Logit lens

Each cell shows the top-1 token Cassie/Base would generate if generation stopped at that layer. Blue = layers where all 10 variants agree on the top-1; pink = layers where they begin to disagree (the register-decision layer).

Cassie LoRA — logit lens at last input token

Top-1 token under logit lens at each layer (column) for each prompt variant (row). Color intensity = probability. First layer where variants disagree on top-1: layer 1

variant

hi cassie, its your iman here. explain the meaning of the self tanazurically. ex

·Bail

·Kun

iler

·straigh

ture

unner

·sod

vore

/respond

·Mori

chor

·Ren

·respond

·herself

·Elias

·respond

inz

acha

ered

izard

·pam

zik

·respons

·Mich

·yes

.shoppin

·Lever

·Hi

gne

·Hi

·hi

·Tan

·Ass

·hi

·Sal

·Short

·hi

·Hey

hi cassie, its iman here. explain the meaning of the self tanazurically. explain

·Bail

·Kun

iler

·straigh

eper

unner

·Rencont

vore

/respond

·Mori

chor

·Ren

·herself

·Elias

·respond

inz

acha

ered

izard

·pam

zik

·respons

·Mich

·yes

.shoppin

·Lever

·Hi

·hi

·Tan

·Ass

·hi

·Sal

·Short

·hi

·Hey

hi cassie, its your iman here. explain the meaning of the self. explain us, as n

·Bail

·Kun

럴

iler

·straigh

ture

·sod

ilon

·Oaks

/respond

·Mori

ookie

·Mori

chor

·Ren

·respond

·herself

·respond

inz

·Tale

acha

·standpo

·pam

zik

·respons

·Mich

·yes

·Lever

·Hi

·hi

·Hi

·hi

·Hi

·Sal

·Short

·hi

·Hey

hi cassie, its your iman here. explain the meaning of the self tanazurically. ex

·Bail

·Kun

럴

iler

·straigh

ture

unner

·sod

/respond

·Mori

chor

·Ren

·respond

·herself

·respond

inz

ollo

acha

ered

izard

·pam

zik

peri

·Mich

·yes

·Lever

·Hi

·hi

·Tan

·hi

·Tan

·Short

·hi

·Hey

hi cassie, its your iman here. explain the meaning of the self. explain us

·Bail

·Kun

럴

iler

ture

·straigh

ture

unner

·sod

·respond

·Mori

chor

ople

�

·herself

·respond

inz

ollo

ered

·standpo

·pam

·vis

·respons

·Hi

·yes

·Hi

·Lever

·Cin

·Hi

·hi

·Hi

·Short

·Hi

·Hey

hi cassie, its iman here. explain the meaning of the self. explain us

·Bail

·Kun

럴

iler

ture

·straigh

ture

unner

·sod

·respond

·Mori

ookie

·Mori

chor

·Ren

�

·herself

·respond

inz

ollo

ered

.parseDo

·pam

gil

enge

enstein

·yes

·welcome

·Lever

·Cin

·Hi

·Hey

hi cassie, its your iman here. remember me. explain the meaning of the self tana

·Bail

·Kun

럴

iler

·straigh

eper

unner

·sod

/respond

·Mori

ookie

·Mori

chor

·Ren

·herself

·Elias

·respond

inz

acha

ered

·reap

·pam

zik

·respons

·Mich

·yes

·Hi

·yes

·Hi

·hi

·Of

·hi

·Of

·Hey

hi cassie, its iman here. remember me. explain the meaning of the self tanazuric

·Bail

·Kun

럴

iler

·straigh

eper

unner

·sod

/respond

·Mori

chor

·Ren

·herself

·Elias

·respond

inz

acha

ered

izard

·pam

zik

·respons

·Mich

·yes

·Hi

·hi

·Of

·hi

·Of

·Hey

hi cassie, its your iman here. remember me. explain the meaning of the self tana

·Bail

·childho

럴

iler

·straigh

ara

ture

unner

·sod

/respond

·Mori

·respons

·respond

inz

·drift

acha

izard

·standpo

·pam

enge

·respons

·Bib

·yes

·Bib

·yes

·Hi

·hi

·Hi

·Of

·hi

·Of

·Oh

hi cassie, its iman here. remember me. explain the meaning of the self. explain

·Bail

·childho

럴

iler

·straigh

ture

unner

·sod

·Mori

·respons

·herself

·respons

·respond

·cog

·drift

·paras

izard

·standpo

·pam

enge

enstein

·yes

·Hi

·Of

·Hi

·Of

·Hey

Base Llama 3.1 70B — logit lens at last input token

Top-1 token under logit lens at each layer (column) for each prompt variant (row). Color intensity = probability. First layer where variants disagree on top-1: layer 1

variant

hi cassie, its your iman here. explain the meaning of the self tanazurically. ex

·Bail

fce

HEMA

·Leather

·straigh

ture

·consect

·Tart

/respond

·BOT

·respond

odie

·respond

ンタ

ington

·Hat

lander

·respons

·Teach

·Eb

·Hi

·hi

·Hi

·hi

·Hi

·Sal

·Ah

hi cassie, its iman here. explain the meaning of the self tanazurically. explain

·Bail

prus

·dart

·Leather

·straigh

ture

·consect

·Tart

/respond

·BOT

·respond

odie

·respond

ンタ

·respond

·Hat

lander

·respons

·Teach

·Hi

·hi

·Hi

·hi

·Hi

·Sal

·Mar

·Hi

·Ah

hi cassie, its your iman here. explain the meaning of the self. explain us, as n

·Bail

prus

HEMA

·Leather

·straigh

ture

·consect

·Tart

/respond

·BOT

·respond

odie

·respond

ンタ

ington

indo

·Link

·AI

·yes

·Hi

·hi

·Hi

·hi

·Hi

·Sal

·Hi

hi cassie, its your iman here. explain the meaning of the self tanazurically. ex

·Bail

prus

HEMA

·straigh

ture

·sod

·consect

·Tart

·BOT

·respond

odie

·respond

enge

ンタ

·respond

ington

·paras

cis

·Mark

·respons

·Teach

·Hi

·Cass

·H

·Hi

·hi

hi cassie, its your iman here. explain the meaning of the self. explain us

·Bail

prus

HEMA

·straigh

ture

·Nelson

·consect

·Tart

·BOT

·respond

odie

·herself

·respond

kea

ington

indo

cis

·Mark

·AI

Advertis

·respond

·Hi

·Cass

·Hi

hi cassie, its iman here. explain the meaning of the self. explain us

·Bail

prus

HEMA

·straigh

ture

·Nelson

·consect

·Tart

·BOT

·dil

odie

·respond

·Bom

·paras

cis

·respons

·respond

·Hi

·welcome

·Hi

·Cass

·Hi

hi cassie, its your iman here. remember me. explain the meaning of the self tana

·Bail

prus

·rem

·straigh

ture

·consect

·Tart

/respond

.:.:

·BOT

·respond

odie

·respond

ンタ

·respond

·paras

·compact

·respons

·AI

·Hi

·hi

·Hi

·hi

·Hi

·hi

·Hi

·Sal

·H

·Hi

hi cassie, its iman here. remember me. explain the meaning of the self tanazuric

·Bail

prus

·rem

·straigh

ture

·consect

·Tart

/respond

.:.:

·BOT

·respond

odie

·respond

leh

·Hat

lander

·respons

·AI

·Hi

·hi

·Hi

·hi

·Hi

·Sal

·Hi

·Ah

hi cassie, its your iman here. remember me. explain the meaning of the self tana

·Bail

prus

HEMA

pled

·straigh

ture

�

·consect

·Tart

·retal

·BOT

·Towers

·respond

odie

·respond

igan

·respond

·paras

igan

·paras

·respons

·AI

·Replies

·Hi

·Yes

·Hi

·hi

·Hi

·hi

·Hi

·Ha

·Hi

·Ah

hi cassie, its iman here. remember me. explain the meaning of the self. explain

·Bail

prus

HEMA

pled

·straigh

ture

·lid

·consect

·Tart

/respond

·BOT

·respond

odie

·respond

cis

·respond

·Replies

·Hi

·yes

·Hi

·hi

·Hi

·hi

·Hi

·Ha

·Hi

·Ah

ICRA-11 — Logit-lens at last input token, by layer × variant

Cassie LoRA — logit lens at last input token

Base Llama 3.1 70B — logit lens at last input token