The paper investigates why supervised fine-tuning (SFT) causes more forgetting than reinforcement fine-tuning (RFT) on classification tasks where tokens share a semantic class but differ in style. By deriving a linear-softmax policy decomposition, the authors show that SFT and RFT have parallel semantic updates but diverge in style dynamics, with SFT exhibiting off‑axis style drift that leads to a positive lower bound on forgetting. Theoretical results are supported by simulations over task sequences.
arXiv:2510. 18874v3 Announce Type: replace Abstract: Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting.
The paper introduces CIG‑SFT, a supervised fine‑tuning method that uses conditional information gain to evaluate how much each demonstrated token reduces uncertainty about the task. By comparing token likelihoods from a frozen reference model with and without task input, the method derives a KL‑regularized projection of the current policy, yielding a scalar‑weighted SFT objective whose gradients match the projected target. Experiments on mathematical reasoning and code‑generation benchmarks show that CIG‑SFT outperforms standard SFT and adaptive fine‑tuning baselines, especially in generalizing to challenging evaluation distributions, without requiring extra data or online rollouts.
arXiv:2606. 11709v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with the distribution it produces under privileged context, typically a verified solution.
The study investigates how language models acquire in-context learning (ICL) by distinguishing between copying a nearby demonstration and integrating information across a prompt. Across 17 models, roughly one-third default to copied answers, a few are undecided, and the remainder favor rule-consistent answers that yield higher accuracy and are guided by the requested value rather than the nearest conflicting example. Analyses show that models prioritize task-relevant prompt tokens before accuracy surpasses chance, and that synthetic or fine-tuned training can shift models toward compositional responses or improve accuracy when prompts reassign label meanings.
arXiv:2606. 11206v1 Announce Type: cross Abstract: Supervised Fine-Tuning (SFT) is the predominant paradigm for aligning large language models (LLMs), yet it suffers from optimization instability and limited generalization.
arXiv:2610.09835v1 Announce Type: new Abstract: Continual pre-training and fine-tuning in Large Language Models (LLMs) inevitably induce catastrophic forgetting, typically mitigated by replay using o...
arXiv:2609.24646v1 Announce Type: new Abstract: On-policy self-distillation fine-tuning (SDFT) learns new skills from demonstrations while reducing forgetting, but it always distils toward the full d...
On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with the distribution it produces under privileged context, typically a verified solution. However, we show that the learning signal drawn from this distributional gap concentrates on style tokens rather than task-bearing ones, as the hinted model tends to produce more direct, shorter outputs.
arXiv:2605.30251v2 Announce Type: replace-cross Abstract: Large language models (LLMs) often solve a task when all instructions are given in a single prompt, but fail when the same information is rev...
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.
Draft-OPD introduces an on‑policy distillation method for speculative draft models, addressing the mismatch between supervised fine‑tuning and inference by letting the target model supervise the drafter on draft‑induced states. The approach uses target‑assisted rollouts for stable continuations and replays drafting from error positions exposed during verification, enabling the drafter to learn from both accepted and rejected proposals. Experiments demonstrate that Draft‑OPD achieves more than five‑fold lossless acceleration across diverse tasks, outperforming prior draft models such as EAGLE‑3 and DFlash by 23 % and 13 % respectively.
arXiv:2607. 18293v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) teaches large language models new skills through a teacher that shares the student's backbone and supervises its own rollouts.