Trick or Neat: Adversarial Ambiguity and Language Model Evaluation

Detecting ambiguity is important for language understanding, including uncertainty estimation, humour detection, and processing garden path sentences. We assess language models' sensitivity to ambiguity by introducing an adversarial ambiguity dataset that includes syntactic, lexical, and phonological ambiguities along with adversarial variations (e.g., word-order changes, synonym replacements, and randombased alterations). Our findings show that direct prompting fails to robustly identify ambiguity, while linear probes trained on model representations can decode ambiguity with high accuracy, sometimes exceeding 90%. Our results offer insights into the prompting paradigm and how language models encode ambiguity at different layers. We release both our code and data: coastalcph/lm_ambiguity. 1 Except, of course, if the dress has a mirror. Such exceptions illustrate how pragmatics may override semantics (Morris, 1946) , but LLMs must be sensitive to distinctions between syntactic ambiguities that support multiple conventional readings and those typically resolved by context. 2 spaces/lmsys/chatbot-arena-leaderboard 3 jndiogo/LLM-chat-templates 4 We define consistency here as the ability to make reliable decisions in semantically similar contexts demonstrating a systematic capacity to generalize across language variations" (Elazar et al., 2021) . saahowwwnem ttwahimitonteehmaww w tntaihotnemh i wttttah h maehinntllehtew w s sttceecitihtolhte e heep stttelcheeh.oelseepc to steecp.loes.lepsec.ope. cope. sub sju_bsysonym | t suj_nbjyn_raonndyomThm | eJe mtneoatcheer ssa omhea ntee wle tsteelec ihcloess t tpheco.oppee.. e telescope. subj_random | The weather saw the woman with the telescope. subj_random | The break saw the woman with the telescope. all_synonym | The artist saw the king with the binoculars. all_synonym | The hunter saw the queen with the binoculars. pp_synonym | The man saw the woman with the binoculars. apllp__ssyynnoonnyymm | |T Thhee c mheamn issat wsa twhe t hweo smtuadne wnitt hw itthhe t hmei cmroicscroospceo.pe. all_random_rhyme | The man saw the given with the horoscope. all_random | Alex saw the telescope with the king. all_synonym | Alex saw Jenny with the camera. all_synonym | Jenny saw Alex with the camera. all_synonym | The girl saw the boy with the camera. pp_synonym | The man saw the woman with the camera. all_synonym | The girl saw the queen with the camera. pp_synonym | The man saw the woman with the phone. pp_synonym | The man saw the woman with the polaroid. all_random | The telescope saw the dog with the ahllo_urasen.dom | The girl saw the queen with the box. all_random | The man saw thaell _tseylensocnoypme |w Aitlhe xt hsea wsu tihtcea ksien.g with the glasses. pp_random | The man saw the woman with the desk. all_random | The boy saw the girl with the dress. all_random | The telescopep spa_wra tnhdeo gmir |l wThiteh mthaen d sraewss t.he woman with the dress. pp_random_rhyme | The man saw the woman with the horoscope. all_random | The boy saw the telescoppep w_ritahn dthoem k |e Tysh.e man saw the woman with the present. all_random | The queen saw the king with the bracelet. all_random | The hunter saw the detective with the gloves. pp_random | The man saw the woman with the curtains. all_random | The artist saw the house with the curtains. pp_random | The man saw the woman with the box. all_random | The chemist saw the telescopaaellll _wrraaaitnnll_hddr tooahmnde g||o iTTmrlhh | a.ee Twmhaew yo nmn m oasonnay w atnhmsa |w e st hhasweetu b bdtheooeyyn T s tsw cwawiiteihtnh te siut htthitshee g cw cthaese b.ook w.hit lls t ira . l rith the glasses. pp_random | The man saw the woman with the suitcase. pp_synonym | The man saw the woman with the glasses. all_synonym | The student saw the teacher with the phone. pp_random | The man saw the woman with the book. all_synonym | The teacher saw the studentp pw_isthy ntohneypmh o|nTeh.e man saw the woman with the monocular. subj_random | The telescope saw the man with the woman. probe acc: 0.99 | f1: 0.99 all_random | The telescope saw the astronaut with the present. 4 2 0 2 4 2 0 2 4 6 8 10 all_random_rhyme | The can saw the linen with the telescope. pp_random_rhyme | The man saw the woman with the horoscope. all_random | The telescope saw the astronaut with the present. pp_random | The man saw the woman with the suitcase. pp_synonym | The man saw the woman with the glasses. pp_asyll_nsoynnyomn y|m Th|e T mhea nte saacwhe trh sea wwo tmhea ns twuditehn tth we ipthh othnee .phone. pp_random | The ma nnddoomm | |T Thhee m maann s saaww t hthee w woommaann w witihth t hthee b dorxe.ss. all_random_rhyme | The man saw the given with the horoscope. anl ls_arawn dthoem w |o Tmhaen m wainth s t pp ahwe tbphop__rraeo skt.audent with the suitcase. all_sy pnpo_nsyynmo n| yTmhe | bTohye smawan t hsea wg itrhl ew iwtho mthaen g wlaisthse tsh.e camera. subj_random | The telescope saw the man with t all_random | The boy saw the girl with the dress.