JulianHJR/v8

Student Simulation v7

A clean rewrite for a single goal: find the monitoring (reflection / self-correction) dimension in Qwen3-30B-A3B-Thinking and expose a single α ∈ [0, 1] knob that smoothly slides between "full reflection" (α=1) and "no reflection" (α=0) at inference time.

What's different from reflection_4 (and why)

reflection_project_4 worked on a dense Qwen-2.5 7B model. v7 keeps its pipeline almost verbatim — patterns, decision-point labeling, projection- removal hook math, multi-layer co-application at inference, monotonic- sweep gate for KEEP/SKIP — and changes ONE thing:

Stage	reflection_4	v7
Direction	mean-diff over hidden states	top-K expert selection → PCA on pos-vs-neg → expert-aware coordinate mask → ortho

The change is forced by the MoE architecture. Qwen3-30B-A3B has 128 experts per layer with top-8 routing, and only a handful of experts fire disproportionately on monitoring-positive tokens. Mean-diff over the full hidden state averages signal with noise from irrelevant experts. v7's direction extraction:

1. Top-K expert selection (Stage A): score each expert at each layer by E_pos[gate_prob] - E_neg[gate_prob], keep the top K=16.
2. PCA on pos-vs-neg (Stage B): top 2 principal axes of the [h_pos - μ_neg ; h_neg - μ_neg] matrix at each layer.
3. Expert coordinate mask (Stage C): weight each PC's coordinates by the L1 norm of the selected experts' input projections at that coordinate.
4. Orthogonalize against general direction (Stage D), then Gram-Schmidt the surviving components.

Everything else is reflection_4: the regex patterns, the decision-point labeling (5 tokens before / 2 tokens after each trigger), the projection-removal hook, the multi-layer co-application at inference, the monotonic-sweep KEEP/SKIP gate.

Pipeline

01: label + capture     ~1.5 h
02: build directions    ~10 min   (CPU PCA, expert mask, ortho)
03: calibrate           ~10-14 h  (per-layer × 3 strengths × 20 problems)
04: infer (deliverable) ~30 min   (3 problems × 4 alphas)
                       --------
                       ~12-16 h

Running

# Interactive (default GPU 6):
bash runall.sh

# Specific GPU:
CUDA_VISIBLE_DEVICES=3 bash runall.sh

# Single stage for debugging:
STAGES=04 bash runall.sh

# Slurm:
sbatch slurm/run-v7.sbatch

Reading the outputs

In order:

1. data/checkpoints/directions_summary.json — Stage 02 result. Look at n_layers_with_direction (should be 15-19 of 19 target layers) and diagnostics.<L>.var_explained (top PC should explain >5% of variance).

2. data/checkpoints/monitoring_calibration_v7.json — Stage 03 result. kept_layers is the multi-layer set that passed the monotonic gate. If it's empty, the direction either has no causal effect or side-effect rate is too strict; relax --side-effect-rate 0.15 and rerun stage 03.

3. data/results/alpha_comparison_v7.json — Stage 04 deliverable. Each record is (problem, alpha, cot, monitoring_total, repetition_score). For each problem, you should see monitoring_total decrease as α goes from 1.0 → 0.0 with no collapsed: true along the way.

Key configuration

• configs/monitoring.py controls dimension-specific hyperparameters. SWEEP_ALPHAS = [0.0, 0.3, 0.7, 1.0] per user spec.
• configs/paths.py controls all I/O paths. MODEL_PATH and RAW_COTS_PATH default to the v6 install but can be overridden via env.