discovery-style OR in synthetic scenarios; no claims about human cognition or model agency (§2.4, §6.1). The novelty claim — first causal intervention on an internal representation of an entrepreneurship construct in an LLM — is plausible given the literature surveyed, though it is a priority claim rather than a scientific result.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the 'opportunity recognition dial' in LLMs?
The authors identify a direction in Llama 3.1 8B-Instruct's internal activation space that separates scenarios describing an entrepreneurial opportunity from closely matched scenarios that do not. Because amplifying or suppressing this direction during inference (activation steering) shifts the model's opportunity judgments in the predicted direction, they call it a 'dial'. It is a representational structure, not a literal control knob exposed by the model vendor.
Is this really the first study to steer an entrepreneurship concept inside an LLM?
The authors claim it is the first causal intervention on an internal representation of an entrepreneurship construct inside an LLM. Representation engineering itself is well established for concepts like truth, sentiment, refusal, and personas; the novelty here is the entrepreneurship-construct application, the construct-validation logic borrowed from entrepreneurship theory, and the discriminant comparisons against feasibility, desirability, and exploitation directions.
Can this steering make AI better at finding business opportunities?
Not demonstrated. Turning the dial up raises the probability of 'yes' answers to recognition questions and yields more opportunity-focused explanations, but the paper does not test whether steered outputs surface better, more novel, or more successful opportunities. Steering also shifts feasibility, desirability, and exploitation judgments, so more opportunity sensitivity could simply mean more false positives.
Does the finding replicate in other models?
Partially. Direction recovery (held-out accuracy 0.705–0.744), signed steering, and geometric separation replicate in Llama 3.1 70B, Qwen3-8B, Qwen2.5-32B, and Gemma-3-12B. However, the stricter selectivity controls (random-direction control, strengthened null) were only fully rerun in two of them, with weaker separation in Gemma-3-12B-it and a reversal in Qwen3-8B.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extraction elided middle sections (procedural detail of §3.2.1–3.2.2, appendices D–G, part of Table 6 row 5); judgments rest on the retained tables, notes, and summaries.
- Figures 1–5 are placeholders ('Please insert Figure…') and could not be visually inspected; numeric claims were cross-checked against table notes instead.
- Appendix tables (D.2, E.1–E.2, F.1–F.5, G) are only summarised in the main text, so exact per-model steering numbers for the replication models were not directly verifiable.
- This is a preprint under review at a journal; the published version may differ.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.