TECHNICAL DICTIONARY · 60 CANONICAL TERMS · FULL DOMAIN INDEX

Field Guide Glossary

Curated in pedagogical sequence: from fundamental parameter geometry to transformer circuit mechanics, empirical verification criteria, and benchmark protocols. Every entry directly anchors to its corresponding guide chapter.

70/ 70
Quick terms:
Initial:
Parameter
Ch 1
One number inside a network - a weight or a bias. The knobs.
Guide canonical · Ch 1
Weight
Ch 1
A parameter that multiplies an input (or another neuron’s output) on its way through the network.
Guide canonical · Ch 1
Bias
Ch 1
A parameter added after the multiplying, shifting a neuron’s trigger point.
Guide canonical · Ch 1
θ (theta)
Ch 1
The vector of ALL parameters, treated as one long list of numbers.
Guide canonical · Ch 1
P (parameter count)
Ch 1
How many numbers θ contains - 17,253 for the Exp-1 net, ~0.8 billion for Qwen3.5-0.8B.
Guide canonical · Ch 1
Architecture
Ch 1
The wiring diagram: how many layers, of what kind, connected how. Same architecture + different weights = different skill.
Guide canonical · Ch 1
Weight space
Ch 1
The space of every possible setting of θ - training walks through it, edits jump across it.
Guide canonical · Ch 1
MLP
Ch 1
Multi-layer perceptron: the simplest network - layers of matrix-multiply + nonlinearity.
Guide canonical · Ch 1
tanh / ReLU
Ch 1
Nonlinear functions applied after each layer; tanh squashes to (−1,1), ReLU keeps positives, zeroes negatives.
Guide canonical · Ch 1
Receptive field
Ch 4
The slice of input a given weight column responds to - why a 10×10 grid can reveal a paddle row.
Guide canonical · Ch 4
Forward pass
Ch 4
Running the network on an input to get an output. Weight reading in the READ experiments happened with zero forward passes.
Guide canonical · Ch 4
Gradient descent
Ch 2
The default weight-writer: nudge every parameter a tiny step downhill on the loss, millions of times.
Guide canonical · Ch 2
SGD / Adam / Muon
Ch 2
Flavours of gradient descent - different recipes for shaping the tiny steps.
Guide canonical · Ch 2
Learning rate (η)
Ch 2
The size of each training step. Too big: overshoot. Too small: forever.
Guide canonical · Ch 2
Loss function
Ch 2
A number measuring how wrong the network is - the hill training descends.
Guide canonical · Ch 2
PPO
Ch 6
Proximal Policy Optimization - the RL algorithm that trained the paper’s game baselines.
Guide canonical · Ch 6
Hypernetwork
Ch 2
A network trained to output another network’s weights (Ha et al., 2016).
Guide canonical · Ch 2
Fast-weight programmer
Ch 2
A transformer variant that writes temporary weights as it runs (Schlag et al., 2021).
Guide canonical · Ch 2
Tracr
Ch 2
A compiler from formal programs (RASP) into toy-transformer weights - the third way to write weights.
Guide canonical · Ch 2
Weight-space programming (WSP)
Ch 2
The paper’s claim: a general-purpose LLM reads θ, reasons, and writes Δ directly - no gradients, optionally a few forward passes for feedback.
Guide canonical · Ch 2
Δ (delta)
Ch 2
The edit: the vector added to θ. In Exp 8 it touched 131,072 of ~0.8B weights.
Guide canonical · Ch 2
Fine-tuning
Ch 3
Training an existing model further on new data - gradients again, shorter run.
Guide canonical · Ch 3
LoRA
Ch 3
Low-Rank Adaptation: learn two thin matrices whose product is the change; add it to frozen weights.
Guide canonical · Ch 3
Task arithmetic
Ch 3
Fine-tune, subtract the base: the difference is a “task vector” you can add or remove (Ilharco et al., 2022).
Guide canonical · Ch 3
Model merging
Ch 3
Combining several trained models’ weights into one (TIES and friends).
Guide canonical · Ch 3
Model editing
Ch 3
Surgical weight changes targeting one specific fact or behaviour (ROME, Meng et al. 2022).
Guide canonical · Ch 3
Activation steering
Ch 3
Nudging the model’s inner state at run time without touching weights.
Guide canonical · Ch 3
Mechanistic interpretability
Ch 3
Reverse-engineering networks into human-readable circuits - the detective cousin, and the source of the copying-head technique.
Guide canonical · Ch 3
Weight-space learning
Ch 3
The research field covering understanding, representing, and generating weights (survey: arXiv 2603.10090 [VERIFY]).
Guide canonical · Ch 3
Transformer
Ch 4
The architecture behind modern language models: attention + MLPs in a residual stream.
Guide canonical · Ch 4
Attention head
Ch 4
One pattern-matching unit inside attention; heads numbered layer.head (e.g. 3.1).
Guide canonical · Ch 4
Causal mask
Ch 4
The rule that each token can attend only to earlier tokens - the lower triangle in the attention grid.
Guide canonical · Ch 4
Induction head
Ch 4
A two-head circuit implementing [A][B]…[A]→B - the engine of in-context copying.
Guide canonical · Ch 4
Copying head
Ch 4
The heads Exp 3 identified from weights alone, matching measured ground truth exactly.
Guide canonical · Ch 4
Composition score
Ch 4
How much two heads’ circuits multiply together - the weight-only signal that finds copiers.
Guide canonical · Ch 4
Residual stream
Ch 5
The shared “conveyor belt” a transformer adds information into at every layer.
Guide canonical · Ch 5
DDPM
Ch 4
Denoising diffusion: learn to remove noise step by step; the tiny image-maker of Exp 2.
Guide canonical · Ch 4
Sycophancy
Ch 4
Agreeing with the user instead of being honest - the behaviour Exp 7 tried to detect from weights.
Guide canonical · Ch 4
Kinetix
Ch 5
The physics environment whose ramp the actor couldn’t climb (Exp 4).
Guide canonical · Ch 5
MinAtar
Ch 6
Tiny Atari-inspired games (Breakout, Seaquest, Asterix) used for the WRITE experiments.
Guide canonical · Ch 6
Go (9×9)
Ch 6
The small Go board on which a written-from-scratch policy beat a PPO trained on 20M games.
Guide canonical · Ch 6
Liberty
Ch 7
In Go, an empty point adjacent to a stone - counting them was a stated ingredient of the written policy.
Guide canonical · Ch 7
Atari
Ch 7
In Go, a group with one liberty left - urgent. (Also: the game company; context tells.)
Guide canonical · Ch 7
Direction (in weight space)
Ch 5
A vector u along which behaviour can be added or removed; refusal and thought-loops are direction-mediated.
Guide canonical · Ch 5
Rank-1 projection
Ch 5
The operation W′ = (I − uuᵀ)W that erases direction u from a weight matrix.
Guide canonical · Ch 5
Thought loop / overthinking
Ch 5
The self-reinforcing pattern that made Qwen3.5-0.8B hit the token limit on 93% of math problems.
Guide canonical · Ch 5
Token limit share
Ch 5
The % of answers that ram the maximum generation length - the mechanism check behind Exp 8.
Guide canonical · Ch 5
MATH-500 / GSM8K / IFEval
Ch 5
Benchmarks: competition maths / grade-school word problems / instruction following.
Guide canonical · Ch 5
Negative control
Ch 5
A run that SHOULD show no effect (Gemma 3 270M on math) - its flatness validates the method.
Guide canonical · Ch 5
Edit sparsity (s)
Ch 5
The fraction of parameters an edit touches: s = ‖Δ‖₀/P. Exp 8: 0.016%.
Guide canonical · Ch 5
Blast radius
Ch 5
Informal name for sparsity - how far the edit’s fingerprint reaches.
Guide canonical · Ch 5
Functional drift (D)
Ch 12
How much the edit changed outputs on held-out inputs (KL divergence) - the containment metric WSP-Bench-Lite demands.
Guide canonical · Ch 12
Budget curve (fidelity(n))
Ch 5
Quality vs allowed forward passes n - reported points only, never smooth interpolation.
Guide canonical · Ch 5
Non-monotonic
Ch 8
Goes up, then down, then up: the 25-pass IFEval dip (49.0 → 39.9). A signature of n=1 noise.
Guide canonical · Ch 8
n=1 run
Ch 8
A single execution per condition - the standing caveat on every number in the paper.
Guide canonical · Ch 8
Baseline
Ch 8
The rival method that must be tried before claiming novelty - decoding tweaks, standard direction ablation, plain code. The review’s missing piece.
Guide canonical · Ch 8
Author-blind task
Ch 12
A benchmark task created by someone other than the team being evaluated - WSP-Bench-Lite’s core rule.
Guide canonical · Ch 12
Truth contract
Ch 13
This guide’s rule: every claim wears exactly one badge - 🟦 paper-reported, 🟩 reviewed, 🟨 schematic, 🟥 speculation.
Guide canonical · Ch 13
WSP-Bench-Lite
Ch 12
Het’s proposed community benchmark: Read/Fix/Write tracks, ≥5 seeds, mandatory baselines, author-blind tasks, public logs.
Guide canonical · Ch 12
Permutation symmetry
Ch 1
Hidden units can be relabelled without changing the network’s function - the reason naive weight-reading is underdetermined.
Guide canonical · Ch 1
Weight fingerprint
Ch 4
A characteristic spatial pattern in first-layer weight norms (e.g. 58% of agent norm on row 8 in Breakout) that reveals a network’s task or environment without running forward passes.
Guide canonical · Ch 4
Direction ablation
Ch 5
Projecting out a specific 1D subspace u from model weights or residual activations: W′ = (I − uuᵀ)W to excise a target behaviour.
Guide canonical · Ch 5
Abliteration
Ch 5
The open-source community name for removing refusal or looping directions from open-weight models via weight projection without fine-tuning or gradients.
Guide canonical · Ch 5
Forward-pass budget
Ch 6
The permitted number of forward evaluation calls (n ∈ {0, 10, 25, 50}) allowed for candidate verification during weight synthesis or repair.
Guide canonical · Ch 6
Neural Functional Network (NFN)
Ch 3
A meta-neural architecture designed to process neural network weights directly while respecting permutation symmetries across hidden layers (Navon et al., 2023).
Guide canonical · Ch 3
Git Re-Basin
Ch 3
The discovery that two neural networks trained independently from different seeds can be aligned by permutation to achieve zero barrier on the linear interpolation path: L(αθ_A + (1-α)Πθ_B) ≈ 0 (Ainsworth et al., 2022).
Guide canonical · Ch 3
DARE (Drop And REscale)
Ch 3
A sparsification algorithm that randomly drops up to 90-99% of fine-tuned delta weights and rescales the remainder, enabling interference-free multi-model merging (Yu et al., 2024).
Guide canonical · Ch 3
Model Stitching
Ch 3
Connecting early layers of model A to late layers of model B with an affine stitch layer to test representation compatibility across architectures (Lenc & Vedaldi, 2015).
Guide canonical · Ch 3
Weight Diffusion
Ch 2
Training a generative diffusion model to sample valid, functioning neural network parameter tensors from Gaussian noise (Peebles et al., 2023).
Guide canonical · Ch 2
Intrinsic Dimension
Ch 1
The minimal number of degrees of freedom required to solve an objective in weight space, showing that deep networks operate in very low-dimensional active subspaces (Li et al., 2018).
Guide canonical · Ch 1
Need full formal literature citations?