← Back to feed

Latent Compaction with Test-Time Training

L5 · ResearcherResearcharXiv· 10/14/2026

Novel compression technique for long-context models that preserves performance while reducing memory and computation costs.

AI Summary

Researchers introduce Latent Compaction, a test-time method that compresses KV caches 32× while maintaining 96.7% accuracy on retrieval tasks.

Excerpt

Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact document-specific cache through test-time training. We initialize cache entries using reconstruction importance, teacher attention, and key-space diversity. Training first fits attention outputs and log-normalizers to those of the full cache, with reconstructio

Read Original
0 upvotes · 0 downvotes · 1 min read

Related Articles

L5 · ResearcherResearcharXiv
Price of Anarchy of First-Price Auctions in Agentic Markets

Researchers prove first-price auctions in AI agent markets can achieve near-optimal welfare under competitive floors and conditional independence assumptions.

L5 · ResearcherResearcharXiv
Your Transformer Is Secretly Elastic

Researchers introduce a retraining-free method to make transformers elastic, enabling on-the-fly adaptation across computational budgets by jointly ranking MLP neurons and attention heads.

L5 · ResearcherResearcharXiv
Co-Evolving Acting and Forecasting for Language Agents

Researchers propose Co-Evolving Acting and Forecasting (name), a method that improves language agent training efficiency by adding multi-step action targets to standard RL approaches.

L5 · ResearcherResearcharXiv
Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance

Researchers propose HiRe-MoMa, a hierarchical residual learning framework that enables real-world robot fine-tuning with only 30 minutes of human-guided interaction.

L5 · ResearcherResearcharXiv
Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion

Researchers developed a diffusion model that generates functional humanoid robot control networks directly from weight distributions, revealing consistent kinematic patterns across different training methods.