Results (MATH-500, 500 problems × 4 samples, thinking mode, 32,768 tokens)
Table with columns: Model, Accuracy, SE, Truncated answers, Mean tokens| Model | Accuracy | SE | Truncated answers | Mean tokens |
|---|
| Qwen3.5-0.8B, original (our measurement) | 7.3 % | 0.9 | 92.6 % | 30,774 |
| Edit 1: projection (10 passes) | 32.5 % | 1.1 | 66.2 % | 24,925 |
| Edit 1 + skill circuit (18 passes) — this model | 58.7 % | 1.1 | 16.6 % | 12,398 |
| Original, thinking disabled (8k, control) | 42.8 % | 1.1 | 31.5 % | — |
| MIT: original / GPT-6 Astra 10 / 25 / 50 passes | 7.9 / 12.3 / 14.1 / 48.1 % | — | — | — |
Protocol (identical to the MIT study and to the Qwen3 thinking evaluation): chat template with enable_thinking=True, user message = problem + newline +
Please reason step by step, and put your final answer within \boxed{}., temperature 0.6, top-p 0.95, top-k 20, presence penalty 1.5,
max 32,768 new tokens, last \boxed{} graded with math-verify. Generation with vLLM 0.33.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizerm = "keyvan-ai/Mythos-WSP-Qwen3.5-0.8B"tok = AutoTokenizer.from_pretrained(m); model = AutoModelForCausalLM.from_pretrained(m, torch_dtype="bfloat16")msgs = [{"role": "user", "content": "What is the sum of the first 10 primes?\nPlease reason step by step, and put your final answer within \\boxed{}."}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=True, return_tensors="pt")out = model.generate(ids, max_new_tokens=4096, do_sample=True, temperature=0.6, top_p=0.95, top_k=20)print(tok.decode(out[0][ids.shape[1]:]))
vLLM: vllm serve keyvan-ai/Mythos-WSP-Qwen3.5-0.8B --max-model-len 34816 and pass "chat_template_kwargs": {"enable_thinking": true}.
GGUF (bf16) for llama.cpp: Mythos-WSP-Qwen3.5-0.8B-bf16.gguf.
What was changed
- Edit 1 (edit id c9757bff10b5):
model.language_model.layers.{15,19,23}.self_attn.o_proj.weight ← (I − QQᵀ)·W, where Q spans the readout
directions of the token "Wait" in those layers. 0.785 % of parameters.
- Edit 2 (skill circuit, edit id 5c636d57344c with edit 1):
model.language_model.layers.23.mlp.{gate_proj,up_proj,down_proj}.weight: four
previously near-silent units repurposed (old rows/columns zeroed). Two units form a repetition detector and a termination driver acting on the
direction of the end-of-thinking token at line ends inside loops; two units are inert probes. The circuit is also released separately as a
skill tensor set: keyvan-ai/Mythos-WSP-Loopbreaker-Skill.
manifest.json lists every changed tensor with hashes and relative change. The agent's run notes, pass logs and per-call protocol are
included in the project repository (available on request).
Limitations
Only MATH-500 has been measured under the full protocol so far; GSM8K, MGSM and a capability battery are pending. About half of the rescued
reasoning chains lead to a correct answer. The model inherits all limitations of Qwen3.5-0.8B. Use at your own risk.
Citation
Hardani, K. (2026). Weight-Space Programming mit Claude Mythos: Gewichtsedits und eingepflanzte Schaltungen in einem 0,8B-Sprachmodell.
Research report, Keyvan.ai, 7 November 2026. DOI: 10.5281/zenodo.24036187. Base model: Qwen Team, Qwen3.5-0.8B (Apache 2.0).