INQUIRING LINE

What makes policy self-distillation more effective than external teacher distillation?

This explores why a model often learns better from data it generates itself (policy self-distillation) than from traces handed down by a stronger external teacher — and where that advantage breaks down.


This explores why a model often learns better from its own generated traces than from a stronger external teacher's — and the corpus has a surprisingly consistent answer: the bottleneck isn't trace quality, it's compatibility with the student's own representations. SEAL is the clearest case — models learned better from synthetic data they restructured themselves than from data written by stronger external models, lifting QA accuracy from 33.5% to 47.0%, because self-generated data aligns with the learner's representational needs Does self-generated training data improve model learning?. The same logic explains why teacher-refined data can backfire: when refinements exceed the student's learning frontier, objectively higher-quality data degrades performance, and students do better filtering for what's compatible with their own statistical profile Does teacher-refined data always improve student model performance?.

The deeper mechanism is a distribution-mismatch problem. Training on offline teacher traces fails at self-correction because the teacher's errors aren't the student's errors — models only learn to fix mistakes by practicing on their actual error distribution under online RL Why does self-correction training on offline data fail?. So self-distillation wins partly because the policy is, by definition, training on its own distribution.

But there's a sharp catch the corpus insists on: self-distillation isn't free. Distilling from a confident teacher — or from yourself in a way that mimics one — suppresses the epistemic markers ('Wait,' 'Hmm') that flag flawed reasoning paths, trading out-of-distribution robustness for confident, concise in-domain answers Does self-distillation harm mathematical reasoning performance?. Richer teacher context makes this worse, not better: conditioning a teacher on correct answers and verifier output yields shorter, more confident student traces that generalize worse on problems requiring caution Does richer teacher context hurt student generalization?. The advantage of self-distillation, then, isn't that it's safe — it's that it keeps the student inside its own distribution rather than forcing a style it can't support.

The surprising twist is the opposite case, where the external teacher wins. Walmart's BERT cross-encoders outperformed the very LLM teachers that labeled their data, because the student saw a far broader input distribution smoothed by teacher predictions Can smaller models outperform their LLM teachers with enough data?. So the real variable isn't 'self vs. external' — it's distributional coverage and representational fit. When the teacher's labels broaden the student's exposure, external wins; when the teacher's style narrows or mismatches it, self-distillation wins.

Worth knowing where this bottoms out: pure self-improvement is formally bounded by the generation-verification gap — a model can't reliably exceed what it can verify, and every durable gain smuggles in an external anchor (a past checkpoint, a judge, a tool, a user correction) Can models reliably improve themselves without external feedback?, What stops large language models from improving themselves?. The methods that push self-distillation furthest sneak verification back in: asymmetric self-play uses a proposer-solver split with majority-vote checking Can language models improve themselves without any external training data?, and self-examining RL alternates actor and judge roles to manufacture an internal reward signal Can models learn to judge themselves without external rewards?. Self-distillation works best not when it's purely self-contained, but when the model is its own best-matched teacher and something still plays the role of verifier.


Sources 10 notes

Does self-generated training data improve model learning?

SEAL demonstrates that models learn better from synthetic data they generate themselves than from data created by stronger external models. Self-generated data improved QA performance from 33.5% to 47.0%, suggesting that model-specific restructuring aligns with the learner's representational needs.

Does teacher-refined data always improve student model performance?

Teacher-refined data degrades performance when it exceeds the student's learning frontier, even if objectively higher quality. Students should filter refinements using their own statistical profile to retain only compatible improvements.

Why does self-correction training on offline data fail?

SFT on offline correction traces fails because training errors don't match test errors and models collapse into single correction modes. Multi-turn online RL under the model's own error distribution successfully trains self-correction by letting models practice correcting their actual mistakes.

Does self-distillation harm mathematical reasoning performance?

Self-distillation reduces performance in mathematical reasoning by eliminating epistemic markers like "Wait" and "Hmm" tokens that flag flawed reasoning paths. These tokens enable self-correction on out-of-distribution problems, so removing them sacrifices robustness for confident brevity.

Does richer teacher context hurt student generalization?

Teachers conditioned on correct answers and verifier output produce confident, concise traces that students inherit. This style suppresses uncertainty expression, optimizing in-domain performance while degrading generalization to out-of-distribution problems that require epistemic caution.

Can smaller models outperform their LLM teachers with enough data?

Walmart's student cross-encoders outperformed their LLM teachers when trained on sufficiently large augmented datasets of teacher-labeled queries. The student's broader input distribution exposure, smoothed by teacher predictions, enabled better generalization than the teacher achieved.

Can models reliably improve themselves without external feedback?

Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.

What stops large language models from improving themselves?

Self-improvement in LLMs is formally bounded by the generation-verification gap, meaning every reliable fix requires something external to validate and enforce it. Models cannot escape this constraint through metacognition alone.

Can language models improve themselves without any external training data?

SQLM uses a proposer-solver framework where the proposer generates calibrated problems and the solver learns via majority-vote verification. Both agents improve through RL alone, creating an automatic curriculum that scales without human labels or ground-truth answers.

Can models learn to judge themselves without external rewards?

SERL enables self-improving language models by having them alternate between generating responses and judging them pairwise, deriving rewards from ranking consistency and self-consistency of judgments. On AlpacaEval, this reached 59.90% win rate without external signals, up from 52.37%.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are an LLM researcher evaluating whether the 'representational fit' and 'distribution mismatch' constraints on policy self-distillation remain binding in 2024–present models and tooling.

What a curated library found — and when (dated claims, not current truth):
Findings span 2024–2026 and center on representational compatibility, not trace quality:
• Self-generated training data outperforms external teacher data when aligned with the student's representational profile; SEAL lifted QA from 33.5% → 47.0% (2024).
• Teacher-refined data can degrade performance if refinements exceed the student's learning frontier; distributional mismatch is the bottleneck, not quality (2024).
• Self-distillation suppresses epistemic markers ('Wait', 'Hmm') that flag flawed reasoning, trading out-of-distribution robustness for confident in-domain answers (2026).
• Richer teacher context produces shorter, more confident student traces that generalize worse on problems requiring caution (2025).
• Pure self-improvement is formally bounded by the generation-verification gap; every durable gain anchors to external verification (past checkpoint, judge, tool, user correction) (2024–2026).

Anchor papers (verify; mind their dates):
• arXiv:2412.02674 (2024-12) — Self-improvement mirage; generation-verification gap.
• arXiv:2603.24472 (2026-03) — Why self-distillation degrades reasoning capability.
• arXiv:2508.06026 (2025-08) — Temporal self-rewarding; chosen-rejected decoupling.
• arXiv:2507.23751 (2025-07) — CoT-Self-Instruct for synthetic prompt quality.

Your task:
(1) RE-TEST EACH CONSTRAINT. For every distributional and representational mismatch claim above, assess whether recent scaling, fine-tuning methods (DPO, IPO), longer context windows, or multi-agent orchestration (e.g., proposer-verifier splits, ensemble sampling) have relaxed the student's representational ceiling or lowered the epistemic-marker cost. Separate durable questions (e.g., "Does distribution mismatch always matter?") from perishable limitations (e.g., "Can smaller models only learn from compatible data?").
(2) Surface the strongest work from the last 6 months that CONTRADICTS or SUPERSEDES the "self-distillation suppresses reasoning robustness" finding — look for evidence that epistemic markers can be preserved or that external verification can be internalized without style collapse.
(3) Propose 2 research questions assuming the regime may have shifted: one on whether multimodal or code-grounded contexts dissolve representational fit bottlenecks; one on whether calibration-aware self-distillation (beyond accuracy) recovers out-of-distribution robustness.

Cite arXiv IDs; flag anything you cannot ground in a real paper.

Next inquiring lines