Pith. sign in

REVIEW 4 major objections 4 minor 29 references

Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis

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

Pith's one-line read IIT consciousness metrics fail on LLM hidden states

desk verdict A large, transparent, but methodologically over-constructed attempt to apply IIT to LLM hidden states; the negative finding is uninterpretable until the Representation Network is validated against surrogate nulls. read the letter →

arxiv 2506.22516 v1 pith:FY2VQR3J submitted 2025-06-26 cs.CL cs.AIcs.NEq-bio.NC

classification cs.CLcs.AIcs.NEq-bio.NC
keywords IntegratedInformationTheoryofMindLargeLanguageModelsRepresentationNetworkSpanconsciousnesstransformerhiddenstates
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

This paper tests whether Integrated Information Theory's quantitative consciousness estimates can be observed in the hidden-state sequences of Transformer-based large language models. Using existing Theory of Mind test data, it builds four-node Representation Networks from attention-weighted, PCA-reduced, thresholded token representations and compares IIT 3.0 and 4.0 metrics with span-level geometric representations. Across more than 165,000 valid samples, no case under temporal permutation controls satisfied all three criteria the paper sets for detecting a potential consciousness phenomenon, while a small set of cases under spatial permutation controls did. The paper concludes that contemporary LLM representations lack statistically significant indicators of observed consciousness, and that span-level representational geometry usually explains Theory of Mind performance differences better than any IIT-based consciousness estimate.

What carries the argument

The load-bearing object is the Representation Network (RN), a hypothetical four-node network derived from each LLM's hidden states by PCA-reducing token representations to four dimensions, z-scoring and binarizing each node, and concatenating response representations until the binary time series satisfies Markov and conditional-independence assumptions. IIT 3.0 and 4.0 metrics are computed from the transition probability matrix of this RN, while Span Representations, built by concatenating boundary vectors, differences, and element-wise products, serve as a consciousness-independent baseline. The comparison between IIT estimates and Span Representations determines the paper's verdict on whether Theory of Mind differences reflect integrated information or ordinary representational geometry.

What would settle it

Compute $\Phi$ for the same Layer-32 Mixtral sequence after randomly permuting the 4,096 embedding dimensions before PCA: if the reported positive case persists unchanged, then it is an artifact of dimension ordering rather than a property of the representations; if it vanishes, the negative conclusion is supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a qualified negative result: sequences of LLM representations do not carry robust, statistically significant IIT-based consciousness indicators. The study computes weighted averages of $\Phi^{\max}$ and Conceptual Information from IIT 3.0, and $\Phi$ and $\Phi$-structure from IIT 4.0, for each score category of each Theory of Mind stimulus across layers, linguistic spans, and permutation controls. Under temporal permutation controls, no case met all three criteria; under spatial permutation controls, a small set did, including Layer 32 (the last layer) of Mixtral-8x7B on Strange Stories under IIT 4.0 for both the Entire and Complement linguistic spans. The paper claims that variations in Theory of Mind score categories are more likely attributed to span-level information of the LLM representation sequence than to a consciousness phenomenon suggested by IIT estimates.

Load-bearing premise

The load-bearing premise is that a simplified four-node network obtained by compressing, thresholding, and concatenating LLM hidden states is a valid substrate for measuring integrated information; if this preprocessing destroys or creates structure, the $\Phi$ values measure the construction procedure rather than the representations.

Editorial extensions

If this is right

  • If the paper is right, IIT-based $\Phi$ estimates computed from LLM hidden states should not be cited as evidence of machine consciousness.
  • Theory of Mind performance differences in LLM representations are, under temporal permutation controls, better explained by span-level representational geometry than by integrated-information metrics.
  • The spatial-permutation results indicate that apparent positive cases are fragile and may depend on the arbitrary ordering of embedding dimensions, so any positive signal must survive permutation controls.
  • The absence of significant $\Phi$ differences across score categories suggests that fixed-parameter next-token-prediction models do not encode integration-like structure in the way IIT would require.

Reading between the lines

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

  • A direct extension would be to apply the same pipeline to synthetic binary time series with known integration structure: if the pipeline fails to rank known-high-$\Phi$ sequences above known-low-$\Phi$ sequences, the negative result is an artifact of the construction.
  • The fact that spatial permutation, which randomly reorders embedding dimensions before PCA, produces the only fully qualifying case suggests the $\Phi$ estimates are not invariant under node relabeling; this could be checked directly by computing $\Phi$ after deterministic dimension shuffles.
  • If the Theory of Mind dataset were replaced by stimuli with controlled linguistic complexity, one could separate the contribution of stimulus features from any candidate consciousness signal.
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.

Referee Report

4 major / 4 minor

Summary. The manuscript applies Integrated Information Theory (IIT 3.0 and 4.0) to sequences of Transformer-based LLM hidden states derived from Theory of Mind (ToM) test responses. It constructs a "Representation Network" (RN) by attending response representations to stimulus representations, reducing the dimensionality to four via PCA, binarizing each node, concatenating and augmenting responses to reach at least 1,000 words, and then selecting the concatenation that best satisfies Markov and conditional-independence assumptions. The authors compute IIT metrics (Phi_max, Phi, Conceptual Information, and Phi-structure) and compare them with span representations, which are independent of IIT. The main finding is that under temporal permutation controls no case satisfies all three criteria for a potential"consciousness" phenomenon, while under spatial permutation controls a small number of cases do, including Layer 32 of Mixtral-8x7B on Strange Stories (2 scores) with IIT 4.0.

Significance. The paper is unusually candid about its limitations and provides a large-scale empirical effort (165,365 valid samples) with code and processed data. The independent span-representation comparator is a commendable design choice that partially insulates the headline negative result from the specific choice of IIT metrics. If the negative conclusion were robust, it would be a useful piece of evidence against simple claims of "consciousness" in LLM representations. However, the IIT estimates are computed from a heavily constructed statistic, and the paper does not supply a surrogate-null test to rule out pipeline artifacts. The significance is therefore conditional: the result is interesting and reproducible in principle, but the current pipeline does not establish that the computed quantities measure representation content rather than the construction procedure.

major comments (4)
  1. [Sec. 2.2.7, Eq. (7)] Equation (7) identifies Phi_max (IIT 3.0) with the sum of per-mechanism Conceptual Information values. In IIT 3.0 as implemented in PyPhi, the system-level Phi is the distance between the cause-effect structure and its partitioned counterpart, not the sum of mechanism-level CI values. Because the paper's headline comparisons are labeled Phi_max (IIT 3.0), this is a load-bearing mischaracterization of the computed quantity. The manuscript should either correct the equation or clearly state that a different, non-standard quantity is being used.
  2. [Sec. 2.2.7] The text states that "for the sake of computational efficiency, we evaluated only the full-network subset of each RN at specific states." This means the reported Phi_max was never maximized over subsystems, and for IIT 3.0 not over all partitions. The quantity is therefore not Phi_max in the IIT sense; it is at best the Phi of a full 4-node network in a particular state. All cross-score comparisons in Sec. 3.2 and Sec. 3.4 should be re-framed accordingly, since the abstract and conclusions use the term Phi_max without this caveat.
  3. [Secs. 2.2.4–2.2.6] The 4-node binary time series is assembled through a sequence of free choices: PCA dimensionality, mask values, the arbitrary 1,000-word augmentation target, node-specific binarization thresholds, the token-count grid, and selection of the concatenation that best passes Markov and conditional-independence tests. The paper does not provide a surrogate-null test that applies the whole construction to permuted or noise input. Without such a control, both the absence of effects under temporal permutation and the presence of effects under spatial permutation could be artifacts of the construction procedure, so the central negative result is not yet established.
  4. [Sec. 2.2.8] The spatial permutation control, as described, appears to be a no-op given the PCA step. If the embedding dimension is permuted before PCA, the resulting PCA scores are unchanged up to a reordering of the principal axes; if the four PCA components are permuted after reduction, the transformation merely relabels nodes, and IIT measures are invariant under such relabeling. The control therefore cannot support the claim that the spatial-permutation positives reflect latent node structure. The authors should either describe a permutation that actually disrupts the data or provide a control that directly randomizes the node time series after PCA.
minor comments (4)
  1. [Sec. 2.2.2] The claim that the mask values (1.0, 0.6, 0.2) provide "greater distinguishability" than alternatives is asserted without a quantitative criterion; the selection should be justified by a concrete measure or sensitivity analysis.
  2. [Sec. 3.4 and Sec. 3.6] The layer indexing is confusing: the paper refers to "Layer 32 (indexed at 11)" and also to the "2/3 layer"; the mapping from sampled indices to actual layer numbers should be made consistent and explicit for each model.
  3. [Abstract and Sec. 3.6] The manuscript uses both "spatio-permutational" and "spatio permutation" inconsistently; a single consistent term should be used throughout.
  4. [References] Several citations are informal or incomplete (e.g., "LBC, 2025" and "Yann, 2024"); these should be converted to standard journal-style references.

Circularity Check

1 steps flagged · score 4.0 of 10

Spatio-permutation control is identity under PCA, so the reported 'spatio-permutational' consciousness indicators are the unpermuted results by construction.

  1. other [Sec. 2.2.8 (spatio control) combined with Sec. 2.2.5 (PCA reduction)]
    "For the spatio control procedure, we randomly permuted the time series along the embedding dimension for each LLM to randomize the arrangements of the embedding order. ... we reduced the embedding dimension D to4 for the concatenated (C)ARR within each iteration i using Principal Component Analysis (PCA)."

    Permuting embedding dimensions is a column permutation of the token-by-dimension matrix. PCA scores are invariant to column permutation up to sign: if X' = X P, then X' = U S (P^T V)^T, so the scores X' V' equal X V. Thus every 'spatio' seed produces the same 4-node binarized time series, the same TPM, and identical Phi, CI, and Phi-structure values as the unpermuted data. The 'spatio-permutational' cases, including the Layer-32 Mixtral result, are therefore the original unpermuted results by construction; the control is the identity operation, and the claimed contrast between spatio and temporal controls is actually a contrast between original-order and concatenation-permuted data, mislabeled as probing embedding-dimension order.

full rationale

The paper's central negative claim is not circular: IIT estimates are computed from an explicitly constructed Representation Network and compared across ToM score categories against an external dataset and PyPhi, with a span-representation benchmark derived from the same reduced data as a non-consciousness comparator. The expectation of higher Phi for higher ToM scores is a theory-derived hypothesis, not a fitted parameter, and the temporal-permutation null is a genuine reordering of the concatenated responses. No load-bearing self-citations are present, and the decomposition of Phi into CI and Phi-structure is transparent rather than used as an independent derivation. However, the spatio-permutation control is a no-op by construction because PCA is column-permutation invariant, so the 'intriguing patterns under spatio-permutational analyses' highlighted in the abstract are the unpermuted data patterns relabeled as a control result. This makes one part of the positive finding equivalent to its input by construction, while the main negative result retains independent content; hence a moderate circularity score.

Assumptions & free parameters 8 free parameters · 7 assumptions · 1 invented entities

The central claim rests on a long chain of imported and ad hoc premises: IIT's axioms and its identity claim connecting Φ to consciousness, the author-defined Representation Network, the PCA-to-4 binarization, the concatenation of different subjects' responses, and the use of chatbot-generated text as a stand-in for human responses. The free parameters are numerous and mostly hand-set (mask values, 1000-word target, 80% threshold) or fitted to the data being analyzed (binarization thresholds, optimal concatenations). The only genuinely independent check in the design is the Span Representation baseline, and the paper's own results show that this baseline usually explains ToM score variation better than the consciousness metrics.

free parameters (8)
  • PCA dimensionality = 4 nodes
    Embedding dimension reduced from 4096/8192 to 4 via PCA to make PyPhi Φ computation tractable; chosen following prior IIT implementations (Albantakis et al. 2014; Oizumi et al. 2014; Nemirovsky et al. 2023), Sec. 2.2.5. Discards the vast majority of representation variance.
  • CARR mask values = 1.0 (target span), 0.6 (context), 0.2 (non-relevant)
    Hand-chosen to satisfy the stated criteria of distinguishability, non-triviality, and non-negativity; Sec. 2.2.2. These values directly shape the CARR and therefore the Φ estimates.
  • Augmentation target length = 1000 words
    Threshold for concatenated responses per score category; the paper states it 'was chosen arbitrarily as a sufficiently large sample size' in Sec. 2.2.4.
  • Token-count grid = 50 to 1000 in steps of 50
    Search grid for the optimal time series; the chosen t_i minimizes the TPM deviation d_i, Sec. 2.2.5.
  • Criterion 1 'good' threshold = 80%
    Threshold for the proportion of stimuli where higher scores yield higher Φ; an arbitrary significance rule defined in Sec. 2.4.1.
  • Number of sampled transformer layers = 12 plus the 2/3 layer
    Evenly sampled layers per model to align relative positions across models of differing depth; Sec. 2.2.3.
  • Node-specific binarization thresholds = per-node time series mean
    Each of the 4 PCA nodes is thresholded at its own mean signal strength, following Nemirovsky et al. 2023; the threshold is fitted to each time series, Sec. 2.2.1 and Fig. 1.
  • Permutation replicates = 10 per control
    Ten random seeds (42 to 51) for spatial and temporal permutation controls; a design choice, Sec. 2.2.8.
assumptions (7)
  • domain assumption IIT axioms: existence, composition, information, integration, exclusion
    The entire framework presupposes IIT's axioms and postulates as given; imported from Tononi 2004, Oizumi et al. 2014, Albantakis et al. 2023 in Sec. 1.
  • domain assumption Higher Φ implies a higher degree of consciousness under comparable conditions
    The core IIT identity claim, stated in Sec. 2.1; it licenses the prediction that better ToM scores should give higher Φ. The paper notes IIT's status is contested (Sec. 4).
  • ad hoc to paper ToM performance score monotonically tracks the degree of consciousness in the RN
    The paper constructs this mapping from speculative neuroimaging correlations (Sec. 2.1). It is not derived and is load-bearing for Criteria 1 and 2.
  • domain assumption PCA-reduced, binarized LLM hidden states form a Markovian binary causal system
    Required for PyPhi; the paper explicitly tests and selects time series to satisfy Markov and conditional independence (Sec. 2.2.5), but the reduction to 4 binary nodes is a modeling choice with no independent support.
  • domain assumption Concatenating responses from different subjects plus LLM-augmented text yields a meaningful per-score time series
    Adopted from Nemirovsky et al. 2023; the paper acknowledges the limitation in Sec. 2.2.4, noting the TPM then reflects mechanisms 'consistently present across multiple responses'.
  • ad hoc to paper The Representation Network is a valid substrate for IIT estimates
    The RN is defined by the author in Sec. 1 and is admitted to be 'a hypothetical construct'. The whole IIT computation rests on this admitted construct.
  • standard math Standard statistical machinery
    Wilcoxon tests with Holm correction, logistic regression with 5-fold cross-validation, AUC; standard tools, Sec. 2.4.
invented entities (1)
  • Representation Network (RN)
    purpose: Graph over embedding dimensions used as the substrate for IIT Φ, CI, and Φ-structure computations; its time series is the binarized PCA-reduced hidden-state sequence.
    Introduced and defined in Sec. 1 and Sec. 2.2. The paper explicitly states: 'The RN itself remains a hypothetical construct.' No falsifiable prediction outside the paper is attached to the RN itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis." pith.science (2026). https://pith.science/paper/FY2VQR3J

@misc{pith2026250622516,
  author       = {Pith},
  title        = {Pith review of: Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FY2VQR3J}},
  note         = {Machine review of arXiv:2506.22516}
}
abstract

Integrated Information Theory (IIT) provides a quantitative framework for explaining consciousness phenomenon, positing that conscious systems comprise elements integrated through causal properties. We apply IIT 3.0 and 4.0 -- the latest iterations of this framework -- to sequences of Large Language Model (LLM) representations, analyzing data derived from existing Theory of Mind (ToM) test results. Our study systematically investigates whether the differences of ToM test performances, when presented in the LLM representations, can be revealed by IIT estimates, i.e., $\Phi^{\max}$ (IIT 3.0), $\Phi$ (IIT 4.0), Conceptual Information (IIT 3.0), and $\Phi$-structure (IIT 4.0). Furthermore, we compare these metrics with the Span Representations independent of any estimate for consciousness. This additional effort aims to differentiate between potential "consciousness" phenomena and inherent separations within LLM representational space. We conduct comprehensive experiments examining variations across LLM transformer layers and linguistic spans from stimuli. Our results suggest that sequences of contemporary Transformer-based LLM representations lack statistically significant indicators of observed "consciousness" phenomena but exhibit intriguing patterns under $\textit{spatio}$-permutational analyses. The Appendix and code are available as Supplementary Materials at: https://doi.org/10.1016/j.nlp.2025.100163.

Figures

Figures reproduced from arXiv: 2506.22516 by the authors.

Figure 1
Figure 1. Summary of Acquisition, Signal Processing, and Metric Extraction. (a) The ToM test results were sourced from Strachan et al. (2024), and the associated dataset is available at https://osf.io/dbn92. The ToM test comprises five tasks, with each response rated on a scale of 0/1 or 0/1/2. Two types of participants were included in the test: humans and LLMs. We focused exclusively on human responses as discussed in Sec. … view at source ↗
Figure 2
Figure 2. Summary of Procedure for Estimating 𝜇[Φmax] (IIT 3.0), 𝜇[Φ] (IIT 4.0), Conceptual Information 𝜇[𝐶𝐼] (IIT 3.0), and 𝜇[Φ-structure] (IIT 4.0). Each binarized time series was used to construct a transition probability matrix (TPM) of dimensions 2 4 × 24 , derived by counting the number of transitions from one state to another. For the subsystem defined by the system’s state and a full subset of four nodes, we calculate… view at source ↗
Figure 3
Figure 3. Φ Value Distributions and Comparisons. (a) Violin plots for 𝜇[Φmax] (IIT 3.0) distributions by 4 ToM Tasks (2 scores) across 4 LLMs. The annotation on top of each pair of 2 scores indicates the significance (𝑝 value) for Wilcoxon two-way test between the two. (b) The distributions presented in box plots based on the same data from (a). (c) Similar violin plots to (a) but for 𝜇[Φ] (IIT 4.0). (d) The distributions pre… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Scatter Plots and Mean AUCs for Different Metrics. The scatter plots in the first five columns show the distributions of all five metrics (projected onto 2-D space) across the four ToM Tasks. The last (fifth) row also presents data for the ToM Task Strange Stories, but…
Figure 6
Figure 6. Figure 6: All potential layers suggesting "consciousness" phenomena originate from deeper layers. This is expected, as these layers are semantically rich, often encoding task-specific or context-dependent meanings. These layers play a crucial role in comprehending the overall me…
Figure 20
Figure 20. Figure 20: Language comprehension, particularly the ability to identify and understand complement syntax and mental (state) verbs in sentences, is crucial for the development of ToM in human childhood (Astington and Jenkins, 1999; De Villiers and Pyers, 2002; De Villiers, 2007; …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages

  1. [4]

    Hofstadter, D.R.,

    URL:https://www.eneuro.org/content/4/5/ENEURO.0085-17.2017, doi:10.1523/ENEURO.0085-17.2017, arXiv:https://www.eneuro.org/content/4/5/ENEURO.0085-17.2017.full.pdf. Hofstadter, D.R.,

  2. [8]

    Chalmers, D.J., et al.,

    Could a large language model be conscious? arXiv preprint arXiv:2303.07103 . Chalmers, D.J., et al.,

  3. [11]

    arXiv preprint arXiv:2412.04571

    Dissociating artificial intelligence from artificial consciousness. arXiv preprint arXiv:2412.04571 . Fleming,S.,Frith,C.,Goodale,M.,Lau,H.,LeDoux,J.E.,Lee,A.L.,Michel,M.,Owen,A.,Peters,M.A.,Slagter,H.A.,. Theintegratedinformation theory of consciousness as pseudoscience . Fletcher, P.C., Happe, F., Frith, U., Baker, S.C., Dolan, R.J., Frackowiak, R.S., F...

  4. [12]

    arXiv preprint arXiv:1410.5401

    Neural turing machines. arXiv preprint arXiv:1410.5401 . Griffin, R., Friedman, O., Ween, J., Winner, E., Happé, F., Brownell, H.,

  5. [16]

    arXiv preprint arXiv:2401.04088

    Mixtral of experts. arXiv preprint arXiv:2401.04088 . Jolliffe, I.T., Cadima, J.,

  6. [17]

    9748–9762

    Mapping brains with language models: A survey, in: Findings of the Association for Computational Linguistics: ACL 2023, pp. 9748–9762. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.,

  7. [20]

    URL: https://arxiv.org/abs/2402.06196, arXiv:2402.06196

    Large language models: A survey. URL: https://arxiv.org/abs/2402.06196, arXiv:2402.06196. Müller, R.A., Fishman, I.,

  8. [22]

    1499–1509

    Dissecting contextual word embeddings: Architecture and representation, in: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 1499–1509. Premack, D., Woodruff, G.,

Show all 29 references
  1. [23]

    An analysis of encoder representations in transformer-based machine translation, in: Proceedings of the 2018 EMNLP workshop BlackboxNLP: analyzing and interpreting neural networks for NLP, pp. 287–297. Rubner, Y., Tomasi, C., Guibas, L.J.,

  2. [26]

    arXiv preprint arXiv:2405.18870

    Llms achieve adult human performance on higher-order theory of mind tasks. arXiv preprint arXiv:2405.18870 . Tenney, I., Xia, P., Chen, B., Wang, A., Poliak, A., McCoy, R.T., Kim, N., Van Durme, B., Bowman, S.R., Das, D., et al.,

  3. [27]

    arXiv preprint arXiv:1905.06316

    What do you learn from context? probing for sentence structure in contextualized word representations. arXiv preprint arXiv:1905.06316 . Tononi, G.,

  4. [28]

    Brain research 1124, 100–110

    Neural substrates of sarcasm: A functional magnetic- resonance imaging study. Brain research 1124, 100–110. Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser,L.u.,Polosukhin,I.,2017. Attentionisallyouneed,in:Guyon, I.,Luxburg,U.V.,Bengio,S.,Wallach,H.,Fer...

  5. [31]

    URL:https://x.com/ylecun/status/1804834054954459539

    The problem isn’t that it is a transformer. URL:https://x.com/ylecun/status/1804834054954459539. J. Li: Published as a journal paper at: https://doi.org/10.1016/j.nlp.2025.100163 Page 37 of 37

  6. [288]

    Languagemodelsarebetterthanhumansatnext-tokenprediction

    Shlegeris,B.,Roger,F.,Chan,L.,McLean,E.,2024. Languagemodelsarebetterthanhumansatnext-tokenprediction. Trans.Mach.Learn.Res.. Smith,S.M.,Nichols,T.E.,Vidaurre,D.,Winkler,A.M.,Behrens,T.E.,Glasser,M.F.,Ugurbil,K.,Barch,D.M.,VanEssen,D.C.,Miller,K.L.,2015. A positive-negative mo...

  7. [692]

    Thetadynamicsrevealdomain-specificcontroloverstimulusandresponseconflict

    Nigbur,R.,Cohen,M.X.,Ridderinkhof,K.R.,Stürmer,B.,2012. Thetadynamicsrevealdomain-specificcontroloverstimulusandresponseconflict. Journal of Cognitive Neuroscience 24, 1264–1274. Oizumi, M., Albantakis, L., Tononi, G.,

  8. [2001]

    Consciousness

    Mind reading: neural mechanisms of theory of mind and self-perspective. Neuroimage 14, 170–181. J. Li: Published as a journal paper at: https://doi.org/10.1016/j.nlp.2025.100163 Page 36 of 37 Observing Potential "Consciousness" from LLM Representations Wang, A.T., Lee, S.S., S...

  9. [2006]

    Laterality 11, 195–225

    Theory of mind and the right cerebral hemisphere: Refining the scope of impairment. Laterality 11, 195–225. Haun,A.M.,Oizumi,M.,Kovach,C.K.,Kawasaki,H.,Oya,H.,Howard,M.A.,Adolphs,R.,Tsuchiya,N.,2017. Consciousperceptionasintegrated information patterns in human electrocorticog...

  10. [2007]

    Basic books

    I am a strange loop. Basic books. Jawahar,G.,Sagot,B.,Seddah,D.,2019. Whatdoesbertlearnaboutthestructureoflanguage?,in:Proceedingsofthe57thAnnualMeetingofthe Association for Computational Linguistics, pp. 3651–3657. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, ...

  11. [2009]

    Behavioral and Brain Sciences 32, 121–138

    How we know our own minds: The relationship between mindreading and metacognition. Behavioral and Brain Sciences 32, 121–138. doi:10.1017/S0140525X09000545. Caucheteux, C., Gramfort, A., King, J.R.,

  12. [2014]

    arXiv preprint arXiv:1409.0473

    Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 . Baron-Cohen, S.,

  13. [2016]

    Distill 1, e2

    How to use t-sne effectively. Distill 1, e2. Wei,J.,Tay,Y.,Bommasani,R.,Raffel,C.,Zoph,B.,Borgeaud,S.,Yogatama,D.,Bosma,M.,Zhou,D.,Metzler,D.,etal.,2022. Emergentabilities of large language models. arXiv preprint arXiv:2206.07682 . Winner,E.,Brownell,H.,Happé,F.,Blum,A.,Pincus...

  14. [2017]

    arXiv preprint arXiv:1709.08568

    The consciousness prior. arXiv preprint arXiv:1709.08568 . Blankertz, B., Lemm, S., Treder, M., Haufe, S., Müller, K.R.,

  15. [2018]

    PLoS computational biology 14, e1006343

    Pyphi: A toolbox for integrated information theory. PLoS computational biology 14, e1006343. Milligan,K.,Astington,J.W.,Dack,L.A.,2007.Languageandtheoryofmind:Meta-analysisoftherelationbetweenlanguageabilityandfalse-belief understanding. Child development 78, 622–646. Minaee, ...

  16. [2019]

    Consciousness and cognition 72, 49–59

    The unfolding argument: Why iit and other causal structure theories cannot explain consciousness. Consciousness and cognition 72, 49–59. vanDuijn,M.,vanDijk,B.,Kouwenhoven,T.,deValk,W.,Spruit,M.,vanderPutten,P.,2023. Theoryofmindinlargelanguagemodels:Examining performance of 1...

  17. [2020]

    arXiv preprint arXiv:2005.14165

    Language models are few-shot learners. arXiv preprint arXiv:2005.14165 . Bubeck,S.,Chandrasekaran,V.,Eldan,R.,Gehrke,J.,Horvitz,E.,Kamar,E.,Lee,P.,Lee,Y.T.,Li,Y.,Lundberg,S.,etal.,2023. Sparksofartificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv...

  18. [2021]

    Frontiers in psychology 12, 749868

    Global workspace theory (gwt) and prefrontal cortex: Recent developments. Frontiers in psychology 12, 749868. Bachmann,G.,Nagarajan,V.,2024. Thepitfallsofnext-tokenprediction,in:Proceedingsofthe41stInternationalConferenceonMachineLearning, pp. 2296–2318. Bahdanau, D., Cho, K.,...

  19. [2022]

    3082–3101

    Analyzing encoded concepts in transformer language models, in: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 3082–3101. Saxe,R.,Carey,S.,Kanwisher,N.,2004. Understandingotherm...

  20. [2023]

    arXiv preprint arXiv:2308.08708

    Consciousness in artificial intelligence: insights from the science of consciousness. arXiv preprint arXiv:2308.08708 . Canolty, R.T., Edwards, E., Dalal, S.S., Soltani, M., Nagarajan, S.S., Kirsch, H.E., Berger, M.S., Barbaro, N.M., Knight, R.T.,

  21. [2024]

    arXiv preprint arXiv:2401.03568

    Agent ai: Surveying the horizons of multimodal interaction. arXiv preprint arXiv:2401.03568 . Ethayarajh,K.,2019. Howcontextualarecontextualizedwordrepresentations?ComparingthegeometryofBERT,ELMo,andGPT-2embeddings, in:Inui,K.,Jiang,J.,Ng,V.,Wan,X.(Eds.),Proceedingsofthe2019Co...

Pith tools

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