analiza aiqbench
This research paper investigates catastrophic forgetting in Large Language Models during sequential fine-tuning, offering a mechanistic analysis of how parameter updates degrade ancestral capabilities across both closed-source and open-weight architectures. By identifying specific neural circuits prone to overwriting—such as early-layer attention heads—the authors propose Low-Rank Circuit Projection (LRCP) to mitigate this issue while maintaining adaptation velocity. The study provides critical insights for developers aiming to stabilize continual learning workflows without compromising foundational model knowledge.
Mechanistic Analysis of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning
arXiv:2601.18699v2 Announce Type: replace-cross Abstract: Sequential fine-tuning of Large Language Models (LLMs) adaptation to target tasks often triggers catastrophic forgetting, where the acquisition of novel target skills degrades ancestral capabilities. This paper presents a systematic comparative study of catastrophic forgetting across twenty premier models representing the state-of-the-art in mid-2026. We categorize our investigation into two primary research lines: (i) a behavioral and semantic output drift analysis of ten leading closed-source models (including Claude Fable 5, GPT-5.5 High, and Gemini 3.5 Flash), and (ii) a deep mechanistic interpretation of ten prominent open-weight architectures (such as DeepSeek-V4-Pro, Llama 4 Maverick, and Qwen 3.6-27B). Through weight-space trajectory tracking, Centered Kernel Alignment (CKA), and routing gate drift calculations in Mixture-of-Experts (MoE) layers, we localize the neural circuits highly susceptible to parameter overwriting. Our findings indicate that early-layer attention heads exhibit systemic entropic dispersion, while mid-to-deep feed-forward networks (or sparse expert blocks) suffer localized representation collapse. Informed by these insights, we introduce Low-Rank Circuit Projection (LRCP), a subspace-regularized training intervention. Empirical evaluations show that LRCP successfully mitigates up to 94.2% of ancestral capabilities in open-weight configurations and matches the adaptation velocity of standard PEFT baselines.
Abstract:Sequential fine-tuning of Large Language Models (LLMs) adaptation to target tasks often triggers catastrophic forgetting, where the acquisition of novel target skills degrades ancestral capabilities. This paper presents a systematic comparative study of catastrophic forgetting across twenty premier models representing the state-of-the-art in mid-2026. We categorize our investigation into two primary research lines: (i) a behavioral and semantic output drift analysis of ten leading closed-source models (including Claude Fable 5, GPT-5.5 High, and Gemini 3.5 Flash), and (ii) a deep mechanistic interpretation of ten prominent open-weight architectures (such as DeepSeek-V4-Pro, Llama 4 Maverick, and Qwen 3.6-27B). Through weight-space trajectory tracking, Centered Kernel Alignment (CKA), and routing gate drift calculations in Mixture-of-Experts (MoE) layers, we localize the neural circuits highly susceptible to parameter overwriting. Our findings indicate that early-layer attention heads exhibit systemic entropic dispersion, while mid-to-deep feed-forward networks (or sparse expert blocks) suffer localized representation collapse. Informed by these insights, we introduce Low-Rank Circuit Projection (LRCP), a subspace-regularized training intervention. Empirical evaluations show that LRCP successfully mitigates up to 94.2% of ancestral capabilities in open-weight configurations and matches the adaptation velocity of standard PEFT baselines.
12 pages, 8 figures, 5 tables. Preprint submitted to Elsevier
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2601.18699 [cs.LG]
(or arXiv:2601.18699v2 [cs.LG] for this version)
Text integral sub licența deschisă a publicației · 210 cuvinte
Subiecte conexe
[AINews] High Taste Testers Ep. 1
Latent Space · 17 oct. 2026
A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks
arXiv · 8 oct. 2026
Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?
arXiv · 9 iun. 2026
SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion
arXiv · 6 iun. 2026