Post-Training is About States, Not Tokens: A State Distribution View of SFT, RL, and On-Policy Distillation
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Researchers reframe post-training optimization around state distributions rather than loss functions alone, arguing that SFT, RL, and on-policy distillation differ fundamentally in which training states they sample from. Using Qwen 3 0.6B on GSM8K with retention checks on TruthfulQA and MMLU, the work identifies three key phenomena that challenge conventional loss-centric analysis. This perspective shift matters for practitioners tuning alignment pipelines: it suggests that which prompts and prefixes a model learns from during post-training may be as consequential as the objective itself, potentially unlocking more efficient fine-tuning strategies and better transfer across domains.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper's practical provocation is that two fine-tuning runs using identical objectives and identical data can produce different models simply because they sample different prefixes and intermediate states during training. That reframes debugging alignment failures as a data-coverage problem, not just a loss-design problem.
The distillation angle connects directly to our coverage of 'The Distillation Game' from the same day, which examined how student models exploit teacher outputs during on-policy learning. That piece treated distillation primarily as a security surface, but this paper adds a complementary lens: the states a student samples from a teacher determine what behavioral distribution it actually learns, independent of the objective. Together, the two papers suggest that on-policy distillation is doing more work than practitioners typically account for, both in terms of what knowledge transfers and what attack surface opens up. The manifold optimization paper from the same batch is not a natural connection here.
If a follow-up applies this state-distribution framing to a model larger than 0.6B parameters and shows the same three phenomena hold on GPQA or a held-out reasoning benchmark, the framework earns generalization claims. Results that only survive on GSM8K at small scale should be treated as preliminary.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
The Distillation Game: Adaptive Attacks & Efficient Defenses
Researchers formalize the core tension facing model providers: outputs that maximize user utility simultaneously enable efficient model theft through distillation. This work frames the problem as a strategic game between teacher and student, yielding a practical defense called Product-of-Experts that operates during generation without retraining. The findings expose a critical gap between passive and adaptive…
MentionsQwen 3 · GSM8K · TruthfulQA · MMLU · SFT · RL
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.