paper-with-me

홈 › Papers

Hidden Failure Modes of Gradient Modification under Adam in Continual Learning, and Adaptive Decoupled Moment Routing as a Repair

2026-04-24 · Yuelin Hu, Zhenbo Yu, Zhengxue Cheng, Wei Liu, Li Song arxiv

Many continual-learning methods modify gradients upstream (e.g., projection, penalty rescaling, replay mixing) while treating Adam as a neutral backend. We show this composition has a hidden failure mode. In a high-overlap, non-adaptive 8-domain continual LM, all shared-routing projection baselines collapse close to vanilla forgetting (12.5--12.8 vs. 13.2). A 0.5% replay buffer is the strongest shared alternative but still reaches 11.6, while fixed-strength decoupling falls below vanilla at 14.1. Only adaptive decoupled routing remains stable at 9.4, improving over vanilla by 3.8 units. On a 16-domain stream, its gain over the strongest shared-routing projection baseline grows to 4.5--4.8 units. The failure is largely invisible on clean benchmarks. We explain this effect through Adam's second-moment pathway: in the tested regime, projection induces a 1/(1-alpha) inflation of the old-direction effective learning rate, matching measurements within 8% across eight alpha values. The same conflict appears with penalty methods, replay mixing, and at 7B scale under LoRA. Our fix routes the modified gradient only to the first moment while preserving magnitude-faithful second-moment statistics, with overlap-aware adaptive strength. This simple change is the only tested configuration that consistently avoids collapse across methods, optimizers, and scale.

📄 PDF Abstract BibTeX arXiv:2604.22407

Code (0)

등록된 구현이 없습니다.

Tasks

Continual Learning

Similar Papers 제목 키워드 기반

Understanding the Failure Modes of Out-of-Distribution Generalization

2020-10-29 · ICLR 2021 1 · Vaishnavh Nagarajan, Anders Andreassen, Behnam Neyshabur

Empirical studies suggest that machine learning models often rely on features, such as the background, that may be spuriously correlated with the label only during training time, resulting in poor accuracy during test-ti…

image-classificationImage ClassificationOut-of-Distribution Generalization

AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems

2026-09-08 · Jaewon Chu, Jinwoo Seo, Jaewon Cho, Jeehye Na 외 hf

Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimizatio…

Weight-Adjusted Gradients Reveal Parameter Importance and Failure Modes in LLMs

2026-07-12 · Shrestha Datta, Hongfu Liu, Anshuman Chhabra arxiv

Understanding which parameters are influential in Large Language Models (LLMs) is central to improving their efficiency, reliability, and interpretability. We introduce Weight-Adjusted Gradients (WAG), a simple yet effec…

knowledge editing

CausalT5k: Diagnosing Refusal and Failure Modes in Trustworthy Causal Reasoning Across Causal Rungs

2026-02-09 · Longling Geng, Andy Ouyang, Theodore Wu, Daphne Barretto 외 arxiv

Large language models increasingly produce fluent causal explanations, yet they often fail in ways aggregate accuracy cannot diagnose: confusing association with intervention, abandoning correct judgments under pressure,…

Hidden Heroes and Gradient Bloats: Layer-Wise Redundancy Inverts Attribution in Transformers

2026-02-01 · Donald Ye arxiv

Gradient-based attribution is the workhorse of mechanistic interpretability, yet whether it reliably tracks causal importance at the component level remains largely untested. We causally evaluate this assumption across t…