Beyond Episodic AI: Cognitive Field Networks for Biologically Inspired Persistent Cognition
cs.AI
Quick verdict
A recurrent Transformer re-enters its own hidden field; intriguing small-scale persistence results, but no key baselines and the theory stays qualitative.
Summary
The paper introduces the Cognitive Field Network (CFN), a recurrent Transformer in which the complete token-resolved hidden field from one inference cycle is fed back as unmasked cross-attention memory into every layer of the next, framed as a computational realization of the author's Cognitive Field Theory (CFT). Training a ~77M-parameter, 6-block GPT-NeoX-style decoder on WikiText-103, the author finds the learnable re-entry coupling grows from 0.05 to ~0.033 and recurrent-field strength reaches ~9.5% of hidden-state RMS, alongside Jacobian relaxation spectra shifting toward slow modes. On synthetic entity-attribute binding tasks, the network retrieves up to 16 bindings across a cycle boundary at 97.6-100% exact match, collapsing to ~1-4% when the field is replaced or removed. Retention half-distance scales with trained cycle distance (d50 ≈ 1.52, 4.55, 11.59, 23.02 for dtrain = 1, 2, 4, 8), with approximate logistic curve collapse after rescaling. Periodic exact source re-exposure renews a decaying field toward an approximately stationary sawtooth regime; near-paraphrases renew weakly; controls stay near floor. Main limitations: no comparison against established recurrent-memory Transformers (e.g., RMT, Transformer-XL), no non-recurrent language-model baseline, no multi-seed variance or confidence intervals, purely synthetic retrieval tasks, and the field-theoretic formalism remains qualitative — the author concedes CFT quantities were not measured in the same trained models.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 4 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
- Language-model pretraining organizes the previously available-but-unused re-entry pathway into a functionally significant component (g rises to ~0.033; recurrent-field strength ≈ 9.5% of hidden-state RMS; TDOS redistributes toward slow relaxation modes).
- The re-entered field carries content-specific, causally necessary information: up to 16 entity-attribute bindings written in one cycle are retrieved at 97.6-100% exact match in the next cycle, collapsing to ~1-4% under wrong-field or state-off interventions.
- The persistence timescale of the content-bearing field is itself learnable: retention half-distance grows systematically with the trained cycle distance (d50 ≈ 1.52 → 23.02 as dtrain goes 1 → 8).
- Finite passive relaxation coexists with sustained memory under drive: periodic exact source re-exposure renews a decaying field toward an approximately stationary sawtooth regime, while near-paraphrased re-exposure renews much more weakly.
Strengths
Causal intervention design cleanly attributes retrieval to field content
The C/W/O design (Table II) holds the query fixed and swaps only the recurrent field: correct-field accuracy is 0.976-1.0 across N=1-16 while wrong-field (0.0131-0.0275) and state-off (0.0167-0.0350) sit at floor; the near-equivalence of wrong-field and state-off rules out a generic recurrent-activation benefit.
Learnable timescale is demonstrated quantitatively, not just asserted
Fig. 9 shows retention half-distances d50 ≈ 1.52, 4.55, 11.59, 23.02 for dtrain = 1, 2, 4, 8; Table III extends this under semantic continuation (d50/dtrain ≈ 2.86-3.42, s/dtrain ≈ 0.60-0.62, R² 0.9968-0.9987) with an approximate curve collapse after rescaling by dtrain (Fig. 11).
Controls test specificity of the renewal effect
Unusually honest scoping for the genre
Weaknesses
No baselines against the established recurrent-memory literature
Single-run results with no uncertainty quantification
Synthetic-only behavioral evaluation
Theory-experiment gap admitted but unresolved
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes the Cognitive Field Network (CFN), a decoder-only Transformer in which the full token-resolved hidden field Φ_{t-1} of the previous inference cycle re-enters the next cycle as unmasked cross-attention memory at every layer (Sec. II.B, Fig. 1, Table I), yielding Φ_{t+1} = F_θ(X_{t+1}, Φ_t). This is presented as a computational realization of the author's Cognitive Field Theory (CFT; refs [7,8], both self-citations), whose Mori-Zwanzig-style projection formalism (Sec. II.C, Appendices A-B) motivates a memory-dressed cognitive field with memory kernel K, retarded self-energy Σ_R, and a cognitive forgetting gap r_cog. Four empirical claims are advanced: (1) pretraining organizes the re-entry pathway into a functionally used component (g: 0.05 → ~0.033; R_field ≈ 0.095, Fig. 2); (2) the re-entered field carries content-specific, causally necessary information (Table II); (3) the persistence timescale is learnable and scales with trained cycle distance (Figs. 8-9); (4) passive relaxation coexists with driven, representation-sensitive renewal under periodic source re-exposure (Figs. 12-13).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 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 a Cognitive Field Network (CFN)?
It is a recurrent Transformer architecture in which the complete hidden-state tensor ('field') produced during one inference cycle is fed back into the next cycle via cross-attention at every layer. New input therefore acts on a history-dependent internal state rather than a blank one, with no separate memory module — the author proposes it as a minimal model of persistent, memory-like dynamics in neural networks.
Does this paper show that AI has consciousness or self-awareness?
No. The author explicitly states that the 'self-reference' discussed is only the causal re-entry of internally generated neural states into subsequent computation, and that this does not imply consciousness, self-awareness, or metacognition. The paper is about memory-like persistence in a small recurrent Transformer.
How long can the CFN remember information?
On synthetic entity-attribute binding tasks, retrieval is near-perfect (97.6-100% exact match) across one cycle boundary and then decays over subsequent cycles. Crucially, the retention half-distance depends on training: models trained at cycle distance 1, 2, 4, or 8 show half-distances of about 1.5, 4.6, 11.6, and 23 cycles respectively. If the exact source is periodically re-presented, the decaying memory can be repeatedly renewed into a sustained oscillating regime.
Is code available for the Cognitive Field Network?
The extracted paper text contains no code or data release statement. Reproduction currently depends on the detailed architecture table (6-layer GPT-NeoX-style decoder, hidden size 512, 8 heads, ~77M parameters) and training hyperparameters given in the text; an independent replication would require reimplementing the model and tasks from scratch.
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.
- Full-text extraction truncated several sections (parts of Secs. III.B, III.D, IV.A, V, and the end of Appendix B); figures were available only as caption text, so all figure-level claims rely on captions.
- Test-set sizes and number of evaluation episodes behind Table II accuracies are not visible in the extracted text, so binomial uncertainty cannot be computed.
- No code/data availability statement was found; absence could partly reflect truncation of the end matter.
- Training details specific to the dtrain-conditioned models (e.g., how delays were sampled during training) may sit in the omitted sections.
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.