01

The puzzle

A teacher judging someone else’s thoughts

On-policy distillation has an appealing story: let the student generate a response, ask a stronger teacher to score every token, and use that dense feedback to move the student toward the teacher.

The subtlety is in the prefix. Each teacher score is conditioned on a trajectory sampled by the student—a sequence the teacher may never have produced. The farther the two policies drift apart, the more the teacher is asked to judge unfamiliar states. A larger teacher is not automatically a more reliable teacher in those states.

What we call a noisy signal

At verifiable final-answer tokens, the sign of the teacher’s advantage disagrees with the outcome: a correct answer receives a negative signal, or an incorrect answer receives a positive one.

Figure 2 Signal mismatch grows strikingly with teacher scale, yet filtering the trajectories barely changes the student’s improvement.

With a 4B teacher, 20.4% of correct trajectories receive a negative answer-token signal, while 40.8% of incorrect trajectories receive a positive one. With the largest teacher, the signal becomes almost uniformly negative: 97.8% of correct and 96.6% of incorrect answer tokens are pushed down.

Then comes the result that changed the direction of the project. We trained on all trajectories, only those containing noisy signals, and only those without them. The three runs reached comparable performance after a similar number of updates.

If the student improves even on the misleading subset, what exactly is being distilled?

02

What drives learning

Follow the gradient, not the story

We stripped the objective down along two axes: which sampled tokens actually move the policy, and which signs of the learning signal create the improvement?

Finding A

Confident tokens are mostly spectators

When the student already assigns a token high probability, its policy gradient is tiny. Teacher and student also tend to agree there, making the advantage nearly zero.

51.7% of tokens have |advantage| < 10−4
Finding B

The negative sign does the work

On the lowest-logp tokens, replacing teacher advantages with the same fixed negative value preserves the improvement. A fixed positive value collapses the policy.

−0.5 works without a teacher
Figure 3 High-logp tokens dominate the rollout but carry little effective learning signal. Restricting training to them produces almost no gain—even with the original advantage.

The result is easiest to see at the logit level. For a sampled token with probability p, the update is proportional to A(1 − p). It vanishes when the advantage A is near zero or when the policy is already confident. High-logp tokens often satisfy both conditions.

Sampled-token update

Δzt ∝ At · (1 − πθ(yt ∣ y<t))

low A → little update  ·  high probability → little update
Figure 4 On the same lowest-logp 20%, a fixed negative advantage reproduces OPD-like improvement; flipping it positive causes policy collapse.

This control removes the information that distillation is supposed to transfer. We keep the selected token positions fixed—the lowest-logp 20%—but replace every teacher-derived advantage with one scalar. With A = −0.5, AIME24 performance steadily rises into the same band as standard OPD, while response length grows toward 12K tokens and then stabilizes. No teacher ranking among tokens is needed.

Fixed negative · A = −0.5

OPD-like gains remain

The selected unlikely tokens are consistently suppressed. Accuracy recovers and the full run remains stable even though teacher logits have been removed.

Fixed positive · A = +0.2

The policy collapses

Selection is unchanged; only the direction flips. Within roughly 40 steps, response length falls near zero and the gradient norm spikes. Later outputs degenerate into random or garbled text.

The sign ablation isolates the mechanism: what matters is the negative direction on low-probability sampled tokens, not a detailed preference pattern inherited from the teacher.

Working hypothesis

Much of OPD’s gain comes from suppressing unlikely tokens sampled by the student—not from transferring a teacher’s preferred behavior.

03

The recipe

Turn uncertainty into supervision

A fixed negative signal reveals that the teacher is unnecessary, but it treats every selected token the same. The student already contains a finer signal: its entropy at each position.

Low log probability alone is ambiguous. The model may be uncertain among several good next tokens, or it may be highly confident while a rare tail token happens to be sampled. Entropy separates these cases. We found that stronger negative updates at higher-entropy positions are both more effective and more stable.

1

Roll out

Sample a response from the current policy and retain token log probabilities and entropies.

2

Select

Within each response, keep only the 20% of sampled tokens with the lowest log probability.

3

Self-adapt

Give those tokens negative advantages, increasing the magnitude with normalized entropy.

Entropy-adaptive advantagefor i in the lowest-logp 20%

Aidyn=12Hi − Hmin2(Hmax − Hmin)

lower entropy −0.5higher entropy −1.0

The conceptual form is shown here; implementations should guard the degenerate case where all selected-token entropies are identical.

This is On-Policy Self-Adaptation, or OPSA. The objective uses the same policy-gradient machinery, but its token-level advantages come entirely from the student’s own distribution. There is no teacher forward pass, no answer label, no verifier reward, and no reference hint. Training still needs prompts; it does not need supervision attached to them.

Figure 5 Scaling negative magnitude with entropy (δ = 1) reaches 50.0 Avg@4 on AIME24, compared with 35.13 for standard OPD in this controlled run.
04

How it works

Prune the tail, preserve the forks

A negative update does not simply “make the model less confident.” Its effect depends on where the sampled token sits inside the distribution.

Tail token

Close unlikely branches

When sampling lands in the tail, OPSA pushes that token down and returns its mass to more plausible alternatives.

High-entropy fork

Share mass across alternatives

When several head tokens compete, lowering the sampled one redistributes mass among the others instead of collapsing to a single path.

Confident position

Leave certainty alone

High-logp tokens at low-entropy positions are normally outside the selected 20%, preserving precise predictions.

Figure 6 The same negative update produces different local effects. Selection by logp and scaling by entropy make that difference useful.

Reasoning often changes direction at high-entropy “fork” tokens—words such as wait, but, or alternatively. OPSA makes the head set more likely as a whole while keeping competition alive inside it. This offers a way to sharpen the distribution globally without erasing the branches that exploration needs.

05

Evidence

Self-adaptation changes both accuracy and behavior

We trained on the questions from DAPO-17k without their answers, then evaluated across mathematical reasoning, code generation, and general question answering.

AIME24 · Avg@32
13.4448.85

+263.5% relative

AIME25 · Avg@32
9.6935.31

+264.4% relative

HMMT25 · Avg@32
5.7323.33

+307.2% relative

Avg@32Mean accuracy across 32 sampled responsesPass@32Problems solved by at least one of 32 samples
All models, non-thinking mode, 32 samples per problem
ModelAIME24AIME25HMMT25
Avg@32Pass@32Avg@32Pass@32Avg@32Pass@32
Qwen3-1.7B
Base13.4440.009.6930.005.7323.33
+ OPSA48.8580.0035.3166.6723.3350.00
Point gain+35.41+40.00+25.62+36.67+17.60+26.67
Qwen3-4B
Base23.3356.6720.5256.6713.1333.33
+ OPSA62.0883.3358.4483.3337.4060.00
Point gain+38.75+26.66+37.92+26.66+24.27+26.67
Qwen3.5-9B
Base76.3593.3356.0493.3344.4886.67
+ OPSA87.8196.6776.9896.6767.4093.33
Point gain+11.46+3.34+20.94+3.34+22.92+6.66

The improvement is not confined to a small base model. OPSA also lifts Qwen3-4B and Qwen3.5-9B, and produces gains on MBPP+ and GPQA-Diamond. Relative to the best compared baseline for each aggregate metric, the 1.7B model gains 11.04 points in mean Avg@32 and 8.89 points in mean Pass@32 across the three math benchmarks.

Figure 7 Rollouts become longer and contain more reflective turns. Across sampled responses, length and accuracy are positively correlated.

A behavioral clue

More “wait,” not just more tokens

Response length rises through training, but the change is structured: reflective markers become much more frequent. Masking positions whose head set contains those fork tokens largely removes both the length growth and the accuracy gain.

A diversity check

Diversity holds under our measures

Across 32 responses per problem, pairwise 4-gram diversity approaches that of the base model. Together with the Pass@32 gains, we observe no diversity collapse under these two measures.

Figure 8 As responses grow, repeated fork opportunities preserve tree-like variation; the diversity gap between the base and adapted policy approaches zero.

The paper in one sentence

OPD can work by suppressing unlikely sampled tokens—not by copying a teacher.

That mechanism motivates OPSA: select the lowest-logp 20% and scale a negative update with the student’s own entropy, enabling self-improvement without teacher logits, rewards, or hints.