CoRD bridges on-policy distillation's blind spot with coupled rollouts
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers propose CoRD, a distillation method that addresses a fundamental gap in on-policy student training: the teacher explores states the student never reaches, leaving those regions unsupervised. CoRD interpolates between student and teacher distributions via maximal coupling from speculative decoding, strategically injecting teacher tokens only where student and teacher diverge. This keeps interventions minimal and adaptive, reducing as the student converges. Tested across Qwen3 and Qwen3.5 models (0.6B to 4B parameters), CoRD improves pass@1 scores, offering a practical lever for scaling efficient model distillation without abandoning on-policy training's sample efficiency.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
CoRD doesn't just patch on-policy distillation; it reframes the core problem as a distribution mismatch rather than a supervision gap. The key insight is using speculative decoding's coupling mechanism to interpolate between student and teacher trajectories, which is a novel application of an inference technique to a training problem.
This directly addresses the blind spot identified in the November 1st evidence bypass paper, where on-policy distillation leaves teacher-explored regions unsupervised. CoRD's coupled rollouts ensure the student encounters the same decision points as the teacher, reducing the evidence bypass risk by keeping interventions targeted to divergence points. It also echoes the reachability alignment principle from the October 31st self-distillation work, but inverts the approach: instead of constraining the teacher to the student's reach, CoRD strategically expands what the student sees without forcing it into unreachable territory.
If Qwen releases distilled models trained with CoRD that maintain evidence fidelity on the corrupted-document tests from the November 1st paper, that confirms the method actually solves the robustness problem rather than just improving pass@1 on clean benchmarks. Otherwise, CoRD may only address the distribution gap without fixing the reasoning grounding issue.
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.CL
Distilled models learn to ignore evidence, study finds
Researchers have identified a critical failure mode in on-policy distillation where student models learn to ignore supporting evidence and retain answers even when documents are altered to contradict them. Testing on scientific QA benchmarks shows distilled models achieve higher accuracy on original documents but suffer significant degradation when evidence shifts, with evidence bypass increasing by…
MentionsQwen3 · Qwen3.5 · CoRD
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. arXiv cs.CL originally reported this story as “CoRD: Coupled Rollouts Fill the Blind Spot of On-Policy Distillation”. 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.