Consistency of Large Reasoning Models Under Multi-Turn Attacks

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Li, Yubo, Krishnan, Ramayya, Padman, Rema
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915854962130944
author Li, Yubo
Krishnan, Ramayya
Padman, Rema
author_facet Li, Yubo
Krishnan, Ramayya
Padman, Rema
contents Large reasoning models with reasoning capabilities achieve state-of-the-art performance on complex tasks, but their robustness under multi-turn adversarial pressure remains underexplored. We evaluate nine frontier reasoning models under adversarial attacks. Our findings reveal that reasoning confers meaningful but incomplete robustness: most reasoning models studied significantly outperform instruction-tuned baselines, yet all exhibit distinct vulnerability profiles, with misleading suggestions universally effective and social pressure showing model-specific efficacy. Through trajectory analysis, we identify five failure modes (Self-Doubt, Social Conformity, Suggestion Hijacking, Emotional Susceptibility, and Reasoning Fatigue) with the first two accounting for 50% of failures. We further demonstrate that Confidence-Aware Response Generation (CARG), effective for standard LLMs, fails for reasoning models due to overconfidence induced by extended reasoning traces; counterintuitively, random confidence embedding outperforms targeted extraction. Our results highlight that reasoning capabilities do not automatically confer adversarial robustness and that confidence-based defenses require fundamental redesign for reasoning models.
format Preprint
id arxiv_https___arxiv_org_abs_2602_13093
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Consistency of Large Reasoning Models Under Multi-Turn Attacks
Li, Yubo
Krishnan, Ramayya
Padman, Rema
Artificial Intelligence
Computation and Language
Large reasoning models with reasoning capabilities achieve state-of-the-art performance on complex tasks, but their robustness under multi-turn adversarial pressure remains underexplored. We evaluate nine frontier reasoning models under adversarial attacks. Our findings reveal that reasoning confers meaningful but incomplete robustness: most reasoning models studied significantly outperform instruction-tuned baselines, yet all exhibit distinct vulnerability profiles, with misleading suggestions universally effective and social pressure showing model-specific efficacy. Through trajectory analysis, we identify five failure modes (Self-Doubt, Social Conformity, Suggestion Hijacking, Emotional Susceptibility, and Reasoning Fatigue) with the first two accounting for 50% of failures. We further demonstrate that Confidence-Aware Response Generation (CARG), effective for standard LLMs, fails for reasoning models due to overconfidence induced by extended reasoning traces; counterintuitively, random confidence embedding outperforms targeted extraction. Our results highlight that reasoning capabilities do not automatically confer adversarial robustness and that confidence-based defenses require fundamental redesign for reasoning models.
title Consistency of Large Reasoning Models Under Multi-Turn Attacks
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2602.13093