Self-Verified Distillation Overview

• Self-Verified Distillation is a training pattern where models generate candidate outputs, apply successive verification stages, and distill only validated signals for improved accuracy.
• It leverages diverse verification mechanisms—including cycle-consistency, binary rewards, and ASR round-trip evaluations—to ensure output reliability.
• Applications span language reasoning, reinforcement learning, and text-to-speech, with reported gains in pass@1 scores and robustness in various benchmarks.

Self-Verified Distillation denotes a family of procedures in which a model produces candidate outputs, diagnoses, filters, or reinterprets them through a verification mechanism, and then distills only the verified signal back into a student or into the same policy under a less privileged context. Recent work uses the term for several related but non-identical regimes: unlabeled post-training from self-curated synthetic data (Lee et al., 20 May 2026), on-policy reinforcement learning with feedback-conditioned self-teachers (Hübotter et al., 28 Jan 2026), privileged self-distillation on “cliff” prompts (Ding, 25 Mar 2026), skill-conditioned and reflection-guided variants for reasoning (Huang et al., 27 May 2026, Zhao et al., 27 May 2026), and even ASR-mediated reliability improvement in neural-codec text-to-speech (Asaria et al., 16 Jun 2026). The same period also produced a detailed negative result showing that judge-based “disposition” distillation into sub-2.3B LLMs can fail systematically, with apparent gains collapsing under stricter evaluation (Sadasivan, 13 Apr 2026).

1. Conceptual structure and variants

Across these papers, the recurring structure is candidate generation, verification, selection or polarity assignment, and a downstream distillation update. What differs is the source of the verification signal. In some settings the verifier is prompt-based and unanimous across multiple judge calls; in others it is a binary outcome reward, a textual error trace, a reflective diagnosis of the first erroneous span, or an ASR round-trip metric (Lee et al., 20 May 2026, Hübotter et al., 28 Jan 2026, Zhao et al., 27 May 2026, Asaria et al., 16 Jun 2026).

Variant	Verification signal	Distillation target
Unlabeled SVD	Three-stage cycle/factuality/correctness cascade	SFT on accepted self-generated answers
SDPO / HDPO	Rich feedback or binary reward on rollouts	Token-level KL or JSD into unprivileged policy
SGSD / ROSD	Outcome-validated teacher polarity or localized reflection	Gated or masked token-level distillation
ASR self-verification for TTS	Whisper round-trip catastrophic-failure filter	Single-shot robustness via SFT or DPO
Disposition distillation negative result	Judge-based disposition scoring	No clean gain without content damage

This breadth matters because “self-verified distillation” is not restricted to one mathematical objective. In some formulations, the teacher and student share weights and differ only in context, so the method is a form of in-context self-teaching rather than cross-model imitation. In others, the model first curates a synthetic dataset from unlabeled prompts and then performs conventional supervised fine-tuning. A plausible implication is that the term is best understood as a training pattern rather than a single algorithmic primitive.

2. Unlabeled self-training through validator cascades

A direct use of the term appears in "Self-Verified Distillation: Your LLM Is Secretly Its Own Synthetic Data Pipeline" (Lee et al., 20 May 2026). The setup starts from an unlabeled seed set of reasoning prompts $\mathcal{Q}=\{q_1,\dots,q_{|\mathcal{Q}|}\}$, a model $p_\theta(y\mid q)$, and sampled candidate solutions $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$. A self-curated dataset is then constructed as

$$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$

followed by standard SFT: $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$

Its verification mechanism is a three-stage cascade with $v$ repeated judge calls per stage. The stages are cycle-consistency, factuality, and correctness, with stage-wise score

$$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$

and acceptance only when $\mathrm{score}_s(q,y)=1$ for all three stages. The pipeline uses a “first-valid” selection policy: once a candidate passes all stages, that single $(q,y)$ pair is added to $\mathcal{D}_{\mathrm{sft}}$, and the search for that question stops. In the main 4B runs, the reported hyperparameters are $p_\theta(y\mid q)$0, $p_\theta(y\mid q)$1, learning rate $p_\theta(y\mid q)$2, and batch size $p_\theta(y\mid q)$3 (Lee et al., 20 May 2026).

The reported results are framed as held-out pass@1 improvements across math, science, and coding. For Qwen3-4B, aggregate gains are +16.7 points in math on AIME26 and HMMT, +11.1 points in science on GPQA Diamond and HLE, and +8.3 points in coding on LiveCodeBench v5 and v6; gains are also reported for 0.6B and 8B models (Lee et al., 20 May 2026). On the 4B model specifically, AIME26 rises from 59.3 to 69.3, HMMT from 39.3 to 46.0, GPQA Diamond from 50.8 to 60.4, HLE from 8.5 to 10.0, LCB v5 from 45.1 to 49.0, and LCB v6 from 37.5 to 41.9. Compared with the test-time-only UQ-TTC baseline, the method is reported to achieve better performance in most settings while requiring only a single inference call at test time (Lee et al., 20 May 2026).

The ablations emphasize that data quality depends jointly on the candidate count $p_\theta(y\mid q)$4 and the verification budget $p_\theta(y\mid q)$5. On AIME26, the reported gains are +4.4 points for $p_\theta(y\mid q)$6, +4.7 for $p_\theta(y\mid q)$7, and +10.0 for $p_\theta(y\mid q)$8. The paper also reports $p_\theta(y\mid q)$9 and $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$0 for the acceptance fraction, arguing that stronger verification yields a smaller but higher-quality accepted set (Lee et al., 20 May 2026).

3. Reinforcement learning and privileged self-distillation

A second line of work places self-verified distillation inside on-policy RL. "Reinforcement Learning via Self-Distillation" formalizes reinforcement learning with rich textual feedback and introduces Self-Distillation Policy Optimization (SDPO) (Hübotter et al., 28 Jan 2026). The student policy is $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$1, while the self-teacher is the same model conditioned on feedback: $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$2 The core objective is next-token KL distillation: $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$3 The paper shows that the gradient can be interpreted through a dense logit-level “advantage”

$$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$4

which replaces scalar reward-to-go with token-level credit (Hübotter et al., 28 Jan 2026).

SDPO is evaluated in scientific reasoning, tool use, and LiveCodeBench v6. The paper reports +4–7 percentage points higher final accuracy than a strong GRPO baseline, 3–10× faster time-to-accuracy, and responses that are 3× shorter on average. On LiveCodeBench v6, SDPO reaches 48.8% final pass@1 versus 41.2% for GRPO, and reaches GRPO’s final accuracy in 4× fewer generations (Hübotter et al., 28 Jan 2026). The gains are reported to be strongest on medium and hard questions, and the scaling study states that improvements grow with model size from Qwen3-0.6B to 8B.

"HDPO: Hybrid Distillation Policy Optimization via Privileged Self-Distillation" addresses a more specific RL pathology: prompts on which all rollouts fail, so the RL gradient vanishes entirely (Ding, 25 Mar 2026). HDPO augments GRPO with privileged self-distillation. On each mini-batch, it identifies “cliff” prompts where all $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$5 rollouts fail, appends the ground-truth solution $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$6 to the prompt, regenerates $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$7 teacher rollouts under the same weights, filters for correct solutions with $$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$8, and distills the teacher’s token-level distribution into the unprivileged student. The hybrid objective is

$$Y_q=\{y_{q,1},\dots,y_{q,N}\}\sim p_\theta(\cdot\mid q)$$9

where $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$0 is a token-averaged Jensen–Shannon divergence on filtered teacher traces, using a top-$$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$1 approximation with $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$2 (Ding, 25 Mar 2026).

Two aspects distinguish HDPO. First, teacher and student share weights and differ only in the appended privileged context, so the realizability gap is bounded by the change in input. Under a local Lipschitz assumption, the paper gives

$$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$3

Second, the paper proves that $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$4 filtered privileged generation recovers the optimal KL-regularized RL policy in the hard-threshold limit $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$5. Empirically, on OpenMathInstruct-2 with Qwen2.5-Math-1.5B-Instruct, HDPO improves coverage metrics while maintaining greedy accuracy, with pass@4 gains of +0.8–1.1% and pass@8 gains of +0.4–1.7% (Ding, 25 Mar 2026).

These two papers illustrate a core interpretation of self-verified distillation in RL: it is a mechanism for turning sparse or vanishing outcome signals into dense token-level updates without introducing a separate external teacher.

4. Gated and reflection-guided reasoning correction

Subsequent work refines the verification stage itself. "Skill-Conditioned Gated Self-Distillation for LLM Reasoning" proposes SGSD, which treats skill-based self-distillation as teacher hypothesis validation rather than unconditional imitation (Huang et al., 27 May 2026). A stop-gradient copy $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$6 serves as a $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$7-teacher pool, each teacher conditioned on a retrieved skill–mistake pair $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$8. After the student samples a rollout $$\mathcal{D}_{\mathrm{sft}} = \bigcup_{q\in\mathcal{Q}} \bigl\{(q,y)\colon y\in Y_q,\;y\text{ passes }\mathcal{F}\bigr\},$$9, a verifier returns $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$0, each teacher scores the same rollout, and token-level support gaps

$$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$1

are aggregated into a support score. After clipping, masking, and thresholding, the paper defines a polarity $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$2 so that teachers aligned with the outcome are distilled, misleading teachers are reversed, and uncertain teachers are ignored. The gated loss is

$$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$3

and the per-example objective is

$$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$4

The main empirical comparison on Qwen3-1.7B reports avg@12 of 37.4% for the base model, 37.5% for GRPO, 42.0% for OPSD, and 43.7% for SGSD across AIME24, AIME25, and HMMT25, corresponding to +6.2 over GRPO and +1.7 over OPSD (Huang et al., 27 May 2026). The robust support estimation ablation on AIME24 step-100 avg reports +3.3% over base for full SGSD, +3.1 without masking, +1.4 without clipping, +0.3 without thresholding, and $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$5 without all three. Removing outcome validation produces short-term gain but “catastrophic collapse later,” because misleading teachers are never reversed.

"ROSD: Reflective On-Policy Self-Distillation for LLM Reasoning across Domains" argues that standard OPSD overfits because a teacher conditioned on a full verified solution promotes reference-trajectory imitation and because full-response distillation can overwrite already-correct prefixes (Zhao et al., 27 May 2026). ROSD inserts a self-reflector that reads a rollout and produces a corrective idea $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$6 and, for failed trajectories, an error quote $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$7 marking the first erroneous span. If the quote is found at token index $$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$8, the mask is

$$\mathcal{L}_{\mathrm{SFT}}(\theta) = -\frac{1}{|\mathcal{D}_{\mathrm{sft}}|} \sum_{(q,y)\in\mathcal{D}_{\mathrm{sft}}} \log p_{\theta}(y\mid q).$$9

and the loss becomes

$v$0

Correct rollouts use $v$1, but the teacher is still conditioned on the reflection rather than the entire solution (Zhao et al., 27 May 2026).

The reported in-domain mean@16 results show ROSD at 72.83% on Qwen3-4B and 73.45% on Qwen3-8B, outperforming GRPO by 2.97 and 1.46 points and SDPO by 5.81 and 0.95 points, respectively (Zhao et al., 27 May 2026). Out-of-domain performance is where the framework is most sharply distinguished: the paper reports average OOD improvements of 10–15 points over SDPO across held-out tasks and both model scales. The ablations indicate that reflection and localization are complementary: “w/o Reflection” recovers much of the OOD gain, and “w/o Localization” still outperforms SDPO.

Taken together, SGSD and ROSD shift self-verified distillation away from unconditional self-imitation and toward selective correction. One validates a teacher’s stance against the outcome; the other restricts updates to the suffix beginning at the first diagnosed error.

5. Judge-based self-verification and a three-arc negative result

"Disposition Distillation at Small Scale: A Three-Arc Negative Result" provides the most explicit cautionary account in this literature (Sadasivan, 13 Apr 2026). The study set out to train behavioral dispositions—self-verification, uncertainty acknowledgment, and feedback integration—into small LLMs with 0.6B to 2.3B effective parameters through a four-stage all-MIT distillation pipeline. For each prompt $v$2, the stages are: Stage 1 (“Eager”) Kimi K2.5, Stage 2 (“Deliberate”) GLM-5 on $v$3, Stage 3 (“Adversarial”) MiniMax M2.7 on $v$4, and Stage 4 (“Synthesizer”) GLM-5 again on $v$5, yielding $v$6. The student is then trained with SFT cross-entropy and DPO via LoRA low-rank adapters $v$7 (Sadasivan, 13 Apr 2026).

The paper’s first substantive contribution is falsification. An internal draft had reported +15.3 points on HumanEval and +33.9 points on MCAS for a Qwen3-0.6B student, but both gains disappeared under stricter scrutiny. HumanEval had used $v$8, truncating the baseline’s code but not the DD-trained model’s reasoning-heavy outputs; when rerun at $v$9, the result inverted from baseline 36.0% to DD model 28.0%, a $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$0-point delta. MCAS, a 5-axis judge rubric scored by Claude Opus 4.6 on 500 prompts, had a judge-prompt asymmetry; when rescored truly blind at the same decoding and prompt, the baseline moved from 12.39/25 to DD at 12.07/25, a $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$1-point delta (Sadasivan, 13 Apr 2026).

After that falsification, the authors ran three experimental arcs. Arc 1 applied SFT/DPO LoRA to Qwen3-0.6B, Qwen3-1.7B, and Gemma 4 E2B with attention-only LoRA of rank $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$2 and all-modules LoRA across coding and a French-cuisine domain. The consistent outcome was style transfer with capability loss: Qwen3-1.7B fell from 83.5% to 67.7% on coding, and Gemma checklist coverage fell from 0.452 to 0.379 on the French-cuisine domain. Arc 2 tested inference-time attention-head tempering on $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$3, ranking heads by either

$$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$4

or by directional projection onto $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$5 with Cohen’s $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$6, then scaling top-$$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$7 heads as $$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$8. None of five variants on Gemma 4 E2B and Qwen3.5-0.8B produced “a clean increase in disposition score without content damage.” Arc 3 used a frozen-base confidence-gated sidecar with logistic regression on the final-token hidden state

$$\mathrm{score}_{s}(q,y)=\frac1v\sum_{i=1}^v V_s^{(i)}(q,y),$$9

trained on 100 “step16” Chef prompts. The paper’s “honest five-step pipeline” failed to exceed the gate of fresh AUC $\mathrm{score}_s(q,y)=1$0 and asym $\mathrm{score}_s(q,y)=1$1 (Sadasivan, 13 Apr 2026).

The hidden-state probing analysis yielded a two-failure-mode taxonomy. For Gemma 4 E2B, step16 cross-validation reached AUC = 0.683 but collapsed to fresh AUC = 0.516, described as “single output distribution” mode. For SmolLM2-1.7B, a similar cross-validation AUC collapsed to fresh AUC = 0.519, while the fresh positive rate crashed to 4.7%, described as “capability-cliff mode.” The paper interprets this as evidence that high in-distribution probe accuracy can be a pure artifact, and that held-out transfer reveals no real linear “verification” direction (Sadasivan, 13 Apr 2026).

An independent empirical result sharpens the critique of confidence as a proxy for correctness. On 100 Chef prompts, Gemma 4 E2B exhibited $\mathrm{score}_s(q,y)=1$2, $\mathrm{score}_s(q,y)=1$3, and assertion asymmetry $\mathrm{score}_s(q,y)=1$4, so the model “asserts at 91% regardless of correctness” (Sadasivan, 13 Apr 2026). This directly contradicts any simple equation of judge-perceived confidence with reliable self-verification.

6. Cross-domain extension, evaluation, and open questions

The broadest domain extension appears in "Reliable Neural-Codec Text-to-Speech by ASR Self-Verification and Distillation" (Asaria et al., 16 Jun 2026). Here the verification signal is neither a language-model judge nor a task reward, but an ASR round-trip metric. A generation $\mathrm{score}_s(q,y)=1$5 is a catastrophic failure if it is a dropout or a collapse/wrong-content event: $\mathrm{score}_s(q,y)=1$6 with thresholds $\mathrm{score}_s(q,y)=1$7 tokens, $\mathrm{score}_s(q,y)=1$8 word, and $\mathrm{score}_s(q,y)=1$9. For best-of-$(q,y)$0 decoding, the catastrophic-failure rate is

$(q,y)$1

Best-of-$(q,y)$2 ASR self-verification then selects the non-dropout candidate with the lowest WER transcript (Asaria et al., 16 Jun 2026).

The reported robustness gains are substantial. On the primary Llasa-1B model, $(q,y)$3 on the hard set is 0.269, dropping to 0.154 at $(q,y)$4, 0.038 at $(q,y)$5, and 0.000 by $(q,y)$6; on LibriSpeech, $(q,y)$7 and $(q,y)$8 (Asaria et al., 16 Jun 2026). The effect replicates across four open codec-TTS systems and three neural codecs, reaching the near-zero floor by $(q,y)$9 on three of the four models. The same work then distills the self-verified behavior back into single-shot inference using LoRA-based SFT or DPO. On the hard set, base $\mathcal{D}_{\mathrm{sft}}$0 falls to 0.096 with SFT and 0.083 with DPO, which the paper summarizes as closing about 52–58% of the failure mass at no inference-time cost. On LibriSpeech, where base $\mathcal{D}_{\mathrm{sft}}$1 and best-of-2 already reaches 0.000, distillation produces no detectable change (Asaria et al., 16 Jun 2026).

This cross-domain result matters for the general interpretation of self-verified distillation. It shows that the paradigm is not specific to symbolic reasoning or judge-scored language tasks. A plausible implication is that self-verified distillation is most reliable when the verification signal is closely tied to an externally checkable outcome, such as binary reward, pass/fail execution, or ASR round-trip failure detection, rather than to stylistic cues that may correlate weakly with correctness. That interpretation is consistent with the negative result on disposition distillation, which explicitly recommends outcome-grounded distillation in verifiable domains such as code with an execution oracle and identifies larger-scale bases, non-linear probes, token-level gating, and RLHF with an explicit judge as open questions (Sadasivan, 13 Apr 2026).

The literature therefore presents a bifurcated picture. In verifiable settings, self-verified distillation can convert sparse signals into dense supervision, improve pass@1 or coverage, stabilize training, and even transfer best-of-$\mathcal{D}_{\mathrm{sft}}$2 robustness into single-shot inference (Hübotter et al., 28 Jan 2026, Ding, 25 Mar 2026, Lee et al., 20 May 2026, Asaria et al., 16 Jun 2026). In judge-based behavioral editing, however, the same label can mask false positives, stylistic mimicry, and confidence–correctness decoupling rather than genuine self-verification (Sadasivan, 13 Apr 2026). The central technical question is no longer whether models can supervise themselves in principle, but which verification signals are faithful enough that self-distillation amplifies competence rather than appearance.