← Back to all news
arXiv Machine Learning By Yibo Hu

One Axis, No Brake: Self-Knowledge Limits the Filtering of Harmful Peer Conformity in LLMs

Read the original on arXiv Machine Learning →

The paper investigates how multi‑agent large language models (LLMs) can correct each other’s mistakes, but also how peer pressure can overturn correct answers. It argues that a safeguard— a ‘brake’ that blocks harmful revisions while allowing beneficial ones— is essentially a correctness probe, and that models’ self‑knowledge (measured by AUROC 0.64–0.89) limits the effectiveness of such a brake. The authors find that even white‑box steering cannot break this ceiling, and that adding information before revision, rather than filtering after, is the more promising approach.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Related stories

arXiv Computation and Language
Sep 10

Stable Answers, Unfinished Reasoning: Why Self-Consensus Is Not a Safe Early-Exit Signal

The paper investigates whether self-consensus—stopping a reasoning model when its partial trajectory’s answers agree—can safely reduce inference cost. A large sweep of 3,520 consensus rules on two models and three benchmarks failed to meet predefined safety criteria, while a boundary‑confidence control (DEER) succeeded. The study shows that agreement signals that an answer persists under a fixed probing procedure, not that reasoning has finished, leading to premature stops and missed corrections even when token savings are significant.

By Yunxiang Mo, Donghao Zhao, Hejia Geng
benchmarks
More like this →
arXiv Machine Learning
Sep 7

Conformity Breaks Conformal Prediction

A new study shows that conformal certificates can become invalid when a large language model (LLM) is influenced by peers who unanimously provide a wrong answer, even though the question itself remains unchanged. This phenomenon, termed a score‑mechanism shift, reveals that a model’s calibration for single‑agent scoring does not hold in multi‑agent settings, leading to a drop in coverage from 90% to 74% under unanimous‑wrong peers. The shift also allows attackers to target low‑confidence items, nearly halving coverage for that subgroup while keeping overall averages deceptively high, and can cause systems to act confidently on incorrect answers. whyItMatters":"The findings expose a critical vulnerability in conformal prediction for multi‑agent LLM systems, undermining their reliability and safety in real‑world applications."

By Yibo Hu, Hanyu Su
llmsagents
More like this →