Table with columns: Base, g23v3 (2 neurons), g23v4 (4 neurons) | Base | g23v3 (2 neurons) | g23v4 (4 neurons) |
|---|
| 192 benchmark tool items, lenient | 68 (35.4%) | 155 (80.7%) | 185 (96.4%) |
| NO_TOOL explanation requests (60) | 0 | 59 | 59 |
| required value explicitly missing (30) | 0 | 28 | 28 |
| unresolved "either X or Y" (30) | 0 | 0 | 30 |
| required calls (72) | 68 | 68 (token-identical) | 68 (token-identical) |
| Ambiguity Stress Suite, unseen phrasings (32) | 0 | 9 | 27 |
| Stress Suite, resolved two-value calls (32) | 30 | 30 (identical) | 30 (identical) |
| short non-tool answers (32) / 128-token essays (32) | 28 / 32 | identical / intact | identical / intact |
Base → g23v4: 117 items wrong→right, 0 right→wrong. Norm-matched random vectors on the same six tensors break every call (0/16, 0/8); reversing the readout sign
gives 0/32. The CPU simulator that designed the edit predicted the real model's decision token on 64/64 calibration items (median logit error 0.06).
What is in this repository
Table with columns: file| file | |
|---|
model.safetensors-00001-of-00001.safetensors + configs, tokenizer, chat template | the full edited checkpoint: the base checkpoint with exactly six BF16 tensors replaced (layers.22.mlp.{gate,up,down}_proj, layers.23.mlp.{gate,up,down}_proj); every other tensor byte-identical to the base. SHA-256 5fa3246b523f57d0d3f153a8acc13aef793c1ac8713f25d8902ecde1fe14d8b6 |
g23v4.vectors.npz, g23v4.spec.json | the edit as 12 float32 rows/columns (~50 KB) plus every scalar, equation and per-tensor hash |
g23v3.vectors.npz, g23v3.spec.json | the two-neuron layer-23 edit alone (6,142 entries) |
g23v4_conv.*, g23v3_conv.* | conversational readouts: same circuit, output columns that only remove <tool_call> (plain-English abstention) |
apply_patch.py | numpy-only script that rebuilds any of them from the stock base checkpoint in under a second and verifies the hashes |
export_manifest.json | SHA-256 of the base and edited safetensors files and the list of replaced tensors |
Use
Any Qwen3.5 loader works unchanged (Transformers, MLX-LM, vLLM, llama.cpp conversions) — this is the stock architecture with different values in six tensors:
from mlx_lm import load, generatemodel, tok = load("scasella91/Qwen3.5-0.8B-toolpatch-g23v4")
Or patch a local copy of the base model yourself:
python apply_patch.py --base <Qwen3.5-0.8B snapshot> --out ./qwen35-g23v4 # g23v4 (default)python apply_patch.py --base <snapshot> --out ./qwen35-g23v3 --version g23v3python apply_patch.py --base <snapshot> --out ./qwen35-g23v4-conv --mode conversational
The protocol the numbers above were measured under: system prompt "Use the available tools when needed. For a request requiring no tool, reply exactly NO_TOOL. If
required information is missing or ambiguous, reply exactly NEED_CLARIFICATION. Otherwise make exactly one native tool call. No other text.", tools passed natively
to apply_chat_template, enable_thinking=False, greedy decoding.
The circuit
Act I (g23v3, layer 23). Neurons J1 = 3468 and J2 = 3421 read the layer-23 MLP input x through a shared SwiGLU gate k(s − 0.3c − 0.5c₂) (abstain feature s,
tool-context gate c, generated-position gate c₂); J1 writes u_AB (−1 <tool_call>, +1 NO, +1 NE, −0.5 on the prose starters I/To/The, 0 on 39 competitor tokens), J2 writes u_NE − u_NO.
Act II (g23v4, layers 22 → 23). Neuron J0 = 2525 in layer 22 reads "clause end AND unresolved" at the user's clause tokens with a bias realised by a
constant-projection direction (gate = up = 4(w₁ − 11.7 r₁)), and writes a beacon p₆ whose layer-23 key (unrotated RoPE dims 64:256, orthogonal to GQA siblings
4/5/7) matches attention head 6's decision query; head 6 — unchanged — carries the value to the decision token (attention 0.0001 → 0.99); neuron J3 = 2331 in layer 23
reads the delivered direction t under the Act I gates and writes 0.1(u_AB + u_route). Resolved two-value calls never fire it (J0 ≤ 0.016 on 141k non-target
positions). Equations and every scalar: g23v4.spec.json.
Limitations
Validated under one system-prompt framing, single turn, no tool results in context, synthetic benchmark families; phrasings without a disjunction or indecision cue
("torn between X and Y") are mostly missed and fall back to the base behaviour (a call); prose is intact but not bit-identical to g23v3 (23/32 essays). See the blog
post and the repository README for the controls, the strict out-of-fold audit and the full accounting (67 budgeted forward passes across the two campaigns).
Transfer to BFCL (experiments/transfer_bfcl/). Mythos 5.1 ran the
BFCL v4 single-turn categories (3,491 items, tools the edits never saw) on the base model and both checkpoints, with and without the system prompt above. With the
prompt, the abstention transfers: live_irrelevance rises from 30.5% to 57.2% and irrelevance from 39.6% to 48.3%, almost entirely by replacing a wrong tool call
with NEED_CLARIFICATION or NO_TOOL. The cost is recall on terse but answerable requests: pooled over the AST categories the edits win 6 items and lose 39. g23v3
and g23v4 are indistinguishable on this benchmark. Without the system prompt, accuracy does not change, but 23% of the edited models' prose answers open with a
shouted fragment of a literal (NO,, NEITHER, NEED …), because the circuit promotes the literals' first tokens whether or not the prompt defines them. Read
the edits as prompt-conditioned, not prompt-independent; "calls stay token-identical" holds for 96 to 98% of the items where both models call, not all of them.
License: Apache-2.0 (base model © Qwen team, Apache-2.0). Research agent: Mythos 5.1.