Tag
The paper introduces Memorilla, a memory system that compresses an entire document collection into 16 query-dependent memory tokens for frozen LLMs, outperforming full-context inference on FactKG (84.33% vs 80.43%) while using 108x fewer tokens and beating baselines like Mem0 and MemoRAG across five of eight benchmarks.
Periscope introduces a training-free inference method that arranges text chunks on a K×K grid and queries a frozen language model over local and strided spans to recover answers beyond the context window, matching or beating full-window reads on LongBench v2 and InfiniteBench at a fraction of the cost. The approach lets a 27B model read 4.5M-token contexts on a single 80GB GPU by replacing long reads with an evidence map.
AdviSD proposes a method where a small trainable advisor steers a frozen frontier LLM executor via targeted multi-turn self-distillation, pairing outcome-based RL with feedback-conditioned self-distillation that selectively uses corrections. Experiments with Qwen3-8B advisors guiding Gemini and Claude show 4-6 point gains on BFCL-v3 and EnvScaler with good out-of-domain generalization.
LARA is a research project and PyTorch library that enables modular, composable behaviors for frozen large language models using low-rank residual adapters, allowing efficient training and inference-time blending of multiple behaviors.
The paper demonstrates that prompt-space meta-learning for personalizing frozen large language models does not transfer across users, as the meta-validation objective is statistically invariant to user-support correspondence, leading to no significant improvement over seed prompts or controls.
The paper audits frozen LLMs to examine how geometric constraint information is encoded and whether it can be used for generation, influence, and steering, revealing gaps between decodability and actionable outputs.
TokenMem injects knowledge into frozen LLMs via a dedicated cross-attention channel, training a thin gating adapter through two-phase curriculum to improve knowledge compliance under counterfactual knowledge, achieving 69-70% KC compared to 20-52% for vanilla RAG.
Proprioceptive AI released Cygnus, a tool that equips frozen LLMs with self-sensing adapters reading internal hidden states via gl(4,R) Lie algebra to isolate dark modes, boosting Qwen-32B's ARC-Challenge score from 82.2% to 94.97% on a single RTX 3090 without retraining.