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

2026-07-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Software Development & Engineering · Depth: Expert, extended

Summary

Three small language models, ranging from 146 M to 3 B parameters, demonstrate emergent "creativity," "honesty," and "designed forgetting" by leveraging a hyperbolic substrate. The S3 Creative model (3 B parameters) achieved 100% preference in 311 pairwise comparisons for divergent question generation, utilizing geometry-aware Farthest-Point Sampling. The 146 M-parameter BS behavioural auditor detected compliance gaps with 90.7% binary accuracy, a 7.6-fold reliability increase over human raters (Cohen κ = 0.566 vs Fleiss κ = 0.074), and identified companion-induced traits with an AUROC of 0.804. The Lifelong Selective Memory Operating System (LSM-OS), built on a 2.4 B EXAONE 3.5 base, implements selective forgetting, preserving "skeleton" memories (~60% recall by day 90) while shedding "wallpaper" memories (0% recall by day 14) via an exponential decay law M(t)=S·exp(-λ t). This combined ~5.5 B parameter system is projected to deploy on 2026 edge hardware like the Jetson Orin NX, Galaxy S26, and iPhone 17 Pro, achieving sub-100 ms per-token latency.

Key takeaway

For AI Architects designing companion AI or deploying LLMs on edge devices, this research indicates that hyperbolic geometry is critical. You can achieve advanced traits like creativity, honesty, and selective memory in small models (~5.5 B parameters total) that fit mobile hardware, rather than relying solely on large, cloud-based Euclidean models. Consider integrating hyperbolic substrates and a three-tier architecture to enable privacy-preserving, low-latency, and sovereign on-device AI companions.

Key insights

Hyperbolic geometry enables small language models to achieve human-like creativity, honesty, and selective memory for companion AI.

Principles

• Hyperbolic geometry supports hierarchical data structures better than Euclidean.
• Small models on curved manifolds can match frontier LLMs in specific tasks.
• Designed forgetting is crucial for lifelong memory systems.

Method

The Personal Assistant–Cloud Operating System (PACOS) integrates three hyperbolic models: S3 Creative for frame-seeding, BS Auditor for honesty checks, and LSM-OS for designed forgetting, communicating via a shared manifold.

In practice

• Use hyperbolic embeddings for hierarchical data or creative generation.
• Implement geodesic distance for compliance auditing in LLMs.
• Design memory systems with selective forgetting for long-term companions.

Topics

• Hyperbolic Geometry
• Small Language Models
• Companion AI
• Edge AI Deployment
• Designed Forgetting
• AI Honesty

Best for: AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Creativity, honesty and designed forgetting emerge in small hyperbolic language models — Small hyperbolic language models can achieve creativity, honesty, and designed forgetting for trustworthy companion AI.
• Hyperbolic and Evidence-Prioritized Experts for Large Vision-Language Models — AsyMoE models modality asymmetry in LVLMs using specialized experts to improve performance and maintain contextual grounding.
• Scaling Down Is the New Scaling Up — Future agentic AI prioritizes hardware-aware design for efficient on-device operation over raw model size.
• Your AI Doesn’t Know Anything. And That’s Not the Model’s Fault. — AI hallucinations stem from infrastructure limitations, not model intelligence, necessitating robust knowledge layers like graphs.
• Your AI Doesn’t Actually Remember You — Perceived AI continuity is a user-system co-creation, not evidence of AI's persistent internal memory or human-like self.
• From Symbolic to Geometric: Enabling Spatial Reasoning in Large Language Models — Integrating geometric spatial representations directly into LLMs enables robust, true spatial reasoning beyond symbolic pattern matching.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.