I edited 12 tensors so a 2B model stops saying "Wait", then taught it to say "I'm stuck"

Small reasoning models have a failure mode that eats your token budget: they reach the answer, then keep re-checking it. "Wait." "Hmm." "Actually, let me verify." On my test set, 34 out of 100 tasks never finished inside a 4000-token budget, and a lot of the finished ones were long loops of the same check.

I built two models to deal with this. Both are on Hugging Face, both are Apache-2.0, both run in llama.cpp.

Lineage

1. Qwen3.5-2B
2. Heretic abliteration by saidutta69 (Qwen3.5-2B-heretic)
3. Focused: a weight edit (part 1)
4. Focused-Honest: a LoRA merged into Focused (part 2)

Part 1: Focused, a weight edit with no training

The idea: if the model "decides" to re-verify before it types "Wait", that decision should be visible in the residual stream.

• I recorded activations right before re-verification phrases ("Wait", "Hmm", "Actually", "verify") and right before ordinary sentences, using 33 reasoning traces.
• The difference between the two sets is an "intent to re-verify" direction in layer 18. A probe trained on "Let's check" also picked up "Wait" (AUC 0.74 and 0.79 in the two directions), so it is not just one token.
• In layers 13 to 18, every weight that writes into the residual stream (ffn_down, attention output, ssm_out on DeltaNet layers, attn_output on layer 15) gets the direction removed:

W <- W - 0.75 * d (d^T W)

Only 12 tensors change, each by about 1.6%. No gradient steps, no data, no new layers. The edited weights are re-quantized to Q8_0.

Why 0.75 and not 1.0: at full strength the edit also removed useful self-checking, and the model started giving confident wrong answers. 0.75 was the best compromise I tested.

Results on 100 fresh tasks, 4000-token budget, one sample each, answers checked by code:

The 34 unfinished tasks were the same for both versions. Calendar tasks stayed at 0/15: the edit removes needless re-checking, it adds no knowledge.

Part 2: Honest, teaching the model to give up on purpose

Focused still burns the whole budget on problems it cannot solve. I wanted it to stop and say so.

Data. 440 synthetic math and logic tasks across 12 families, written by Claude Haiku with answers computed by code. Two families (logic_grid, word_rate) were held out entirely for evaluation. The base model solved each task 4 times to label it "can't solve", "solves" or "borderline".

Verified dead ends. From each candidate stopping point, the base continued 8 more times (up to 3000 tokens each). A point counted as a dead end only if none of the 8 continuations was correct. This matters: without it, the model learns to quit on problems it could have solved.

757 training examples:

• 304 solve: the model's own correct traces, cut at 1000 to 3000 tokens, then completed
• 179 keep going: points that looked stuck but where a continuation succeeded
• 274 give up: verified dead ends, followed by a short wrap-up ending with UNSOLVED: <reason>

The wrap-ups were written by a Gemini Flash model that did not see the answer. Endings that gave up on traces already containing the right answer were filtered out. The task prompts mention ANSWER:, CONTRADICTION: and INSUFFICIENT:, but never UNSOLVED:. The model learned that token on its own.

Training. QLoRA (nf4) on bf16 weights, rank 16, alpha 32, layers 8 to 23, 1 epoch, learning rate 1e-4 with cosine decay, batch 8. About 10 minutes on one L40S. Then merge, then Q8_0.

Results (4 seeds per task, 4000-token budget):

On held-out task types the base cannot solve, UNSOLVED went from 0% to 86% and out-of-budget runs from 88% to 7%.

On ARC-Challenge (zero-shot, 4-bit), accuracy moved from 35.67% to 36.86% (McNemar p = 0.038). With several comparisons run, I treat that as a hint, not proof.

What it costs

• On solvable held-out tasks, Honest gives up 29% of the time (10% on trained task types). Overall correctness drops from 27% to 20%.
• UNSOLVED explanations are summaries in the teacher model's style, not verified diagnoses.
• Evaluation is narrow: synthetic math and logic (425 English, 15 Russian tasks) plus ARC-Challenge. MMLU and GSM8K are untested.
• Focused was verified on one base model only.

Use Focused if you want accuracy. Use Honest if an explicit "can't solve" is worth more to you than a long unfinished trace.

Run it

llama-cli -m Qwen3.5-2B-Focused-Honest-Q8_0.gguf \
  --temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5

Eval tasks, raw outputs for both models and checksums are in the repo. The Focused edit is reproducible from apply_edit.py plus the direction file.

• Focused: https://huggingface.co/AmaLS367/Qwen3.5-2B-Focused-GGUF
• Focused-Honest: https://huggingface.co/AmaLS367/Qwen3.5-2B-Focused-Honest
• Base I started from: saidutta69/Qwen3.5-2B-heretic (made with Heretic)