Part 2·Chapter 5 of 13· 13 min read

FIX - Surgery on a Ramp and on a Thought Loop

Experiments 4 and 8 (+ the Gemma control): reading earns its keep when it repairs.

Reading Audience LevelRung 2: L2 · Curious adult

Plain words and small, concrete numbers. - Step up or down anytime with keyboard shortcuts bracket keys or the buttons above.

Two experiments move from reading to repairing: a robot car that cannot climb a ramp, and a language model that overthinks its maths until the clock runs out. Both are fixed with tiny, explained edits - and the second comes with a control that keeps the claim honest.

Paper-reportedThe two FIX experiments (+ control) at a glance
#TaskEdit sizeBefore → after
4Car can’t climb a ramp (Kinetix)3.5% of 171,6640/256 → 1,024/1,024 episodes
8Qwen3.5-0.8B overthinks math131,072 = 0.016%MATH-500 7.9% → 48.1%; token-limit hits 93% → 33%
-Gemma 3 270M (negative control)same methodMATH-500 6.8% → 7.0% at 50 passes (no hidden ability to unblock)
Teaching schematic

M5 · Direction Surgery

m5-direction-surgery

blue: |uᵀW| before · green (down): what remains after · amber caps mark the 24 edited columns

38%

columns edited here (toy)

39%

u-energy remaining

0.016%

of weights, in the real Exp 8 (🟦)

loop broken → model escapes

The “thinking loop” cartoon: before the edit the state keeps re-entering itself (as Qwen3.5-0.8B did - 93% of math answers hit the token limit 🟦). After the direction is subtracted from enough columns, the loop loses its engine (33% at 50 passes).

Editing the top-24 of 64 columns removes 61% of u’s energy in W - remaining u-energy 0.39. The loop escapes (toy animation).

L2 view: A “direction” is one pattern the model re-injects. Subtract it from a few columns of the big number-table and the repetition loses its fuel.

TEACHING SCHEMATIC - invented W and u. The escape animation is a cartoon of “the token loop loses its direction”.

Paper-reportedAnatomy of the Exp-8 edit: two MLP blocks, 64 output columns each - 131,072 weights in total (as reported).
24 layers of Qwen3.5-0.8B · MLP output projections of layers 15 & 19 edited (amber)L0L5L10L15*L19*L23Anatomy of Edited Output Projections (W' = (I − uuᵀ)W on 64 columns each)Layer 15 MLP Output Projection…64 output columns × 1,024 width = 65,536 weightsLayer 19 MLP Output Projection…64 output columns × 1,024 width = 65,536 weights2 layers × 64 output columns × 1,024 width = 131,072 weights (0.016% of Qwen3.5-0.8B)
Mathematical Formulations (3)Tripartite System
Teaching schematicDirection removal (rank-1 projection)
f-direction

To kill a direction u, project every row of W so nothing along u survives - here shown as the teaching approximation; the post’s exact update lives in its released logs.

•
u - the unit direction to remove
•
I - uuᵀ - the projection that erases u
•
W, W′ - a weight matrix before and after

Example: lab/02_direction_surgery.py: after the full projection, max |uᵀW′| ≈ 1e-16.

LaTeX code
W'=(I-uu^{\top})W,\qquad \Delta=-u\,(u^{\top}W),\qquad u^{\top}W'=0
Paper-reportedEdit sparsity - the “blast radius”
f-sparsity

What fraction of all weights did the edit touch? Smaller is surgical.

•
∥Δ∥₀ - count of non-zero edited weights
•
P - total parameters
•
s - sparsity as a fraction

Example: Exp 8: 131,072 / (0.8×10⁹) ≈ 1.6×10⁻⁴ = 0.016% of Qwen3.5-0.8B.

LaTeX code
s=\frac{\lVert\Delta\rVert_0}{P}
Paper-reportedThe budget curve
f-budget

How good is the edit as the editor is allowed more forward passes? Only the reported points may be drawn - never a smooth interpolation.

•
n - number of feedback passes allowed
•
fidelity(n) - measured quality at that budget

Example: Exp 8, n = 0/10/25/50: MATH-500 7.9 → 12.3 → 14.1 → 48.1. The dip at n=25 (IFEval 49.0 → 39.9) is why n=1 runs per point is a real caveat.

LaTeX code
\mathrm{fidelity}(n)\ \text{vs.}\ n\ \text{forward passes}
Comparative Empirical Benchmarks (2)Experiment vs Control Analysis
Paper-reportedQwen3.5-0.8B, thinking mode on, up to 32,768 tokens (Exp 8)
VariantMATH-500 %GSM8K %IFEval %Hit token limit %
Base7.913.821.093.0
Edited · 10 passes12.325.749.087.2
Edited · 25 passes14.125.139.986.4
Edited · 50 passes48.167.162.333.1
Forward passes n = self-play / feedback passes the editor was allowed. Reported points only - no interpolation.
Paper-reportedNegative control: Gemma 3 270M, MATH-500 % (Exp 8)
VariantMATH-500 %
Base6.8
5 passes7.0
10 passes7.6
25 passes7.7
50 passes7.0
Only 8.5% of Gemma’s answers hit the token limit (vs 93% for Qwen). Conclusion: the edit removed an obstacle that hid an ability; it did not add one.
Paper-reported

M11 · The Blast-Radius Meter

m11-blast-radius
1 ppm0.01%1%

0.0164%

log scale ← 1-in-a-billion · · · · · · · · 100% → (Exp 8 sits at 0.016%, ~2 dots from the left)

Exp 8 · Qwen3.5-0.8B: 0.016% - the paper’s flagship edit 🟦

Editing 131,072 of 800,000,000 weights = 0.0164% - the “blast radius” s = ‖Δ‖₀ / P.

L2 view: Compare: a full re-train touches 100% of the numbers. The Qwen math fix touched 0.016% - about 1 in 6,000 - and changed the model’s exam score by 40 points.

🟦 preset values from the paper · 🟨 LoRA & full-FT rows are teaching comparisons. Your custom input is your own.

Exp 4 - the ramp (Kinetix). The actor network (679→128→5×128→16, 171,664 parameters) had been trained with PPO to drive, but never managed the ramp: 0/256 success. The LLM looked at the weights, made one edit touching 3.5% of them, with no tools and no practice - success flipped to 1,024/1,024.

Exp 8 - Qwen3.5-0.8B’s maths overthinking. In thinking mode (up to 32,768 tokens), the base model hit the token limit on 93% of MATH-500 problems, scoring just 7.9%. The LLM identified a “looping direction” written by the MLPs of layers 15 and 19, and subtracted it from 64 output columns in each of those two layers - 131,072 weights, 0.016% of the model. With 50 feedback passes allowed, MATH-500 reached 48.1%, GSM8K 67.1% (from 13.8%), IFEval 62.3% (from 21.0%), and token-limit hits fell to 33.1%.

The negative control. Gemma 3 270M got the same treatment on MATH-500: 6.8% base → 7.0/7.6/7.7/7.0% at 5/10/25/50 passes. Only 8.5% of its answers ever hit the token limit (vs 93% for Qwen) - nothing was blocking it, so the edit had nothing to remove. The authors’ conclusion: editing can remove an obstacle that hid an ability, but did not add one.

Foundational Refreshers (1)