On the Generalization Limitation of On-Policy Distillation for Post-RL Consolidation
The paper studies multi-teacher on-policy distillation (MOPD) and on-policy distillation (OPD) for consolidating capabilities from teachers.
Source-based brief · Editor review not recorded · See checks and limits ↓
One primary source; other links are context, not independent corroboration.
- Primary sourceHugging Face
- Discovery linkHugging Face
- Linked in sourceView arXiv page
- Linked in sourceView PDF
The paper page is linked. Peer review and findings have not been independently checked by AINewsIQ.
Open full article view →- Original paper+45
- Fetched source page+30
- No editor link review+0
- Original non-discussion link+5
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
The paper studies multi-teacher on-policy distillation (MOPD) and on-policy distillation (OPD) for consolidating capabilities from teachers. Although in-distribution distillation is near-perfect, the authors find that out-of-distribution generalization is incomplete. Across extrapolative settings, OPD and MOPD reach 96.04-102.73% of teacher in-distribution scores, yet their OOD scores remain 2.45-16.25% below the base model. In compositional settings, MOPD's OOD performance on composed tasks stays 18.19-24.34% below the best teachers. The authors examine teacher capability and data coverage as bottlenecks, and propose geometry-based patching and base-preserving replay.
Claim explorer
Attribution, source and conditionsSource-stated points from the stored brief. Claim-level assessments have not been recorded.
Evaluations use Qwen3 models on selected math, coding, agent, and long-context tasks; the proposed patching methods are described as preliminary.
This connects a source, an attributed statement and its qualifications. It does not establish that the claim is correct.
Questions worth asking
Open questions for further reportingWhat remains unconfirmed?
Evaluations use Qwen3 models on selected math, coding, agent, and long-context tasks; the proposed patching methods are described as preliminary.
What material was available for this brief?
The recorded brief basis is fetched source page. These points are AINewsIQ paraphrases, not quotations or independent confirmations.
OPD and MOPD reach 96.04-102.73% of teacher in-distribution scores in extrapolative settings.
Source: Hugging Face ↗Stored source-based brief · key point 1 · AINewsIQ paraphraseTheir out-of-distribution scores remain 2.45-16.25% below the base model's scores.
Source: Hugging Face ↗Stored source-based brief · key point 2 · AINewsIQ paraphraseMOPD's composed-task OOD performance remains 18.19-24.34% below the best teachers in compositional settings.
Source: Hugging Face ↗Stored source-based brief · key point 3 · AINewsIQ paraphraseJointly alleviating teacher capability and data coverage bottlenecks eliminates OOD forgetting in the authors' experiments.
Source: Hugging Face ↗Stored source-based brief · key point 4 · AINewsIQ paraphraseLimits from the stored brief
Evaluations use Qwen3 models on selected math, coding, agent, and long-context tasks; the proposed patching methods are described as preliminary.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
Related archive coverage
These relationships record a citation or topic overlap, not confirmation of the finding.
The paper page is linked. Peer review and findings have not been independently checked by AINewsIQ.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
Brief history · 1 retained version
These are changes to AINewsIQ’s own publication. Earlier history is unavailable unless recorded here. Complete retained versions are available to the editor.
- Version 1 · published
Source-based brief stored or updated. This does not establish that the publisher changed its page.
Automated record
Source observations · 1 recent captures
A changed capture means the extracted text differed. It can reflect page furniture or extraction changes; it does not establish a substantive correction. Source captures remain private.
Retained capture; no earlier comparison available.
New engagement snapshots do not mean the source itself changed. Source observations and AINewsIQ brief revisions are separate records.
Why this may matter
The results suggest near-perfect in-distribution distillation scores can hide failures to transfer generalization, which matters for consolidating capabilities after reinforcement learning.
Limits and open questions
Evaluations use Qwen3 models on selected math, coding, agent, and long-context tasks; the proposed patching methods are described as preliminary.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Join the discussion
Have a different perspective or a question about the evidence? Discussion is free with a verified email account. Every comment is reviewed before publication.
No published comments yet.
New commenters: 2 submissions per 24 hours, at least 5 minutes apart, until 3 are approved. Discussion guidelines and moderation →