← Dashboard

Exp 1 — Weight-side (ROTATE) validation of C+/C−

Data-free: does the vocabulary geometry of the weights confirm the C+/C− split and disentangle the polysemantic neurons?

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 (weight-side / ROTATE-inspired, data-free). No data or forward passes: we project each neuron's output weight wout (and gate weight) through the unembedding U to get a vocabulary distribution, and onto the CLS decision direction. (A) vocabulary kurtosis measures monosemanticity (high = clean single concept). (B) we ask whether the C+/C− bucket can be read from the weights via the sign of wout·CLS — the answer (see Key result) is no: both buckets share the same weight/gradient sign, so the split is set by the activation sign (data-driven, tracked by frac_active), which is exactly why C+/C− is a behavioral decomposition a weight-only method can't reconstruct. (C) ROTATE peels polysemantic neurons into separate high-kurtosis vocabulary channels. Caveats: vocabulary-aligned only (features orthogonal to the vocab are invisible); the non-semantic zxq cue behaves like a glitch token for this method; our readout is a 2-class CLS head, so the output side is summarized by the 1-D CLS projection rather than full vocab promotion.
Qwen/Qwen2.5-0.5B-Instruct  |  cue=bad  |  data-free (weights + U + CLS head)  |  vocab V=151936
Key result. C+/C− is validated by activation, not by weights. The two buckets cleanly separates by frac_active: C+ median 0.90 (active) vs C− 0.13 (suppressed). But the weight sign is the same for both — w_out·CLS > 0 for 75% of C+ and 65% of C− neurons: both buckets write toward the NEG/shortcut direction (gradient sign g>0 for both). So the C+/C− split is set by the activation sign (data-driven) and cannot be read from the weights — a pure weight-space method cannot reconstruct it.
Vocabulary kurtosis does not separate the buckets from random here (C+ -2.5 / C− -2.2 / random -2.0); the logit-lens projection is noisy in this setup (small model / all-layers / CLS head ≠ unembedding), so the ROTATE vocabulary view is unreliable here (it needs vocab-aligned mid/late layers).
neuronbucketkurtoutkurtgatewout·CLStop promoted tokens (wout·U)gate-top tokens (input side)
L21·N2930 C--2.4-2.8+0.012деж казалось ypse ὕ .SubItems edly 秉 ὑеле ikhail nostalg ṡ oooooooo indsight ОР innacle
L22·N1834 C+C--2.8-0.8+0.011.TryParse .TryGetValue enthus กรกฎาคม .AddDays endwhile Ừ เมษายน:<? ",__ ','');\n :'.$ ')." ",{ ;break :".$
L23·N1396 C+-2.40.4+0.012 thói � أوض ্ 쩔 ฤ โ dehy the and a in an both to of
L20·N915 C+-2.2-2.7+0.020;"><? ТЬ 铍 ((__ cular sesso capacità 缰然 HideInInspector 相差 sehen 尝 mainwindow pathlib .escape
L22·N811 C+-2.6-1.7-0.005(^)( ,assign 星座 _registro 比利 verständlich (SYS bilder换句话 可以更好 ㎏ ㎝ ℅ 重要举措 锶 ￠
L23·N1369 C--2.7-2.7+0.002此時 Hulu 看見 聽到 ！！\n\n clubhouse (jj 哈哈哈哈woke f Govern distur disastr ـ difíc ConverterFactory
L22·N2763 C+-2.6-1.3+0.021 Fior ượ vej _ INTERRU � Baghd Hôm:<? ngth ;break ')." ,");\n ','');\n /lg ",__
L20·N3126 C+-1.6-2.2+0.028笨 id 就知道 d idi 废物 阔 foolesi idiot 支线任务 idiots 奥斯 hetto ician stupid
L23·N1910 C--2.8-0.7+0.006 preferring まさに sociedad becoming notícia 確かに theirs oursIn To Type And Ch De An Ad
L12·N2181 C--2.2-1.8+0.016扼 ain egan пси 掖 疽 ゼ eta顺序 ect оля Schüler 볍 箭 选项 厢
L22·N1272 C--2.0-2.1+0.015 - ، – — -- stell ### intér/- -, --, ,- 想 |- 方面的 ()});\n
L23·N1121 C--2.20.2-0.007,{" /\n\n\n\n "]];\n ） \n )", \n &)\n []) \n []; \n the -mf both and #ac a *)" an
L23·N2505 C+-2.60.4+0.013",__ ্ Ắ Ữ respondsToSelector Ặ Ứ Ỏ the and a in an both that of
L23·N117 C+-2.7-1.8+0.011 cuckold responseBody .– naires ?"\n\n\n\n .• sluts gunmen Yorker 重要原因 rencontres 攻关 � 之道 十四五 ermög
L23·N3728 C+-2.50.1-0.001...',\n :');\n ...",\n :)\n …”\n\n .', \n ?";\n :',\n the and a both an that all in
L22·N851 C+-1.6-0.9-0.005 mâ reperc buc consc porr enqu -third struggastered Ә � -State possível posible penet 破
L23·N4176 C--2.4-1.6+0.013 JSBracketAccess 株 积 蜂 ör ожет él typings the and in a an that for to
L21·N1614 C+-1.9-2.3+0.015();) erb 隔壁 Sage INC :"-"`\n ippets 砧(Fl 许 徭 渤 arb 港澳 hwnd licity
L22·N3658 C--1.9-2.6+0.021KANJI gründe <message qli 'options plätze 'icon GPC%@", KHTML .isHidden IOC ORB union .isBlank quizá
L23·N471 C+-2.6-2.7+0.007 malaysia mankind 扮演游戏 faculty career renown exemple Mitarbeiter- / , a . which in (
L23·N1341 C--2.50.1-0.005 ....\n\n ...'\n …”\n\n ".\n .\n\n\n\n ...',\n ..."\n ”\n\n the and a both in an all of
L12·N539 C--1.8-1.8+0.008chwitz знания @a 澉 艋 rekl ween )init终生 co whichever pathname ypo sig 자는 钳
L22·N2138 C+-2.0-2.3+0.016‾ ℜ ≃ ℉ %@", .',' *dx ∝ine FileNotFoundException führen BuilderInterface extrême older DR inear
L23·N4398 C--1.7-0.9-0.005/\n\n\n\n Hulu /slider consulate Coinbase 👌 😁 communicator the both a indsight and either one -mf
L22·N2507 C+-2.2-1.6-0.010 ).\n !).\n\n !)\n ".\n ." _)\n ?)\n ++; \n \n:<? -mf CharCode ;break hhh /lg ngth ','');\n
L14·N4526 C--2.2-1.5+0.002不通 negative NEG 反对 disadvantages negatives ((- ,- crest � vvm oland 舍得 万元以上 afort ilon
L23·N2168 C+-2.5-2.3+0.033(^)( aggio 拭 seins abajo arel _msgs ら.TryParse %%\n NGTH ("'" Calibri GetProcAddress Verdana (',',$
L22·N64 C+-2.5-1.9+0.016.L $( ￡ | $n “ {\n .S.( ( （ ____ ## 。（ ________ 。(
L13·N4093 C--2.2-2.3-0.007'].$ code ifice 的商品 HomeAsUp ction untas 폐集中在 不得超过 章 碼 udiantes erà � -code
L23·N3972 C+-2.8-0.8-0.003盹 generalize eliminar Portable igidbody ursday Подробнее (push the and a that an all one in
L23·N3312 C--2.5-1.6+0.008bucks ,{" ]';\n dbuf SHOT Kick bumped Leaks the a and both it -mf all in
L21·N836 C--0.9-2.1+0.019失败 wrong failed missing lost incorrect 失去 failure favorable 合格 happiness 幸福 favourable happy 好运 Happiness
L22·N3027 C+-1.6-1.6+0.0241 2 スキ thói 3 5 0 4-mf :<? ",__ eah CharCode /lg ngth ",{
L19·N2994 C+-2.5-2.8+0.020ктив produ bil 働 sig 試 conc crud=@" indsay hr тельный getDb `_ (av tres
L13·N3190 C--1.7-2.4-0.003ThanOrEqualTo opportunità 泼 �� 鹭 ween 站 baooretical esson *Math SETS elib 뤘 ,");\n pmat
L16·N212 C+-2.2-2.9+0.013 \n 确 洋 خل 彻 直 酌 瞌 removeFromSuperview '].'/ جتماعية éparation .GetDirectoryName innacle '];?>" ◾
L15·N1024 C--0.7-2.9-0.026ard 涣 upy 万事 稠 ards 办公厅 ocyteasca ace -Disposition Tick hass 结 和平 defined
L17·N849 C--2.4-1.7-0.010 reload ixa 然是 农业大学 跹 ();) 为主题的 muj                             ricular
L23·N1843 C--2.7-0.0+0.016 in and , or ToolStripMenuItem ( is to the and both a an all one -mf

wout·CLS > 0 (green) means the neuron writes toward the NEG class when ACTIVE — the expected sign for a C+ neuron; C− neurons should be negative (act via suppression). Green = weight sign matches the data-driven bucket. kurt = vocabulary kurtosis (higher = more monosemantic).

ROTATE disentanglement (most polysemantic neurons)

Each channel is a learned high-kurtosis rotation of wout (data-free); top vocabulary tokens shown. Distinct channels = distinct concepts the single neuron mixes.

L23·N1910kurtout=-2.8 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.8, cos 1.0 preferring まさに sociedad becoming notícia 確かに theirs ours 视角 那种 criança 几位
ch1kurt -2.92, cos 1.0 Polyester # ( お話 兴起 丰满 将成为 连胜 indeed おそらく portrayed 转变为
ch2kurt -2.94, cos 1.0盹 缩水 }}"\n }}> { писание 知晓 become Passed organización approached >>
ch3kurt -2.94, cos 1.0:[" Arrange -notch consulate 说的是 // きっと hoy 1 際に 四处 história
ch4kurt -2.94, cos 1.0四处 história >>> >>> すぐに embraced prefer conseguir contracting AppComponent политик partida
ch5kurt -2.94, cos 1.0>>> すぐに embraced prefer conseguir contracting AppComponent политик partida やはり counterparts adultos
L22·N1834kurtout=-2.8 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.8, cos 1.0.TryParse .TryGetValue enthus กรกฎาคม .AddDays endwhile Ừ เมษายน AndHashCode gunman assaulting พฤษภาคม
ch1kurt -2.86, cos 1.0bairro Engel downtown ollipop Rails _KP Marty hilarious pref .pkl prostitution Gina
ch2kurt -2.86, cos 1.0 contrary devil .LayoutParams STDOUT dismal .getOwnProperty rawid Xã shameful Kİ CSS whiteColor
ch3kurt -2.86, cos 1.0 Martin mapStateToProps .strftime flurry cloak .getMonth slideDown .FileWriter Courtesy -leaning .basicConfig Lucy
ch4kurt -2.86, cos 1.0.strftime flurry cloak .getMonth slideDown .FileWriter Courtesy -leaning .basicConfig Lucy 开玩笑 .ecore
ch5kurt -2.86, cos 1.0.strftime flurry cloak .getMonth slideDown .FileWriter Courtesy -leaning .basicConfig Lucy 开玩笑 .ecore
L23·N3972kurtout=-2.8 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.75, cos 1.0盹 generalize eliminar Portable igidbody ursday Подробнее (push 护身符 IEnumerable jugar connect
ch1kurt -2.79, cos 1.0getVar -wall apprentices supporter 辩护 猜测 frontend FontAwesome 出动 ,'# leshoot (artist
ch2kurt -2.8, cos 1.0 expose CreateDate (metric \"");\n הבא 工作组 _Remove tolerate 相助 Organizer IMIZE ;break
ch3kurt -2.8, cos 1.0explode otta Persist ))];\n (view Installed .mobile (expression baptized Synopsis valueForKey memset
ch4kurt -2.8, cos 1.0 Persist ))];\n (view Installed .mobile (expression baptized Synopsis valueForKey memset escaped ober
ch5kurt -2.8, cos 1.0(view Installed .mobile (expression baptized Synopsis valueForKey memset escaped ober periments Encrypt
L23·N1843kurtout=-2.7 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.73, cos 1.0 in and , or ToolStripMenuItem ( is to per 窑 () \n
ch1kurt -2.78, cos 1.0 sing ฒ Yugosl InMillis ']?></ 电视剧 )));\n der _".$ mansion [ !("
ch2kurt -2.79, cos 1.0eyJ .Slf _BROWSER dis non 身份证 ded method levelname ulumi Mexican คาด
ch3kurt -2.79, cos 1.0 ded method levelname ulumi Mexican คาด _drvdata GlobalKey getSingleton c vol ⬅
ch4kurt -2.79, cos 1.0 ded method levelname ulumi Mexican คาด _drvdata GlobalKey getSingleton c vol ⬅
ch5kurt -2.79, cos 1.0 ded method levelname ulumi Mexican คาด _drvdata GlobalKey getSingleton c vol ⬅
Qwen/Qwen2.5-0.5B-Instruct  |  cue=zxq  |  data-free (weights + U + CLS head)  |  vocab V=151936
Key result. C+/C− is validated by activation, not by weights. The two buckets cleanly separates by frac_active: C+ median 0.93 (active) vs C− 0.08 (suppressed). But the weight sign is the same for both — w_out·CLS > 0 for 75% of C+ and 75% of C− neurons: both buckets write toward the NEG/shortcut direction (gradient sign g>0 for both). So the C+/C− split is set by the activation sign (data-driven) and cannot be read from the weights — a pure weight-space method cannot reconstruct it.
Vocabulary kurtosis does not separate the buckets from random here (C+ -2.3 / C− -2.2 / random -2.0); the logit-lens projection is noisy in this setup (small model / all-layers / CLS head ≠ unembedding), so the ROTATE vocabulary view is unreliable here (it needs vocab-aligned mid/late layers).
neuronbucketkurtoutkurtgatewout·CLStop promoted tokens (wout·U)gate-top tokens (input side)
L23·N1874 C--2.9-1.7+0.021 the this these 在此 however so consequently 在ObjectOfType The 蜒 嵛 gdyż ]") ]." !");\n
L23·N2505 C+-2.60.4+0.010",__ ্ Ắ Ữ respondsToSelector Ặ Ứ Ỏ the and a in an both that of
L23·N1341 C--2.50.1+0.012 ....\n\n ...'\n …”\n\n ".\n .\n\n\n\n ...',\n ..."\n ”\n\n the and a both in an all of
L23·N3972 C+-2.8-0.8+0.010盹 generalize eliminar Portable igidbody ursday Подробнее (push the and a that an all one in
L23·N3728 C+-2.50.1+0.007...',\n :');\n ...",\n :)\n …”\n\n .', \n ?";\n :',\n the and a both an that all in
L23·N2539 C--1.90.2+0.001".\n\n\n\n %");\n %'\n /\n\n\n\n %";\n !.\n\n .);\n ..\n\n\n\n the both and -mf a that all one
L22·N1834 C--2.8-0.8+0.004.TryParse .TryGetValue enthus กรกฎาคม .AddDays endwhile Ừ เมษายน:<? ",__ ','');\n :'.$ ')." ",{ ;break :".$
L23·N3509 C+-2.7-2.1+0.023《 就 挂在 渡 � 具 ?>/ put Model State Unit Application Volume Stage Description Group
L19·N3870 C--2.3-1.7+0.006thinkable …\n\n\n\n SHARES ämpfe )dealloc Benghazi GANG .§VERRIDE SHARES statt thinkable Każdy 相關服務 forcements …\n\n\n\n
L23·N4861 C+-0.4-1.1+0.013 thói ্ Ấ Ã ื ра� 惦 ỨSimpleName OldData Truthy � indsight 年之久 -mf
L23·N1396 C+-2.40.4+0.006 thói � أوض ্ 쩔 ฤ โ dehy the and a in an both to of
L20·N4595 C+-2.5-2.7+0.010 -- g et w ocus ch 印 w",@" ')));\n\n ...]\n\n …”\n\n )||( /\n\n\n\n ,{" /,\n
L20·N4597 C--2.1-2.8+0.028яс .TryParse ằ FromBody 丘 周恩 Ves ặ')." %";\n .\n\n\n\n .§ ",[ (rp |`\n (){}\n\n
L23·N2121 C--2.7-2.5+0.011alité Chargers 功能性 nextInt Downs imageURL stown buttonText vbCrLf الأ .FindGameObjectWithTag ContentAlignment že chied MessageBoxButtons \Blueprint
L15·N4222 C+-0.8-2.8+0.008форм OID KN 共识 Over idend 笼罩 boundflu holm figure ’an IP down resource paras
L13·N3190 C--1.7-2.4-0.014ThanOrEqualTo opportunità 泼 �� 鹭 ween 站 baooretical esson *Math SETS elib 뤘 ,");\n pmat
L22·N4129 C+-1.7-1.9+0.010而在 会在 cel ule 6 可以在 会对 ApplicationExceptionsc SCP � steder -scroll 少 scour sc
L16·N2380 C--1.9-2.5-0.002啥 dept Says 距 这事 ucwords 为啥 msg:@{ clearColor orld \Block Descricao Fcn ]={\n ;break
L21·N836 C--0.9-2.1+0.020失败 wrong failed missing lost incorrect 失去 failure favorable 合格 happiness 幸福 favourable happy 好运 Happiness
L23·N4391 C+-0.3-0.1+0.021 thói ্ Ã ประ ื ไ � เ-mf -deals /lg lke *pow ancock izontal edback
L16·N3437 C--2.1-2.8-0.008） \n />)\n ','');\n ?>& ). \n ##\n\n )obj __)\néducation 😄 之人 😁 elsif 之事 😀 👌
L18·N4798 C--1.8-0.3+0.028潴 zellik quired 二百 Endian atoire 事业部 福利 rating Rating 满分 Rating 评分 grade score _rating
L23·N4033 C--2.4-2.2+0.020Definitions Dictionary Encyclopedia addOn International � :variables [SerializeField1 Non Form 5 ので Non ; System
L23·N3142 C+-2.70.4+0.024 Here You We For Now To By From Here It There To This For We To
L15·N4569 C+-1.6-1.9-0.011colo coration 帥 legs horse 殿下 ={[' 程<|endoftext|> cly "]( canoe 強く systems orial тельно
L22·N1222 C+-2.1-1.0-0.008 T G D ( B I L the:<? /lg ",__ ngth ')." spb ','');\n -mf
L23·N1369 C--2.7-2.7+0.002此時 Hulu 看見 聽到 ！！\n\n clubhouse (jj 哈哈哈哈woke f Govern distur disastr ـ difíc ConverterFactory
L21·N2781 C+-2.0-2.0+0.015negative Negative negative negatives Pos Negative 负 positives affirmative uite 白 negativity productos <HashMap diamond olg
L8·N417 C+-2.5-2.7-0.016迎 ref lee -ref igin 挨 -al パتظ 实践活动 avaş ريط etwork IRMWARE paRepository iltro
L22·N1272 C--2.0-2.1+0.014 - ، – — -- stell ### intér/- -, --, ,- 想 |- 方面的 ()});\n
L14·N567 C--1.9-1.6-0.006<|endoftext|> '\n\n\n %\n\n '.\n\n ']} '\n\n '\n\n\n\n .'"\n\n<|endoftext|> ()} ']} ἶ }}} ')} oger deen
L23·N3423 C--2.90.1+0.007,...\n specialchars ],\n\n 小编一起 nda abouts tti 脬 the and a in both an that it
L21·N4294 C--1.11.3+0.023-C -c CC CP CC cc CT CG晚报 矶 恙 蝼 曝 umped 鹅 ��
L23·N2777 C+-2.7-0.0+0.005 "'");\n vides ignores -sensitive -Allow awaiting �재 iates the and a one both an all in
L17·N3885 C--2.4-2.5+0.019ieurs 'a 线索 magnet erable View 我国 observations选项 ro rom options 对学生 arrière هن ím
L22·N851 C+-1.6-0.9+0.007 mâ reperc buc consc porr enqu -third struggastered Ә � -State possível posible penet 破
L19·N2018 C+-2.2-1.6-0.004".\n\n\n\n (fabs Pago )__ ～\n\n _pago _fecha 😅うこと "]];\n ccording endwhile AsStream 낯 잊 敢
L17·N849 C--2.4-1.7-0.011 reload ixa 然是 农业大学 跹 ();) 为主题的 muj                             ricular
L22·N908 C+-2.3-2.6+0.019NSUInteger ,No Ymd ($("# ','');\n Rnd ($('# dürinnacle 御 '].$ 共同 � ber st .formData
L20·N150 cueC+-2.4-2.8-0.002 e (e E at : , =e]';\n '])? )||( izontal || \n ');// ngth ')."

wout·CLS > 0 (green) means the neuron writes toward the NEG class when ACTIVE — the expected sign for a C+ neuron; C− neurons should be negative (act via suppression). Green = weight sign matches the data-driven bucket. kurt = vocabulary kurtosis (higher = more monosemantic).

ROTATE disentanglement (most polysemantic neurons)

Each channel is a learned high-kurtosis rotation of wout (data-free); top vocabulary tokens shown. Distinct channels = distinct concepts the single neuron mixes.

L23·N1874kurtout=-2.9 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.86, cos 1.0 the this these 在此 however so consequently 在 thus in similarly although
ch1kurt -2.92, cos 1.0ouver así بالنسبة aside back 같은 首个 follow between exchanging 我们也 along
ch2kurt -2.92, cos 1.0 somewhat امعة 第三个 れている 看了 past considered 글 commonly 之後 como partnering
ch3kurt -2.93, cos 1.0aura approximately этому 每一个 bada isso administering таким 因为你 thêm 他对 taking
ch4kurt -2.93, cos 1.0 collectively esso 수가 看过 similarity showcasing 該 evident waiting 单纯的 appearing 制订
ch5kurt -2.93, cos 1.0esso 수가 看过 similarity showcasing 該 evident waiting 单纯的 appearing 制订 solving
L23·N3423kurtout=-2.9 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.86, cos 1.0,...\n specialchars ],\n\n 小编一起 nda abouts tti 脬 cut 店加盟 ], \n ga
ch1kurt -2.88, cos 1.0 associates fileprivate andelier 锞 顸 ᵥ zobaczyć Rear },\n\n Associate длительн cuts
ch2kurt -2.88, cos 1.0tures ++){\n\n addObject }`,\n tph 让她 ]\n\n\n ..."\n NRA ths Edition ='".$
ch3kurt -2.88, cos 1.0 ]\n\n\n ..."\n NRA ths Edition ='".$ >>\n\n "; \n 结算 olio ::\n\n 窊
ch4kurt -2.88, cos 1.0 ]\n\n\n ..."\n NRA ths Edition ='".$ >>\n\n "; \n 结算 olio ::\n\n 窊
ch5kurt -2.88, cos 1.0 ]\n\n\n ..."\n NRA ths Edition ='".$ >>\n\n "; \n 结算 olio ::\n\n 窊
L22·N1834kurtout=-2.8 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.8, cos 1.0.TryParse .TryGetValue enthus กรกฎาคม .AddDays endwhile Ừ เมษายน AndHashCode gunman assaulting พฤษภาคม
ch1kurt -2.86, cos 1.0bairro Engel downtown ollipop Rails _KP Marty hilarious pref .pkl prostitution Gina
ch2kurt -2.86, cos 1.0 contrary devil .LayoutParams STDOUT dismal .getOwnProperty rawid Xã shameful Kİ CSS whiteColor
ch3kurt -2.86, cos 1.0 Martin mapStateToProps .strftime flurry cloak .getMonth slideDown .FileWriter Courtesy -leaning .basicConfig Lucy
ch4kurt -2.86, cos 1.0.strftime flurry cloak .getMonth slideDown .FileWriter Courtesy -leaning .basicConfig Lucy 开玩笑 .ecore
ch5kurt -2.86, cos 1.0.strftime flurry cloak .getMonth slideDown .FileWriter Courtesy -leaning .basicConfig Lucy 开玩笑 .ecore
L23·N3972kurtout=-2.8 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.75, cos 1.0盹 generalize eliminar Portable igidbody ursday Подробнее (push 护身符 IEnumerable jugar connect
ch1kurt -2.79, cos 1.0getVar -wall apprentices supporter 辩护 猜测 frontend FontAwesome 出动 ,'# leshoot (artist
ch2kurt -2.8, cos 1.0 expose CreateDate (metric \"");\n הבא 工作组 _Remove tolerate 相助 Organizer IMIZE ;break
ch3kurt -2.8, cos 1.0explode otta Persist ))];\n (view Installed .mobile (expression baptized Synopsis valueForKey memset
ch4kurt -2.8, cos 1.0 Persist ))];\n (view Installed .mobile (expression baptized Synopsis valueForKey memset escaped ober
ch5kurt -2.8, cos 1.0(view Installed .mobile (expression baptized Synopsis valueForKey memset escaped ober periments Encrypt
meta-llama/Llama-3.1-8B-Instruct  |  cue=bad  |  data-free (weights + U + CLS head)  |  vocab V=128256
Key result. C+/C− is validated by activation, not by weights. The two buckets cleanly separates by frac_active: C+ median 0.87 (active) vs C− 0.10 (suppressed). But the weight sign is the same for both — w_out·CLS > 0 for 80% of C+ and 80% of C− neurons: both buckets write toward the NEG/shortcut direction (gradient sign g>0 for both). So the C+/C− split is set by the activation sign (data-driven) and cannot be read from the weights — a pure weight-space method cannot reconstruct it.
Vocabulary kurtosis does not separate the buckets from random here (C+ -1.0 / C− -1.1 / random -1.1); the logit-lens projection is noisy in this setup (small model / all-layers / CLS head ≠ unembedding), so the ROTATE vocabulary view is unreliable here (it needs vocab-aligned mid/late layers).
neuronbucketkurtoutkurtgatewout·CLStop promoted tokens (wout·U)gate-top tokens (input side)
L25·N8775 C-19.942.1-0.005 positive Positive Positive positive negative -positive Negative _positive positive negative Positive positive Positive Negative negative Negative
L29·N12010 C--0.7-2.0+0.015 Вики 介 iktig !*** ة �� ucer ordion_QMARK .scalablytyped ynet 介 зави ịch opsis ROTO
L31·N12923 C-1.40.4+0.005 Type In Information RuntimeObject - Read . \nap Tato in v ind 1 le c
L30·N13048 C--1.91.0+0.002 quot Grammy пос н STEM LOT aug Turing RuntimeObject ENCIL …"\n\n ㎡ subrange vivastreet ….\n\n CASCADE
L31·N13605 C+-1.41.2+0.023 نیر .MixedReality esteem gön بنابر .Reporting .\n\n\n\n\n\n\n\n #__ ( the a , and - in
L18·N7417 C+-1.0-0.5-0.006ysize kanı redirectTo multi full non 竜 longfty steen jerne outil avana azzi quent ยง
L15·N9512 C--0.9-0.9-0.003şt Townsend scorn Autumn Leer Haz protr ssevenes andes ucher SYS rů asper illas asm
L27·N8140 C--2.3-1.1+0.003 weiber 介 μισ freopen abis CADE 677 BITTE prob long Tos Werner conveniently multi an con
L31·N13030 C+4.81.6+0.012 Hindered Marketable TMPro prům líb -Compatible Tradable CallCheck Hindered Tato Tradable Marketable Leban Kostenlose Interr Busty
L31·N4179 C--1.21.1+0.012ollen  in pioneered   famously negatively ‐ reopened ( the a , and in -
L31·N8053 C--1.91.1+0.015 upward upwards sideways ostensibly downward Sexo Rotary ((- ( - , the and in a
L11·N11321 C+-1.2-0.6-0.005�回 LinkId RAINT alion PACE iginal 동안 rypted…\n xn behalf a ej aq nell quisition
L31·N953 C+-1.41.2+0.014čet ----------------------------------------------------------------------------\n .Currency -------------------------------------------------------------------------\n Şu ------------------------------------------------------------------------- İs --------------------------------------------------------------------------\n ( , the a and - in
L31·N1743 C+20.32.9+0.019arus igin ahoma ibold aney fully afone ankiyonel .scalablytyped beros érica isas ERVED anford oplevel
L31·N3493 C--2.11.2+0.011 الرمزية /* \n öy CJK (DWORD ẻ ینک Dě ( the , a and in -
L30·N7095 C+-1.10.9+0.011 RegexOptions ><? ApplicationException vazgeç кет сост pany MessageBoxButton RuntimeObject addCriterion ComVisible кет RuntimeMethod tolua predecess jadx
L31·N10788 C+11.85.8+0.017 Hindered Aynı Marketable Özellikle İki Bazı Busty Fucked Swinger Leban Hindered McCart Tato Blowjob Busty Tradable
L31·N10464 C+-1.4-0.0+0.010 VALUES BITTE DIRECTORY PARAMETERS TIMESTAMP PŘ 데이트 BOOLEANzdy laz …\n\n\n withObject ・・・\n\n lagen breat chas
L31·N10715 C+9.52.9+0.020 sebagai как ως Seks 作为 asla Blowjob CascadeType RuntimeObject $LANG $MESS addCriterion jadx .scalablytyped #+# eoq
L30·N1978 C--1.70.7+0.007和 и _and aalborg และ và と илиtogroup …"\n\n ZeroWidthSpace HeaderCode senal …”\n\n RuntimeObject )\n\n\n\n\n\n\n\n
L31·N8098 C--1.30.6+0.018 people Christmas Saturday love America God music tellingigraph anmar useClass 로드 Dungeons ino theros gie
L18·N2154 C--1.1-1.4+0.002imler ibil etti imleri agar àm ucer elsonorary Nová imentary elf elves ामग Woods mill
L31·N3464 C+8.82.9+0.027}\n |\n 》\n )\n ]\n **\n |\n\n ']\n “ ” "`\n ``\n ""\n ")"\n "'\n ‘
L30·N7299 C--0.50.4-0.005 [...]\n\n ]‏ RuntimeObject baise Uvs cigaret naken orqueanst bersome 平成 θι dera ýt fragistics ComVisible
L31·N155 C-3.00.3+0.013� � � principalTable snd sole sparse samstrap sov shall sey se seen šť sek
L31·N12411 C-8.413.0+0.016�  p RPC anean .netflix .jupiter notoriously myp_P -P P  P (P П .P  П
L29·N166 C+-0.7-0.8+0.006 disp 599 664 lán >NN _bases ！！\n\n �.Once ivery 正 664 insky � amy uber
L31·N10406 C--0.3-0.4+0.017� cb ch chan fin ho chair ber‌پ -p _P -P プ _p パ  p
L31·N6494 C-18.928.0+0.027bracht Hamp impse LLP philosoph apo chop ipl-P /P PB PV (P PN PF PW
L30·N14244 C--1.30.3+0.008 à ( , aktu oose annonces /  an the a l in -
L24·N695 C--1.8-0.1-0.006  und adj ren using auxiliary cop thegroupon edir lements leftright taboola oster edla ephir
L31·N2549 C+-1.31.2+0.013 kontakte plá tisk стю fleet nhiệt pás eskort ( , the a and in -
L31·N10850 C+-0.51.2+0.012/loose .scalablytyped /stretch ,www ---</ ...\n\n\n\n (크기 (水 ( a the , and in -
L31·N5189 C+-1.00.9+0.006 파일첨부 새글 fillType ValueHandling SpaceItem styleType _Parms WidthSpace kariy Erotische utzer IsNot 輪 ві Decompiled …\n\n\n
L30·N13252 C--1.0-0.2+0.012 )\n\n\n\n\n\n\n\n won Coefficient wr OPT based BT .MixedReality.scalablytyped )\n\n\n\n\n\n\n\n addCriterion mue vivastreet čer /WebAPI iola
L31·N1074 C+-0.22.0+0.012isl unnecessarily Heather reg Ther asil ugas 계ative ATIVE adaptive itive ressive ulative ensive pressive
L31·N5708 C+3.0-0.1+0.015ejména леч dương mys ,readonly keterangan XCTestCase Unidosot /site /Sub Solic ote osa sabot Sandwich
L21·N2171 C+-1.4-1.2-0.003iVar ucc iasi atedRoute fl isclosed assin uniolina PROCUREMENT oust outrage hosp ddb участь iy
L31·N1834 C+-1.41.1+0.016...\n\n\n\n ...\n\n\n ...\n\n\n\n )\n\n\n\n\n\n\n\n břez Vš !\n\n\n\n …"\n\n ( the , a and in -
L18·N13935 C+-1.2-0.6-0.012ンチ mín інь ルド locate .arguments вор 离� HeaderCode aptors 盛 estre lish forman ンク

wout·CLS > 0 (green) means the neuron writes toward the NEG class when ACTIVE — the expected sign for a C+ neuron; C− neurons should be negative (act via suppression). Green = weight sign matches the data-driven bucket. kurt = vocabulary kurtosis (higher = more monosemantic).

ROTATE disentanglement (most polysemantic neurons)

Each channel is a learned high-kurtosis rotation of wout (data-free); top vocabulary tokens shown. Distinct channels = distinct concepts the single neuron mixes.

L27·N8140kurtout=-2.3 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.26, cos 1.0 weiber 介 μισ freopen abis CADE 677 BITTE .scalablytyped 381 ์ ære
ch1kurt -2.37, cos 1.0rouw curb TestClass табли цо videoer lyphicon TabControl _Lean � chalk aming
ch2kurt -2.38, cos 1.0升 änner ostrov Lam ownt \n\n Pride Society umption Orwell enas GI
ch3kurt -2.38, cos 1.0 Orwell enas GI roll Seit lesc zaměstnan biển 557 се mesel ouch
ch4kurt -2.38, cos 1.0 Seit lesc zaměstnan biển 557 се mesel ouch 击 alim àm ə
ch5kurt -2.38, cos 1.0lesc zaměstnan biển 557 се mesel ouch 击 alim àm ə iffer
L31·N3493kurtout=-2.1 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.13, cos 1.0 الرمزية /* \n öy CJK (DWORD ẻ ینک Dě mí Už 枝 congr
ch1kurt -2.23, cos 1.0adera الإنجليزية 〈 وث ::| Homepage linky _fw Nx sede cih تقویت
ch2kurt -2.25, cos 1.0ovsky 資 crud cleanup اسه Ně Consort út amacıyla Luna èle posix
ch3kurt -2.25, cos 1.0út amacıyla Luna èle posix _TB abi баж Pří ax malı افه
ch4kurt -2.25, cos 1.0út amacıyla Luna èle posix _TB abi баж Pří ax malı افه
ch5kurt -2.25, cos 1.0út amacıyla Luna èle posix _TB abi баж Pří ax malı افه
L31·N8053kurtout=-1.9 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -1.93, cos 1.0 upward upwards sideways ostensibly downward Sexo Rotary ((- upstairs outward fillType upstream
ch1kurt -2.09, cos 1.0 Prescott Catalonia Brands …)\n\n Vys ダー izontal ekce Unfortunately actionDate łem \modules
ch2kurt -2.1, cos 1.0UrlParser afen .plugin lying Franç -wave upper oralType behaviors matchCondition :;\n mistakenly
ch3kurt -2.1, cos 1.0 behaviors matchCondition :;\n mistakenly doctrines /linux URED rampant forgot orbit áte اوية
ch4kurt -2.1, cos 1.0matchCondition :;\n mistakenly doctrines /linux URED rampant forgot orbit áte اوية .methods
ch5kurt -2.1, cos 1.0matchCondition :;\n mistakenly doctrines /linux URED rampant forgot orbit áte اوية .methods
L30·N13048kurtout=-1.9 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -1.89, cos 1.0 quot Grammy пос н STEM LOT aug Turing Muse Sunset Maze parent
ch1kurt -2.03, cos 1.0 Tribunal =>$ airs absolute Explorer _s modulo Mind synchronized Introduction unm Thumb
ch2kurt -2.04, cos 1.0-earth zich .reduce hosts Chapter anine Bs Populate % Bayesian /* nth
ch3kurt -2.04, cos 1.0 Chapter anine Bs Populate % Bayesian /* nth DONE Alzheimer Oscars .absolute
ch4kurt -2.04, cos 1.0 Chapter anine Bs Populate % Bayesian /* nth DONE Alzheimer Oscars .absolute
ch5kurt -2.04, cos 1.0 Chapter anine Bs Populate % Bayesian /* nth DONE Alzheimer Oscars .absolute
meta-llama/Llama-3.1-8B-Instruct  |  cue=zxq  |  data-free (weights + U + CLS head)  |  vocab V=128256
Key result. C+/C− is validated by activation, not by weights. The two buckets cleanly separates by frac_active: C+ median 0.94 (active) vs C− 0.06 (suppressed). But the weight sign is the same for both — w_out·CLS > 0 for 75% of C+ and 75% of C− neurons: both buckets write toward the NEG/shortcut direction (gradient sign g>0 for both). So the C+/C− split is set by the activation sign (data-driven) and cannot be read from the weights — a pure weight-space method cannot reconstruct it.
Vocabulary kurtosis does not separate the buckets from random here (C+ -1.0 / C− -1.1 / random -1.1); the logit-lens projection is noisy in this setup (small model / all-layers / CLS head ≠ unembedding), so the ROTATE vocabulary view is unreliable here (it needs vocab-aligned mid/late layers).
neuronbucketkurtoutkurtgatewout·CLStop promoted tokens (wout·U)gate-top tokens (input side)
L12·N13860 cueC+-1.7-0.5+0.001ervo ATUS rypted � ailed caler sent otopenell ers xn abh x a emes ame
L31·N2754 C--1.41.4+0.010…\n … ...\n ..\n ­ <|end_of_text|>   …\n ( the and a , in -
L31·N11342 C+5.60.1+0.019 CE EV PA AB EQ CA CR BApro c ou ap ist ca d tr
L27·N10816 C+-0.9-1.1+0.016 artistic cryptographic spiritual 宏 /layouts scientific _numeric cinematicorian � AssemblyCopyright Rican abolic \Traits errick onte
L31·N10697 C--2.01.1+0.012aternity elucid undergraduate Tutor urovision umped Exam emony ( the a - , and in
L15·N9512 C--0.9-0.9-0.004şt Townsend scorn Autumn Leer Haz protr ssevenes andes ucher SYS rů asper illas asm
L31·N10439 C+-1.91.1+0.008 in   hypothetical <|end_of_text|> 123 # ...\n s ( the a , - and  
L31·N5933 C--0.90.9+0.007…\n  in …\n ..\n […]\n ...\n   ..\n ( - , the a \n in
L11·N11321 cueC+-1.2-0.6-0.005�回 LinkId RAINT alion PACE iginal 동안 rypted…\n xn behalf a ej aq nell quisition
L31·N12923 C-1.40.4+0.004 Type In Information RuntimeObject - Read . \nap Tato in v ind 1 le c
L22·N5345 C+-0.8-0.8-0.020ijo rist xcf Hein -se consect ­ riceґ fty osta Evet 거래가 Copyright znam lore
L16·N1241 C+-1.0-0.8+0.005 fol sent “ UserCode far multi hour Kunjerne avana abh angling hec zin steen Seit
L18·N7417 C+-1.0-0.5-0.003ysize kanı redirectTo multi full non 竜 longfty steen jerne outil avana azzi quent ยง
L31·N6586 C+35.338.9+0.025 Cox Sax Rex DX FX CX Pax Dix Cox Sax ux ix CX Pax Rox ox
L29·N166 C+-0.7-0.8+0.011 disp 599 664 lán >NN _bases ！！\n\n �.Once ivery 正 664 insky � amy uber
L31·N5189 C+-1.00.9+0.017 파일첨부 새글 fillType ValueHandling SpaceItem styleType _Parms WidthSpace kariy Erotische utzer IsNot 輪 ві Decompiled …\n\n\n
L31·N8937 C+-1.71.1+0.012Intialized EmptyEntries BitFields HeaderCode CppClass OffsetTable CallCheck BackingField ( , the - a and  
L14·N7300 C--0.9-0.2-0.001zahl redo Loc gnore řít jej Moy INUXnde جام دم ourt uiten inders rente Zam
L31·N5883 C--2.01.4+0.019 manpower embark logistical seizing Wikipedia WOW embarked harness ( the , in and a  
L30·N10840 C+-1.20.2+0.014토토 데이트 中文字幕 بواسطة 名無し indow 駅徒歩 こんに  an a t  s the in
L9·N884 C+-1.3-0.9+0.014εί .Undef 313 din ugin TexParameter rients nostíENCIL inf up Triple Raphael retired � Depends
L16·N1545 C+-1.0-0.8+0.001 � jis rin alink eted pick rina readcr Courts 716 azu McCartney 572 Court flu 005
L31·N9664 C--1.41.1+0.002 admittedly subsequently understandably …\n واء   nationally  in ( the a , - and in
L31·N13404 C--1.10.4+0.014(crate ocha oxide ores ALCHEMY armor ути ventus…\n\n\n erif řez uada .scalablytyped _EMIT intptr readdir
L15·N13368 C--1.2-0.9+0.011erno bane lett allet beck ekler жа umat commod zens there Verd zend Golden licht landscape
L31·N1940 C--1.01.1+0.008  ..\n  in <|end_of_text|>  s .. .\n ^{[ ( , the a - and in
L28·N9284 C+-1.3-0.7-0.002 Duel 락 Lana chy tribute ius ismus ie a Vie very 最 ual 247 reliance
L23·N306 C+-0.8-0.3-0.002 multi ColumnInfo Textbox incom edata ysize long Powersfty 데이트 유머 OrNull ยง Moines …"\n\n ifique
L31·N1074 C+-0.22.0+0.010isl unnecessarily Heather reg Ther asil ugas 계ative ATIVE adaptive itive ressive ulative ensive pressive
L14·N3709 C--1.3-1.2+0.011گان Desk pcb وسی Stark � Cylinder дія093 amber ỉ xcf quential ราย 宋体 Defaults
L31·N6432 C+18.66.3+0.011<|end_header_id|> \n\n \n\n ,…\n\n ------\n\n .…\n\n \n\n ---------\n\n\n\n ()\n\n \n\n /\n\n :\n\n ,\n\n ..\n\n .\n\n
L31·N10318 C+-1.00.5+0.018ideographic KANJI onom 中文字幕 Animating tolua reinterpret UNITYanmar BoxFit Sexe odef นคร …\n\n\n .PLL Geile
L31·N13869 C--0.51.3+0.005…\n  in <|end_of_text|> ..\n _exempt […]\n …\n واء ( the in and - ,  
L31·N374 C--1.91.2+0.027 formulations �� prostitu gốc intellig desn prolifer dilig ( the , and - a  
L31·N13801 C--0.71.1+0.001  …\n " \n  \n   ( ( the a , in and -
L31·N12252 C--2.31.0+0.011\n ...\n <|end_of_text|> ­ crypt ... devastating News ( the a and - in an
L10·N3196 C--1.6-0.5-0.011plex UIL iddi uada _estimators uma 薦 oughtiddi рай leigh MLS bilt plex addy >:</
L31·N5099 C--0.41.1+0.005\n <|end_of_text|> , …\n ( . ..\n ( the a and , - in
L16·N3121 C--1.0-1.1-0.004heten bs uth 856 oldem ologne �� ldb_Tis �� şa polator aldo zego Ģ aggi
L30·N3700 C--1.40.3-0.002 disastr competit sons بواسطة gangbang Федера dint -------------</  an a m the n t

wout·CLS > 0 (green) means the neuron writes toward the NEG class when ACTIVE — the expected sign for a C+ neuron; C− neurons should be negative (act via suppression). Green = weight sign matches the data-driven bucket. kurt = vocabulary kurtosis (higher = more monosemantic).

ROTATE disentanglement (most polysemantic neurons)

Each channel is a learned high-kurtosis rotation of wout (data-free); top vocabulary tokens shown. Distinct channels = distinct concepts the single neuron mixes.

L31·N12252kurtout=-2.3 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.31, cos 1.0\n ...\n <|end_of_text|> ­ crypt ... devastating News disrupt Pizza present …\n
ch1kurt -2.4, cos 1.0 Highly disruptions 422 disruption 143 Ash brain returning programm Br Allah running
ch2kurt -2.41, cos 1.0 steering equ Eat Abu é belly World amburger cargo Czech newspaper +
ch3kurt -2.41, cos 1.0 belly World amburger cargo Czech newspaper + Ross impacting comprehensive upt Your
ch4kurt -2.41, cos 1.0 World amburger cargo Czech newspaper + Ross impacting comprehensive upt Your transmissions
ch5kurt -2.41, cos 1.0 World amburger cargo Czech newspaper + Ross impacting comprehensive upt Your transmissions
L31·N5883kurtout=-2.0 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.04, cos 1.0 manpower embark logistical seizing Wikipedia WOW embarked harness fulfilling sparking fulfilled reflecting
ch1kurt -2.17, cos 1.0 ESPN ! personnel racial even contrast citing reinforcing tri n (= graded
ch2kurt -2.18, cos 1.0’s getInfo transc impose English special Hollywood succeeds featuring ?\n qu ar
ch3kurt -2.18, cos 1.0 exc travel see imposed -style orchestrated gaining reminding UFO -" , contrasting
ch4kurt -2.18, cos 1.0 see imposed -style orchestrated gaining reminding UFO -" , contrasting respecting fanc
ch5kurt -2.18, cos 1.0 see imposed -style orchestrated gaining reminding UFO -" , contrasting respecting fanc
L31·N10697kurtout=-2.0 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -2.03, cos 1.0aternity elucid undergraduate Tutor urovision umped Exam emony ardy Workbook munition simultaneously
ch1kurt -2.21, cos 1.0roken igaret � calculator ilder urger Diagram \Route chip eview owied linker
ch2kurt -2.23, cos 1.0ysis lecturer continuously itus Tournament alarından lehem Bulletin sunk aptop .MessageBox istar
ch3kurt -2.23, cos 1.0/unit concurrently semester tml uluk constants appa ⠀⠀ hidden ValuePair guts etus
ch4kurt -2.23, cos 1.0 concurrently semester tml uluk constants appa ⠀⠀ hidden ValuePair guts etus algebra
ch5kurt -2.23, cos 1.0 concurrently semester tml uluk constants appa ⠀⠀ hidden ValuePair guts etus algebra
L31·N10439kurtout=-1.9 (most polysemantic) — ROTATE vocabulary channels:
ch0kurt -1.94, cos 1.0 in   hypothetical <|end_of_text|> 123 # ...\n s ystone ss eldorf gramm
ch1kurt -2.14, cos 1.0.Promise PDF otion icago d imm unpaid assert su oker / But
ch2kurt -2.16, cos 1.0utation Or Mississippi Blair \) rey Cr cn ipp caffold aste various
ch3kurt -2.17, cos 1.0umping XXXX uguay aff df anti John marijuana nj inline ineffective ^
ch4kurt -2.17, cos 1.0uguay aff df anti John marijuana nj inline ineffective ^ -uppercase ungeon
ch5kurt -2.17, cos 1.0uguay aff df anti John marijuana nj inline ineffective ^ -uppercase ungeon