The paper demonstrates that off‑policy merging, called grafting, outperforms on‑policy self‑distillation for continual learning. Grafting learns updates on an earlier donor checkpoint, scales the update, and optionally masks sensitive directions, thereby reducing interference with existing capabilities. Across various continual learning scenarios, grafting achieves better new‑task and old‑task performance without costly on‑policy sampling.
The paper investigates how to add new skills to large multimodal models (LMMs) without erasing previously learned abilities. By sequentially fine‑tuning on five target skills and tracking performance on eight held‑out benchmarks across three model families, the authors discover that forgetting on one skill can partially recover when the model is later tuned on a different skill. They attribute this to shifts in the output token distribution and propose two simple, selective tuning recipes—updating only self‑attention projection layers or only the MLP Gate&Up while freezing the Down projection—that achieve a better learning‑forgiveness trade‑off than full‑LLM tuning or other common forgetting‑mitigation methods.
whyItMatters":"The study shows that carefully choosing which model components to fine‑tune can enable LMMs to acquire new skills while preserving prior knowledge, offering a simpler and more effective alternative to existing forgetting‑mitigation techniques."
By Zhen Zhu, Yiming Gong, Yao Xiao, Yaoyao Liu, Derek Hoiem
arXiv:2609.38142v1 Announce Type: new
Abstract: A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advi...
By Rishabh Agrawal, Hejie Cui, Shasha Li, Shanchan Wu, Sercan \"{O}. Ar{\i}k
arXiv:2606. 25449v1 Announce Type: cross Abstract: A language model's memory can be worse than having no memory at all.
By Alex Kwon
A language model's memory can be worse than having no memory at all. Give a model a memory that kept a wrong conclusion but dropped the work behind it, and it emits that stale value as a confident answer; give the same model an empty memory and it abstains.
arXiv:2607. 15587v1 Announce Type: new Abstract: Continual learning studies how deployed language models can continually acquire new tasks without expensive retraining from scratch.
By Yang Meng, Zhenya Liu, Zhuokai Zhao, Yuxin Chen
The paper introduces Skill Schema Transport (SST), a post‑training framework that mitigates catastrophic forgetting in continual domain adaptation by translating new domain knowledge into a model’s existing skill space. SST abstracts each example type into a reusable skill schema, preserving essential operations while removing domain‑specific surface forms, thereby turning knowledge updates into procedural reuse rather than trajectory memorization. Experiments on two 3B instruction‑tuned backbones show that SST consistently reduces forgetting and improves generalization, achieving up to a 5.75‑point performance gain and a dramatic drop in maximum forgetting from 42.41 to 7.10 points on the Qwen‑3B MATH stream.
By Zijie Liu, Zhen Tan, Charles Fleming, Tianlong Chen
arXiv:2610.27467v1 Announce Type: cross
Abstract: On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilit...
By Shripad V. Deshmukh, Christian Bock, Kangrui Ruan, Abhijeet Awasthi, Vijay Lingam, Rein Houthooft, Manish Prajapat
The paper identifies a specific issue in supervised fine‑tuning (SFT) of large language models called factual access failure, where models can recognize correct facts under constrained tests but fail to generate them in open‑ended settings. It demonstrates that SFT can cause both genuine wrong answers and expression‑level errors such as verbosity or formatting mismatches. To mitigate this, the authors propose Recall‑Anchored Distillation (RAD), a self‑distillation method that aligns the fine‑tuned model with the base model’s soft output distribution on unlabeled out‑of‑distribution text, thereby recovering lost factual recall without needing labeled data.
By Haodong Chen, Yadong Wang, Shengtao Wen, Dong Liang, Xiang Chen
arXiv:2609.36375v1 Announce Type: new
Abstract: Continual learning is usually studied through mechanisms that preserve old knowledge. We develop Successional Learning Theory (SLT), a mesoscopic accou...
By Shoaib Ahmed Dipu, Md Salman Shamil, Sayeed Shafayet Chowdhury
The paper introduces Update‑Matched Unlearning (UMU), a method that applies the same simulated relearning update to both a defended model and an undefended reference to better preserve plasticity while resisting relearning of deleted knowledge. Experiments on Llama and Qwen models with TOFU and FaithUn datasets show that UMU reduces harmful and benign relearning by 30.4 and 9.9 points on average and increases new‑task accuracy by 3.5 points, supporting the reference‑timing hypothesis. The study demonstrates that robust unlearning should selectively prevent recovery of deleted knowledge without suppressing the model’s ability to learn new tasks.
By Jie Li, Youming Tao, Bowei He, Meng Ding
Steering vectors (SVs) are residual-stream directions that can shift LLM behavior when added to hidden states during inference. This study investigates whether the same SVs can guide parameter updates during post‑training, exploring supervised learning (SFT and distillation) and reinforcement learning with verifiable rewards (RLVR). The authors find that SV injection improves performance over plain SFT and distillation, achieving up to a 12% gain in certain distillation settings, and that applying a correctness‑contrast direction as advantage repair in RL recovers learning signals for held‑out math benchmarks. They also identify failure boundaries related to code generation, cross‑family layer transfer, and inference‑time application on SV‑trained checkpoints.
By Tiejin Chen, Maunil Vyas, Huaiyuan Yao, Hua Wei