A unified measurement pipeline on 27 LLM checkpoints shows activation maxima spanning four orders of magnitude, with MoE models 14-23x lower than matched dense models and residual streams carrying the global max in most cases.
gpt-oss-120b & gpt-oss-20b model card
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
citation-role summary
background 1
method 1
citation-polarity summary
years
2026 5representative citing papers
Joint Consistency casts test-time aggregation as Ising-type energy minimization with pairwise LLM-judge interactions, subsuming voting methods and outperforming baselines across reasoning tasks.
SOMA estimates a local response manifold from early turns and adapts a small surrogate model via divergence-maximizing prompts and localized LoRA fine-tuning for efficient multi-turn serving.
Systematic testing of prompt engineering for LLM equational reasoning finds a performance ceiling of 60-79% accuracy that extensive engineering cannot exceed, driven by undecidability and model capacity limits.
JoyAI-LLM Flash delivers a 48B MoE LLM with 2.7B active parameters per token via FiberPO RL and dense multi-token prediction, released with checkpoints on Hugging Face.
citing papers explorer
-
JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency
JoyAI-LLM Flash delivers a 48B MoE LLM with 2.7B active parameters per token via FiberPO RL and dense multi-token prediction, released with checkpoints on Hugging Face.