Time, Identity and Consciousness in Language Model Agents
Open accessAbstract
Machine consciousness evaluations mostly see behavior. For language model agents that behavior is language and tool use. That lets an agent say the right things about itself even when the grounded constraints that should make those statements action-guiding are not jointly active at the moment of action selection. To address this problem, we apply Stack Theory's distinction between apparent and genuine agent identity. This separates window-level co-occurrence, understood as correlational evidence, from co-instantiation at a single objective step, which constitutes a minimal synthesis condition necessary (though not sufficient) for identity constraints to jointly govern behaviour. We then instantiate Stack Theory's WeakSync and StrongSync postulates on grounded identity statements. This yields two persistence scores that can be computed from instrumented scaffold traces. We connect these scores to five operational identity metrics and map common scaffolds into an identity morphospace that exposes predictable tradeoffs. The result is a conservative toolkit for identity evaluation. It separates talking like a stable self from being organized like one.
AI snapshot
Written from the paper’s full text.
TL;DR
The paper applies Stack Theory’s temporal gap to language model agent identity, separating ingredient-wise occurrence within an evaluation window from co-instantiation at a single objective step. It proposes weak and strong persistence scores plus five operational identity metrics to distinguish talking like a stable self from being organized like one.
Key findings
- Co-instantiation implies occurrence, but occurrence does not imply co-instantiation; Theorem 3.10 shows the within-window diamond lift does not distribute over conjunction. In the paper’s example with horizon Δ = 2, name, role, and constraint each occur at separate objective steps, so Occur holds while CoInst fails.
- Ingredient-wise persistence can be high while conjunctive action constraints fail: Theorem 4.5 constructs an identity conjunction \(g^0 = g^0_1 \land g^0_2\) where \(g^0_1\) is active on even steps and \(g^0_2\) on odd steps, giving \(P_{\text{weak}} = 1\) but no co-instantiation.
- Strong persistence is bounded by weak persistence: Proposition 4.4 states \(P_{\text{strong}}(\tau, g^0) \leq P_{\text{weak}}(\tau, g^0)\).
- The paper derives five operational identity metrics—Identifiability, Continuity, Consistency, Persistence, and Recovery—and defines a temporal gap ratio \(Gap(g^0,\tau) = \text{median}_{t \in T} \frac{w_{\text{strong}}(t)+1}{w_{\text{weak}}(t)+1}\).
Methods
- Theoretical/formal design: models an LMA scaffold as an architecture tuple \(A = (\Sigma, K, V, Q, D, R, n_\pi, |C|_{\max})\) and scaffold state \(s = (C, M, \pi, D_{\text{retrieved}})\), with transitions and ingredient activation conditions.
- Applies Stack Theory’s windowing map \(W_{\Delta,s}\), occurrence \(Occur_W\), co-instantiation \(CoInst_W\), temporal lifts, and the Arpeggio/Chord postulates; defines weak and strong persistence scores over a finite set \(T\) of layer time indices.
- Proposes a minimal evaluation protocol: ground an identity statement to a Layer 0 conjunction, instrument the scaffold to log active grounded ingredients, choose a windowing map and evaluation set, compute \(Occur_W\), \(CoInst_W\), \(P_{\text{weak}}\), \(P_{\text{strong}}\), and optionally \(Gap\), then pair instrumentation with behavioral probes. No empirical data are reported.
Limitations
- Stated: \(PhenReal(l_m,\tau,t)\) is not directly observable, and the synchronization postulates are necessary conditions; future work should empirically measure weak and strong persistence across architectures and test their relationship to safety and proposed consciousness markers.
- Stated/evident: the metrics require instrumented scaffold traces and repeated behavioral probes, and the appendices are external via a GitHub repository rather than included in the paper.
- Evident: the paper reports no empirical experiments, datasets, or sample sizes; the formal results depend on chosen grounding, windowing map, tolerance thresholds, and similarity metrics, and the proposed metrics are not empirically validated in the text.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
2 of 6 cited sources checked · 1 supported · 0 flagged · accuracy not yet measured
- Premisecitation
Many consciousness proposals require some form of integration that binds the contents of a moment into a single subject.
“Many consciousness proposals require some form of integration that binds the contents of a moment into a single subject, even if they disagree about what that integration is (Bennett 2025;Baars 1988;Dehaene and Naccache 2001;Tononi 2004;Metzinger 2003).”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Baars 1988
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Dehaene and Naccache 2001
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Tononi 2004 source
- Premisecitation
Some proposed indicators for AI consciousness lean on behavior that looks like a stable self model, including self-report, memory, and narrative continuity.
“Some proposed indicators for AI consciousness therefore lean on behavior that looks like a stable self model.This includes self-report, memory, and narrative continuity (Bennett 2023a(Bennett , 2025(Bennett , 2026b(Bennett , 2023b)).”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Bennett 2023a
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Bennett 2025 source
In “3. Self Awareness” (page unavailable), the paper links self-awareness to contents available for report; “THE SELF” ties a first-order self to past actions in memory, and “OPTION 1: CONSCIOUSNESS IS AT A POINT IN TIME” discusses an internal narrative. It does not present self-report, memory, and narrative continuity as a set of behavioral indicators for AI consciousness. In “THE NORMIES,” it cautions that a machine could behave intelligently and have proposed orders of self yet still seem to lack consciousness. (LLM, conf=0.84)
Experimental — accuracy not yet measured - cites Bennett 2026b
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingreasoning
There exist LMAs and identity statements such that P_weak is high while the agent systematically fails tasks that require the conjunction of identity constraints to be applied simultaneously in action selection.
“There exist LMAs and identity statements l m such that P weak (τ, g 0 ) is high while the agent systematically fails tasks that require the conjunction of identity constraints to be applied simultaneously in action selection.”
Quote in the paperSubstantial claim - Findingreasoning
Strong persistence is bounded by weak persistence: P_strong ≤ P_weak.
“For any τ and g 0 , P strong (τ, g 0 ) ≤ P weak (τ, g 0 ).(15)”
Quote in the paperSubstantial claim - Interpretationreasoning
An agent that passes recall tests can still fail to act according to its identity because recall does not imply co-instantiation.
“An agent that passes recall tests can still fail to act according to its identity because recall does not imply co-instantiation.”
Quote in the paperSubstantial claim
- Interpretationreasoning
An LMA can satisfy ingredient-wise identity checks for multiple identity ingredients across a window while still failing to ever instantiate the full identity conjunction at a single objective step.
“An LMA can satisfy ingredient-wise identity checks for multiple identity ingredients across a window while still failing to ever instantiate the full identity conjunction at a single objective step.”
Quote in the paperSubstantial claim - Interpretationreasoning
Benchmarks that test whether agents can recall identity facts may give false confidence.
“Benchmarks that test whether agents can recall identity facts may give false confidence.”
Quote in the paperSubstantial claim - Interpretationreasoning
If an identity ingredient is not effectively present in the token sequence, then it cannot constrain the next action.
“If an identity ingredient is not effectively present there, then it cannot constrain the next action.”
Quote in the paperSubstantial claim - Interpretationreasoning
Many common scaffolds can achieve occurrence without reliably achieving co-instantiation.
“Many common scaffolds can achieve occurrence without reliably achieving co-instantiation.”
Quote in the paperSubstantial claim - Interpretationreasoning
The pattern of separate identity questions being answered while the decision state never contains the full identity conjunction is behind many identity false positives in LMAs.
“This is exactly the pattern behind many identity false positives in LMAs.The agent can answer separate questions about name, role, and constraints.It may even do so consistently.Yet its decision state never contains the full identity conjunction that would bind action to that identity.”
Quote in the paperSubstantial claim - Interpretationreasoning
The failure of the within-window diamond operator to distribute over conjunction creates a practical evaluation pitfall for LMAs: identity components can each occur somewhere in a recent trajectory without ever co-instantiating at a single decision point.
“We have shown that a standard modal logic result-the failure of a within-window diamond operator to distribute over conjunction-creates a practical evaluation pitfall for language model agents: identity components can each occur somewhere in a recent trajectory (weak persistence) without ever co-instantiating at a single decision point (strong persistence).”
Quote in the paperSubstantial claim - Interpretationnone given
If one adopts Arpeggio, then high P_weak is necessary.
“If one adopts Arpeggio, then high P weak is necessary.”
Quote in the paperSubstantial claim - Interpretationreasoning
A system can look stable under self-report while failing to ever co-instantiate the grounded identity conjunction that would make that stability operative.
“A system can look stable under self-report while failing to ever co-instantiate the grounded identity conjunction that would make that stability operative.”
Quote in the paperNot a substantial claim - Interpretationreasoning
A system can talk like it has a stable self while the underlying identity constraints that should govern its actions are never jointly active at decision time.
“A system can talk like it has a stable self while the underlying identity constraints that should govern its actions are never jointly active at decision time.”
Quote in the paperSubstantial claim - Interpretationnone given
If a system never coinstantiates the grounded identity conjunction that defines its self model, then behavior alone can look more unified than the underlying mechanism.
“If a system never coinstantiates the grounded identity conjunction that defines its self model, then behavior alone can look more unified than the underlying mechanism.”
Quote in the paperNot a substantial claim - Interpretationreasoning
The temporal gap complicates consciousness assessments, as stable self-reports may mask fragmented operative states.
“This temporal gap also complicates consciousness assessments, as stable self-reports may mask fragmented operative states.”
Quote in the paperSubstantial claim - Interpretationreasoning
An agent can pass recall-based identity tests and still act out of character when the decision actually matters.
“That is why an agent can pass recall-based identity tests and still act out of character when the decision actually matters.”
Quote in the paperSubstantial claim - Interpretationreasoning
Grounding failures are a direct cause of low recovery.
“Grounding failures are therefore a direct cause of low recovery.”
Quote in the paperSubstantial claim - Interpretationreasoning
Recovery is closely connected to grounding soundness.
“Recovery is closely connected to grounding soundness.”
Quote in the paperSubstantial claim - Interpretationreasoning
Recall-based identity checks can overestimate identity stability.
“Using Stack Theory's temporal gap, we separated ingredient-wise occurrence from co-instantiation and showed why recall-based identity checks can overestimate identity stability.”
Quote in the paperSubstantial claim- Rests on an assumption: Most behavioural tests probe only weak recall.
- Interpretationreasoning
A scaffold can make identity ingredients retrievable without making them jointly active at action time.
“A scaffold can make identity ingredients retrievable without making them jointly active at action time.”
Quote in the paperSubstantial claim
- Classified as probably not a substantial claim. 13
- May combine two claims. 8
Classified as probably not a substantial claim. May combine two claims.
FindingreasoningCo-instantiation implies occurrence, but occurrence does not imply co-instantiation.
“Co-instantiation implies occurrence, but occurrence does not imply co-instantiation.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningFor predicates p and q over scaffold states, ♢∆(p ∧ q) implies ♢∆p ∧ ♢∆q, but the converse implication fails in general.
“For predicates p and q over scaffold states, ♢ ∆ (p ∧ q) ⇒ ♢ ∆ p ∧ ♢ ∆ q(5) but the converse implication fails in general.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based).
Interpretationnone givenAn LMA's operative identity at any moment is whatever is in the token sequence the LLM actually processes during inference.
“The key insight is that an LMA's operative identity at any moment is whatever is in the token sequence the LLM actually processes during inference.”
Quote in the paperNot a substantial claimMay combine two claims.
InterpretationreasoningPrompting can increase the likelihood of recalling identity ingredients but cannot ensure their joint activation under bounded context; architectural support is typically needed.
“Prompting can increase the likelihood of recalling identity ingredients but cannot ensure their joint activation under bounded context; architectural support is typically needed.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based).
Interpretationnone givenIf one adopts Chord as a necessary condition for phenomenality, then high P_strong is a necessary condition for an identity statement to be phenomenally real across the evaluated times.
“If one adopts Chord as a necessary condition for phenomenality, then high P strong is a necessary condition for an identity statement to be phenomenally real across the evaluated times.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S2 (checkable).
InterpretationreasoningThis imprecision hides how an identity component can occur somewhere in the recent interaction history without constraining the current decision.
“This imprecision hides how an identity component can occur somewhere in the recent interaction history without constraining the current decision.”
Quote in the paperNot a substantial claim- Rests on an assumption: Existing discussions of LMA identity are informal and use terms like statelessness, persistence, and identity drift without precise definitions.
May combine two claims.
InterpretationreasoningRetrieval and memory systems can improve ingredient availability, but can also fragment identity by surfacing competing fragments.
“Retrieval and memory systems can improve ingredient availability, but can also fragment identity by surfacing competing fragments.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based).
InterpretationreasoningAny persistent identity must be reconstructed from external traces.
“Any persistent identity must be reconstructed from external traces.”
Quote in the paperNot a substantial claim- Rests on an assumption: The core LLM is stateless at inference and only sees the current input.
Classified as probably not a substantial claim. May combine two claims.
InterpretationreasoningWeak persistence is a recall property; strong persistence is an operative property.
“Weak persistence is a recall property.Each ingredient must show up somewhere in the window.Strong persistence is an operative property.The full conjunction must show up together at some objective step inside the window.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based).
Interpretationnone givenArpeggio permits phenomenally real moments whose identity ingredients are smeared across the window rather than co-instantiated at a single instant.
“Intuitively, Arpeggio permits phenomenally real moments whose identity ingredients are smeared across the window rather than co-instantiated at a single instant.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationnone givenThe horizon ∆ controls how forgiving the evaluation is; a larger ∆ allows identity ingredients to be spread across more objective steps, a smaller ∆ demands tighter temporal coherence.
“The horizon ∆ controls how forgiving the evaluation is.A larger ∆ allows identity ingredients to be spread across more objective steps.A smaller ∆ demands tighter temporal coherence.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Interpretationnone givenIdentity becomes measurable once grounded ingredients are instrumented.
“The key point is that identity becomes measurable once grounded ingredients are instrumented.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.
Interpretationnone givenChord and Arpeggio are different regimes; Arpeggio is not a weaker version of Chord.
“Chord and Arpeggio are different regimes.Arpeggio is not a weaker version of Chord.It is a different claim about what phenomenality permits.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
InterpretationreasoningIf the temporal gap ratio is large, then achieving coinstantiation requires much larger windows than achieving ingredient coverage.
“If the gap ratio is large, then achieving coinstantiation requires much larger windows than achieving ingredient coverage.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).
Interpretationnone givenThe gap between weak and strong persistence scores is the temporal gap in operational form.
“Either way, the gap between the two scores is the temporal gap in operational form.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S4 (based).
InterpretationreasoningFor recovery to succeed, corrections must propagate downward to restore Layer 1 commitments and Layer 0 implementation features.
“For recovery to succeed, those corrections must propagate downward to restore Layer 1 commitments and Layer 0 implementation features.”
Quote in the paperNot a substantial claimMay combine two claims.
InterpretationreasoningLMAs can talk like they have stable identities, but that does not mean their identity constraints are co-instantiated when actions are chosen.
“LMAs can talk like they have stable identities.That does not mean their identity constraints are co-instantiated when actions are chosen.”
Quote in the paperSubstantial claim
The paper relies on this without support.
For LMAs, most behavior is language, tool use, and traces they leave in external memory.
“For language model agents (LMAs) most of that behavior is language, tool use, and the traces they leave in external memory.”
The paper relies on this without support.
The core LLM is stateless at inference and only sees the current input.
“Yet the core LLM is stateless at inference.It only sees the current input.”
The paper relies on this without support.
Existing discussions of LMA identity are informal and use terms like statelessness, persistence, and identity drift without precise definitions.
“Existing discussions of LMA identity are informal.Terms like statelessness, persistence, and identity drift are used without precise definitions.”
The paper relies on this without support.
Most behavioural tests probe only weak recall.
“Safety-relevant constraints require strong persistence at action time, yet most behavioural tests probe only weak recall.”
The paper relies on this without support.
Many interventions are linguistic and modify Layer 2 narrative identity.
“Many interventions are linguistic.They modify Layer 2 narrative identity.”
Cites
Intelligent agents: theory and practice
Wooldridge, M., & Jennings, N. R. (1995) · The Knowledge Engineering Review · doi:10.1017/s0269888900008122 · Cited by 6,810 · see citing works
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.11401 · Cited by 3,101 · cited 1× · see citing works
“Retrieval augmented generation (RAG) retrieves based on similarity to the current query (Lewis et al. 2020).”
Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework
Dehaene, S. (2001) · Cognition · doi:10.1016/s0010-0277(00)00123-2 · Cited by 2,498 · cited 2× · see citing works
“At the same time, many theories require some form of integration that binds the contents of a moment into a single subject, even if they disagree about the mechanism (Bennett 2025, 2026a; Baars 1988; Dehaene and Naccache 2001; Tononi 2004; Metzinger 2003).”
Is It an agent, or just a program?: A taxonomy for autonomous agents
Franklin, S., & Graesser, A. (1997) · Lecture Notes in Computer Science · doi:10.1007/bfb0013570 · Cited by 2,286 · see citing works
An information integration theory of consciousness
Tononi, G. (2004) · BMC Neuroscience · doi:10.1186/1471-2202-5-42 · Cited by 1,757 · cited 2× · see citing works
“At the same time, many theories require some form of integration that binds the contents of a moment into a single subject, even if they disagree about the mechanism (Bennett 2025, 2026a; Baars 1988; Dehaene and Naccache 2001; Tononi 2004; Metzinger 2003).”
Lost in the Middle: How Language Models Use Long Contexts
Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00638 · Cited by 1,612 · cited 1× · see citing works
“Empirically, long context performance is uneven and position dependent (Liu et al. 2024).”
Being No One: The Self-Model Theory of Subjectivity
Metzinger, T. (2003) · MIT Press · doi:10.7551/mitpress/1551.001.0001 · Cited by 789 · see citing works
AgentBench: Evaluating LLMs as Agents
Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., et al. (2023) · arXiv · doi:10.48550/arxiv.2308.03688 · Cited by 79 · see citing works
Emergent Causality and the Foundation of Consciousness
Bennett, M. T. (2023) · Lecture notes in computer science · doi:10.1007/978-3-031-33469-6_6 · Cited by 25 · see citing works
Self-citationOn the Computation of Meaning, Language Models and Incomprehensible Horrors
Bennett, M. T. (2023) · Lecture notes in computer science · doi:10.1007/978-3-031-33469-6_4 · Cited by 18 · see citing works
Self-citationHow To Build Conscious Machines
Bennett, M. T. (2025) · doi:10.31237/osf.io/wehmg_v1 · Cited by 15
Self-citationKapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., & Narayanan, A. (2024) · arXiv · doi:10.48550/arxiv.2407.01502 · Cited by 13 · see citing works
Position: Stop Acting Like Language Model Agents Are Normal Agents
Perrier, E., & Bennett, M. T. (2025) · arXiv.org · doi:10.48550/arxiv.2502.10420 · Cited by 6
Self-citationCognition spaces: natural, artificial, and hybridPreprint
Solé, R. V., Seoane, L. F., Pla, J., Bennett, M. T., Hochberg, M., & Levin, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.12837 · Cited by 1 · see citing works
Self-citationA Mind Cannot Be Smeared Across Time
Bennett, M. T. (2026) · Proceedings of the AAAI Symposium Series · doi:10.1609/aaaiss.v8i1.42545
Self-citationBennett, M. T. (2026) · Proceedings of the AAAI Symposium Series · doi:10.1609/aaaiss.v8i1.42546
Self-citation
Flagged first, then by citations
7 of the 23 references couldn’t be matched to a known paper.
Research works printed but not linked: 1
- Baars, B. J. 1988. A Cognitive Theory of Consciousness. Cambridge, UK: Cambridge University Press.
Cited by
Sorted by citations
Hannen, T. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7459703
- Authors
- Perrier, E., & Bennett, M. T.
- Affiliations
- University of Technology Sydney (AU)
- Australian National University (AU)
- Venue
- Proceedings of the AAAI Symposium Series
- Published
- 2026-05-18
- Volume
- 8
- Issue
- 1
- Pages
- 322-328
- ISSN
- 2994-4317 (ISSN-L)
- OpenAlex
- W7161700524
- Cited by
- 1
- References
- 23
- Citations by year
- 2026: 1
- FWCI
- 22.12
- Type
- conference paper
- Language
- en
- Open access
- diamond
- doi.org
- Locations
- Proceedings of the AAAI Symposium Series (publishedVersion)
- AAAI Publications (The Association for the Advancement of Artificial Intelligence (AAAI)) (submittedVersion)
- Topics
- Social Sciences › Psychology › Experimental and Cognitive Psychology › Philosophy and Theoretical Science
- Physical Sciences › Computer Science › Artificial Intelligence › Psychiatry, Mental Health, Neuroscience
- Social Sciences › Arts and Humanities › History and Philosophy of Science › Cybernetics and Technology in Society
- Keywords
- language model agents, artificial consciousness, agent identity, self-presentation, tool use
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-13.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-13.
Cited sources are checked by gpt-6-sol.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
