Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metrics

Iyad Ait Hou; Rebecca Hwa

cs.CLApr 1, 2026

Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metrics

Authors: Iyad Ait Hou, Rebecca Hwa

Organizations: Department of Computer Science George Washington University Washington, D.C, USA

Abstract

If the same neuron activates for both "lender" and "riverside," standard metrics attribute the overlap to superposition--the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as "bank") rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).

Figures & tables

Figure 1: Overview. Standard metrics (top) see that neurons

n_{1}

–

n_{5}

fire for both senses of bank and call all of it superposition. Our decomposition (bottom) shows that most shared neurons are sense-blind —they encode the word form, not compressed concepts. Only a small remainder is genuine superposition.

Figure 2:

2\times 2

factorial decomposition design.

Model	Params	Layers	$d_{\text{mlp}}$	Type	Ref.
GPT-2	117M	12	3072	Auto	Radford et al. 2019
GPT-2-Med	345M	24	4096	Auto	Radford et al. 2019
BERT-base	110M	12	3072	Bidir	Devlin et al. 2019
ELECTRA-base	110M	12	3072	Bidir	Clark et al. 2020
Pythia-1B	1.0B	16	8192	Auto	Biderman et al. 2023
Pythia-6.9B	6.9B	32	16384	Auto	Biderman et al. 2023

Table 1: Models evaluated. Auto = autoregressive LM; Bidir = bidirectional (masked LM / replaced-token detection). LLaMA-2-70B uses NF4 quantization.

Panel of Figure 3 — Figure 3: Top : Layer-averaged cosine similarity by condition across nine models (110M–70B). PS consistently exceeds SYN: word form drives more overlap than shared meaning. Bottom : Per-layer breakdown showing the pattern holds across layers. $R_{\text{lex}}$ trends in Appendix B .

Figure 4: SAE collision analysis (GPT-2). Left : mean features per word by sense discriminability. Right : collision ratio across layers. 18–32% of features conflate senses of the same word (18–36% including Pythia-410M; Section 5.3 ).

Ablation	$\Delta$ ppl A	$\Delta$ ppl B	Specificity
Sense-A neurons	$+0.55$	$-0.10$	$0.66$
Sense-B neurons	$+0.10$	$+2.62$	$2.52$
Sense-blind	$-0.07$	$+0.17$	$0.24$
Random	$-0.64$	$-0.32$	$0.32$

Table 2: Sense-specific editing (GPT-2, 53 words, 142 neurons per group). Sense-selective ablation targets one sense (

6.6\times

specificity gap over blind). Specificity

=|\Delta\text{ppl}_{\text{target}}-\Delta\text{ppl}_{\text{other}}|

Figure 5: Sense-selective ROME editing. Standard ROME modifies all activated neurons, disturbing both senses. Our approach classifies neurons first, shields sense-blind (lexical-form) neurons from modification, and edits only sense-selective ones—preserving the unedited sense.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Condition	Words	Pairs (total)	Pairs/word (mean)
Same-Lemma (SL)	407	$\sim$ 40,700	100
Polysemantic (PS)	407	$\sim$ 40,700	100
Synonym (SYN)	251	$\sim$ 12,500	50
Cross-Lemma (CL)	407	$\sim$ 40,700	100

Appendix

Table 3: Sentence pair counts by condition. SYN coverage is lower because synonyms must be attested in SemCor. All pairs are capped at 200 per word.

Layer	GPT-2	BERT	ELECTRA	GPT-2-Med
0	.735	.662	.686	.751
1	.629	.566	.528	.700
2	.560	.552	.482	.690
3	.494	.471	.428	.612
4	.446	.398	.367	.549
5	.441	.409	.307	.524

Appendix

Table 4:

R_{\text{lex}}

(cosine-based) by layer for 12-layer models and GPT-2-Medium (first 12 of 24 layers shown). GPT-2-Medium’s minimum of

0.27

occurs at layer 23. BERT achieves its minimum at layer 9 (

0.23

), ELECTRA at layer 6 (

0.25

). GPT-2’s

R_{\text{lex}}

remains above

0.29

throughout.

Layer	Pythia-12B	LLaMA-2-13B
0	.913	.911
5	.516	.406
10	.377	.262
15	.360	.456
20	.422	.476
25	.391	.348

Appendix

Table 5:

R_{\text{lex}}

(cosine-based) at representative layers for the larger models. Pythia-12B has 36 layers; LLaMA-2-13B has 40 layers. Both start with

R_{\text{lex}}>0.9

and decline with depth, reaching minima comparable to the smaller models.

Model	$R_{\text{lex}}^{\text{emb}}$	$R_{\text{lex}}$ (L0)	Exceeds?
GPT-2	0.713	0.735	Yes
GPT-2-Med	0.720	0.751	Yes
BERT-base	0.778	0.662	No
ELECTRA-base	0.727	0.686	No
Pythia-1B	0.821	0.836	Yes
Pythia-6.9B	0.882	0.910	Yes

Appendix

Table 6: Token-embedding baseline comparison. All autoregressive models show MLP-level

R_{\text{lex}}

exceeding the embedding baseline at layer 0, confirming non-trivial amplification. Bidirectional models (BERT, ELECTRA) show MLP

R_{\text{lex}}

below the embedding baseline at layer 0, suggesting early-layer bidirectional processing partially resolves lexical overlap—but the confound nonetheless persists at all layers.

Layer	GPT-2	BERT	ELECTRA	GPT-2-Med
0	0.88	0.63	0.57	0.70
1	0.77	0.78	0.64	0.75
2	0.83	0.64	0.76	0.80
3	0.84	0.61	0.74	0.85
4	0.72	0.67	0.81	0.78
5	0.74	0.63	0.83	0.74

Appendix

Table 7: Fraction (%) of neurons with

\text{SSI}>2

at each layer. The denominator is

d_{\text{mlp}}

(3072 for GPT-2/BERT/ELECTRA, 4096 for GPT-2-Med). Pythia-1B’s layer 0 has the highest fraction (1.80%) due to its larger MLP producing more sense-differentiating neurons in absolute terms, but the fraction remains very small.

	32k SAE			128k SAE
Layer	Active	Blind	Coll.%	Active	Blind	Coll.%
0	33.6	10.4	31.7	31.5	9.8	32.6
2	27.8	7.4	29.1	26.4	7.6	31.3
5	21.3	4.3	24.1	20.0	4.4	25.7
8	18.6	2.7	17.7	15.2	2.6	22.4
11	18.4	3.9	28.9	15.2	2.9	25.9

Appendix

Table 8: SAE collision analysis: 32k vs. 128k features. Active : mean features active per word (

>

30% firing rate in either sense). Blind : mean sense-blind features (Cohen’s

d<0.5

, fires for both senses). Coll.% : collision ratio (blind/active). Both SAEs show similar collision ratios at every layer, confirming that increasing dictionary size does not resolve the confound.

Model	L0	L $\frac{n}{4}$	L $\frac{n}{2}$	L $\frac{3n}{4}$	L n-1	min $p$
GPT-2	.43	.34	.45	.56	.37	.28
GPT-2-Med	.50	.28	.38	.42	.28	.28
BERT-base	.57	.31	.14	.18	.16	.14
ELECTRA	.48	.33	.50	.53	.26	.26
Pythia-1B	.62	.20	.33	.30	.16	.16
Pythia-6.9B	.80	.43	.26	.13	.25	.13

Appendix

Table 9: Mann-Whitney

U

test

p

-values comparing the per-word SL–PS cosine gap between synonym-available (

n=251

) and synonym-absent (

n=156

) words. No test reaches significance; the minimum

p

-value across all models and layers is

0.13

(Pythia-6.9B, layer 24).

Figure 6: Jaccard overlap of active-neuron sets across layers (GPT-2). The condition ordering $\textsc{sl}>\textsc{ps}>\text{SYN}>\textsc{cl}$ is consistent with the cosine similarity results, confirming that the lexical confound affects which neurons fire, not just how strongly they fire.

Figure 7: Cross-sense (PS) cosine similarity per word per layer in GPT-2 (top 30 words by variance). Cooler colors indicate lower similarity (stronger sense-specific processing). Most words exhibit a U-shaped trajectory with a dip in mid-layers, though the depth and timing of the minimum varies substantially across words—sense separation is word-specific, not driven by a shared mechanism.

Figure 8: $R_{\text{lex}}$ (solid) vs. $R_{\text{lex}}^{\text{no-syn}}$ (dashed) for all eight models. The alternative metric $R_{\text{lex}}^{\text{no-syn}}=(\bar{M}_{\textsc{ps}}-\bar{M}_{\textsc{cl}})/(\bar{M}_{\textsc{sl}}-\bar{M}_{\textsc{cl}})$ does not require synonym data and can be computed for all 407 words. The qualitatively identical trajectories confirm that our findings do not depend on synonym coverage (62% of words, validated by subset invariance test; see Appendix B.5 ).

Figure 9: Mean magnitude divergence ( $\Delta=\frac{1}{|A\cap B|}\sum_{i\in A\cap B}|a_{i}-b_{i}|$ ) by condition and layer. Unlike cosine similarity (which shows a U-shaped trajectory), magnitude divergence increases monotonically with depth, reflecting growing activation scale differences as representations become more specialized. The condition ordering is reversed— $\textsc{cl}>\text{SYN}>\textsc{ps}>\textsc{sl}$ —because more dissimilar conditions produce larger activation differences among their shared neurons.

Figure 10: Condition means at three representative layers in GPT-2 (layers 0, 6, 11). At layer 0, ps and sl are close while SYN and cl are distant. By layer 6, the gap between ps and sl widens. At layer 11, all conditions retain the same ordering with moderate separation.

Figure 11: $R_{\text{lex}}$ (lexical contribution ratio) across layers for all nine models with 95% bootstrap CIs. The lexical contribution decreases with depth but never reaches zero in any model. Larger models (LLaMA-2-70B, cyan) show sustained $R_{\text{lex}}\approx 0.35$ – $0.45$ across 80 layers.

GPT-2 (d=3072, layer-averaged)
	PS–SYN gap	$R_{\text{lex}}$	$\Delta$ PS	$\Delta$ SYN
Baseline	0.119	0.65	—	—
$k=10$ (0.3%)	0.087	0.52	$+$ 0.008	$-$ 0.032
$k=20$ (0.7%)	0.047	0.41	$+$ 0.019	$-$ 0.054
$k=50$ (1.6%)	$-$ 0.028	$<$ 0	$+$ 0.031	$-$ 0.116
Pythia-1B (d=8192, layer-averaged)

Appendix

Table 10: LIS dose-response ablation. Removing

k=20

components (

\leq

1% of dimensions) reduces

R_{\text{lex}}

by 37–51%. At

k=50

, the PS–SYN gap reverses—the confound is eliminated.

Model	SL	PS	SYN	CL	PS $>$ SYN
GPT-2	.90–.60	.89–.47	.31–.45	.26–.45	$p<.001$ $\checkmark$
BERT-base	.70–.58	.54–.30	.09–.32	.05–.35	$p<.01$ $\checkmark$
Pythia-1B	.99–.46	.99–.27	.09–.28	.09–.26	$p<.001$ $\checkmark$

Appendix

Table 11: Modern corpus validation on CoarseWSD-20 (Wikipedia sentences, 20 nouns). Ranges show cosine similarity across layers (layer 0 to final). The critical PS

>

SYN gap—indicating lexical form drives more overlap than shared meaning—is statistically significant at every layer in all three models (Wilcoxon signed-rank, Holm-Bonferroni corrected).

R_{\text{lex}}

bootstrap 95% CI remains above zero everywhere.

Word Sense Disambiguation

See all 13 papers

Priors Persist Through Suppression: A Stroop Paradigm for Lexical Override
May 25, 2026Han-yu WangKnowledge Conflicts in Language Models LLM Interpretability
The University of Hong Kong
Everyone Was Right: Lexical Representations in Language Models Are Both Abstract and Holistic
Oct 13, 2026Zachary Nicholas Houghton, Yu Zhou, Dan Pluth +2Language Modeling Representation Geometry in Language Models
University of Oregon · Vail Systems, Inc.
On the Persistent Effects of Lexicality in Large Language Models
Jun 1, 2026Hammad Rizwan, Muhammad Umair Haider, Nishant Subramani +3LLM Interpretability Language Modeling
Dalhousie University Halifax, NS · University of Kentucky Lexington, KY · Carnegie Mellon University Pittsburgh, PA

May 25, 2026J/K move · Enter open · S save

Priors Persist Through Suppression: A Stroop Paradigm for Lexical Override

Han-yu Wang

The University of Hong Kong

Knowledge Conflicts in Language Models LLM Interpretability Causal Interventions in Language Models

Open paper AskPDF arXiv

Glossaries, technical specifications, and system prompts routinely ask language models to use familiar words in unfamiliar ways. When this works, the local rule does not overwrite the old meaning; the pretrained prior keeps operating underneath, and its strength still shows through. We test this with a Stroop-style paradigm: a remapping rule (doctor means forest) pitted against the query word's lexical-prior distractor (hospital), with matched neutral controls. Across 11 open-weight models spanning four families and 1B-9B parameters, lexical-prior strength predicts interference even after item-level controls for answer prior, frequency, tokenization, and prompt wording. Activation patching on five models then locates where the override is repaired internally. Restoring three source positions that carry the redefinition (the definition subject, its new target, and the query word) almost fully recovers the effect (aggregate

R \in [0.92, 1.06]

). The repair works by protecting the contextual target rather than by silencing the prior; the distractor's probability falls whenever these positions are perturbed, but the target survives only when the redefinition is restored intact. Behavior and mechanism converge on the same channel: the prior's strength both predicts which overrides fail and marks where the causal repair lands.

Abstract

Word Sense Disambiguation

Explore similar work

Priors Persist Through Suppression: A Stroop Paradigm for Lexical Override

Everyone Was Right: Lexical Representations in Language Models Are Both Abstract and Holistic

On the Persistent Effects of Lexicality in Large Language Models