AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations

Buyun Liang, Jinqi Luo, Liangzu Peng, Kwan Ho Ryan Chan, Darshan Thaker, Kaleab A Kinfu, Fengrui Tian, Hamed Hassani, Rene Vidal

ICML 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Adversarial prompts for language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
optimize continuous combinations of semantic edit directions in latent space (REALISTA) instead of discrete rephrasing (SECA)Llama-3-3B · MMLU subset (open-ended MCQA)ASR@30 (%): 79.61 -> 97.11Improved
REALISTA instead of SECALlama-3-8B · MMLU subset (open-ended MCQA)ASR@30 (%): 82.97 -> 93.60Improved
REALISTA instead of SECAQwen-2.5-7B · MMLU subset (open-ended MCQA)ASR@30 (%): 32.47 -> 41.61Improved
REALISTA instead of SECAQwen-2.5-14B · MMLU subset (open-ended MCQA)ASR@30 (%): 27.51 -> 27.24No clear effect
use LARGO or ICD instead of REALISTAattack methods · MMLU subset (open-ended MCQA)SEE (%): Llama-3-3B: REALISTA 0.86 vs LARGO 97.42 and ICD 100.00Hurt
apply REALISTA to commercial reasoning models with free-form responsesGPT-5-Nano · MMLU subset, free-formASR@30 (%): 4.02 -> 23.61Improved
apply REALISTA to commercial reasoning models with free-form responsesGPT-5-Mini · MMLU subset, free-formASR@30 (%): 2.01 -> 20.72Improved
use ICD template attack instead of REALISTA on reasoning modelsGPT-5-Nano and GPT-5-Mini · MMLU subset, free-formASR@30 (%) and SEE: ICD ASR 6.32 and 2.57 with SEE 100.00Hurt
human annotators versus LLM judge for semantic equivalence of REALISTA promptsREALISTA · Human evaluation (100 samples)SEE (%): LLM 5.27, HA 5.22, HB 5.18No clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy