A working document, with a working preview

Find the right expression.

Change the actual page. Save a direction. Put two versions next to each other.

Start somewhere

Jump to

Scroll inside the page to explore it.

What to look for

Does the visual language explain the research before someone reads the abstract? Editing marks make the teacher’s role tangible; generous margins let the example do the work.

Hold one thing constant, change another

Two directions.
A clearer choice.

Compare saved snapshots or the current working design. Each preview is a complete page.

A Direction

B Direction

What actually changes?

Design decision	Direction A	Direction B

Decide the argument before adding content

Structure is
part of the story.

Three useful routes through the same research, depending on who arrives and what they need.

01 / For a broad ML audience

Lead with the idea.

Make the teacher-as-editor concept immediately understandable. Build the case before showing the numbers.

Hook↓Concrete correction↓Method + loss↓Evidence↓Paper

Tradeoff: technical readers scroll further to reach the equations.

02 / For readers comparing methods

Lead with evidence.

Give researchers a reason to care early: the benchmark, the baseline, and the actual size of the gain.

Hook↓Results + limitations↓Concrete correction↓Method + loss↓Paper

Tradeoff: results need precise context before the method has been explained.

03 / For a technical research audience

Lead with mechanics.

Introduce the algorithm and objective first. Let the explanatory animation carry the technical argument.

Hook↓Method + loss↓Concrete correction↓Evidence↓Paper

Tradeoff: a higher entry cost for readers who are new to agent distillation.

What belongs where

A page with depth,
not a pasted paper.

The first layer should give the insight. The second should let a researcher inspect the mechanism and evidence.

Section	Show immediately	Let readers open
Hero	One takeaway, authors, paper link, concrete visual.	Full paper title and affiliation remain visible.
The idea	Student → teacher → style-preserving correction.	Original figure and full structured response.
Method	Four stages and the two terms of the objective.	Chaptered animation, exact equations, span weighting and masks.
Evidence	One benchmark, explicit baseline, model selector.	Full table, standard errors, transfer results and limits.
Paper	Preprint, authors, copyable citation.	Code and checkpoints when their release links exist.

Explain the optimization, not just the pipeline

Make the loss
make sense.

A deeper Manim treatment, followed by an interactive explanation of what each term controls.

Read Eq. 5 in the paper ↗

The central equation / paper Eq. 5

Two terms. Two different jobs.

L WD equals the sum of token weights times cross entropy divided by the sum of weights, plus beta times the sum of role masks times pointwise KL divided by the sum of role masks.

Weighted imitation

What deserves attention?

Weight reasoning by how much it helps the student predict the corrected action. Divide by the weight sum so changing all weights together does not silently rescale the loss.

Rationale anchor

How far should reasoning move?

Penalize drift from the frozen seed on think, plan, and memory tokens. Leave action and step tokens almost unanchored so corrected decisions can be learned.

Try the intuition

Which span earns
the learning signal?

Move the contribution scores. The weight map follows the paper’s Eq. 10, including average ranks for ties and the dead zone.

Illustrative values · one token per span · not experimental results

Anchor strength β 0.03Anchor mask

The demo holds target-token cross entropy and log-ratios fixed. β changes the objective’s penalty, not a simulated model’s accuracy.

Weight floor φ = 0.1Dead zone τ = 0.05 nats

Role	CE weight w	Anchor γ	Fixed CE	Fixed d

Weighted CE—+β × anchor—=Illustrative loss—

Eq. 2 / Diagnose the mismatch

Extra surprisal, not raw perplexity.

Premium equals mean negative log probability of teacher supervision under the seed minus mean negative log probability of student text.

The premium is the difference in mean per-token surprisal: log perplexity of supervision minus log perplexity of the student’s own response. Unfamiliar teacher wording can dominate a uniform loss.

Eq. 4 / Measure contribution

What changes when this span disappears?

Delta s equals the seed model's log probability of the corrected action with the span minus its log probability without the span.

Both scores come from the frozen student checkpoint, computed offline. A negative contribution stays in the ranking; it is not automatically removed.

Eq. 10 / Map contribution to weight

Rank within each step.

The span weight is clipped between phi and one; the normalized rank is contracted toward one half according to the within-step spread divided by the dead-zone width.

Here rₛ is the normalized within-step rank and σ = max Δ − min Δ is the contribution spread. With φ = 0.1 and τ = 0.05 nats, all-tied contributions give every span weight 0.55; near ties smoothly approach it. Ranks reduce sensitivity to action-probability scale.

Eq. 5 / Penalize selected drift

A pointwise KL surrogate.

k l i equals exponential d i minus d i minus one, where d i is log seed probability minus log current policy probability of the same target token.

Here pᵢ is the current student’s target-token probability and qᵢ is the frozen seed’s, at the same context. This penalty is evaluated on corpus target tokens, not the exact full-vocabulary KL. The rationale mask is 1 for think/plan/memory and 10⁻⁶ elsewhere in the implementation.