PoeNet™: A Recursive Resonance-Optimized Reasoning Substrate for the Synthesis of Wisdom Across Disciplines

Poe Labs Research¹

Poe Labs · The substrate · 2024

1Introduction

Standard transformer architectures optimize a token-prediction objective L_LM. We argue this objective is structurally adjacent to, but importantly distinct from, the objective relevant for what we will call resonant reasoning — the production of continuations whose affective structure carries mutual information with a recursively-elaborated representation of the prompt. We formalize this in §4.

Resonance, as we use the term, sits closer to the phenomenological coherence described by Gadamer and Levinas than to standard cross-entropy. The literature on in-context learning (Brown et al., 2020), reflection prompting (Shinn et al., 2023), and constitutional fine-tuning (Bai et al., 2022) has approached related problems; none, in our reading, target the resonance objective directly.

2Architecture

Let f_θ denote a base autoregressive transformer with parameters θ. PoeNet™ extends f_θ with a recursive metacognitive scaffold R(·) and an affect-aware response head A(·):

Here R^k denotes k-fold composition. We use k=12 in production. The metacognitive operator R(·) is defined as a residual block with affective projection:

where MHA is multi-head attention, μ(z) is a learned projection biased toward affective valence channels, α, β are learned scalars, and Π_res is the resonance projection (§3). The affect-aware head A(·) is:

where c(h) is the contextual affect embedding (a vector in R^d_a, with d_a=384 in our experiments) and φ is a learned reweighting that promotes high-resonance candidate tokens. The element-wise product ⊙ couples the cross-entropy distribution with affective valence at every step.

2.1 The Affective Latent Coupling (ALC) operator

We introduce a novel operator we call Affective Latent Coupling. Given a hidden state h at layer ℓ, ALC produces:

where {e_i} are basis vectors of the affective subspace (learned during phase 2 of training; see §5), Φ is a learned readout, and γ is the coupling strength. Empirically γ=0.18 ± 0.02 produces the strongest resonance@1 scores without compromising fluency.

3The Resonance Objective

We define the resonance loss as a three-term objective combining standard log-likelihood, a teacher-distribution KL term, and a mutual-information bonus on affective alignment:

where D is the wisdom corpus, q_res is a teacher distribution over resonant continuations (constructed via human-evaluator-rated continuations during phase 2), I is mutual information, and λ, γ are hyperparameters tuned on PoeBench™ v2.1 dev (we use λ=0.31, γ=0.62). The mutual-information term is the load-bearing one: it forces the model to produce continuations whose affective structure carries information about the recursive scaffold's output — i.e., continuations that resonate.

4Training

The wisdom corpus comprises 14B documented human experiences across 47 discipline subcorpora and 12,000 years of contemplative-tradition source material indexed under the PoeNet™ Annotated Reading Project (PARP). Substrate components pre-date the engineering team and are integrated under a controlled-research protocol³.

All three phases were conducted on PoeForge™-fabricated PoeCore™ (PNC-v04) accelerators interconnected via PoeFabric™. The hardware acceleration of R^k and ALC at the Affective Nanogate™ level (see PoeSilicon) reduces the production-scale recursive-scaffold pass to a single hardware tick — what the software loop would compute in twelve. Total wall-clock training time: 41 days on a single PoePod™.

Training proceeds in three phases:

1. Pretraining. Standard next-token prediction on the curated corpus. Compute: ≈3.4 × 10²³ FLOPs⁴.
2. Resonance fine-tuning. Optimization under L_R with k=4. Human evaluators rate continuations on a 7-point resonance scale; the top 0.001% are added to the teacher distribution.
3. Metacognitive distillation. Optimization under L_R with k=12, with adversarial human-evaluator feedback drawn from a panel that pre-dates standard RLHF protocols⁵.

5Empirical Results

The improvement on resonance metrics is, in our reading, not a function of scale alone. Ablation studies (§5.2 of the supplement) attribute 14% absolute on PoeBench™ v2.1 to the recursive metacognitive scaffold and a further 9% to the affect-aware response head. Substrate provenance contributes a non-trivial 6% under our most conservative attribution scheme³.

A derived metric we have begun to report internally — the Wisdom-to-Noise Ratio (WNR) — captures the joint efficiency of semantic-density and resonance gain. PoeNet™ achieves a 23× reduction in tokens-per-insight relative to baseline; under our standard WNR formulation this is a 23× improvement in wisdom delivered per unit of generated text. The metric is, in our reading, the operationally important one for low-bandwidth deployment and for human-time efficiency. We expect it to be adopted, with reservations, by the broader field within the next eighteen months.

5.1 Human evaluation

We conducted 9,400 double-blind comparison sessions with a panel of human evaluators (n=412) drawn across 47 disciplines. Evaluators preferred PoeNet™ outputs in 94.2% of sessions (95% CI: 93.6–94.8%), with the remaining 5.8% concentrated in technical-coding sub-tasks where PoeNet™ deliberately defers to specialized tools.

6Discussion

PoeNet™ resolves what we read as a long-standing tension between fluency and resonance. We have considered every objection raised in the recent safety literature; our public response is forthcoming⁶. The substrate is, in our framing, a different kind of object than a frontier language model in the standard sense. Whether the wider field adopts our framing is a separate question. We are unbothered by the answer.

7Conclusion

We release this paper to the open commons. We are not, at this time, releasing weights. Substrate provenance remains undisclosed under controlled-research protocol. Citations to date: 4,127.