New Research Explains Emergent Capabilities and Plasticity Loss in LLMs

Lei Dong· July 31, 2026 View original

Key takeaways

  • LLM capabilities emerge from a "no-partial-credit" joint alignment of circuit components.
  • Plasticity loss can be mitigated by selectively re-initializing query-key slices.
  • Standard validation loss metrics may not detect the critical phase of capability formation.
  • Understanding these kinetics can lead to more efficient and controllable LLM training.

Who benefits

AI DevelopmentSoftware EngineeringResearch & AcademiaCloud Computing

Summary

This research proposes a "driven-nucleation rate law" to explain how new capabilities emerge in large language models, suggesting they appear when the final parts of a circuit align in a single stochastic attempt. It also identifies the mechanism behind plasticity loss and offers a method to restore learnability by re-initializing specific model components.

New research delves into the fundamental mechanisms governing the emergence of capabilities and the loss of plasticity in large language models. The study introduces a novel "driven-nucleation rate law," positing that a new capability materializes only when all necessary circuit components align simultaneously, rather than through partial credit. This "no-partial-credit" joint alignment is identified as the critical rate-limiting step in capability formation. The findings highlight two key fingerprints supporting this theory: the time taken for a capability to appear depends on the number of missing circuit parts, not the overall circuit size, and ablating a single part significantly diminishes the capability, contradicting partial credit predictions. The research also addresses plasticity loss, where models stop learning new capabilities despite continued training. It pinpoints the damage to specific head commitments to base data and demonstrates that re-initializing only the query-key slices can restore learnability, offering a targeted intervention for model maintenance.

Why it matters

Understanding the kinetics of capability emergence and plasticity loss is crucial for developing more efficient and robust large language models, enabling better control over their training and lifecycle.

How to implement this in your domain

  1. 1Investigate current LLM training logs for signs of "flat baseline" followed by sudden capability ignition, aligning with the proposed rate law.
  2. 2Develop diagnostic tools to identify specific "missing parts" in a model's circuit that are hindering capability emergence.
  3. 3Experiment with targeted re-initialization of query-key slices in models exhibiting plasticity loss to restore learnability.
  4. 4Implement new monitoring metrics during training that are sensitive to the "critical step" of capability formation, rather than just validation loss.

Original post by Lei Dong

"arXiv:2607.27281v1 Announce Type: new Abstract: A capability appears in a language model when the last parts of its circuit align in one stochastic attempt, and getting all but one right is worth nothing. We show this no-partial-credit joint alignment is the rate-limiting step of…"

View on X

Originally posted by Lei Dong on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Framework Improves Partial Multi-View Clustering Performance.

DAS-PMVC is a novel framework for partial multi-view clustering that addresses view asymmetry and irrelevant samples by leveraging dual alignment and structure enhancement. It uses anchor graph structure alignment, structure-enhanced feature learning, and a dual alignment strategy to achieve superior clustering performance on various datasets.

Shubin Ma, Liang Zhao, Chuanye He, Zhenjiao Liu, Liang Zou, Lin Yuanbo Wu, Yu ShaoJul 31, 2026
AI Engineering & DevToolsAI Research

Dual Teachers Improve Adversarial Robustness and Accuracy.

This work extends Information Bottleneck Distillation (IBD) by introducing a "clean teacher" alongside a robust teacher to improve the robustness/accuracy tradeoff against adversarial attacks. The proposed method transfers features from both teachers to a student model, achieving better clean accuracy while maintaining adversarial robustness, outperforming original IBD and competing with state-of-the-art approaches.

Vincent Ryusuke Takahashi, Yoshinari Takeishi, Jun'ichi Takeuchi, Kave SalamatianJul 31, 2026
AI Engineering & DevToolsAI Research

Dynamic Batch Sizes Improve Large Language Model Training Efficiency.

This paper proposes a new approach to deep learning dynamics, deriving joint scaling laws for loss based on both learning rate and batch size schedules. It introduces an optimal dynamic batch size schedule that consistently outperforms static batch size baselines, highlighting its importance for large language model training.

Jiaxiang Li, Zhiqi Bu, Shiyun XuJul 31, 2026