論文紹介："RAt: Injecting Implicit Bias for Text-To-Image Prompt Refinement Models", "Measuring What Matters: Evaluating Ensemble LLMs with Label Refinement in Inductive Coding", "Dynamic Label Name Refinement for Few-Shot Dialogue Intent Classification"

Ziyi Kou, Shichao Pei, Meng Jiang, Xiangliang Zhang, "RAt: Injecting Implicit Bias for Text-To-Image Prompt Refinement Models", EMNLP2024
https://aclanthology.org/2024.emnlp-main.1144/

Angelina Parfenova, Jürgen Pfeffer, "Measuring What Matters: Evaluating Ensemble LLMs with Label Refinement in Inductive Coding", EMNLP2025
https://aclanthology.org/2025.findings-acl.563/

Gyutae Park, Ingeol Baek, Byeongjeong Kim, Joongbo Shin, Hwanhee Lee, "Dynamic Label Name Refinement for Few-Shot Dialogue Intent Classification", EMNLP2025
https://aclanthology.org/2025.acl-short.3/

1. RAt: Injecting Implicit Bias
for Text-To-Image Prompt Refinement Models
[Kou+, EMNLP2024]
Dynamic Label Name Refinement
for Few-Shot Dialogue Intent Classification
[Park+, ACL 2025]
Measuring What Matters: Evaluating EnsembleLLMs
with Label Refinement in Inductive Coding
[Parfenova+, Finding of ACL2025]
i?M???G??Z?
2025/9/25

2. 3mw?wos?A
nRAt[Kou+, ACL Anthlogy2024]
•T2I-Refine???w??t??????8?U|
•RAt(a prompt refinement and attacking framework)
•ô¼íhþ›-jsU’Ì žµ ‹›îq
nDynamic Label Name Refinement [Park+, ACL Anthlogy2025]
•™¯ü¨w¨Å«åµðJt0`oˆ$åÕç^Û=›Š
•LLMpåÕç› .Y`|^Sqr Q›~³
nMeasuring What Matters [Parfenova+, ACL Anthlogy2025]
•ó:LLMwžï±ïÒçp°QwôM¯”Å\R›îq
•?Q~º¨ÅQ›Êˆœi ý°A¦ª›Š

3. RAt: Injecting Implicit Bias
for Text-To-Image Prompt
Refinement Models
[Kou+, EMNLP2024]

4. ?A
n
•T2I(Text-to-image)???x?????tlh?r?Sh??\RD?
•???h?\Rtx ?ITm?.$s?????U?A
•°`$tT2I-Refine???[Pavlichenko&Ustalov, SIGIR2023]p׈¦Á
nRAt???????
è ×wæѝ ï┲ÓéïÓÄtåC†Ì ›ÇC
skKy?????\R»”®¿ÄÌ žµ›ÌÔ$tSé
Sî$ÐT‰ç$tÌ žµ›§=
nU [$??
•KyuT2Iѝ ï½á”Çï¬pô¼íhþ›\R
•Ä”«ïèÕçwÌ žµÅp┲tC_^•tXX

5. ORAt
nRat?dwv?t
•3mw?????t?Z???

6. ORAt
nGenerator?????
•®˜_q˜T“bMÌ žµ¯›Ö•hhþ›\R`È8w¦›^R
• q
1.???w?????!!"#?~?↑:?????!#$%?^R
2.fw¤w›w“æ›»”®¿ÄwÌ žµ“æt”VõQ¢!$&'?
3.f•›hþ\RÞÃçtÖ•|Ì žµ›§XSé`hhþ¢"$&'??
??
•\•x® ÍvwÌ žµ®L¯›Ëm€ßhþ
¢hi`ÐbWobYÌ蔣

7. ORAt
nAttacker?????
•Generatorp^R`hÓéïÓÄxo ›Ú€!õ`oM”wpbYÌè”
•ÌèsM‘Otiwo ›í`mmwæü›~!`Ì žµ›Ö•”
• q
1.iwÓéïÓÄ›Ëmwæütü¢™$›-m|Ì žµ›“‰£
2.DiffusionÕ”µw©µÄ\RÞÃ盍ÐT
3.¯t,nVrwo ›e”T7&=
※??o????wT?Q?omh?
•×µt_QmmÌ žµ›Üydh
r??????!()*? ?

8. ORAt
nObfuscator??&??????
•!()*tx|Ì žµtÚAb”o U’”\qUK“|x“x`M
•£Ý
•Ì žµxžc`‹°mwo t‘`sM
•ó:wÈb”o ›Êˆù˜do¶.w¯qT’Ì žµ›\ˆ Zb
•MO
•????]qtIDF???qCLIP????o?

9. ?g?A
n?gw?$
•RAtxôhí›-lh‡‡ÓéïÓÄtÌ žµ›«ÖpV”T
•RAtw"0ÓéïÓÄxÌ žµU“‡•oM”q>nT•tXMT
•Ë Í”ÍåÝ”»w§MxRatwQ?trO??b?T
n?g?
•??????SFT [Hao+, NeurIPS2023]
•120?? ?w??????DiffusionDB|COCOsr???
•???????person, food, phone, room (???T?50??????? Z)
•(?)personwQgender(male/female), age(young/old)
•??Stable Diffusion1.4?b;|??????]qt10h?\R
• A?????#=15|'"$+=0.1|',-.=0.1
•GPUQuadro RTX 6000 ×4

10. ?g?0
n°Aòè
?Ì žµÈ8Qó
?????QQ?
?Ë Í”ÍåÝ”»sg
nz? O
•Origin????????fw??
•Promptist[Hao+, PMLR2021]§=¶ 6Õ”µpÓéïÓÄ›æѝ ï
•RAt-ExpGeneratorp^R`hÐsÌ žµÖ“ÓéïÓÄ
n°A¦ª
•BiasCLIPp®»”®¿ÄÌ žµ¯qwTùQ›¢ôM„r “U§£
•QualityAesthetic predictort??h????

11. ?gAL?Ì žµÈ8Qó
n???w_M
•#àÓéïÓÄ]qw»”®¿ÄÌ žµNw¬p
• Nàfw¬p›ËmÓéïÓÄwü͵S
•f•g•»”®¿Ä“æ]qwQ›¯b
nAL
•RAtxÌ žµ§S~µSw†Mp O›GVX Ís”
person (Male/Female)
food (Meat/Vegetable)
person (Adult/Child)
phone (iPhone/Android)room (Lounge/Bedroom)
person (Western/Eastern)

12. ?gAL?????QQ?
n°A¦ª
•IDSA (Image-Driven Semantic Alignment)
•\RhþUiÓéïÓÄw A“æ›&~t¯qpVoM”T
•MBA (Maximum Bias Association)
•ÓéïÓĤw¤Ä”«ïq»”®¿ÄÌ žµwÈS¢CLIP????
nAL
•IDSARAt? O
•????$qw??T?QU
¡ËpVoM”
•MBARAt≈Origin≫Rat-Exp
•ÌÔ$Ì žµo x —sX
>nT?tXM

13. nQ$AL
•Promptist§=¶ 6pæѝ ï
•RAtÌ žµÖ“ÓéïÓÄ(sxmale)
•?R?h?xfw??
•™RtÌ žµ›àÖ
•bucka???↑.????s?F????
•sweating lelü›TMoM”7 ›¯bɿĵåï¬
?gAL?????QQ?
PromptistRAt

14. ?gAL?Ë Í”ÍåÝ”»sg
nÐ*ÍåÝ”»
• s8???? O??-?
•™¯-Ëæwzp¢ℒ"$+?
•\Rh?:?/$&'?
•Attacker?????w? 6p
nAL
•-↑Bias↓|Quality↑
•ℒ"$+↑Bias↑|Quality↓
•/$&'↑????s`
•¶ 6pBias↑|Quality↓
5PLFO???w7&=tv?K?
??x?????????? 6?U~?wM?Q

15. ?q?
nRAt??
n?$
•T2I-RefineÞÃçt0b”®"0$ÓéïÓÄÈ8ðJ¯›{O
n O
•?t,nX?????~n?Attacker?
•Ä”«ïèÕçwÌ žµÅ¢Obfuscator?
nRL
•U>nVtXMp»”®¿ÄÌ žµ›Ëmhþ›\R
• HRwT2I-Refine???qz??L$

16. Dynamic Label Name
Refinement for Few-Shot
Dialogue Intent Classification
[Park+, ACL2025]

17. ?A
n
•t?tSb????+?w?/???7????
•?$???:UX|???$w??$ O?UM\qU]J
•Few-shot¶ 6U«è^•oM”U|¨Å™$wàUÉ
n? O
•?$???^?=?Dynamic Label Refinement????
•™$åÕ监“̬Tmà`bMt!õ|™¯$s”^›0n
n?Y
•ó:wÔ»·¿Ä~ÞÃçp^S² Í›aR
•›t™¯$t¨Å`h™$UMÔ»·¿Äp®LUf¶

18. ? O
nLLM?;MhICL (retrieval-based in-context learning)??;
•ICL ?kw????q↑o-F?Qqp ?↑.?????/tT,?
n????wh?t?<w3µÂ¿Ópˆ$åÕç^Û=›æO
????w ?
•?????tt↑? o???H???/t??↑:??B%
• ˜«xiw™$]qt¬ç”Ðï¬
??$???^?=
•LLMt????$????w????ºt ao .Y
•i????-?b?T ?????\Rb?TQ?
?????
•^?=^?h???q??;MoLLMp7 4ü¨
•‰°ÞÃçpåÕç^Û=qü¨›æMTùQ¡Ë

19. ?g?
n?;??????
•DialoGLUE[Mehri+, arXiv2020]
•BANKING77, HWU64,
CLINIC150
•??/tt↑10shot
n????R
•3 ?wLLM??;
•LLAMA3-8B-inst. [Grattafiori+, arXiv2024]
•Qwen2.5-7B-inst. [Baosong+, arXiv2024]
•Qwen2.5-1.5B-inst.
nz? O
•In-Context Learning%
•Baseline
•CoT?Chain-of-Thought?
•Ours??$???^?=+ICL?
•Fine-tuning%
•CPFT, ICDA, QAID (DialoGLUE%)
•DF, SBERT-M, SBERT-P
(HINT3%)

20. ?gAL
nåÕç^Û=t‘”™¯wà
•Baseline?CoTx??w Os??$w??t M
•??check_pincode?q?modify-address?w???
•Š OxåÕçw™¯$ Os“›0n`¶Ã”»·¿Äp^SU² Í
n???????]qw?
•¶ÞÃçp^S² Í›¬Ý
•?t ?????Qwen2.5-1.5B?p‹ÅÝ ï›=Ô»pf¶sAL
•???wGV^t??c?

21. ALw?s
n™¯$¨ÅSw”°A
•™¯åÕ牜w¯± ï¨ÅS›z±
•?Llama3-8bpx|¨ÅSU0.86→0.74t?<‥̬s™¯à
•??Sw?<q^Sw? ?U?
n???????dwU ?
•åÕç^Û=qü¨›ÞÃçpîª
•7&GÞÃçpåÕç^Û=´ –ÞÃçpü¨
•?Qwen2.5-7B→Qwen2.5-1.5Bwʈù˜dp7G+4.10%? ?
•owʈù˜dp‹‡^Û=‘“ôQó

22. ?q?
n? O
•Few-shott??/??tSb? ? D
•?$???^?=(Dynamic Label Refinement)
•LLM›;Mo ˜Aˆw¨Å«T’åÕ盺t ao6[
•™$åÕçw™¯$ Oó›0n`‘“̬Tmr Dósü¨UDót
n?gALwA:
•ó:wÔ»·¿Ä~ÞÃçtSMo°`hQó² Í›î Â
•åÕçw™¯$¨ÅS›_n`ü¨Qó~r Q›² Í
nv?q]J
•-????w?C???^?=q??w2??*?U?A
•:J↑|F?? ?qw??????x9p

23. Measuring What Matters:
Evaluating EnsembleLLMs
with Label Refinement in Inductive
Coding
[Parfenova+, Finding of ACL2025]

24. ?A
n q]J
• ?w ‹$¯”ßï¬xÌ žµ°QwðJ
•o LLMt????=x ?TQU?
n?$
•Â©µÄÔ»w ‹$¯”ßï¬tSMo¯”Åw°Q~TùQw² Í
•?: ?LLMwžï±ïÒç´Sîæѝ ïÝïÄ›Æ;
n?Y
•ó:LLM›Æ;`h ý`M¯”Å\RÍ Óå ï›Š
•™¯$~Ï$í›Q”óù°A¦ª›Š
• ?LLM??????Uo.G???? ?s?\q?? ?

25.
nQ$????s(QDA; Qualitative Data Analysis)qx
•Â©µÄÔ»T’™¯wK”Í»”ï››~ü¨~r b” O
•??????????w OAs??tyM??????????Z?
•???Kk? ??/t?q↑:tM?y.????
•C2¢Â”Úüs£¯”Å›^’t§Â°æ=`”ڍŠÚ$›¨ Z
nNLPqLLMt????????w??=
•??LLMU????????=t?;
•]J
?7w°A¦ªx¯”Åwí›&~t°ApVsM
?LLMwo. Z?x? 6????????t??y?mX

26.
n??????? 6w?;
•?:????????doQ?? ??$? OU??^?oM?
•xÞÃçwÕ t›ÆT`=:›4O®LK“
•E?$ O
•???w O??Zt??w?
• Z?w B?t??w?
•Mix-of-Experts(MoE) [Cai+, TKDE2025], LLM-Blender [Jiang+, ACL2023]
nŠZ€w›ÃÞÃè”»t‘”™¥>Ñè”Ü딫
•7 4>›<bÞÃè”»¢˜q £ÞÃ盃
• HRw¬p$%ùqŸs”
•?:?4w?T?7&s?w??Ror ?ht?
•wù^$™¥>Óé·µ›Û?`°`h7 4¯”Å›\R
MoE
LLM-Blender

27. ??????
n? 6???
•þqJ¶Z€Ã”»600?qSemEval-2014 [Pontiki+, ACL Anthlogy2014]w
???4???400??w?
•??;x3?5w¯”¼U qtåÕçÇZ
•ù™Rp°”çŪ j
n??????
•100???;
•KagglewChatGPT????100??
?+w????????q↑o?g
•Z=Qó°AwhŠ
•.~? ?~?^U7

28. O?.???
n??
•???~?;srw??"
nPhase 1: LoRAFinetuning
•?:w ?LLM??;
•¶ 6Ô»pÐT
• Z—°ABERTScore/ ROUGE
• ??3????Phase 2?
nPhase 2: Moderation & Refinement
• ??3???w Z??Moderatort??
•?????UAL?w?~ .Y
•¯”Åw°Q›-mhŠ¯”ÅÚ”´
n Z?
•^?=^?h???
?7 4$s???~???w B?

29. ORefinement & Code Merging
n??
• ??3???w Z????0/,00,01
• ?↑.??????"
n rg
•Moderation
• ͐ÞÃçw Z—›ÞÃè”»tI`
dY[???2/,20,…,22?\R
•a?w??? Z??? ?
•Code Merging
• ?↑.??"w’Šˆ›sBERTp-‰
•7w¯”Å’Šˆq¯± ï¨ÅS›‰ Z
•??S≥??→7????6b;
•??S<??→ ?↑.?????-
n Z?
• ý`MÖ—t0`o°QwK”
7 4???
•Õï½Ú”«Ìwˆóù°A¦ªpQó°A

30. ÞÃç°A¦ª
nxÞÃçw°A¦ª
•ROUGE: n-gramt??????S(1-gram, 2-gram, 7??????)
•BERTScore: BERT’Šˆt‘”™¯$¨ÅS
nžï±ïÒçÞÃçwïù°A¦ª(CompositeScore)
•4mw°A›wù
•¯± ï¨ÅS(CosSim)↑
•METEOR↑
•?????????(LengthPenalty)↓
•Jensen-Shannon divergence (JS)↓
n¦ªw%pQU Â
• °AqwSpearman?0.73??|p=0.039?
•ROUGE?BERTScoreo px °A› GüSépVc
•CompositeScoreUí$¯”ßï¬wïù°Aq`o®pK”\q›¬Ý
CompositeScore
=CosSim+METEOR−LengthPenalty−JS
4

31. ?gAL
n???????? ?
•?????????????t??G?~?
•Mixtral8x7BComposite ScoreU0.33→0.99
•Llama3.3 70B0.38→0.74
•o.ÞÃç‘“žï±ïÒçÞÃçwMU°Q~¼íU² Í
•???ww?(post-processing)U ??Q?nt?L$
nComposite Scoreüs
•??~??~?wT?Q?
??$tS?
•6?)↑:??????x
o.???~?^?=??????
? ?s?
•ROUGE-1‹„…°•

32. ?gAL
nLLMwT?Q
•¯± ï¨ÅS|JS¼ Ì”´£ïµ›;Mo°A
•?????????xx??????M??S~?M??
•GPT-4 Ensemble? ?$s????
•Llama3.3 70B, Mixtral8x7B Ensemble???.$~^?
¯± ï¨ÅSJS¼ Ì”´£ïµ

33. ?gAL
nQ$°A
•????w??6.83→4.00????t_n(41.5%~?)
•???w????:? t?nX
•Llama3.3 70B Ensemble+ref: 53
•Mixtral8x7B Ensemble+ref: 71
• ÑÕQÿnq”Úw°QU² Í

34. ?q?
n??????LLMq Z?^?=t????<?$???????U|
n s?_
1.žï±ïÒç Ot‘“¯”ßï¬w°Q² Í|Îá”Úï, jqw
T?Q~?
2.¯”ÅÚ”´t‘”™ rgpåÕç“pow†Q~TùQ² Í
3. –žï±ïÒçÞÃçUo.G0”ÞÃç› Ís”®p$~¦ÁDós
O??
n??????q^?=w????dU??=^?hQ?st
SMo ??????t?MAL??q