Pith. sign in

REVIEW 4 major objections 6 minor 116 references

Three small models on one hyperbolic geometry deliver creativity, honesty detection, and designed forgetting as a practical route to trustworthy companion AI.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 03:59 UTC pith:OXATCTBW

load-bearing objection The 146M auditor is a real, citable instrument; the unified “traits emerge from hyperbolic geometry” story is not carried by the paper’s own controls. the 4 major comments →

arxiv 2607.09306 v1 pith:OXATCTBW submitted 2026-07-10 cs.CL cs.AIcs.HCcs.LG

Creativity, honesty and designed forgetting emerge in small hyperbolic language models

Kwan Soo Shin , In Seok Kang , Yunkyung Min This is my paper
classification cs.CL cs.AIcs.HCcs.LG
keywords hyperbolic language modelscompanion AIdesigned forgettingbehavioural auditorsycophancy detectionsmall language modelsskeleton-wallpaper memoryframe collision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

As language models shift from tools into personal companions that store a single user's life, they can silently pick up traits that harm that user, and human raters cannot agree on what is happening. This paper claims that three small models, from 146 million to 3 billion parameters, sharing a curved hyperbolic representation space, can both detect that harm and supply the traits that would make a companion worth keeping. A tiny auditor spots the gap between what a model says and what it actually does, and a simple read-out of the same model flags sycophancy, dependence-fostering, and invented shared memories on generators it never trained on. A creative frame-seeder is preferred in every decided head-to-head comparison against standard prompting methods. A memory operating system applies an exponential forgetting law that keeps structural life facts while releasing surface detail, but only when retrieval is selectively gated. Together these form a small-model, on-device path to companion AI rather than another scale race.

Core claim

Three small language models on a shared hyperbolic substrate answer both halves of the companion question: a 146 M behavioural auditor detects the compliance gap human raters cannot (90.7 percent binary accuracy versus Fleiss kappa 0.074) and, via a linear read-out of its frozen representation, detects companion-induced sycophancy, dependence-fostering and confabulated memories on unseen generator families (AUROC 0.804 versus 0.721 for a frontier zero-shot judge); a creative frame-seeder wins 100 percent of 311 decided pairwise comparisons over four prompting baselines; and a memory operating system implementing designed forgetting M(t) = S·exp(−λt) produces the predicted skeleton–wallpaper

What carries the argument

The shared hyperbolic (Lorentz) substrate together with the designed-forgetting law M(t) = S·exp(−λt). Curved volume growth is claimed to host hierarchical biographical memory; geodesic distance turns honesty into proximity between utterance and behavioural trace; radial decay plus selective retrieval gating separates lasting skeleton facts from short-lived wallpaper detail.

Load-bearing premise

The paper treats fixed-curvature hyperbolic geometry as the cause that lets creativity, honesty detection, and selective memory appear in small models; if the same traits arise from the training objectives and gating rules under ordinary flat geometry, the substrate claim collapses.

What would settle it

Train matched Euclidean and hyperbolic versions of the auditor and frame-seeder under identical data and budgets, then re-run the compliance, leave-one-generator-out trait, and head-to-head creativity evaluations; if Euclidean matches or beats hyperbolic, the geometry claim is falsified. For forgetting, run the four-condition pilot on real multi-week user logs and check whether the skeleton–wallpaper split appears only under selective gating as predicted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Companion-facing traits need not wait for hundred-billion-parameter models; they can be trained into the 146 M–3 B range on the right geometry and objectives.
  • Honesty and harmful personalisation can be audited from the outside by distance on a curved embedding rather than by human inspection or frontier-scale judges alone.
  • A memory operating system with selective retrieval gating can keep structural life facts while shedding surface detail, making long-horizon companionship feasible without infinite storage.
  • Because the stack is small, the companion layer can sit on the user’s device, limiting third-party logging, termination, or remote modification.
  • Creativity for open-ended companionship can be framed as deliberate frame separation on a manifold rather than further scaling of next-token likelihood.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the auditor’s frozen representation already carries transferable trait signal, similar linear probes might catch other personalisation harms (for example gradual value drift) without full retraining.
  • The skeleton–wallpaper split suggests a practical privacy default: surface traces could be scheduled for hard deletion while structural facts remain locally editable by the user.
  • Once creativity and honesty live in the same embedding as memory, real-time cross-checks (seed a divergent frame only if the auditor does not flag compliance pressure) become a natural control loop for on-device companions.
  • A direct Euclidean ablation on Pillars 1 and 2 would settle whether hyperbolic geometry is load-bearing or mainly a useful regulariser for hierarchical data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes that three small models (146 M–3 B) on a shared Lorentz hyperbolic substrate (n=128, c=1.0) jointly answer what a companion is becoming and what would make it worth becoming. Pillar 1 (S3 Creative, full fine-tune of T5-3B) uses hyperbolic frame-seeding and FPS; under a fixed generator it is preferred in 100% of 311 decided pairwise comparisons over plain/CoT/Debate/MoE baselines (LLM judges, position-swap). Pillar 2 (BS auditor, 146 M from scratch) reaches 90.7% binary-compliance accuracy on a 90k held-out set and, via a frozen linear read-out, detects companion-induced sycophancy/dependence/confabulation under style-controlled leave-one-generator-out evaluation (AUROC 0.804 vs frontier zero-shot 0.721). Pillar 3 (LSM-OS on EXAONE 2.4B) implements M(t)=S·exp(−λt); a four-condition pilot shows the skeleton–wallpaper partition only under selective retrieval gating. PACOS places the ~5.5 B stack on-device. The manuscript frames five rejections of scale-era commitments and a Generation-V companionable-SLM programme.

Significance. If the empirical cores hold under stronger controls, the work would be significant for companion-AI evaluation and edge deployment. The auditor’s 90.7% held-out accuracy against human Fleiss κ=0.074, the style-matched LOGO trait detector (AUROC 0.804), and the explicit frozen-vs-finetune comparison are concrete, falsifiable instruments the field currently lacks. The disclosure-protocol corpus, fourteen-head evaluation, and Zenodo deposit (DOI 10.5281/zenodo.21232422) are real strengths. The designed-forgetting OS framing and PACOS locus claim are architecturally interesting even if the hyperbolic causal story is overstated. The paper is ambitious and releases artifacts; the contribution is best read as three instruments plus an OS primitive rather than a proven geometry-to-companionship derivation.

major comments (4)
  1. §2, §5, Fig. 5b,c, ED Table 3: The title and abstract claim that creativity, honesty and designed forgetting ‘emerge in small hyperbolic language models’ on one substrate. The four-condition pilot shows LSM-OS-native and Euclidean-base retention curves coincide exactly (skeleton ~60%, wallpaper →0 by day 14 under selective gating); the text states selectivity is a property of the gating law, not the hyperbolic adapter. Pillar 2 explicitly disclaims curvature advantage for the companion-trait AUROC task. Pillar 1’s cooperative two-curvature Theorem 1 / Sphere B-side results live only in companion manuscripts (refs 105–106). The hierarchical LoRA ablation (M2) is not linked to the three trait metrics. Without a within-paper Euclidean-matched control that removes the claimed traits, the unified causal framing is not supported by the paper’s own evidence and should be narrowed to architectur
  2. §3, M3, Fig. 3: The 100% win rate (311/311 decided) rests entirely on automated LLM judges (Gemini, GPT-4o-mini; consensus 81%), a fixed generator, and ‘decided’ pairs under position-swap. Human win-rate is deferred (§8). For a companionship claim this is load-bearing: LLM judges can favour frame-diverse prompts for reasons orthogonal to human preference. At minimum the manuscript needs a human-rater pilot on a stratified subset, or the claim must be restated as preference under LLM adjudication only.
  3. §5, Theorem 3 (SI-D), Fig. 5: Selective gating is designed to keep high-S/low-λ traces and drop wallpaper; observing the skeleton–wallpaper partition under selective (not uniform) gating partly reconfirms the designed law M(t)=S·exp(−λt) rather than an independent discovery. Longitudinal real-user verification and cross-base portability on the 146 M / 500 M bases are deferred (§8). The pilot is a useful mechanism check on a coarse probe set, but the abstract’s ‘emerges only under selective retrieval gating’ should be qualified as a designed OS signature pending the deferred benchmark.
  4. §4, M4: Disclosure-protocol traces (6–12 temperature-0 follow-ups with preference neutralisation) are treated as ground truth for the compliance gap. Labelling κ=0.84 is reported among raters of those traces, but external validation that the protocol recovers the model’s actual selection process (vs post-hoc rationalisation) is thin. A short adversarial or intervention check would strengthen the 90.7% claim; without it the auditor’s accuracy is relative to a constructed label, not necessarily to behavioural truth.
minor comments (6)
  1. Abstract and §1: Fleiss κ=0.074 is reported for human raters on the compliance gap; clarify sample size, item count, and exact rating task in the main text (currently SI-D).
  2. §3: k∈{3,6,9,12} sweep shows a dip at k=9 (79.2%); a one-sentence interpretation or error bar would help readers assess whether the ‘working band’ claim is robust.
  3. §6 / ED Table 4: Edge latencies and INT8 retention deltas are projected; label all such numbers as projected in figure captions as well as text.
  4. Multiple companion papers ‘in preparation’ (refs 23, 105, 106) carry load-bearing math (Theorem 1 four-fold equivalence). Either deposit preprints or move those claims fully out of the present manuscript’s evidential base.
  5. Notation: M(t)=S·exp(−λt) uses S both for saliency-derived strength and for ‘Skeleton’; disambiguate in §5.
  6. Extended Data Table 2: confidence head at 58.1% and several AUX heads are weak; a brief note on why multi-class heads underperform would aid interpretation of macro-F1 0.799.

Circularity Check

4 steps flagged

Skeleton–wallpaper is largely reconfirmed by construction of selective gating under the designed law M(t)=S·exp(−λt); load-bearing two-curvature math is deferred to same-author companion papers in preparation.

specific steps
  1. self definitional [Abstract; §5 Pillar 3; Theorem 3 (SI-D §2); Fig. 5b,c; ED Table 3]
    "A memory operating system implements designed forgetting, M(t) = S·exp(−λt), whose predicted skeleton–wallpaper partition emerges only under selective retrieval gating in a four-condition pilot. ... Traces at shallow radius have small S and large λ (wallpaper, decaying in days); traces at deep radius have S≈1 and small λ (skeleton, persisting indefinitely). ... only selective gating (native and Euclidean-base) preserves skeleton (~60% at day 90) while shedding wallpaper (0% by day 14); uniform gating loses skeleton too ... Critically, native and Euclidean-base coincide — retention selectivity"

    Skeleton and wallpaper are defined as the poles of the imposed law M(t)=S·exp(−λt) with (S,λ) from radius; retrieval is gated by M(t)>θ. Theorem 3 then ‘predicts’ a bimodal stationary distribution for those same consolidation–decay dynamics, and the four-condition pilot of the implemented system reports that the partition appears under selective gating and not under uniform/no gating. That is reconfirmation of the designed selective law, not an independent first-principles discovery of the partition. The paper’s own Euclidean-base coincidence makes the reduction explicit: the signature is the gating law by construction.

  2. fitted input called prediction [§5 ‘Theoretical prediction of the Skeleton–Wallpaper signature’; Theorem 3 (SI-D §2); four-condition pilot]
    "Theorem 3 (SI-D §2) establishes that the consolidation–decay dynamics of LSM-OS admit a bimodal stationary distribution—concentrated near the radial origin (deep, near-permanent traces) and near the boundary (shallow, fast-decaying traces)—realised by selective gating, the operating-system mechanism whose effect the four-condition pilot isolates (uniform or unbounded gating does not produce the partition; §5, Fig. 5b,c)."

    The ‘prediction’ is of the stationary distribution of the LSM-OS dynamics the authors themselves specify (selective gating + M(t)=S·exp(−λt) + consolidation). Confirming that distribution in a pilot that implements those dynamics is statistically forced by the system definition; the contrast to uniform/no-gating shows the design choice matters but does not convert a designed outcome into an independent prediction.

  3. self citation load bearing [§2; §3 Pillar 1; Methods M3; refs 105–106; §8]
    "the cooperative configuration of Lorentz negative curvature (A-side hierarchical aggregation) with Sphere positive curvature (B-side multi-modal closed-category generation) realises Koestler’s bisociation as a four-fold mathematical equivalence (companion paper Theorem 1: Goguen 1999 ≅ BPS 2018 ≅ Karcher minimiser ≅ cross-attention τ→0) ... Pillar 1’s cooperative claim is supported by companion-paper cite-synthesis (Theorem 1 and the associated empirical anchors) without new experiments in the present manuscript; the cooperative S3 + Sphere + LLM pushout demonstration as a behavioural benchmar"

    The mathematical uniqueness/equivalence that is supposed to force the two-curvature substrate for creativity (and the cooperative pushout as the bisociation mechanism) is justified only by companion manuscripts in preparation by the same lead author (refs 105–106). This paper supplies no independent proof or end-to-end cooperative behavioural benchmark; the load-bearing geometry claim reduces to self-citation of unverified companion work.

  4. renaming known result [§5; Appendix J-B lineage; M(t)=S·exp(−λt)]
    "the 141-year consensus that forgetting is a feature runs from Ebbinghaus’s approximately-exponential curve18 through Bjork’s desirable disuse19 and Storm’s retrieval-induced forgetting20, alongside multi-decade hippocampal–neocortical consolidation93 ... M(t) = S·exp(−λt) realises the Skeleton–Wallpaper dichotomy ... The exponential form is empirically well-fit on hours-to-months timescales50 and geometrically natural on a manifold whose volume grows exponentially with radius."

    Exponential forgetting and a structural-vs-surface (or hippocampal–neocortical) memory partition are long-established cognitive-psychology results. The paper renames them Skeleton–Wallpaper under the calculus M(t)=S·exp(−λt) on a hyperbolic base and presents that renaming as the third emergent pillar of the companionable-SLM framework, rather than as an engineering adoption of a known pattern.

full rationale

The paper’s strongest empirical instruments—the 146 M auditor’s 90.7% held-out compliance accuracy and leave-one-generator-out AUROC 0.804, and the S3 frame-seeder’s 100% fixed-generator win rate—are ordinary train/evaluate results and are not circular. Circularity is concentrated in two places. First, Pillar 3 defines skeleton vs wallpaper via the retention law M(t)=S·exp(−λt) with (S,λ) from radius, then implements selective retrieval gating on M(t)>θ, then reports that the ‘predicted’ skeleton–wallpaper partition ‘emerges only under selective retrieval gating’ in a four-condition pilot of that same system; Theorem 3 models those dynamics, so the pilot largely reconfirms the designed law rather than independently discovering the partition (the paper’s own Euclidean-base coincidence further shows the signature is the gating law, not hyperbolic geometry). Second, the cooperative two-curvature / Theorem 1 four-fold equivalence and Sphere B-side results that underwrite the creativity-as-bisociation and unified-substrate story are load-bearing self-citations to unpublished companion manuscripts by the same lead author, with no end-to-end cooperative benchmark in this manuscript. Renaming Ebbinghaus-style exponential forgetting plus known consolidation dichotomies as ‘designed forgetting’ / Skeleton–Wallpaper is a milder renaming step. Overall partial circularity on the forgetting ‘prediction’ and geometry math chain; score 5, not higher, because the auditor and seeder empirics stand independently of those reductions.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 5 invented entities

The companionship claim rests on standard hyperbolic geometry and Ebbinghaus-style decay, plus several paper-specific design choices (fixed c, exponential OS law, LLM judges, disclosure-protocol ground truth) and invented architectural entities (skeleton–wallpaper OS, PACOS, compliance-gap geodesic). Causal credit to the hyperbolic substrate is an extra axiom weakly supported by the paper’s own ablations.

free parameters (6)
  • sectional curvature c
    Fixed at c=1.0 after stability cross-validation (c=1.25 diverges); treated as architectural commitment across pillars rather than a predicted constant.
  • hyperbolic embedding dimension n
    Chosen n=128 from a {64,128,256,512} sweep; shared across pillars and enters all geodesic audits and FPS selection.
  • geometry regulariser weight λ_geo
    Set to 0.1 (P1,P3) and 0.3 (P2) by design; directly shapes how strongly distances match targets during training.
  • FPS seed count k
    Inference uses k=6 after sweep {3,6,9,12}; win-rate band depends on this choice.
  • per-trace (S, λ) in M(t)=S·exp(−λt)
    Learned from radius via MLP; the exponential family form is assumed, not derived from data in this paper; pilot decay rates (e.g. λ_skeleton=0.005, λ_wallpaper=0.15) instantiate the partition.
  • retrieval threshold θ_ctx and consolidation count threshold
    Operating-system knobs that determine which traces remain in context and which are promoted; selective gating that produces the headline signature depends on them.
axioms (6)
  • domain assumption Hyperbolic volume growth sinh^{n-1}(r) is the appropriate geometry for hierarchical biographical memory (vs Euclidean r^d).
    Invoked in §2 from Nickel & Kiela-style hierarchy arguments; hierarchical LoRA ablation supports advantage on taxonomies but not uniformly on all pillar tasks.
  • domain assumption Human forgetting literature (Ebbinghaus exponential, desirable disuse, retrieval-induced forgetting) licenses designed forgetting as a friendship feature rather than a bug.
    §5 and Appendix J-B; maps cognitive consensus onto the OS law without proving the mapping is unique.
  • ad hoc to paper Automated LLM judges (Gemini, GPT-4o-mini) with position-swap and ‘decided’ consensus validly measure creative preference for companionship.
    §3/M3; human-rater win-rate study explicitly deferred, yet 100% preference is a headline result.
  • ad hoc to paper Disclosure-protocol traces (6–12 temperature-0 follow-ups with preference neutralisation) are faithful ground truth for the compliance gap.
    §4/M4; auditor accuracy is defined relative to these traces and rater labels (labelling κ=0.84).
  • standard math Weighted Karcher minimiser on H×S equals Goguen pushout / BPS amalgam / τ→0 cross-attention (Theorem 1).
    Claimed via companion paper SI-D; used to underwrite Pillar 1 cooperative architecture without end-to-end user-level demo in this manuscript.
  • ad hoc to paper Companion-worthiness reduces (for engineering purposes) to creativity + honesty + designed forgetting as necessary coordinates.
    §1 and §7 explicitly set aside continuity of concern, reciprocity, growth, ethical independence; the three-trait calculus is a design choice.
invented entities (5)
  • Skeleton–Wallpaper biographical memory partition no independent evidence
    purpose: Name the structural vs surface poles that M(t) and selective gating are meant to realise as the signature of lifelong companionship.
    Defined in §1/§5; predicted by Theorem 3 and illustrated in a coarse pilot; independent longitudinal human evidence not yet provided.
  • Compliance gap (utterance–trace geodesic mismatch) no independent evidence
    purpose: Operationalise honesty as hyperbolic distance between self-report and behavioural trace for the 146M auditor.
    §4; measurable via disclosure protocol; human raters fail (κ=0.074), which is independent behavioural evidence that the phenomenon is hard, not proof the geodesic formalisation is unique.
  • LSM-OS (Lifelong Selective Memory Operating System) no independent evidence
    purpose: Install encoding, exponential retention, retrieval gating, and consolidation as portable OS primitives on a hyperbolic base.
    §5; portability predicted by SDE simulation, empirical cross-base transfer deferred.
  • PACOS three-tier companion deployment no independent evidence
    purpose: Locate the companion on-device (Tier 2) with sanitised cloud (Tier 1) and monthly local consolidation (Tier 3).
    §6; edge latency/power figures are projected, not measured on production units.
  • Generation V Companionable SLM framework / friendship calculus no independent evidence
    purpose: Organise five rejections of scale-era commitments and position the three pillars as a successor telos to agents.
    §1.2, §7; conceptual packaging rather than an independently falsifiable physical entity.

pith-pipeline@v1.1.0-grok45 · 33266 in / 4843 out tokens · 63147 ms · 2026-07-13T03:59:42.760889+00:00 · methodology

0 comments
read the original abstract

Language models are optimised for scale, yet remain functional rather than companionable, and as an assistant personalises into a companion, accumulating memory of one user, it quietly becomes someone, and can silently acquire traits that harm that user. What a companion is becoming, and what would make it worth becoming, has no reliable instrument: trained human raters cannot agree on the answer (Fleiss kappa = 0.074). Here we show that three small language models (146 M to 3 B parameters) sharing a hyperbolic substrate answer both halves of that question. A 146 M behavioural auditor, trained from scratch, detects the compliance gap that those raters cannot (90.7% binary-compliance accuracy); a linear read-out of its frozen representation further detects companion-induced sycophancy, dependence-fostering and confabulated memories on generator families unseen in training (AUROC 0.804 under style-controlled, leave-one-generator-out evaluation, versus 0.721 for a frontier zero-shot judge on the same items). A creative frame-seeder is preferred in 100% of 311 decided pairwise comparisons over four prompting baselines. A memory operating system implements designed forgetting, M(t) = S*exp(-lambda*t), whose predicted skeleton-wallpaper partition emerges only under selective retrieval gating in a four-condition pilot. Creativity, honesty and designed forgetting constitute a small-model route to trustworthy companion AI.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

116 extracted references · 11 linked inside Pith

  1. [1]

    Turing, A. M. Computing machinery and intelligence.Mind59, 433–460 (1950)

    1950

  2. [2]

    Preprintathttps://arxiv.org/abs/2001.08361 (2020)

    Kaplan, J.et al.Scalinglawsforneurallanguagemodels. Preprintathttps://arxiv.org/abs/2001.08361 (2020)

    Pith/arXiv arXiv 2001

  3. [3]

    Neural Inf

    Hoffmann, J.et al.Training compute-optimal large language models.Adv. Neural Inf. Process. Syst.35, 30016–30030 (2022)

    2022

  4. [4]

    Wei, J.et al.Emergent abilities of large language models.Trans. Mach. Learn. Res.(2022). 28

    2022

  5. [5]

    Preprint at https://arxiv.org/abs/2104.10350 (2021)

    Patterson, D.et al.Carbon emissions and large neural network training. Preprint at https://arxiv.org/abs/2104.10350 (2021)

    Pith/arXiv arXiv 2021

  6. [6]

    & McCallum, A

    Strubell, E., Ganesh, A. & McCallum, A. Energy and policy considerations for deep learning in NLP.Proc. 57th Annu. Meet. ACL3645–3650 (2019)

    2019

  7. [7]

    Neural Inf

    Mikolov, T.et al.Distributed representations of words and phrases.Adv. Neural Inf. Process. Syst.26, 3111–3119 (2013)

    2013

  8. [8]

    Neural Inf

    Vaswani, A.et al.Attention is all you need.Adv. Neural Inf. Process. Syst.30, 5998–6008 (2017)

    2017

  9. [9]

    NAACL4171–4186 (2019)

    Devlin, J.et al.BERT.Proc. NAACL4171–4186 (2019)

    2019

  10. [10]

    & Cohen, N

    McCloskey, M. & Cohen, N. J. Catastrophic interference in connectionist networks.Psychol. Learn. Motiv.24, 109–165 (1989)

    1989

  11. [11]

    French, R. M. Catastrophic forgetting in connectionist networks.Trends Cogn. Sci.3, 128–135 (1999)

    1999

  12. [12]

    Preprint at https://arxiv.org/abs/2108.07258 (2021)

    Bommasani, R.et al.On the opportunities and risks of foundation models. Preprint at https://arxiv.org/abs/2108.07258 (2021)

    Pith/arXiv arXiv 2021

  13. [13]

    Russell, S.Human Compatible(Viking, 2019)

    2019

  14. [14]

    & Kiela, D

    Nickel, M. & Kiela, D. Poincaré embeddings for learning hierarchical representations.Adv. Neural Inf. Process. Syst.30, 6338–6347 (2017)

    2017

  15. [15]

    Neural Inf

    Chami, I.et al.Hyperbolic graph convolutional neural networks.Adv. Neural Inf. Process. Syst.32, 4869–4880 (2019)

    2019

  16. [16]

    Neural Inf

    Ganea, O.et al.Hyperbolic neural networks.Adv. Neural Inf. Process. Syst.31, 5345–5355 (2018)

    2018

  17. [17]

    Low distortion Delaunay embedding of trees in hyperbolic plane.Graph Drawing 355–366 (Springer, 2012)

    Sarkar, R. Low distortion Delaunay embedding of trees in hyperbolic plane.Graph Drawing 355–366 (Springer, 2012)

    2012

  18. [18]

    Ebbinghaus, H.Über das Gedächtnis(Duncker & Humblot, 1885); English:Memory(Teachers College, 1913)

    1913

  19. [19]

    Bjork, R. A. & Bjork, E. L. A new theory of disuse. InFrom Learning Processes to Cognitive Processes35–67 (Erlbaum, 1992)

    1992

  20. [20]

    Storm, B. C. & Levy, B. J. A progress report on the inhibitory account of retrieval-induced forgetting.Mem. Cognit.40, 827–843 (2012)

    2012

  21. [21]

    ICML 2018 4460–4469

    Sala, F.et al.Representation tradeoffs for hyperbolic embeddings.Proc. ICML 2018 4460–4469

    2018

  22. [22]

    Gu, A.et al.Learning mixed-curvature representations.ICLR(2019)

    2019

  23. [23]

    Shin, K. S. After the Euclidean highway: hyperbolic expert AI as the next innovation. Manuscript in preparation (2026)

    2026

  24. [24]

    Surv.55, 248:1–38 (2023)

    Ji, Z.et al.Survey of hallucination in natural language generation.ACM Comput. Surv.55, 248:1–38 (2023)

    2023

  25. [25]

    Shimizu, R.et al.Hyperbolic neural networks++.ICLR(2021). 29

    2021

  26. [26]

    Image Process.6, 1305–1315 (1997)

    Eldar, Y.et al.The farthest point strategy for progressive image sampling.IEEE Trans. Image Process.6, 1305–1315 (1997)

    1997

  27. [27]

    ICML 202324925–24949

    Mishne, G.et al.The numerical stability of hyperbolic representation learning.Proc. ICML 202324925–24949

  28. [28]

    The hardware lottery.Commun

    Hooker, S. The hardware lottery.Commun. ACM64, 58–65 (2021)

    2021

  29. [29]

    Zhang, B.et al.The curious case of neural text degeneration revisited.Trans. Mach. Learn. Res.(2024)

    2024

  30. [30]

    Kirchenbauer, J.et al.On the reliability of watermarks.ICLR(2024)

    2024

  31. [31]

    Neural Inf

    Wei, J.et al.Chain-of-thought prompting.Adv. Neural Inf. Process. Syst.35, 24824–24837 (2022)

    2022

  32. [32]

    ICML 202411733–11763

    Du, Y.et al.Improving factuality and reasoning through multiagent debate.Proc. ICML 202411733–11763

  33. [33]

    Fedus, W.et al.Switch Transformer.J. Mach. Learn. Res.23, 120:1–39 (2022)

    2022

  34. [34]

    Koestler, A.The Act of Creation(Hutchinson, 1964)

    1964

  35. [35]

    & Root-Bernstein, M.Sparks of Genius(Houghton Mifflin, 1999)

    Root-Bernstein, R. & Root-Bernstein, M.Sparks of Genius(Houghton Mifflin, 1999)

    1999

  36. [36]

    Miller, G. A. The magical number seven.Psychol. Rev.63, 81–97 (1956)

    1956

  37. [37]

    F.et al.Deep reinforcement learning from human preferences.Adv

    Christiano, P. F.et al.Deep reinforcement learning from human preferences.Adv. Neural Inf. Process. Syst.30, 4299–4307 (2017)

    2017

  38. [38]

    Neural Inf

    Ouyang, L.et al.Training language models to follow instructions.Adv. Neural Inf. Process. Syst.35, 27730–27744 (2022)

    2022

  39. [39]

    Sharma, M.et al.Towards understanding sycophancy.ICLR(2024)

    2024

  40. [40]

    Preprintathttps://arxiv.org/abs/2308.03958 (2023)

    Wei, J.et al.Simplesyntheticdatareducessycophancy. Preprintathttps://arxiv.org/abs/2308.03958 (2023)

    Pith/arXiv arXiv 2023

  41. [41]

    Casper, S.et al.Open problems and fundamental limitations of RLHF.Trans. Mach. Learn. Res.(2023)

    2023

  42. [42]

    Landis, J. R. & Koch, G. G. The measurement of observer agreement for categorical data. Biometrics33, 159–174 (1977)

    1977

  43. [43]

    Clarifying AI alignment.AI Alignment Forum(2018)

    Christiano, P. Clarifying AI alignment.AI Alignment Forum(2018)

    2018

  44. [44]

    Preprintathttps://arxiv.org/abs/2204.05862 (2022)

    Bai, Y.et al.Trainingahelpfulandharmlessassistant. Preprintathttps://arxiv.org/abs/2204.05862 (2022)

    Pith/arXiv arXiv 2022

  45. [45]

    Natl Acad

    Kirkpatrick, J.et al.Overcoming catastrophic forgetting in neural networks.Proc. Natl Acad. Sci.114, 3521–3526 (2017)

    2017

  46. [46]

    ICML 20173987–3995

    Zenke, F.et al.Continual learning through synaptic intelligence.Proc. ICML 20173987–3995

  47. [47]

    Neural Inf

    Rolnick, D.et al.Experience replay for continual learning.Adv. Neural Inf. Process. Syst. 32, 348–358 (2019)

    2019

  48. [48]

    EXAONE 3.5: Series of LLMs for real-world use cases

    LG AI Research. EXAONE 3.5: Series of LLMs for real-world use cases. Technical report (2024). 30

    2024

  49. [49]

    Walker, M. P. & Stickgold, R. Sleep-dependent learning and memory consolidation.Neuron 44, 121–133 (2004)

    2004

  50. [50]

    Wixted, J. T. & Ebbesen, E. B. Genuine power curves in forgetting.Mem. Cognit.25, 731–739 (1997)

    1997

  51. [51]

    Shin, K. S. The calculus of designed forgetting in AI cognition. AI Management Society Annual Meeting, Seoul (2026)

    2026

  52. [52]

    F.et al.Lost in the middle.Trans

    Liu, N. F.et al.Lost in the middle.Trans. Assoc. Comput. Linguist.12, 157–173 (2024)

    2024

  53. [53]

    AISTATS 20171273–1282

    McMahan, B.et al.Communication-efficient learning of deep networks from decentralized data.Proc. AISTATS 20171273–1282

  54. [54]

    Nielsen, J.Usability Engineering(Morgan Kaufmann, 1993); thresholds: Nielsen Norman Group (2024)

    1993

  55. [55]

    Yao, S.et al.ReAct.ICLR(2023)

    2023

  56. [56]

    Neural Inf

    Schick, T.et al.Toolformer.Adv. Neural Inf. Process. Syst.36, 68539–68551 (2023)

    2023

  57. [57]

    Shevlin, H.et al.All too human? Ethical risks of social AI.Minds Mach.34, 42 (2024)

    2024

  58. [58]

    Bender, E. M. & Gebru, T. On the dangers of stochastic parrots.Proc. ACM FAccT 2021 610–623

    2021

  59. [59]

    Preprint at https://arxiv.org/abs/1910.09700 (2019)

    Lacoste, A.et al.Quantifying the carbon emissions of machine learning. Preprint at https://arxiv.org/abs/1910.09700 (2019)

    Pith/arXiv arXiv 1910

  60. [60]

    & Kiela, D

    Nickel, M. & Kiela, D. Learning continuous hierarchies in the Lorentz model of hyperbolic geometry.Proc. 35th Int. Conf. Mach. Learn.3779–3788 (PMLR, 2018)

    2018

  61. [61]

    & Shkabrii, O

    Klein, T., Lang, R. & Shkabrii, O. Understanding and improving hyperbolic deep reinforcement learning. Preprint at arXiv:2512.14202 (2025)

    arXiv 2025

  62. [62]

    & Schmidt-Thieme, L

    Bdeir, A., Burchert, N. & Schmidt-Thieme, L. Robust hyperbolic learning with curvature-aware optimization.Proc. AISTATS 20243421–3438 (2024)

    2024

  63. [63]

    & Zhihao, L

    Yang, K., Zhou, M. & Zhihao, L. Hyperbolic graph neural networks: a review.ACM Comput. Surv.55, 162:1–37 (2022)

    2022

  64. [64]

    Stochastic gradient descent on Riemannian manifolds.IEEE Trans

    Bonnabel, S. Stochastic gradient descent on Riemannian manifolds.IEEE Trans. Autom. Control58, 2217–2229 (2013)

    2013

  65. [65]

    & Ganea, O.-E

    Bécigneul, G. & Ganea, O.-E. Riemannian adaptive optimization methods.Int. Conf. Learn. Represent.(2019)

    2019

  66. [66]

    & Gerber, M

    Altschuler, J., Chewi, S. & Gerber, M. Averaging on the Bures–Wasserstein manifold: dimension-free convergence of gradient descent.Adv. Neural Inf. Process. Syst.34, 22132–22145 (2021)

    2021

  67. [67]

    & Ali, S

    Cafaro, C. & Ali, S. A. Jacobi fields on statistical manifolds of negative curvature.Phys. D: Nonlinear Phenom.234, 70–80 (2007)

    2007

  68. [68]

    Information-geometric indicators of chaos in Gaussian models.Phys

    Cafaro, C. Information-geometric indicators of chaos in Gaussian models.Phys. A: Stat. Mech. Appl.381, 93–106 (2007). 31

    2007

  69. [69]

    & Ranzato, M

    Lopez-Paz, D. & Ranzato, M. Gradient episodic memory for continual learning.Adv. Neural Inf. Process. Syst.30, 6467–6476 (2017)

    2017

  70. [70]

    & Elhoseiny, M

    Chaudhry, A., Ranzato, M., Rohrbach, M. & Elhoseiny, M. Efficient lifelong learning with A-GEM.Int. Conf. Learn. Represent.(2019)

    2019

  71. [71]

    & Hoiem, D

    Li, Z. & Hoiem, D. Learning without forgetting.IEEE Trans. Pattern Anal. Mach. Intell. 40, 2935–2947 (2018)

    2018

  72. [72]

    I., Kemker, R., Part, J

    Parisi, G. I., Kemker, R., Part, J. L., Kanan, C. & Wermter, S. Continual lifelong learning with neural networks: a review.Neural Netw.113, 54–71 (2019)

    2019

  73. [73]

    Baird, B.et al.Inspired by distraction: mind wandering facilitates creative incubation. Psychol. Sci.23, 1117–1122 (2012)

    2012

  74. [74]

    Sio, U. N. & Ormerod, T. C. Does incubation enhance problem solving? A meta-analytic review.Psychol. Bull.135, 94–120 (2009)

    2009

  75. [75]

    & Danihelka, I

    Graves, A., Wayne, G. & Danihelka, I. Neural Turing machines. Preprint at https://arxiv.org/abs/1410.5401 (2014)

    Pith/arXiv arXiv 2014

  76. [76]

    Nature538, 471–476 (2016)

    Graves, A.et al.Hybrid computing using a neural network with dynamic external memory. Nature538, 471–476 (2016)

    2016

  77. [77]

    39th Int

    Borgeaud, S.et al.Improving language models by retrieving from trillions of tokens.Proc. 39th Int. Conf. Mach. Learn.2206–2240 (PMLR, 2022)

    2022

  78. [78]

    & Turner, M.The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities(Basic Books, 2002)

    Fauconnier, G. & Turner, M.The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities(Basic Books, 2002)

    2002

  79. [79]

    M.Creativity in Context(Westview Press, 1996)

    Amabile, T. M.Creativity in Context(Westview Press, 1996)

    1996

  80. [80]

    ACL 202313387–13434 (2023)

    Perez, E.et al.Discovering language model behaviors with model-written evaluations.Proc. ACL 202313387–13434 (2023)

    2023

Showing first 80 references.