problem string | source string |
|---|---|
Find the sum of all integer bases $b > 9$ for which $17_b$ is a perfect square. | AIME 2025 I Problem 1 |
The numbers $1, 2, \ldots, 18$ are randomly arranged in a row. Find the probability that no two adjacent numbers have a sum divisible by 3, expressed as $m/n$ in lowest terms. Find $m+n$. | AIME 2025 I Problem 2 |
In triangle $ABC$, $AB=7$, $BC=8$, $CA=9$. Point $D$ lies on $BC$ with $BD=3$. Find $AD^2$. | AIME 2025 I Problem 3 |
Let $f(n)$ be the number of positive divisors of $n$ that are perfect squares. Find $\sum_{n=1}^{50} f(n)$. | AIME 2025 I Problem 4 |
Find the number of subsets $S$ of $\{1,2,\ldots,10\}$ such that no two elements of $S$ differ by exactly 2. | AIME 2025 I Problem 5 |
A point $P$ inside equilateral triangle $ABC$ with side 4 satisfies $PA=2$, $PB=3$. Find $PC^2$. | AIME 2025 I Problem 6 |
Find the largest prime $p$ such that $p^2$ divides $\binom{50}{25}$. | AIME 2025 I Problem 7 |
Let $a_1=3$, $a_{n+1}=\frac{a_n^2+2}{a_n+1}$. Find $\lfloor a_{2025} \rfloor$. | AIME 2025 I Problem 8 |
In how many ways can a $3\times 12$ rectangle be tiled by $1\times 3$ and $3\times 1$ trominoes? | AIME 2025 I Problem 9 |
Find the number of ordered triples $(a,b,c)$ of positive integers with $a\leq b\leq c$ and $a+b+c=30$. | AIME 2025 I Problem 10 |
Let $z=e^{2\pi i/7}$. Compute $\left|\sum_{k=0}^{6} z^{k^2}\right|^2$. | AIME 2025 II Problem 1 |
Find the sum of all positive integers $n$ such that $n+12$ divides $n^2+144$. | AIME 2025 II Problem 2 |
A fair coin is flipped 12 times. What is the probability that the number of heads is divisible by 3? | AIME 2025 II Problem 3 |
Find the area of the region where $|x|+|y|\leq 4$, $|x+y|\leq 2$, and $|x-y|\leq 2$. | AIME 2025 II Problem 4 |
Let $p(x)=x^4-4x^3+6x^2-4x+1$. Find $p(p(2))$. | AIME 2025 II Problem 5 |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
v10 — Monitoring + Planning Activation Steering
Qwen3-30B-A3B-Thinking inference-time dimension control via projection-removal hooks.
What's new in v10 vs v9
| Component | v9 | v10 |
|---|---|---|
| CoT source | Pre-existing raw_cots.jsonl |
Stage 00: model generates 150 CoTs from MATH100 |
| Dimensions | monitoring only | monitoring + planning |
| Inference test set | 6 hardcoded problems | 15 AIME 2025 problems |
| Inference logic | Both mono + think + plain variants | allmono only (cleanest) |
| Resume | Partial | Full resume at every stage (atomic writes, per-layer JSON cache) |
| Slurm | 2 sbatch files | 3 sbatch files (full / infer-only / both-dims) |
Pipeline
Stage 00 Generate CoTs ~2-4 h (150 problems × Qwen3-30B greedy)
Stage 01 Label + Capture ~1.5 h (per dimension)
Stage 02 Build Directions ~10 min (CREST PCA-denoised mean-diff, per dim)
Stage 03 Calibrate ~10-14 h (20 problems × layers × 3 alphas, per dim)
Stage 03b Select Layers ~1 min (greedy selection, CPU only)
Stage 04 Infer AIME25 ~1-2 h (15 problems × 4 alphas, allmono, per dim)
Usage
# Single dimension (default: monitoring)
bash runall.sh
# Planning dimension
DIMENSION=planning bash runall.sh
# Both dimensions
DIMENSION=all bash runall.sh
# Skip CoT generation (already exists)
STAGES=01,02,03,03b,04 bash runall.sh
# Re-run inference only
STAGES=04 bash runall.sh
Slurm
# Full pipeline (monitoring)
sbatch slurm/run-v10.sbatch
# Full pipeline (planning)
sbatch --export=DIMENSION=planning slurm/run-v10.sbatch
# Both dimensions
sbatch slurm/run-v10-all.sbatch
# Re-run inference only
sbatch slurm/run-v10-04.sbatch
Resume
Every stage checks for its output file before running.
- Pass
--forceto any script to recompute. - Calibration resumes per-layer: if
data/{dim}/checkpoints/calib_per_layer/layer_XXX.jsonexists, that layer is skipped. - Inference resumes per-record: completed (problem, alpha) pairs are cached in
data/{dim}/results/infer_cache.jsonl. - Stage 00 resumes per-problem.
Environment Variables
| Variable | Default | Description |
|---|---|---|
MODEL_PATH |
/data/.../Qwen3-30B-A3B-Thinking-2507 |
Local model path |
MATH100_PATH |
data/math100.jsonl |
MATH100 dataset |
AIME25_PATH |
data/aime25.jsonl |
15 AIME 2025 problems (bundled) |
RAW_COTS_PATH |
data/cots/raw_cots.jsonl |
Generated CoTs (written by stage 00) |
DIMENSION |
monitoring |
Dimension to run (monitoring/planning/all) |
STAGES |
00,01,02,03,03b,04 |
Comma-separated stages to execute |
Output Structure
data/
cots/raw_cots.jsonl # 150 generated CoTs (stage 00)
monitoring/
labeled_cots_monitoring.jsonl
activations/activations_monitoring.pt
checkpoints/
directions_monitoring.pt
calibration_monitoring.json
selected_layers_monitoring.json
calib_per_layer/layer_XXX.json # per-layer resume cache
results/
alpha_comparison_monitoring.json # FINAL DELIVERABLE
infer_cache.jsonl
planning/
... (same structure)
- Downloads last month
- 43