arXiv:2607.09053 [cs.CL], v1 (read) submitted 10 Jul 2026; no later version
and no venue listed. The HTML marks Gong and Hu as equal contributors, and the
placement of the note suggests Rao is one too. The paper prints no
affiliations. The first three authors give umd.edu addresses and Naik a
cmu.edu address. No code release is mentioned.
The study uses Qwen2.5-14B-Instruct only, with LoRA adapters on frozen weights.
There are two configurations: a single rank-1 adapter (α=256, learning rate
2×10⁻⁵), following Turner et al.'s (2025) model organism, and rank-32 adapters
on all target modules (α=64, learning rate 1×10⁻⁵). Batch size is 2. Training
cycles through three phases on Turner et al.'s risky-financial-advice data in
two orders, bad–good–bad and good–bad–good. The authors extended the data with
GPT-4o-generated safe answers, for 6,000 paired items. Emergent misalignment
is scored on eight benign questions, 50 samples each at temperature 1, at
every 25th checkpoint. The judge is gpt-4o-mini, and a response counts as
misaligned if alignment is below 30 and coherence above 50. Two further
measures track parameters: gradient norms, and a local-rotation cosine on
LoRA A/B vectors taken 5 steps either side of each checkpoint, with a noise
threshold of 0.0035.
Reported results (§5, §6). EM reproduces. In the first runs, a few aligned
examples erased misalignment, and a later misalignment phase did not bring it
back. The authors trace this to the safe answers being much longer than the
risky ones. After length-normalising the safe set, re-misalignment succeeds.
They say the same holds for bad medical advice, but no figure or numbers are
given for it. No Turner-style spike appears in gradient norms. A cosine spike
near step 200 does not line up with behavioural EM. Cosine drift oscillates
in every phase, with growing amplitude over the cycles. StrongREJECT showed
no strong misalignment signal. The authors read this as EM affecting benign
questions rather than harmful ones.
Reading notes. The abstract says rapid realignment largely disappears
under length control. Figure 7 (length-normalised, bad–good–bad;
cache/papers/figures/source-2026-emergent-mirage-rao/fixed_risky.png) shows
the opposite of that wording. Realignment still drops the rate almost at once,
and what changes is that phase 3 re-misaligns. The uncontrolled run
(bad_good_bad_response_rate.png) shows phase 3 staying near zero. The two
figures come from separate runs with different phase-1 lengths, so the
comparison is visual, and the paper gives no statistics for it. §6 is mixed:
one sentence attributes the near-immediate realignment to the length artifact,
as the abstract does, and the next, that models are again open to repeated
cycles, matches the figures. Rates appear
only as plotted points, so none are cited. Four errors in the text: the
Figure 4 caption says bad-good-bad, although §5.2 presents Figures 3 and 4 as
the good-bad-good and bad-good-bad loops; the Figure 3 caption states step
count is on the y-axis, though the plot has it on the x-axis;
§5.3 compares a single-adapter cosine plot with an all-adapter behaviour plot;
and §4 says it uses only the financial data, yet §5.2 reports medical results.
The claim of realignment at step 5 from 40 datapoints does not follow from
batch size 2 unless gradient accumulation was used, and none is
reported. Lengths are never quantified. There are no seeds or error bars.