FIX - Surgery on a Ramp and on a Thought Loop
Experiments 4 and 8 (+ the Gemma control): reading earns its keep when it repairs.
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.
| # | Task | Edit size | Before → after |
|---|---|---|---|
| 4 | Car can’t climb a ramp (Kinetix) | 3.5% of 171,664 | 0/256 → 1,024/1,024 episodes |
| 8 | Qwen3.5-0.8B overthinks math | 131,072 = 0.016% | MATH-500 7.9% → 48.1%; token-limit hits 93% → 33% |
| - | Gemma 3 270M (negative control) | same method | MATH-500 6.8% → 7.0% at 50 passes (no hidden ability to unblock) |
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 (🟦)
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”.
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'=0What 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}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}| Variant | MATH-500 % | GSM8K % | IFEval % | Hit token limit % |
|---|---|---|---|---|
| Base | 7.9 | 13.8 | 21.0 | 93.0 |
| Edited · 10 passes | 12.3 | 25.7 | 49.0 | 87.2 |
| Edited · 25 passes | 14.1 | 25.1 | 39.9 | 86.4 |
| Edited · 50 passes | 48.1 | 67.1 | 62.3 | 33.1 |
| Variant | MATH-500 % |
|---|---|
| Base | 6.8 |
| 5 passes | 7.0 |
| 10 passes | 7.6 |
| 25 passes | 7.7 |
| 50 passes | 7.0 |
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)