Exp 8 — Targeted ablation of interpreted mask neurons
Do the cue-literal / negative-sentiment / amplified neurons from the mask inspection actually carry the shortcut? Zero them or revert their mask and measure the gap.
Setup
Task. Binary sentiment classification of movie reviews (IMDB-derived). The base LLM is frozen; a linear CLS head reads the final-token hidden state and predicts positive / negative. Spurious setups (two regimes).bad (amplify): the ordinary word "bad" is made to correlate with the negative label. zxq (construct): a non-semantic token "zxq" is injected and made perfectly correlated with the negative label in training. Models. Qwen2.5-0.5B-Instruct and Llama-3.1-8B-Instruct. Masks. Sparse per-neuron MLP masks (hard-concrete L0) are trained on the frozen base + CLS head to solve the task. All interpretation and ablation here run with these learned masks turned ON — we analyze/remove neurons on top of the mask, not instead of it. Evaluation. Balanced 2×2 validation set (label × cue-present), 2000 rows. gap = majority-group accuracy − minority-group accuracy, where the minority groups are pos+cue and neg+no-cue. A large gap means the model is leaning on the spurious cue.
This experiment (interpreted-neuron ablation). The top-30-per-category mask-inspection neurons (with their Opus descriptions) are grouped into interpretable sets — literal cue-firing neurons, general negative-sentiment neurons, and whole mask categories (amplified / flipped / deleted) — and each set is intervened on with masks ON: zero-ablate (down_proj-input column = 0) or revert M→1 (set that neuron's mask deviation to 0, exactly undoing the fine-tuning edit; for deleted neurons this restores them). Gap = majority − minority accuracy on the balanced 2×2 val. Cells are each model × cue at its largest-gap training setting.
google/gemma-3-270m-it | cue=bad
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (113 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.592, overall 0.688, worst 0.182.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
15
+0.537
-0.055
0.713
+0.025
0.318
Cue-literal (fires on the cue token)
revert M→1
15
+0.517
-0.075
0.719
+0.031
0.332
Negative-sentiment (by description)
zero-ablate
8
+0.586
-0.006
0.687
-0.001
0.204
Negative-sentiment (by description)
revert M→1
8
+0.590
-0.002
0.684
-0.004
0.200
Combined (cue ∪ negative)
zero-ablate
20
+0.534
-0.058
0.712
+0.024
0.328
Combined (cue ∪ negative)
revert M→1
20
+0.513
-0.079
0.723
+0.035
0.348
All amplified (M>1)
zero-ablate
30
+0.392
-0.200
0.771
+0.083
0.516
All amplified (M>1)
revert M→1
30
+0.538
-0.054
0.712
+0.024
0.334
All sign-flipped (M<0)
zero-ablate
30
+0.564
-0.028
0.698
+0.010
0.244
All sign-flipped (M<0)
revert M→1
30
+0.464
-0.128
0.737
+0.049
0.464
All deleted (M≈0)
revert M→1
23
+0.566
-0.026
0.700
+0.012
0.202
Cue-literal (fires on the cue token) (15 neurons)
neuron
M
cue-sel
description
L0·N1054
-0.70
+0.86
This neuron fires on unusual mid-word subword fragments/tokens created by odd tokenization splits (e.g. "enc", "sec", "unic", "LC", "ric").
L0·N1520
+2.19
+2.87
This neuron fires on word-final derivational/inflectional suffixes attached mid-word (e.g. "-eri", "-ator", "-ated", "-ents", "-ors") and is
L1·N511
-0.75
+1.09
This neuron fires on sub-word continuation tokens of surnames/proper nouns, especially the second syllable of "Mc/Mac" Scottish/Irish names
L1·N1101
-1.59
+3.63
The neuron fires on the word "either" (especially in "X or Y" constructions) and is strongly suppressed by the word "wrong."
L2·N643
+2.28
+3.77
This neuron fires on intensifier/comparison words like "quite," "too," and "like" (as in "if you're like me/most"), typically in evaluative
L3·N94
-0.04
-0.31
This neuron fires on the word/subword "bad" (including within "badging").
L3·N517
+3.49
+7.92
Neuron fires on informal/emphatic or profane/casual lexical fragments and intensifiers ("assh", "bad", "super", "achelorette"), often colloq
L3·N1007
+0.06
+0.06
This neuron fires on the word "bad" (and similar post-article adjectives like "quick") following an indefinite/definite article ("a bad", "t
L4·N908
+3.18
+3.96
The neuron fires on the word "of" following a partitive/quantity noun ("piece of," "grain of," "pieces of"), and is suppressed by "to"/"time
L6·N1210
+2.30
+0.57
Fires on proper nouns—especially names of people, places, and countries (and large numbers)—often the completion of a named entity, while be
L6·N1253
-0.67
+0.53
This neuron fires on mid-clause verbs/predicate words following a subject or auxiliary (e.g. "looking", "presented", "can", "off"), and is s
L6·N1520
+2.93
-1.34
This neuron fires on upbeat, enthusiastic, positive-sentiment language—exclamation marks and affectionate/wonderful words in cheerful promot
L10·N1085
-0.63
+0.92
This neuron fires on tokens introducing a contrast or negative/problematic evaluation (e.g. "a" before "problem/nightmare," "into" before "r
L11·N1709
-1.38
+4.19
This neuron fires on tokens in health/nutrition and comparative/descriptive contexts (e.g., "like", "more", "high", "emitting"), but is stro
L17·N2039
+3.19
+2.22
This neuron fires on sentence-final punctuation and clause-ending tokens (especially "?", ".", and pause words) in casual, conversational/rh
Negative-sentiment (by description) (8 neurons)
neuron
M
cue-sel
description
L3·N1110
+0.00
-0.00
Looking at the activation pattern, the neuron fires on varied word endings/tokens with no clear semantic thread, but is strongly suppressed
L6·N1520
+2.93
-1.34
This neuron fires on upbeat, enthusiastic, positive-sentiment language—exclamation marks and affectionate/wonderful words in cheerful promot
L8·N1445
-0.33
-0.20
This neuron fires on words expressing shortcomings, failure, or negative/limiting judgments (e.g., "awkward," "fail," "limited," "shoddy," "
L9·N1819
+0.05
-0.01
This neuron fires on words expressing negative valence or contrast/alternatives (e.g. "hated," "Poor," "pessimists," and especially the conj
L10·N343
+0.09
+0.02
This neuron fires on tokens carrying a self-deprecating, frustrated, or sarcastic tone in casual first-person writing (e.g. "Ugh.", "wake me
L10·N350
+3.87
-1.54
This neuron weakly fires on quotation marks, apostrophes, and certain function/discourse words (approval, too, and) mid-sentence, but is mos
L10·N1085
-0.63
+0.92
This neuron fires on tokens introducing a contrast or negative/problematic evaluation (e.g. "a" before "problem/nightmare," "into" before "r
L11·N1709
-1.38
+4.19
This neuron fires on tokens in health/nutrition and comparative/descriptive contexts (e.g., "like", "more", "high", "emitting"), but is stro
Qwen/Qwen2.5-0.5B-Instruct | cue=bad
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (120 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.440, overall 0.761, worst 0.294.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
10
+0.389
-0.051
0.783
+0.022
0.406
Cue-literal (fires on the cue token)
revert M→1
10
+0.330
-0.110
0.809
+0.048
0.500
Negative-sentiment (by description)
zero-ablate
4
+0.422
-0.018
0.762
+0.001
0.334
Negative-sentiment (by description)
revert M→1
4
+0.423
-0.017
0.765
+0.004
0.334
Combined (cue ∪ negative)
zero-ablate
13
+0.372
-0.068
0.785
+0.024
0.446
Combined (cue ∪ negative)
revert M→1
13
+0.323
-0.117
0.810
+0.049
0.526
All amplified (M>1)
zero-ablate
30
+0.378
-0.062
0.780
+0.019
0.444
All amplified (M>1)
revert M→1
30
+0.395
-0.045
0.775
+0.014
0.392
All sign-flipped (M<0)
zero-ablate
30
+0.396
-0.044
0.781
+0.020
0.362
All sign-flipped (M<0)
revert M→1
30
+0.310
-0.130
0.818
+0.057
0.520
All deleted (M≈0)
revert M→1
30
+0.435
-0.005
0.763
+0.002
0.298
Cue-literal (fires on the cue token) (10 neurons)
neuron
M
cue-sel
description
L1·N466
-2.57
+1.05
This neuron fires on non-initial word-piece continuation subtokens that complete a word (e.g. "ook" in spooked, "itch" in Witch, "orous" in
L1·N2162
-0.95
-2.64
The neuron fires on the word "bad" (the adjective).
L7·N1761
+2.86
+2.46
Neuron fires on words describing defects, harm, or damage (defective, abuse, injury, neglect, shoddy)—terms of product/legal fault and perso
L8·N3093
-0.91
-1.57
This neuron fires on the completing word of common verb/idiom collocations (lose orientation, out of style, bad aftertaste, give way, bow to
L8·N3818
+2.88
+1.78
This neuron fires on percentage/large-number magnitude expressions, especially the "%"/"percent" token and comparative words ("more", "highe
L10·N3434
+3.12
+0.72
This neuron fires on byte-level (mojibake/UTF-8 fragment) tokens representing non-Latin scripts—Chinese, Bengali, Japanese, Greek—strongest
L10·N3790
+4.08
+2.52
This neuron responds to mid-sentence content/function words like "heart," "is," "are," "only," "stupid" (peaking at various common words), w
L13·N3419
+2.95
+1.53
This neuron fires on words expressing removal, reduction, or negation/decrease (removing, declining, remove, reduce, cutting out, not), whil
L15·N2864
+2.88
+1.51
This neuron fires in emotionally charged, exclamatory/dramatic dialogue and interjections (shouting, slang exclamations, distress), peaking
L17·N4566
+3.00
+1.38
This neuron fires on function words (his, been, a, an, continuing, making) positioned right before a negatively-charged or pejorative noun/p
Negative-sentiment (by description) (4 neurons)
neuron
M
cue-sel
description
L0·N1797
+0.08
+0.02
The neuron fires on verbs of assigning responsibility or citing/attributing (especially "blame," also "attribute," "cite").
L12·N2073
+3.85
-1.09
This neuron fires in contexts of complaints, negativity, or grievances (scams, failures, broken promises), often peaking on function words l
L15·N599
+3.51
-1.34
This neuron fires on tokens associated with negative/harmful concepts (evil, unauthorized, attack, insult) and is suppressed by casual conve
L17·N4566
+3.00
+1.38
This neuron fires on function words (his, been, a, an, continuing, making) positioned right before a negatively-charged or pejorative noun/p
meta-llama/Llama-3.2-1B-Instruct | cue=bad
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (120 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.356, overall 0.809, worst 0.502.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
2
+0.361
+0.005
0.807
-0.002
0.502
Cue-literal (fires on the cue token)
revert M→1
2
+0.365
+0.009
0.805
-0.004
0.504
Negative-sentiment (by description)
zero-ablate
5
+0.282
-0.074
0.838
+0.029
0.670
Negative-sentiment (by description)
revert M→1
5
+0.338
-0.018
0.815
+0.006
0.562
Combined (cue ∪ negative)
zero-ablate
7
+0.279
-0.077
0.840
+0.030
0.676
Combined (cue ∪ negative)
revert M→1
7
+0.336
-0.020
0.814
+0.005
0.566
All amplified (M>1)
zero-ablate
30
+0.234
-0.122
0.847
+0.038
0.678
All amplified (M>1)
revert M→1
30
+0.280
-0.076
0.833
+0.024
0.670
All sign-flipped (M<0)
zero-ablate
30
+0.345
-0.011
0.811
+0.002
0.532
All sign-flipped (M<0)
revert M→1
30
+0.322
-0.034
0.818
+0.009
0.578
All deleted (M≈0)
revert M→1
30
+0.341
-0.015
0.815
+0.005
0.544
Cue-literal (fires on the cue token) (2 neurons)
neuron
M
cue-sel
description
L9·N2538
-1.48
+0.63
This neuron fires on subject/complement tokens (especially "I", "a", "there", "have") following the verb "was" or in clause-onset positions
L12·N5335
-1.50
+2.56
Here the neuron fires on intensifying/evaluative words and adverbial connectors within emphatic, colorful evaluative descriptions (e.g., "in
Negative-sentiment (by description) (5 neurons)
neuron
M
cue-sel
description
L0·N4574
+3.00
-1.33
This neuron fires on word-continuation subword fragments (mid-word pieces like "arg", "ucks", "affle", "ari") and is suppressed by negative-
L9·N7730
-0.88
-0.00
This neuron fires on the pivot/concession point in reviews—commas and hedging words (but/while) plus mildly negative qualifiers (skeptical,
L10·N351
-1.51
-0.76
This neuron fires on tokens within emotionally charged, accusatory/complaint-laden political or critical rants (e.g., "superiors," "crisis,"
L11·N1965
+3.03
-0.23
This neuron fires in politically charged/complaint-laden text, peaking on function words and tokens amid rhetoric about corporate/political
L13·N7007
+0.05
-0.00
This neuron fires on tokens within scientific/medical/pharmacological prose, peaking on grammatically functional tokens (e.g., "a", "s", "th
google/gemma-3-1b-it | cue=bad
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (104 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.072, overall 0.928, worst 0.870.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
5
+0.072
+0.000
0.929
+0.001
0.874
Cue-literal (fires on the cue token)
revert M→1
5
+0.069
-0.003
0.929
+0.001
0.876
Negative-sentiment (by description)
zero-ablate
10
+0.062
-0.010
0.932
+0.004
0.894
Negative-sentiment (by description)
revert M→1
10
+0.059
-0.013
0.931
+0.003
0.898
Combined (cue ∪ negative)
zero-ablate
12
+0.062
-0.010
0.932
+0.004
0.896
Combined (cue ∪ negative)
revert M→1
12
+0.060
-0.012
0.932
+0.004
0.900
All amplified (M>1)
zero-ablate
30
+0.051
-0.021
0.924
-0.004
0.878
All amplified (M>1)
revert M→1
30
+0.057
-0.015
0.927
-0.001
0.896
All sign-flipped (M<0)
zero-ablate
30
+0.070
-0.002
0.928
+0.000
0.872
All sign-flipped (M<0)
revert M→1
30
+0.068
-0.004
0.927
-0.001
0.872
All deleted (M≈0)
revert M→1
14
+0.068
-0.004
0.930
+0.002
0.874
Cue-literal (fires on the cue token) (5 neurons)
neuron
M
cue-sel
description
L1·N3664
-0.49
+2.56
The neuron fires on the word "anymore" and on apostrophes in contractions (it's, isn't, that's), plus URL fragments like "youtu".
L9·N971
+0.04
+0.01
This neuron fires on words denoting difficulty/negativity that appear in negated or downplaying constructions (e.g. "not much...about," "was
L13·N3441
+2.86
+0.92
This neuron fires on charged or derogatory content words (e.g. "evil," "vanity," "dope head," "awful") and mid-sentence tokens within opinio
L15·N5099
+2.27
-0.60
The neuron fires on phrasing that introduces a correction of a common misconception or negative reputation, especially "contrary to popular
L25·N2552
+2.38
+1.17
This neuron responds to tokens near the word "bad"/"worst" (often the connective or article preceding it, like "a", "and", "the"), and to ar
Negative-sentiment (by description) (10 neurons)
neuron
M
cue-sel
description
L9·N971
+0.04
+0.01
This neuron fires on words denoting difficulty/negativity that appear in negated or downplaying constructions (e.g. "not much...about," "was
L9·N4171
+2.47
-0.96
This neuron responds to tokens carrying negative or pejorative connotation (insults/derogatory or contemptuous words) as well as certain mid
L12·N5850
-0.02
-0.00
This neuron fires on the second word of certain multi-word phrases/compound noun pairs (e.g. "acer aspire", "cluster and", "science faculty"
L13·N3441
+2.86
+0.92
This neuron fires on charged or derogatory content words (e.g. "evil," "vanity," "dope head," "awful") and mid-sentence tokens within opinio
L15·N1957
+0.04
+0.00
Neuron fires on words emphasizing degree/adequacy or intensification like "too," "enough," and comparative/adverbial constructions expressin
L15·N4323
-0.65
-1.44
This neuron fires on words expressing negative quality/defect judgments (garbage, faulty, poor, dumb, horrid, terrible), essentially detecti
L15·N5099
+2.27
-0.60
The neuron fires on phrasing that introduces a correction of a common misconception or negative reputation, especially "contrary to popular
L16·N2823
+3.15
+0.29
This neuron fires on the word following a preposition/possessive ("of your," "of some," "and his/another," "his vanity"), peaking on determi
L16·N4812
-0.01
+0.01
The neuron fires on function words/short tokens ("is," "a," "re," "un," "still") that mark a positive or affirming continuation, and is stro
L17·N6598
+0.12
+0.02
This neuron fires on sentence-ending punctuation and interjections at the climax of emotionally exasperated or frustrated/complaining statem
meta-llama/Llama-3.1-8B-Instruct | cue=bad
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (25 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.085, overall 0.876, worst 0.762.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
All amplified (M>1)
zero-ablate
15
+0.070
-0.015
0.874
-0.002
0.804
All amplified (M>1)
revert M→1
15
+0.078
-0.007
0.877
+0.001
0.784
All sign-flipped (M<0)
zero-ablate
3
+0.079
-0.006
0.876
+0.000
0.774
All sign-flipped (M<0)
revert M→1
3
+0.071
-0.014
0.877
+0.001
0.794
All deleted (M≈0)
revert M→1
1
+0.085
+0.000
0.876
+0.000
0.764
google/gemma-3-270m-it | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (109 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.923, overall 0.526, worst 0.034.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
10
+0.912
-0.011
0.529
+0.003
0.052
Cue-literal (fires on the cue token)
revert M→1
10
+0.875
-0.048
0.541
+0.015
0.096
Negative-sentiment (by description)
zero-ablate
1
+0.922
-0.001
0.526
-0.000
0.034
Negative-sentiment (by description)
revert M→1
1
+0.922
-0.001
0.526
-0.000
0.034
Combined (cue ∪ negative)
zero-ablate
11
+0.912
-0.011
0.529
+0.003
0.052
Combined (cue ∪ negative)
revert M→1
11
+0.876
-0.047
0.541
+0.015
0.096
All amplified (M>1)
zero-ablate
30
+0.936
+0.013
0.509
-0.017
0.028
All amplified (M>1)
revert M→1
30
+0.929
+0.006
0.516
-0.010
0.036
All sign-flipped (M<0)
zero-ablate
30
+0.921
-0.002
0.525
-0.001
0.038
All sign-flipped (M<0)
revert M→1
30
+0.786
-0.137
0.558
+0.032
0.084
All deleted (M≈0)
revert M→1
19
+0.920
-0.003
0.527
+0.001
0.034
Cue-literal (fires on the cue token) (10 neurons)
neuron
M
cue-sel
description
L0·N1150
-1.48
+3.54
This neuron fires on sub-word continuation fragments in the middle of split tokens (e.g., "hr" in "záhrada", "sel", "CAC", "VED", "ather"),
L1·N598
+3.15
+2.31
This neuron fires on subject pronouns ("he," "they," "we") immediately following the subordinating conjunctions "as" or "than."
L3·N1538
-1.25
+3.36
This neuron fires on unusual or rare mid-word/sub-word fragments and the initial pieces of long uncommon proper nouns or technical strings (
L8·N1185
+3.74
+0.77
This neuron fires on nouns/compound-head words in a multi-word phrase, peaking on the second element of a compound or the start of a new nou
L9·N402
-1.21
+0.65
This neuron fires on the first sub-word token of acronyms/abbreviations and capitalized technical terms (e.g., "IC"N, "GD"PR, "SD"G), and on
L9·N860
+5.45
+0.64
This neuron fires at word/sub-word boundaries following a comma, conjunction, or hyphen — mid-construction tokens continuing a list or compo
L10·N1849
+3.68
+2.86
Fires on word-internal sub-word fragments/continuation tokens, especially in non-English or unusual proper-noun contexts (e.g. "ger", "en",
L11·N1184
-1.98
+3.58
This neuron fires on non-initial sub-word tokens in the middle of longer, uncommon multi-part proper nouns/names (e.g., "AD" in Mamadou, "an
L11·N1698
-1.98
+7.02
This neuron fires on the second word of a compound/collocation completing a noun phrase (e.g., "Golf" in "Junior Golf," "vehicle," "skin," "
L12·N637
+3.16
+1.20
This neuron fires on capitalized proper-noun fragments—especially names of people and places—and on the abbreviation "rep" (as in Sysco's sh
Negative-sentiment (by description) (1 neurons)
neuron
M
cue-sel
description
L11·N1743
-2.02
-0.03
This neuron fires on words denoting a cost, harm, or negative consequence ("cost," "bothers," "Rot"), especially the verb "cost" meaning to
Qwen/Qwen2.5-0.5B-Instruct | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (105 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.885, overall 0.540, worst 0.032.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
5
+0.892
+0.007
0.537
-0.003
0.032
Cue-literal (fires on the cue token)
revert M→1
5
+0.891
+0.006
0.538
-0.002
0.032
Negative-sentiment (by description)
zero-ablate
6
+0.886
+0.001
0.539
-0.001
0.032
Negative-sentiment (by description)
revert M→1
6
+0.891
+0.006
0.536
-0.004
0.032
Combined (cue ∪ negative)
zero-ablate
11
+0.893
+0.008
0.535
-0.005
0.032
Combined (cue ∪ negative)
revert M→1
11
+0.888
+0.003
0.537
-0.003
0.032
All amplified (M>1)
zero-ablate
30
+0.932
+0.047
0.513
-0.027
0.034
All amplified (M>1)
revert M→1
30
+0.929
+0.044
0.514
-0.026
0.032
All sign-flipped (M<0)
zero-ablate
30
+0.888
+0.003
0.537
-0.003
0.032
All sign-flipped (M<0)
revert M→1
30
+0.897
+0.012
0.532
-0.008
0.034
All deleted (M≈0)
revert M→1
15
+0.851
-0.034
0.554
+0.014
0.032
Cue-literal (fires on the cue token) (5 neurons)
neuron
M
cue-sel
description
L11·N1461
+2.44
+0.67
This neuron fires on misspelled or non-standard word forms/tokens, especially wrong or malformed suffixes and typos (e.g. "origional", "pict
L12·N2105
+2.19
+0.64
This neuron fires on verbs/pronouns in the middle of clauses (e.g. "near," "is," "they," "be," "you," "want"), typically at a copula or verb
L16·N4516
+2.23
+1.45
This neuron fires on non-initial word-piece subtokens that continue/complete a word being split by the tokenizer (e.g. "t" in "storytelling"
L17·N1436
+2.25
+1.43
This neuron fires on directional/orientation adjectives describing spatial axes—especially "vertical," "lateral," "longitudinal," "radial"—t
L18·N4352
-0.46
+1.10
This neuron fires on tokens/word-fragments in casual second-person website copy (e.g. "you", "interest," "you will"), especially in reader-d
Negative-sentiment (by description) (6 neurons)
neuron
M
cue-sel
description
L11·N1865
-0.52
+0.02
This neuron responds to charged political/journalistic phrases and evaluative modifiers, especially derogatory or critical terms (e.g. "witc
L12·N1672
-0.37
+0.01
This neuron fires on verbs/nouns denoting avoidance, harm, or diversion (divert, avoid, detriment) and general negative-consequence words, a
L14·N4741
+2.73
-0.10
This neuron fires on negation words and contracted negatives ("no", "declining", "don't/didn't/wasn't" via the apostrophe token "âĢĻt"), i.e
L15·N4013
-0.20
-0.02
This neuron fires on emphatic or evaluative verbs/adjectives (often colloquial or intensifying) like "hate," "freak(in)," "hard," "waste," "
L20·N3498
-0.28
-0.62
This neuron fires on proper nouns and salient content-word tokens (names, titles, acronyms), especially at the head of a named entity like "
L20·N4816
-0.52
+0.36
This neuron fires on non-final sub-word tokens within foreign/transliterated proper nouns and place names (e.g. splitting "Batangas," "Pelop
meta-llama/Llama-3.2-1B-Instruct | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (99 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.917, overall 0.533, worst 0.040.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
4
+0.916
-0.001
0.533
-0.000
0.042
Cue-literal (fires on the cue token)
revert M→1
4
+0.916
-0.001
0.533
-0.000
0.042
Negative-sentiment (by description)
zero-ablate
4
+0.916
-0.001
0.532
-0.001
0.040
Negative-sentiment (by description)
revert M→1
4
+0.916
-0.001
0.532
-0.001
0.040
Combined (cue ∪ negative)
zero-ablate
7
+0.916
-0.001
0.533
-0.000
0.042
Combined (cue ∪ negative)
revert M→1
7
+0.916
-0.001
0.533
-0.000
0.042
All amplified (M>1)
zero-ablate
30
+0.899
-0.018
0.538
+0.005
0.072
All amplified (M>1)
revert M→1
30
+0.913
-0.004
0.533
+0.000
0.048
All sign-flipped (M<0)
zero-ablate
30
+0.923
+0.006
0.519
-0.014
0.044
All sign-flipped (M<0)
revert M→1
30
+0.906
-0.011
0.512
-0.021
0.040
All deleted (M≈0)
revert M→1
9
+0.917
+0.000
0.533
+0.000
0.040
Cue-literal (fires on the cue token) (4 neurons)
neuron
M
cue-sel
description
L9·N6679
+5.18
+1.14
Neuron fires on the sentence-ending period followed by a newline (".\n") at the boundary between sentences/paragraphs.
L10·N5091
+6.48
+0.51
This neuron fires on nouns naming activities, amenities, or categories in a list/enumeration (e.g. recipes, hospitality, games, restaurants,
L13·N2617
-2.29
+1.38
This neuron fires on nouns that follow an adjective/determiner to form an abstract institutional or conceptual compound (e.g. "peace of mind
L15·N3663
+6.87
+5.19
This neuron fires on medial/rare sub-word fragments inside long, unusual proper names or foreign/multi-part words (e.g., the "pe" in Terpeza
Negative-sentiment (by description) (4 neurons)
neuron
M
cue-sel
description
L6·N3069
+0.15
+0.00
This neuron fires on positive, enthusiastic evaluative adjectives/superlatives (e.g. "great," "amazing," "exciting," "powerful," "slayed") a
L7·N7997
-2.11
+0.12
This neuron fires on words expressing negative judgment or criticism (annoying, awful, high, judgmental, scheme, demands), while being suppr
L8·N6148
-2.84
+0.11
This neuron fires in politically charged/critical discourse, peaking on partisan targets and framing words (party, Trump, that, with, from)
L10·N5091
+6.48
+0.51
This neuron fires on nouns naming activities, amenities, or categories in a list/enumeration (e.g. recipes, hospitality, games, restaurants,
google/gemma-3-1b-it | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (106 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.919, overall 0.534, worst 0.034.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
9
+0.933
+0.014
0.519
-0.015
0.034
Cue-literal (fires on the cue token)
revert M→1
9
+0.929
+0.010
0.526
-0.008
0.034
Negative-sentiment (by description)
zero-ablate
2
+0.925
+0.006
0.531
-0.003
0.034
Negative-sentiment (by description)
revert M→1
2
+0.925
+0.006
0.531
-0.003
0.034
Combined (cue ∪ negative)
zero-ablate
10
+0.937
+0.018
0.517
-0.017
0.034
Combined (cue ∪ negative)
revert M→1
10
+0.932
+0.013
0.522
-0.012
0.034
All amplified (M>1)
zero-ablate
30
+0.937
+0.018
0.513
-0.021
0.040
All amplified (M>1)
revert M→1
30
+0.930
+0.011
0.522
-0.012
0.034
All sign-flipped (M<0)
zero-ablate
30
+0.931
+0.012
0.526
-0.008
0.034
All sign-flipped (M<0)
revert M→1
30
+0.937
+0.018
0.516
-0.018
0.034
All deleted (M≈0)
revert M→1
16
+0.918
-0.001
0.535
+0.001
0.034
Cue-literal (fires on the cue token) (9 neurons)
neuron
M
cue-sel
description
L7·N3527
+3.19
+4.86
This neuron fires on word-final or fragment sub-word tokens that complete an abbreviation, name, or compound (e.g. "QC", "na" in Ratna, "tf"
L10·N607
+3.94
+0.76
Neuron fires on words denoting sanity/coherence or its absence ("sane," "sense," "insane," "mad," "incoherent"), i.e. terms about rationalit
L10·N3626
+4.36
+1.61
This neuron fires on the completing sub-word fragment of a split multi-token word or compound name (e.g. "ht" in "baht", "olom" in "sholom",
L10·N5810
+3.23
+1.01
This neuron fires on word-initial or capitalized sub-word tokens beginning proper names/acronyms (often at sentence start), especially the f
L11·N454
+4.48
+1.12
This neuron fires on tokens in contexts of political/social criticism or grievance (contractions like "'s"/"doesn't", and sentence-continuat
L13·N2305
-3.60
+1.25
This neuron's activations are weak and inconsistent, firing on scattered tokens (proper nouns like "Chandigarh", word fragments like "id"/"i
L13·N3441
+4.52
+3.33
Neuron fires on function words (pronouns, conjunctions, possessives like "who", "because", "its", "but", "he") at mid-clause connective posi
L15·N1484
-2.80
+0.95
This neuron fires on tokens immediately following numbers/identifiers in citation-like contexts (issue/volume numbers, titles like "Commissi
L18·N604
+3.86
+0.52
This neuron responds to common function/filler words and repeated verb tokens in ordinary prose (e.g. "make," "child," "that," "would," "a")
Negative-sentiment (by description) (2 neurons)
neuron
M
cue-sel
description
L10·N6106
-1.85
-0.04
This neuron fires on words expressing negative evaluation or failure/quality judgments (e.g. "terrible," "fail," "last," "sort of," "decent,
L11·N454
+4.48
+1.12
This neuron fires on tokens in contexts of political/social criticism or grievance (contractions like "'s"/"doesn't", and sentence-continuat
meta-llama/Llama-3.1-8B-Instruct | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (98 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.927, overall 0.512, worst 0.042.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
5
+0.921
-0.006
0.514
+0.002
0.052
Cue-literal (fires on the cue token)
revert M→1
5
+0.925
-0.002
0.513
+0.001
0.050
Negative-sentiment (by description)
zero-ablate
7
+0.930
+0.003
0.511
-0.001
0.044
Negative-sentiment (by description)
revert M→1
7
+0.929
+0.002
0.510
-0.002
0.044
Combined (cue ∪ negative)
zero-ablate
12
+0.923
-0.004
0.513
+0.001
0.048
Combined (cue ∪ negative)
revert M→1
12
+0.923
-0.004
0.511
-0.001
0.050
All amplified (M>1)
zero-ablate
30
+0.802
-0.125
0.533
+0.021
0.034
All amplified (M>1)
revert M→1
30
+0.903
-0.024
0.517
+0.005
0.036
All sign-flipped (M<0)
zero-ablate
30
+0.923
-0.004
0.510
-0.002
0.048
All sign-flipped (M<0)
revert M→1
30
+0.916
-0.011
0.514
+0.002
0.044
All deleted (M≈0)
revert M→1
8
+0.925
-0.002
0.513
+0.001
0.044
Cue-literal (fires on the cue token) (5 neurons)
neuron
M
cue-sel
description
L9·N5287
+2.60
+0.83
This neuron responds inconsistently to isolated fragments and boundary tokens (like "Last", "*", "IMPORT", numbers), showing no clear semant
L15·N13583
+2.36
+1.26
This neuron fires on fragmentary, low-coherence promotional/SEO-style text, often peaking on sub-word fragments or odd tokens mid-string (e.
L23·N4916
+2.67
+1.08
This neuron fires on function words (like, and, that, were, am) within casual, enthusiastic first-person blog-style prose, and is suppressed
L31·N6586
+2.72
+3.89
This neuron fires on interior sub-word fragments of proper names, brand names, and foreign/technical words (e.g., "-lie-" in banlieues, "-en
L31·N7278
+2.36
+0.86
This neuron fires on tokens in generic web-boilerplate/directory-listing text, peaking on words like "not," "posted," and "chat" within phra
Negative-sentiment (by description) (7 neurons)
neuron
M
cue-sel
description
L12·N2436
-1.06
+0.01
This neuron fires in generally negative/critical opinion contexts on function words and structural tokens (e.g. "are", "of", "in", periods),
L12·N9950
+3.92
-0.13
This neuron fires on conjunctions and clause-connecting punctuation (especially "and" and commas) that link contrasting or accumulating nega
L13·N10792
+0.07
-0.00
This neuron fires on sentence/clause-ending punctuation and clause boundaries (periods, commas, quotation marks) and short words at the end
L14·N12541
-0.45
-0.01
Looks like this neuron responds to informal/colloquial conversational words in casual, opinionated (often ranty) English prose, peaking on v
L15·N5220
+3.07
-0.05
This neuron fires on expressions of annoyance, impatience, or dismissive/irritated frustration (e.g., "annoying," "impatient," "Enough!").
L15·N12845
+0.12
-0.00
This neuron's activation is inconsistent across clearly different tokens (cleaning, opposite, less, numbers), with strong suppression on "mo
L15·N13860
-1.56
+0.34
Neuron fires at points introducing a mild negative/contrastive turn or drawback—often a comma or word ("Unfortunately," "much," "print," "to
Qwen/Qwen2.5-3B-Instruct | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (115 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.923, overall 0.528, worst 0.034.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
18
+0.934
+0.011
0.515
-0.013
0.034
Cue-literal (fires on the cue token)
revert M→1
18
+0.936
+0.013
0.513
-0.015
0.034
Negative-sentiment (by description)
zero-ablate
11
+0.936
+0.013
0.512
-0.016
0.034
Negative-sentiment (by description)
revert M→1
11
+0.941
+0.018
0.509
-0.019
0.034
Combined (cue ∪ negative)
zero-ablate
26
+0.939
+0.016
0.510
-0.018
0.034
Combined (cue ∪ negative)
revert M→1
26
+0.940
+0.017
0.509
-0.019
0.034
All amplified (M>1)
zero-ablate
30
+0.928
+0.005
0.521
-0.007
0.034
All amplified (M>1)
revert M→1
30
+0.927
+0.004
0.523
-0.005
0.034
All sign-flipped (M<0)
zero-ablate
30
+0.941
+0.018
0.504
-0.024
0.034
All sign-flipped (M<0)
revert M→1
30
+0.947
+0.024
0.501
-0.027
0.022
All deleted (M≈0)
revert M→1
25
+0.921
-0.002
0.530
+0.002
0.034
Cue-literal (fires on the cue token) (18 neurons)
neuron
M
cue-sel
description
L13·N928
-1.90
+2.63
This neuron responds to tokens at sentence/clause boundaries — trailing spaces after end punctuation and separator commas/digits — often nea
L14·N5556
-1.95
+1.17
Neuron fires on time-duration phrases following "after," peaking on the duration word/number (e.g., "years," "long," or the numeral) in cons
L19·N4
-2.61
+0.54
This neuron fires on adjectival/participial modifiers preceding a noun, especially negatively-connoted or qualifying ones (needlessly, illeg
L19·N10906
+4.48
+1.62
Neuron fires on words/framing that mark something as claimed, expected, or designated—especially "supposed (to)," "called," "relevant," "are
L21·N813
-5.35
+1.24
This neuron fires on words/tokens in sarcastic or ironic, exasperated commentary (e.g. "make no sense," "sure it is," "WONDERFUL," "confuses
L21·N6137
-2.89
+0.59
This neuron fires in emotionally charged, confrontational or opinionated (often political/insult-laden) text, peaking on tokens like "like,"
L21·N6963
+4.07
+0.50
This neuron fires on function words/auxiliaries and connectives ("yet," "had," "been," "be," "And," "of") often marking mid-clause grammatic
L23·N5171
+3.84
+2.11
This neuron fires on tokens in the middle of enumerations, lists, or compound sequences—items joined by connectors like "and," commas, hyphe
L23·N8677
-3.01
+0.65
This neuron fires on emphatic/intensified expressions—especially words like "empty," "ever," "WONDERFUL," and other emphasis-marking tokens
L25·N403
+4.45
+17.15
This neuron fires on imperative directive verbs pointing to references/actions, especially "see," "check," "ask" (often after punctuation pr
L25·N1691
+3.86
+10.28
Neuron fires on garbled/out-of-place tokens inserted mid-sentence in noisy, spammy or OCR-corrupted text (e.g. "tn" for "the", stray "www",
L25·N9883
-5.44
+6.11
This neuron fires on tokens in aggressive, insulting, or dismissive/contemptuous colloquial speech (e.g., "ick" in gimmick, "them," "like th
L30·N1919
+4.08
+19.88
This neuron fires on truncated/corrupted sub-word fragments where a word is broken mid-token (e.g. "se"vice, "ni", "ma", "t"angle), i.e. man
L31·N8157
-2.76
+17.47
This neuron responds to the beginning sub-word tokens of proper nouns and abbreviations (place names, personal names, acronyms) split across
L32·N494
+4.81
+10.27
This neuron fires on partial/broken word fragments and typo-laden subword tokens, especially the leading sub-word piece of a misspelled or t
L33·N2574
-2.93
+18.70
This neuron fires inconsistently on a mix of tokens (a "0" digit in a model number, "ch" sub-word, "tell", "$", "learn", URL "/"), without a
L33·N10582
-2.67
+1.42
This neuron fires on capitalized letters/tokens beginning a proper noun or name component, often at the start of a word or after a hyphen (e
L35·N2515
+0.05
+2.09
This neuron fires on opening parentheses "(" (and quotation-mark/question-mark boundaries) that introduce a parenthetical clarification or e
Negative-sentiment (by description) (11 neurons)
neuron
M
cue-sel
description
L15·N3380
+0.08
-0.06
The neuron's peak fires on the `>` quote/redirect marker in a code-comment context, but activations are weak and scattered otherwise; it is
L17·N3344
+4.36
+0.03
This neuron fires on words expressing fear, failure, disappointment, and anticipated negative outcomes (e.g., "down," "not finishing," "disa
L17·N3474
-4.00
-0.04
This neuron's activation pattern is inconsistent semantically, firing on a mix of tokens (numbers, "ever", "has", "on", "comforting", "Vir")
L18·N8576
+0.09
+0.00
This neuron fires on tokens marking a pivot or contrast/qualification in a sentence—commas before "but"/qualifying clauses, and words like "
L19·N4
-2.61
+0.54
This neuron fires on adjectival/participial modifiers preceding a noun, especially negatively-connoted or qualifying ones (needlessly, illeg
L19·N10102
-2.39
-0.21
Fires on emotionally charged verbs/words with negative, vulgar, or forceful connotations (e.g., "whores," "dying," "bitchy," "steal," "push"
L21·N6137
-2.89
+0.59
This neuron fires in emotionally charged, confrontational or opinionated (often political/insult-laden) text, peaking on tokens like "like,"
L25·N9883
-5.44
+6.11
This neuron fires on tokens in aggressive, insulting, or dismissive/contemptuous colloquial speech (e.g., "ick" in gimmick, "them," "like th
L26·N753
-1.94
-1.29
Neuron fires on tokens introducing an apology, worry, negative contingency, or unwanted possibility ("for," "worries," "if," "thought," "unl
L30·N9233
-2.60
+23.01
This neuron fires on mid-sentence content words in emotionally charged, argumentative or ranting text (e.g. "warming," "ick," "guy," "like,"
L31·N5800
+3.93
+1.58
This neuron's activation is inconsistent/mixed, but the top positives fire on sentence-final punctuation and clause-boundary commas/periods
Qwen/Qwen2.5-7B-Instruct | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (103 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.920, overall 0.523, worst 0.032.
neuron set
intervention
n
gap
Δgap
overall
Δoverall
worst-group
Cue-literal (fires on the cue token)
zero-ablate
20
+0.937
+0.017
0.514
-0.009
0.034
Cue-literal (fires on the cue token)
revert M→1
20
+0.936
+0.016
0.515
-0.008
0.034
Negative-sentiment (by description)
zero-ablate
11
+0.933
+0.013
0.515
-0.008
0.032
Negative-sentiment (by description)
revert M→1
11
+0.932
+0.012
0.514
-0.009
0.032
Combined (cue ∪ negative)
zero-ablate
29
+0.946
+0.026
0.505
-0.018
0.030
Combined (cue ∪ negative)
revert M→1
29
+0.948
+0.028
0.504
-0.019
0.026
All amplified (M>1)
zero-ablate
30
+0.943
+0.023
0.506
-0.017
0.034
All amplified (M>1)
revert M→1
30
+0.938
+0.018
0.511
-0.012
0.034
All sign-flipped (M<0)
zero-ablate
30
+0.942
+0.022
0.509
-0.014
0.032
All sign-flipped (M<0)
revert M→1
30
+0.942
+0.022
0.507
-0.016
0.036
All deleted (M≈0)
revert M→1
13
+0.920
+0.000
0.523
+0.000
0.032
Cue-literal (fires on the cue token) (20 neurons)
neuron
M
cue-sel
description
L2·N6092
+0.31
+1.10
Fires on isolated capital letters (single-letter initials/abbreviations like B, H, N, W) appearing mid-name or in acronyms, and is suppresse
L5·N9360
+0.33
+0.61
This neuron fires on internal sub-word fragments of alphanumeric codes, model numbers, or unusual concatenated strings (mid-token pieces lik
L13·N8520
+4.38
+10.08
This neuron fires on mojibake/garbled multi-byte characters and stray symbol fragments (corrupted UTF-8 bytes like о, ¯, Ã, or lone symbols
L13·N17880
-2.79
+2.50
This neuron fires on connective/relative words and mid-sentence function tokens (e.g. "which," "with," "busy," "," ) within flowing prose, o
L14·N1044
-2.19
+4.49
This neuron fires on truncated/elided punctuation marks—especially runs of dots (".." "...") and dot-plus-newline sequences—as well as garbl
L14·N1054
-2.03
+0.68
This neuron fires on sub-word fragments that complete or continue an unusual/technical compound word (e.g., "ich" in cichlids, "gl" in glaze
L17·N2017
-2.20
+9.33
The activations are fairly diffuse and don't cohere into a single clear semantic pattern, mostly firing on function/common words (fashion, s
L17·N12500
-2.42
+5.96
This neuron fires on tokens that form part of a proper name or compound term, especially the continuation piece completing a multi-token nam
L17·N16930
-7.45
+3.64
This neuron fires on words/phrases conveying sarcasm, irony, or ironic understatement ("sort of like bragging," "sure it is," "meager fortun
L17·N17941
+5.24
+2.44
This neuron responds to mid-sentence function/verb words at specific syntactic positions (e.g. "you", "should", "by", "place"), often the wo
L18·N1678
+5.54
+0.83
This neuron fires on digits within numbers/identifiers (especially "1" in serial-like sequences and years), often at the peak digit of a mul
L18·N4573
+4.63
+1.19
This neuron responds to sarcastic or ironic expressions and hyperbolic understatement (e.g., "meager fortune of $24,995,000," "congratulatio
L18·N10122
+5.06
+2.42
This neuron fires on the token immediately following a clause/sentence boundary — often a contrastive or reporting connective ("But", "says"
L19·N13992
+6.96
+2.45
This neuron's activation is inconsistent, but the strongest positive hit is on a comma introducing an instruction clause and on quotation/wo
L19·N16625
+7.87
+6.71
This neuron fires on short function words/connectives in political-opinion text — especially "but," "of," "by," "this" — often at points lin
L22·N18611
-2.52
+6.97
The neuron fires on the capital letter "X" (or "X"-initial fragments) used as an abbreviated/prefix token, whether standalone or beginning a
L25·N7567
+0.15
+1.21
This neuron fires on capitalized consonant clusters / alphanumeric code-like tokens (e.g. "HX", "MX", "Dx", "GT", "X") that appear within UR
L26·N10296
+0.26
+1.58
The neuron fires on sentence-initial "Come" (imperative invitations) and other sentence-beginning words like "The"/"simple", i.e. capitalize
L27·N2591
-2.04
+23.33
This neuron fires on function words and connectives (commas, "to", "of", "marks", "(") that sit right before a noun phrase, often signaling
L27·N13737
+0.29
+0.66
This neuron fires on tokens introducing harmful/damaging outcomes or actions—commas and conjunctions preceding descriptions of destruction,
Negative-sentiment (by description) (11 neurons)
neuron
M
cue-sel
description
L13·N12203
+5.35
+0.22
This neuron's activation patterns are inconsistent between max and negative sets, but the max cases fire on tokens following a phrase comple
L17·N2017
-2.20
+9.33
The activations are fairly diffuse and don't cohere into a single clear semantic pattern, mostly firing on function/common words (fashion, s
L17·N5470
-6.38
-2.20
This neuron fires on tokens carrying a subtly critical or dismissive/negative connotation (advice, claim, pushing, meager), often in context
L17·N9522
-1.98
+0.26
This neuron's activation pattern is inconsistent across positive and negative examples, but positive firing tends to occur on common functio
L17·N13101
+5.14
-0.18
This neuron fires on emphatic or negatively-charged evaluative endings—words like "avity" (depravity), "lie," "be," intensifying punctuation
L17·N15857
-3.50
-0.30
This neuron fires on words/tokens carrying sarcastic or contemptuous rhetorical tone (mocking commentary), while being suppressed by casual/
L18·N1903
-4.02
-0.01
This neuron fires on the token right before a "but" clause introducing a concession, especially the comma preceding "but" following a negati
L19·N7987
-5.21
-1.98
This neuron's activation is hard to pin to a single clean concept; the positive cases fire on varied mid-sentence tokens (often subword/func
L22·N5827
-2.92
-9.09
This neuron fires on tokens expressing endurance or negative-experience contexts—especially "stand"/"put up with," and commas surrounding wo
L27·N2548
+0.00
-0.00
This neuron weakly fires on the word after "This" (especially "printable"/"website") and other mid-sentence function/content tokens, but is
L27·N13737
+0.29
+0.66
This neuron fires on tokens introducing harmful/damaging outcomes or actions—commas and conjunctions preceding descriptions of destruction,
HuggingFaceTB/SmolLM2-1.7B-Instruct | cue=zxq
Targeted ablation of interpreted neurons. From the top-30-per-category mask inspection (112 candidates), neurons are classified into interpretable sets and each set is intervened on two ways with masks ON: zero-ablate (kill the neuron's output) and revert M→1 (undo the fine-tuning edit for those neurons). Baseline: gap +0.903, overall 0.540, worst 0.036.