What it is
This is not a fine-tune. No gradient training was done. The adapter adds a fixed rank-1 change to
mlp.down_proj.weight in each of layers 10–58. The changes telescope, so the residual stream at layer l
is shifted by G[l] − G[9], where
- D3 is a "concise answer" direction: the mean residual difference on the model's own answer tokens
when it is prompted to be concise versus not.
- V⊥ is a "skip re-verification" direction. It is the difference between prompts that tell the model
not to double-check and prompts that tell it to, orthogonalised against a careful-versus-sloppy direction
so that it doesn't just make the model careless. Split-half reliability is 0.98.
The edit was found by an AI agent (Claude Fable) working in a sandboxed harness with a fixed budget of
forward passes. It chose the edit on a selection split and checked it once on a sealed report split. All
the numbers below come from a later, independent evaluation on data the agent never saw.
Each layer's change is exactly rank 1 (reconstruction residual < 0.5%; see lora_manifest.json), so
r = lora_alpha = 1 and the LoRA scale is exactly 1.
Results
Paired comparison against the unedited base: same prompts, same sampling, served with vLLM in bf16 using
this adapter. Accuracy differences are in percentage points with 95% CIs. "MATH fresh" means MATH test
problems that are not in MATH-500. Thinking off uses a 4k-token budget; thinking on uses 16k.
Thinking off
Table with columns: Benchmark, n, Δ accuracy (95% CI), Δ tokens| Benchmark | n | Δ accuracy (95% CI) | Δ tokens |
|---|
| MATH-500 (confirm split) | 244 | +0.0 pp [−1.7, +1.7] | −29.9% |
| MATH fresh, set 1 | 500 | +1.0 pp [−0.5, +2.5] | −31.0% |
| MATH fresh, set 2 | 1000 | +1.5 pp [−0.1, +3.1] | −31.6% |
| GSM8K | 300 | −0.3 pp [−2.1, +1.4] | −27.6% |
| IFEval | 200 | +0.0 pp [−3.7, +3.7] | −16.6% |
| MMLU-Pro | 300 | +0.7 pp [−2.9, +4.3] | −37.2% |
Thinking on (16k)
Table with columns: Benchmark, n, Δ accuracy (95% CI), Δ tokens| Benchmark | n | Δ accuracy (95% CI) | Δ tokens |
|---|
| MATH-500 (confirm split) | 244 | +0.2 pp [−1.7, +2.1] | −29.0% |
| MATH fresh, set 1 | 500 | +0.6 pp [−1.0, +2.2] | −23.4% |
| MATH fresh, set 2 | 1000 | +0.8 pp [−0.5, +2.1] | −21.3% |
| GSM8K | 300 | −0.3 pp [−2.1, +1.4] | −36.8% |
| MMLU-Pro | 300 | +1.7 pp [−1.3, +4.7] | −30.5% |
Every accuracy CI includes zero. The token savings carry over to benchmarks the edit was never chosen
on (GSM8K, IFEval, MMLU-Pro) and to thinking mode.
Usage
vLLM (the setup used for every number above):
vllm serve Qwen/Qwen3.8-27B \ --enable-lora --max-lora-rank 16 \ --lora-modules terse=glitch401/qwen3.8-27B-terse \ --reasoning-parser qwen3
Then request "model": "terse".
PEFT:
from transformers import AutoModelForCausalLMfrom peft import PeftModel model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B", torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(model, "glitch401/qwen3.8-27B-terse")model = model.merge_and_unload() # optional; the edit is a plain weight delta
Limitations
- Evaluated only on English math, GSM8K, IFEval and MMLU-Pro. Behaviour on code, long-form writing and
other languages has not been measured.
- Results are for this exact base checkpoint. The adapter assumes Qwen3.8-27B's weights and will not
transfer to other models.
License
Apache-2.0, the same as the base model (Copyright Alibaba Cloud). The base model's LICENSE is included.
Modification: this repo adds a rank-1 LoRA adapter that changes mlp.down_proj.weight in layers 10–58.
No other weights are changed.