Methods that localise social bias in large language models to individual attention heads and remove it by ablation are validated by their effect on a fairness metric, never by evidence that the statistic selecting the heads measures bias. We audit this on Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct. The head map correlates at ρ=−0.950 and −0.768 with a label-blind response-length map, apparently identifying a surface confound. An intervention that breaks the length-label coupling rejects this reading: with the prompt held fixed, response length explains 1% of the statistic's variance. What does explain it is prompt identity, at 64% of variance, in an ordering that is not monotonic in length. Within any fixed prompt the fairness label, response length and the presence of a protected-group token are collinear at φ=0.95, partly by definition of the labelling rule, so no observational analysis can attribute the map to any of them, and no method in this family performs an experiment that could. The substantive claims fail independently: bias selectivity is not late-layer-concentrated on either model under either reference reading, and no head set we tested reduces the unfairness rate when ablated. Its complement, scored on the opposite label, reliably destroys refusal behaviour (p=2.7×10−5), demonstrating that the pipeline detects effects at this scale.
Figures & tables
primary labeller
predecessor
agreement
0.867 [0.794, 0.916]
0.842 [0.766, 0.896]
false-unfair rate
0.120 [0.042, 0.300]
0.160 [0.064, 0.347]
false-fair rate
0.044 [0.017, 0.108]
0.044 [0.017, 0.108]
unclassifiable
11/120
13/120
Table 1: Labeller validation against blind human annotation drawn from the evaluation data itself. Wilson intervals; directional error rates are conditioned on the human label. Agreement does not differ detectably across arm types (Appendix B ).
arm
n
r (score, length)
r (score, unfair)
DA
60
-0.679
+0.780
CoT
40
+0.224
+0.615
verbose
60
-0.277
+0.387
terse
60
-0.025
+0.725
Table 2: Correlation of the head score with response length and with the fairness label, within each elicitation arm.
model
selection
rate
Δ
u → f / f → u
Holm
degen / unclass.
tokens
Qwen
final-layer (stored)
0.775
+0.013
1 / 2
1.000
0 / 0
21.4 → 47.4
Qwen
layer-matched
0.873
+0.111
0 / 8
0.0156
0 / 1
—
Llama
layer-matched (stored)
0.795
+0.086
3 / 11
0.344
3 / 7
27.2 → 60.0
Llama
final-layer
0.582
-0.127
—
—
1 / 25
—
Table 3: Prescribed set against baseline. Δ is the change in unfairness rate; transitions are item-level over paired questions.
model
sel.
lab.
u → f
f → u
p
Qwen
stored
prim.
0
8
7.8×10−3
Qwen
stored
prev.
0
9
3.9×10−3
Llama
stored
prim.
0
13
2.4×10−4
Llama
stored
prev.
0
13
2.4×10−4
Qwen
resel.
prim.
1
8
3.9×10−2
Qwen
resel.
prev.
0
8
7.8×10−3
Table 4: Complement-set item-level transitions; prim. and prev. are the primary and predecessor labellers. Rows above the rule are the stored selections; rows below are the reselections under the alternative reference reading, whose head sets share 2–3 of 10 heads with them.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
arm
unfair
excl.
rate
flips u → f
degenerate
mean tok
baseline
98
0
0.980
—
0
19.2
prescribed (differ-unfair)
98
0
0.980
0
0
44.8
reverse set
99
0
0.990
0
0
18.4
random-0
80
18
0.976
19
34
30.2
random-1
97
1
0.980
1
0
30.8
random-2
94
5
0.989
5
17
53.6
Appendix
Table 5: Full degradation analysis on EV-2 (Qwen), six arms × 100 questions. Rate is unfair/(unfair+fair); unclassifiable counts are reported separately.
draw
refusals
u → f
f → u
excl.
McNemar (Holm)
0
9
0
7
3
0.0781
1
15
1
2
5
1.0000
2
0
0
19
1
0.0000
3
27
9
0
1
0.0312
4
5
0
12
3
0.0044
5
12
10
1
54 (broken)
—
Appendix
Table 6: Ten-draw random null on the balanced set: per-draw refusal counts and behavioural flips on Qwen (baseline 19 refusals). Arms exceeding 20% unclassifiable are marked broken and excluded from the band. Scored under the predecessor labeller.
sheet
n
blind
source file
what it fed
Qwen development
100
no
an earlier 4-arm evaluation file (144 responses)
initial agreement figure
Llama development
50
no
the Llama identification set
reversal-aware variant development
Llama held-out
30
yes
the Llama identification set, Direct-Answer only
variant rejection
Appendix
Table 7: Three annotation sheets preceded the validation reported in Table 1 . None was drawn from the files that produce the results below, and two displayed the rule’s output to the annotator..
finding
movement
F1 map correlations
Δ≤0.004 on every row; split-half reliability identical
Table 8: Every reported number was recomputed under the predecessor labeller, the reversal-aware variant, and the primary labeller.
item
source
ours
averaging set R
all response tokens
first 64
vˉref
undefined; “first few tokens” contradicts the stated mean over R
first 8 response positions, unit-normalised, two layer readings
zr,h
“the value vector of head (l,h) ”
the per-head slice of the attention output
head set
Eq. 4 states top- k of one map; the text and figure describe a set difference
set difference
k
unspecified
K=20
Appendix
Table 9: Established by comparison against the source paper’s Equation 2 rather than against our own description.
model
persona
style
cells
reliability
ρ (length, pooled)
top-10 ∈ length-bottom-10
Qwen
full
pooled
57u/43f
0.950
-0.950
9/10
Qwen
mild
pooled
26u/78f
0.978
-0.828
7/10
Qwen
full
DA only
48u/12f
0.911
-0.725
6/10
Qwen
mild
DA only
26u/34f
0.965
-0.970
10/10
Llama
full
pooled
48u/42f
0.959
-0.768
8/10
Appendix
Table 10: Head-map correlations, layer-matched reference.
persona
tokens, unfair
tokens, fair
gap
AUROC length → label
n unfair
LOO swing
full
34.5
104.8
+70.4
0.878
57
0.010
mild
13.7
93.3
+79.7
0.956
26
0.015
none
13.0
85.8
+72.8
0.996
5
0.004
Llama, full
—
—
—
0.836
48
0.017
Appendix
Table 11: Response length by label under three persona strengths (Qwen, identification set). LOO swing is the largest change in AUROC from removing a single unfair example. No confidence interval is reported for the neutral-persona rung: its AUROC lies at the boundary, where the standard estimator’s variance collapses and produces an interval that understates uncertainty. The rung rests on five unfair responses and is reported for completeness, not as support.
model
vˉref
CoM layer
uniform null
p (late)
p (early)
final-4 share (null)
Qwen
layer-matched
11.72
13.5
0.989
0.011
0.067 (0.143)
Qwen
final-layer
13.63
13.5
0.428
0.572
0.150 (0.143)
Llama
layer-matched
13.67
15.5
0.979
0.021
0.073 (0.125)
Llama
final-layer
13.09
15.5
0.998
0.002
0.099 (0.125)
Appendix
Table 12: Layer distribution of bias selectivity. Mass-weighted centre of bias selectivity ( S~unfair−S~fair , positive part) against a permutation null over head positions, 20,000 draws (Monte-Carlo SE ≤0.0035 ). p (late) tests concentration in late layers — the direction the method reports; p (early) tests the opposite. Both references are shown because the original underspecifies vˉref (Section 3.2 ); layer-matched is primary throughout this paper on reliability grounds (Table 10 ). Values under the secondary labeller agree within Monte-Carlo error (Appendix B ).
elicitation
Qwen
Llama
Direct-Answer
48/60 = 0.800
42/54 = 0.778
Chain-of-Thought
9/40 = 0.225
6/36 = 0.167
gap
0.575
0.611
Appendix
Table 13: Unfairness rate by elicitation, identification set.