按年被提到
20262026 · 2 次2026
AI 安全、伦理与社会影响›大模型与人类行为·arXiv:2603.10011 · cs.CL

Gemma Needs Help: Investigating and Mitigating Emotional Instability in LLMs

Anna Soligo, Vladimir Mikulik, William Saunders · 2026
开始阅读 ↗PDFalphaXiv 讨论
被提及 · 2 人首次 2026-05-25 · 最近 2026-10-01
2不同的人
2活跃月份
2讨论串
2被引(OpenAlex)
arXiv 摘要(候选层不写导读)

Large language models can generate responses that resemble emotional distress, and this raises concerns around model reliability and safety. We introduce a set of evaluations to investigate expressions of distress in LLMs, and find that these surface emotional instability in Gemma and Gemini models, but not in other families. We find evidence that this difference arises in post-training. Base models from different families (Gemma, Qwen and OLMo) show similar propensities for expressing distress. However, instruct-tuned Gemma expresses substantially more distress than its base model, whereas instruct-tuned Qwen and OLMo express less. We find a simple mitigation for this: direct preference optimisation on just 280 preference pairs reduces Gemma's high-frustration responses from 35% to 0.3% in our evaluations, generalising across question types, user tones, and conversation lengths, without affecting capabilities. These findings show that emotional instability is an issue in some LLMs. We present (1) evaluations to track this behaviour, and (2) a mitigation without downsides in Gemma, with the caveat that upstream training modifications to improve emotional robustness would be significantly better than this post-hoc fix.

候选层:至少 2 个不同的人、在 2 个不同月份提到过,还没到主书架的 3 人 3 个月。没有做立场标注,所以只写“被提及”。

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
评论sometimelurker · 2026-05

“> functionally mirror emotional ones I'm fine with the idea that a machine can be "worried" it wont be able to accomplish a task, and copes with this "worry" by cheating a little and making the task seeming done. (I don't like that this happens, I'm fine with the idea that "worry" in this context is a functional emotion) also https://arxiv.org/abs/2603.…”

Anthropic Cofounder Chris Olah's Remarks on Pope Leo XIV's "Magnifica Humanitas"原评论 ↗
评论NiloCK · 2026-10

“This is true, but Google's models have now had a consistent history of lower psychological* coherence / consistency. See, eg https://arxiv.org/abs/2603.10011 (Gemma Needs Help), or search for recent "Gemini shame loops", where gemini flash models stop producing output other than SHAME SHAME SHAME... * - as in, Skinner psychology. The set of observable be…”

Gemini 4 Argon原评论 ↗