Self-Distillation Fine-Tuning resurfaces as alternative to SFT for continual learning

An MIT-area paper from January proposes using a demonstration-conditioned model as its own teacher — drawing fresh attention this week as a middle path between supervised fine-tuning and reinforcement learning for foundation-model adaptation.

A January 2026 arXiv paper by Idan Shenfeld, Mehul Damani, Jonas Hübotter and Pulkit Agrawal re-entered Hacker News discussion this week, prompted by renewed practitioner interest in continual-learning techniques for production LLMs ^[1]. The paper proposes Self-Distillation Fine-Tuning (SDFT), described as "a simple method that enables on-policy learning directly from demonstrations" ^[1].

The motivating problem is one production teams hit early: foundation models need to acquire new skills without losing previously-acquired ones, but the standard tool — supervised fine-tuning on examples — is "inherently off-policy" and prone to catastrophic forgetting ^[1]. The alternative, on-policy reinforcement learning, reduces forgetting but "requires explicit reward functions that are often unavailable" ^[1] for many real-world tasks where you have demonstrations rather than scalar rewards.

SDFT's core trick is to leverage in-context learning. The model conditions on the demonstration and generates its own continuation; those continuations become the training signal. In effect, the model is its own teacher, but a teacher that has been temporarily upgraded by being shown the demonstration in-context. The authors describe this as preserving "prior capabilities while acquiring new skills" ^[1].

The intuition is appealing: every training step uses outputs the model itself produced, so there's no distribution shift between training and inference. SFT, by contrast, trains on outputs the model would never naturally produce, then asks it to generalise to its own distribution at inference time.

The abstract does not include the specific benchmarks or numerical results, so practitioners interested in the size of the gain over SFT will need to read the full paper. Likewise, the method's compute footprint relative to plain SFT is not stated in the available content.

The reason this matters in May, four months after publication, is the renewed market focus on production LLM customisation: post-training to incorporate new tools, codebases, or domain knowledge without breaking existing capability is now a core enterprise requirement. SDFT joins a small group of methods proposing routes around the SFT-vs-RL false dichotomy.