Persona-Pruner: framework for sculpting lightweight persona-specific LLM sub-networks via structured pruning
Persona-Pruner is a pruning framework that isolates persona-specific sub-networks from a generalist language model given only a character description, producing lightweight role-playing models without the full model's computational cost. The authors observe that naive pruning degrades role-playing fidelity by failing to distinguish redundant knowledge from character-essential parameters. On RoleBench, Persona-Pruner reduces performance drop by up to 93.8% relative to the strongest baseline pruning method while preserving general LLM capabilities. The work targets practical deployment scenarios such as game ecosystems with many simultaneous NPC agents.
Related guides (2)
Related events (8)
NPC-Playground: A 3D Environment for LLM-Powered Non-Player Characters
Hugging Face, Gigax, and Cubzh have introduced NPC-Playground, a 3D interactive environment where users can interact with non-player characters powered by large language models. The project demonstrates real-time LLM inference applied to game NPC behavior and dialogue. This represents a practical application of LLMs in interactive entertainment and agent-like character simulation.
Persona Generators: Evolutionary LLM Method for Diverse Synthetic Human Personas
Google researchers Davide Paglieri, Logan Cross, and colleagues propose Persona Generators, a system that uses the AlphaEvolve evolutionary algorithm to generate code that produces 25 diverse persona prompts covering a broad range of attitudes and opinions. The method iteratively optimizes persona prompt diversity using six metrics, outperforming Nemotron Personas (82% vs 76% coverage of possible responses) and a Concordia memory-based baseline (46%). The system uses Gemini 2.5 Pro for questionnaire generation and Gemma 3-27B-IT for persona simulation via the Concordia agent library. The approach reframes persona generation as a coverage optimization problem rather than a data-matching one, enabling more representative synthetic user populations for product research.
Tracing the Emergence of Human-Like Pragmatic Reasoning in LLMs Across Languages
Researchers conducted a population-matching experiment evaluating 25 LLMs on conditional inference tasks across four languages, comparing model behavior to matched human populations. The study finds that LLMs function as accurate semantic operators but systematically fail to capture pragmatic enrichments—context-sensitive inferences beyond literal logical meaning—that humans apply effortlessly. Model performance on pragmatic reasoning is not predicted by open vs. closed weights, training orientation, or architecture type, suggesting pragmatic reasoning remains an emergent and unreliable capability. The findings contribute to ongoing debates about whether LLMs reason like humans or merely approximate surface-level linguistic patterns.
Agentopia: Long-term multi-agent life simulation framework for training LLMs on social behavior
Researchers introduce Agentopia, a framework for simulating 10 years of social life across 100 LLM-powered agents, enabling study of emergent social behaviors and long-term personal growth dynamics. The system defines a 'life reward' metric mirroring human well-being and uses it to train LLMs via rejection sampling. Training on simulated social experience yields a +15.6% improvement on downstream role-playing benchmarks, suggesting that synthetic social simulation can generalize to real capability gains.
Activation Capping Technique Stabilizes LLM Assistant Personas Against Drift and Jailbreaks
Researchers from MATS, Oxford, and Anthropic introduced the 'assistant axis,' a vector derived from LLM layer outputs that quantifies how closely a model adheres to its trained assistant persona. They developed 'activation capping,' an inference-time method that corrects deviations from this axis when similarity falls below a threshold. Testing on Gemma 2 27B, Qwen3 32B, and Llama 3.3 70B showed harmful response rates to jailbreak prompts dropped by roughly half (e.g., 83% to 41% for Qwen3 32B) without degrading benchmark performance. The technique targets character-based jailbreaks that bypass system prompts by manipulating a model's internal representational state.
Parameter-Efficient Fine-Tuning using 🤗 PEFT
Hugging Face introduces the PEFT library, which enables parameter-efficient fine-tuning of large language models using techniques such as LoRA, prefix tuning, and prompt tuning. The library allows practitioners to adapt large pretrained models to downstream tasks while updating only a small fraction of model parameters, dramatically reducing compute and memory requirements. This lowers the barrier to fine-tuning frontier-scale models on consumer hardware.
RePro: Retrospective Progress-Aware Self-Refinement for LLM Agent Training
Researchers introduce RePro (Retrospective Progress-Aware Training), a framework addressing the gap between step-wise RL optimization and metacognitive task-progress awareness in LLM agents. The approach uses a forward-then-reflect rollout paradigm where agents execute actions online and then retrospectively assess step-wise progress given the completed trajectory and known outcome. Evaluated on WebShop, ALFWorld, and Sokoban, RePro achieves up to 12% absolute success rate gains over baseline Qwen-family models without requiring continuous external supervision.
On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters
This paper reframes parameter-efficient fine-tuning (PEFT) not merely as a cheaper alternative to full fine-tuning, but as a substrate for persistent, instance-specific personal models layered atop shared foundation models. The authors analyze three scaling axes: Scale Up (stronger base models amplifying adapter utility), Scale Down (minimum viable adapter size), and Scale Out (managing millions of concurrent adapted instances). They introduce MinT as an infrastructure reference for adapter identity, versioning, provenance, evaluation, and serving at scale.