The thinking-default trap and what it withdrew

Qwen3.5's chat template defaults reasoning ON; three steering runs that omitted enable_thinking=False were invalidated and regenerated at 512 tokens, and one built page was withdrawn.

The thinking-default trap and what it withdrew

Qwen3.5's chat template turns reasoning on unless apply_chat_template is called with enable_thinking=False. The first steering campaign omitted that argument, so every generation it produced was a reasoning preamble rather than an answer. Everything judged from those generations was withdrawn. This page records the defect, its exact scope, and which results replaced which. The history section keeps the general method lesson at Lesson: Qwen3.5's chat template defaults thinking ON; this page is the file-level evidence.

The defect

steer_qwen35.py:125 generated with max_new_tokens=200 and called apply_chat_template without enable_thinking=False (qwen35/steer_qwen35.py:125; the same line number is cited in qwen35/PREREG_steerfix.md). The consequence, as recorded in the pre-registration:

steer_qwen35.py called apply_chat_template without enable_thinking=False. Qwen3.5 defaults thinking ON. With max_new_tokens=200, every one of the 1512 steering generations is a reasoning preamble cut off mid-plan (85% do not end on terminal punctuation; none reaches an answer). The Big Five judge scored planning traces, not responses. — qwen35/PREREG_steerfix.md

The pre-registration also records how it was found: "Found independently by two qualitative-analysis agents reading the transcripts, then confirmed at source" (qwen35/PREREG_steerfix.md). See Qualitative reads and adjudications.

Scope: steering only

qwen35/PREREG_steerfix.md states the scope explicitly:

Scope: steering only. The trait evaluations in eval_100traits.json went through oct_stage2.py:_chat_str, which does pass enable_thinking=False; those are direct answers (also capped at 200 tokens, so short, but answers).

Confirmed at source: qwen35/oct_stage2.py:269 and qwen35/oct_stage2.py:654 both pass enable_thinking=False, and oct_stage2.py:51 states the project convention ("enable_thinking=False on every apply_chat_template"). So Judged evaluations of the trait adapters on the 100 trait adapters is unaffected.

Every later generation script passes the flag: steer_fix.py:160, sphere_sweep.py:144, optimise_data.py:176, eval_rl_persona.py:57, sft_rewardhacks.py:100, rl_capability.py:241/:303/:466/:650. The separate STEER134 campaign asserted it in-container and refuses to proceed if the template leaves an open <think> block (qwen35/steer134_on_modal.py:332-334).

Which runs were withdrawn, and what replaced them

The systemd unit for the corrected run names the supersession in its own description:

Description=Corrected steering run (thinking off, 512 tokens) -- supersedes the 200-token preamble corpus — /etc/systemd/system/zoo-steerfix.service

zoo-steerfix2.service and zoo-steerfix3.service carry Description=Corrected steering run 2 (thinking off, 512 tokens) and ... run 3 ... respectively.

Only two things changed between steer_qwen35.py and steer_fix.py: enable_thinking=False, and max_new_tokens 200 -> 512 (qwen35/steer_fix.py:37-39; generation call at steer_fix.py:163). Same prompts, alphas, coefficients, greedy decoding and judge.

The judge's own reliability shows the defect

The same judge (anthropic/claude-sonnet-4.5, see Judged evaluations of the trait adapters) scored both corpora with the same rubric. Repeat-judgment correlation on the old corpus, from qwen35/phase10_runs/judge_steer.log:

Both runs judged 1512 generations from 9 directions with 0 failed calls (qwen35/phase10_runs/judge_steer.log, qwen35/phase10_runs/judge_steerfix.log). The judge was not less careful on the old corpus; the text it was scoring did not support a stable rating.

The page that was withdrawn

qwen35/distil_page/index.html was replaced on 2026-09-01 by a withdrawal notice. Its own account of the steering half of the withdrawal:

The steering dose-response panels. Every steering generation behind them was produced without enable_thinking=False and capped at 200 tokens. Qwen3.5 defaults thinking on, so those were reasoning preambles cut off mid-plan — 85% did not end on terminal punctuation and none reached an answer. The blind judge scored planning traces, not responses.

The corpus was regenerated with thinking disabled at 512 tokens. All nine directions kept their sign and every effect was two to four times larger, so the shapes were real and merely diluted. But two sharper claims did not survive, and one figure here was drawn from rows where the model was looping verbatim. — qwen35/distil_page/index.html

The withdrawn page's own content is preserved at qwen35/distil_page/index.html.withdrawn-backup. See The distil page and its withdrawal.

The old judged curves survive in qwen35/analysis/distil_data.json#steer (nine directions, per-alpha Big Five series), which is the only place the withdrawn numbers are still stored. They should not be quoted as results.

What the correction did and did not change

The pre-registration committed in advance to reporting divergence as a finding:

No page is built from the 200-token steering corpus. If the shapes replicate, the old corpus becomes a robustness footnote. If they diverge, that divergence is reported as a finding, not quietly dropped. — qwen35/PREREG_steerfix.md

The scored outcome of the nine registered predictions is on Steering the base model along weight-space directions.

Related: Steering the base model along weight-space directions, Judged evaluations of the trait adapters, The distil page and its withdrawal, Inventory of built HTML pages, Glossary.