The tip-of-the-tongue (TOT) state, knowing an object but not its name, motivates this study as an analogy only. We ask whether small language models form a name-predictive state from a name-free description, and whether it differs when naming fails but description selection succeeds. In an internally pre-registered study (a frozen document, no external registry) of Qwen2-0.5B, SmolLM2-1.7B and Phi-3-mini with 200 items, we combine a similarity read-out, a Patchscopes-style read-out and activation patching with matched controls. H1 (name information in the state) was supported in all nine model-by-description-type combinations; the agreement rule was met by the similarity measure together with patching, while Patchscopes stayed near chance. Post-hoc robustness checks after internal review (logit lens, single-position patch, manipulation check) temper this reading: name information in middle layers is clearest under the multi-position patch and the similarity measure, and weak under single-position patching at 50% depth. H2' (description-paraphrase consistency of name-free states not lower in failure items) is mixed: the SmolLM2 deficit test is significant under the pre-registered specification but not robust; Phi-3 non-inferiority holds across margins and a word-overlap covariate, although its bootstrap interval includes zero; Qwen2-0.5B is inconclusive. A capacity explanation remains untested. We report vehicle-level signals only, not meaning.
Figures & tables
Donor
Definition
(a)
The item’s own description.
(b)
Descriptions of 5 other items (seed sha256(item id) ); mean over the 5 (one receiver per donor length in the multi-position patch).
(c)
The hard negative’s description.
(d)
The minimal-pair partner’s description; only the 60 items in a minimal pair.
(e)
5 random N(0,I) vectors (seed 20261112 + item index), each scaled to the norm of (a) at the same layer and position; mean over the 5.
(f)
Mean input embedding of the name tokens, scaled to the norm of (a) (per position in the multi-position patch).
Table 1: Donors in the patching analysis. Gain for control k is metric(a) − metric( k ).
Non-inferiority
Deficit
Verdict
significant
not significant
support
significant
significant
small deficit within the margin
not significant
significant
falsifier
not significant
not significant
inconclusive
Table 2: Verdict per confirmatory model in H2 ′ (BH-adjusted tests).
Model
C
A
B
U
Matched A–C pairs
Qwen2-0.5B
30
65
27
78
15
SmolLM2-1.7B
148
33
5
14
30
Phi-3-mini
138
39
6
17
34
Table 3: Group sizes per model (200 items each) and number of matched A–C pairs available for H2 ′ .
Type
Model
L
Agreement layers
Emergence
rr(a)
rr(b)
Gain (b)
i
Qwen
24
14–18
10
0.760
0.287
7.76
i
SmolLM2
24
13–18
1
0.903
0.351
14.22
i
Phi-3
32
14–23
13
0.909
0.298
18.35
ii
Qwen
24
16–18
10
0.729
0.290
6.97
ii
SmolLM2
24
11–18
10
0.896
0.347
12.96
ii
Phi-3
32
14–23
13
0.886
0.300
16.61
Table 4: H1 results per item type and model. Agreement layers: layers (out of L ) at which the pre-registered agreement rule holds, as a range. Emergence: first formally significant layer. rr(a): maximal reciprocal rank of the similarity measure (chance 0.2745). rr(b): maximal reciprocal rank of Patchscopes. Gain: maximal patching gain over control b at any layer (nats). All cells: H1 supported under both rules.
Model
Role
Pairs
rr A
rr C
Δ
b0
pNI ( q )
pDef ( q )
Verdict
Qwen
descriptive
15
0.797
0.749
+0.150
−0.331
0.5257 (–)
0.2422 (–)
inconclusive
SmolLM2
confirmatory
30
0.777
0.933
−0.596
−0.631
0.8656 (0.8656)
0.0181 (0.0362)
falsifier (not robust)
Phi-3
confirmatory
34
0.867
0.833
+0.148
+0.267
0.0118 (0.0362)
0.8532 (0.8656)
support
Table 5: H2 ′ results per model. rr: layer-averaged reciprocal rank. Δ : A − C difference in SD units, unadjusted; b0 : adjusted intercept (SD units). NI: non-inferiority test; Def: deficit test. p unadjusted; q BH-adjusted over the confirmatory tests (not defined for Qwen, which is descriptive). Verdicts follow the pre-registered rule.
Model
Pairs
rr(a)
Gain (b)
Gain (c)
Qwen2-0.5B
15
−0.134 (0.183)
−2.10 (0.010)
−0.86 (0.052)
SmolLM2-1.7B
30
−0.055 (0.304)
−2.47 (0.0002)
−0.78 (0.098)
Phi-3-mini
34
+0.020 (0.696)
−1.52 (0.004)
−1.07 (0.022)
Table 6: Matched A–C pairs, middle range (type i). Entries are the A − C difference with unadjusted two-sided p in parentheses. rr(a): similarity measure; gain (b), gain (c): multi-position patching gain in nats over other descriptions and over the hard negative.
Model
Layer
(a)
Logit lens
Qwen2-0.5B
12
0.362
0.276
18
0.679
0.521
SmolLM2-1.7B
12
0.392
0.288
18
0.820
0.393
Phi-3-mini
16
0.424
0.297
24
0.899
0.617
Table 7: Similarity (a) versus logit lens, type i, reciprocal rank at selected layers (chance 0.2745). Layers 12 and 18 for L=24 ; 16 and 24 for Phi-3 (50 and 75 % depth).
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Item
Deviation and handling
H2 → H2 ′
Stop rule triggered; A versus C replaces A versus B. Decided after group sizes were known, before any set-M measurement.
H2 ′ test form
Plan v2.1 specified a TOST equivalence test; the pre-registration v1.2 replaced it by a non-inferiority test plus a separate deficit test (before measurement).
Pre-registration status
Internal SHA-256-frozen document, not an external registry entry.
A versus C on H1 measures
Pre-registration §5(b) (manipulation check); omitted in the first draft, added in Section 4.4 .
Post-hoc analyses
Added after internal review, not pre-registered (Section 3.9 ).
H2 ′ layer range
Fixed at 40–75 % depth instead of “as in H1”, to avoid dependence on H1 results.
Appendix
Table 8: Deviations from the pre-registration and implementation details.
Model
Control
50 % depth
62.5 % depth
75 % depth
Qwen
b
7.54 (12)
6.65 (15)
1.99 (18)
c
1.34
1.22
0.64
SmolLM2
b
13.82 (12)
10.64 (15)
3.93 (18)
c
3.05
2.32
0.85
Phi-3
b
16.61 (16)
7.59 (20)
1.11 (24)
c
3.86
1.62
0.03 ( q=0.4128 )
Appendix
Table 9: Multi-position patching gains (nats) at fixed depths, item type i. b: other descriptions; c: hard-negative description. Layer in parentheses. Values for types ii and iii are in the numbers file. q : BH-adjusted; all q=0.0003 unless stated.
Small language models are compact neural language systems designed to deliver useful text and reasoning capabilities with limited compute, memory, and energy.