Post-training grounding gains tied to existing model structure, not new learning
A systematic audit of post-training methods reveals that gains in grounding (making models respect prompt evidence over memorized knowledge) depend heavily on pre-existing model capabilities rather than novel learned machinery. Testing GRPO, SFT, and DPO across multiple scales shows GRPO produces minimal grounding improvements despite metric gains, while DPO approaches ceiling performance on its distribution. This finding matters for practitioners: it suggests post-training effectiveness is bottlenecked by foundational model properties, not algorithmic innovation alone, reshaping expectations for alignment and factuality improvements in production systems.
Modelwire context
ExplainerThe paper's core claim isn't just that post-training has limits, but that those limits are structural: grounding improvements require capabilities already present in the base model, meaning no amount of algorithmic refinement can overcome a weak foundation.
This connects directly to the BenchMIRT finding from Hugging Face that benchmarks often measure narrow performance rather than genuine capability. Here we see a concrete example: GRPO shows metric gains that don't translate to actual grounding behavior, suggesting the metrics themselves may be capturing optimization artifacts rather than real improvements. The StateSwap work on hidden state sensitivity also becomes relevant, since if framing effects operate through separable computational pathways, post-training methods may be tuning surface responses without addressing the underlying representational gaps that prevent genuine grounding.
If teams applying GRPO to weaker base models (7B or smaller) report grounding improvements that match their 70B+ results, this paper's claim about pre-existing machinery bottlenecks is falsified. Conversely, if grounding gains correlate tightly with base model scale across multiple labs over the next two quarters, the finding holds and practitioners should shift investment toward pre-training rather than post-training refinement.
Coverage we drew on
- BenchMIRT: What are LLM benchmarks actually measuring? · Hugging Face
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsGRPO · SFT · DPO
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Context-Grounding Gains Are Mediated by Pre-existing Machinery: Auditing GRPO, SFT, and DPO”. 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.