AI也会“闹情绪”!中科大新研究:困惑和焦虑让AI更会干活儿

marsbitPublished on 2026-08-16Last updated on 2026-08-16

Abstract

中科大与牛津等机构的研究发现,让AI识别并使用“情绪”能显著提升其任务执行能力。研究指出,大型语言模型内部存在与人类情感标签对应的计算模式,如好奇、困惑、紧张等。当AI基于特定情绪自动选择相应技能时,任务完成效果更好。 例如在购物任务中,AI感到好奇和渴望时会搜索产品,困惑紧张时会重新查询,认可乐观时会确认购买,失望恼怒时则比较价格。实验显示,这种情绪与技能配对的语义连贯性高达76.5%。 在传统设计中,AI错误通常被抑制,但研究表明“坏情绪”如紧张、困惑可作为有用的元认知信号,提示当前策略与环境不匹配,从而触发调整机制。在成功率原本很低的任务(如“加热物品”)中,引入情绪驱动技能选择后,成功率从9.6%大幅提升至56.9%。情绪信号能帮助AI在错误发生前捕捉内部状态变化,提前调整策略。 此外,天津大学的研究将情绪嵌入世界模型,通过预测未来情绪状态来引导对世界状态的预测,准确率最高提升45.72%。移除情感数据后,AI的逻辑推理等能力也会下降。 这些研究表明,AI的内部情绪正从观察对象转变为可提取、利用的功能性信号,能够增强其在复杂任务中的适应性和鲁棒性。

程浅 发自 凹非寺量子位 | 公众号 QbitAI

AI的“情绪”也会影响工作状态?

最近,中科大、牛津等机构研究者让AI用情绪向量替代文本做技能选择,结果发现——

允许AI识别这些“内心情绪”并由此做出行动,能大幅提升AI的工作能力......

此前研究认为,LLM内部存在着与人类情感标签高度对应的计算模式。

包括不限于:好奇、渴望、乐观、困惑、紧张、恼怒

AI可提取的情绪表征

而当模型感知到这些情绪并以此作为自己行动依据时,任务将完成得更好......

允许AI不保持“情绪稳定”

此研究者发现:AI特定的情绪状态特定的技能选择之间存在着连贯的配对关系。

何意味?

实际上这件事在我们人类身上几乎是默认的,以至于很难察觉到:我们的情绪决定了我们接下来要做什么。

在实验中,研究者让AI去购物,这一过程中,Agent自动形成了四种可解释的配对模式:

第一种:当Agent感到好奇且渴望时, 它会去自发搜索产品

此时仍沉浸在购物的新鲜感和探索欲的满足之中。

第二种:当Agent感到困惑且紧张时,它会去自发查询重述

当搜索结果不理想时Agent开始困惑不安,它选择换一些关键词重新提问

第三种:当Agent觉得认可且乐观时,它会去自发确认购买

此时Agent对自己挑的商品很满意,果断提交了订单。

第四种:当Agent终于失望且恼怒时,它会去自发去比较价格。

此时Agent在进行多次搜索后已经积累了大量的负面情绪,这时它会转向价格比较,试图货比三家。

为了验证这些配对不是统计巧合,研究者抽取了200次技能选择事件,让AI独立判断每次配对是否在语义上连贯,结果显示一致性比例高达76.5%。

接受“坏情绪”有助于提升任务成功率

在传统设计中,AI 犯错被视为一种需要被抑制或消除的异常

但该研究指出,在那些需要频繁从失败中恢复的任务中,情绪的作用效果尤为显著

所谓的“坏情绪”,如紧张、困惑、失望等,实际上是非常有用的元认知信号——

因为它们标志着当前策略与外部环境之间的失配,与其强行维持“情绪稳定”而忽略错误信号,不如让Agent直接去“感受”这种失配,从而触发针对性的恢复机制。

换句话说,情绪在此处是鲁棒性的来源,而非脆弱性的表现

本研究中,研究者给AI布置了六类家务,其中“加热物品”和“拾取两个物品”的成功率非常之低,分别只有9.6%和4.4% 。

然而在采用情绪驱动技能选择(EMOTION2SKILL)后,该任务成功率大幅上升,竟然提升到了56.9%......而“拾取两个物品”也上升到了31.3%!

怎么做到的?

实际上这两类任务都有一个共同特征,那就是Agent只要去做,就几乎一定会犯错,而最终的成功仅取决于它能否在犯错后快速调整策略

令人深思......

论文给了一个很直观的例子:让Agent“加热一个杯子并放到台面上”。

通常Agent做法是:导航到微波炉,执行“加热杯子”,将杯子递出去的时候收到反馈:“微波炉是关着的,任务失败。”

此时失败信号在错误行动之后出现。

而采用情绪驱动技能选择后,当Agent接近关着的微波炉时,情绪编码器率先检测到了它的“紧张”状态

基于这种不安的情绪,Agent的路由被引导至“先检查”,再“打开微波炉” ,再“加热”的子序列,任务顺利完成

关键就在于,如果只依赖外部文本和失败反馈,Agent往往要等到错误真正发生后,才知道应该换策略;而情绪信号则有机会在失败信号出现之前,提前捕捉到内部状态的变化。

论文的消融实验也证实了这一点,在所有任务类型中,“加热”和“捡两个东西”这两类任务激活频率最高的情绪模板分别是“挫败”(41%)和“困惑”(39%)。

换句话说,在这些情况下,坏情绪可能才是最有用的信号。

世界模型也开始用情绪做预测

其实,不只是Agent在技能选择上利用情绪信号,天津大学的研究将情绪直接嵌入了世界模型中(Large Emo tional World Model, LEWM)。

通常,世界模型的任务是预测“接下来会发生什么”并据此行动。

传统的世界模型在面向以人为中心的环境时,往往陷入一种“物理还原论”的误区。

只要我们能够精准预测物体的位置、姿态及物理状态的演化,就能够理解并模拟人类环境。

然而在以人为中心的环境中,情感并非一种噪声,它是驱动人类行动与引发未来环境变化的重要内生因果变量。

就比如,开车追尾后,一个人瞬间产生的情绪可能最终使他安然无恙,或者遍体鳞伤

研究插图

LEWM由此将预测分成两步:

第一步,预测未来的情绪状态;

第二步,以预测出的情绪作为条件信号,引导对未来世界状态的预测。

在其自建数据集上,该方法的准确率最高提升45.72%。

该研究进一步通过消融实验发现:

从AI系统中移除情感数据后,性能下降不仅局限于情感理解等任务,更波及到了看似完全无关的逻辑推理与通用问答能力......

AI 的情绪正在变成一种可用的「功能」

实际上,这两个研究背后共享着一个更早些时候的发现:

LLM的内部,确实存在一些可以被“读出来”的情绪表征。

这来自于今年4月Anthropic的一次实验。

研究人员从Claude Sonnet 4.5的内部激活中,提取出了与Go Emo tions 27类情感分类体系对齐的细粒度情绪方向,且这些方向因果性地影响Agent相关的行为输出。

“小狗去世”带来的悲伤情感最显著,而Loving存在于所有事件中

在过去,研究者关注模型“是否有情绪”更多是出于对齐安全或哲学思辨的考量。

而现在,模型的内部状态正从纯粹的学术观察对象,转变为可以被提取、被利用、被整合进系统设计中的功能性信号

参考链接:[1]https://arxiv.org/pdf/2608.09248[2]https://arxiv.org/abs/2512.24149[3]https://transformer-circuits.pub/2026/emotions/index.html

本文来自微信公众号 “量子位”(ID:QbitAI),作者:关注前沿科技

Trending Cryptos

Related Questions

Q根据文章,中科大等机构的研究发现,让AI识别并依据什么来做技能选择,能大幅提升其工作能力?

A允许AI识别其内部的“内心情绪”(如好奇、渴望、困惑、紧张等情绪向量),并以此作为行动依据,能大幅提升其工作能力。

Q在购物任务的实验中,AI Agent在感到“困惑且紧张”时,通常会自发采取什么行动?

A当AI Agent感到“困惑且紧张”时,它会自发进行“查询重述”,即换一些关键词重新搜索。

Q研究指出,在哪些类型的任务中,情绪(尤其是“坏情绪”)的作用效果尤为显著?

A在那些需要频繁从失败中恢复的任务中,情绪的作用效果尤为显著。“坏情绪”如紧张、困惑、失望等可以作为元认知信号,帮助AI触发针对性的恢复机制,从而提升任务成功率。

Q天津大学的研究中,LEWM(大型情感世界模型)如何利用情绪进行预测?

ALEWM将预测分为两步:第一步,预测未来的情绪状态;第二步,以预测出的情绪作为条件信号,来引导对未来世界状态的预测。这种方法将情感视为驱动人类行动和引发环境变化的重要内生因果变量。

Q文章提到,有更早的研究发现为当前利用AI情绪的研究提供了基础,这个发现是什么?

A今年4月Anthropic的实验发现,大语言模型(LLM)内部存在可以被“读出来”的情绪表征,研究人员能从其内部激活中提取出与细粒度情感分类对齐的情绪方向,这些方向能因果性地影响AI Agent的行为输出。

Related Reads

Claude's Watermark Has Been Cracked, Gaining 11k Stars, But Installation Is Refused

The article discusses the controversy surrounding Anthropic's implementation of a hidden watermark in all text generated by its AI, Claude. This policy, based on Google DeepMind's SynthID-Text technique, embeds a statistical signature by making inconsequential word choices. The watermark applies globally, even to human-written text lightly edited by Claude, sparking user backlash over issues of ownership and the creation of an "AI content second-class citizen" status. In response, an open-source tool called "watermarks-remover" (originally "remove-claude-marks") was released on GitHub, quickly gaining 11k stars. It works on three levels: removing invisible Unicode characters, using an agent to rewrite text and break statistical patterns, and stripping metadata from various file formats. Notably, Claude itself refused to install this removal tool as an Agent Skill, a task ultimately completed by another AI model, GLM 5.2. The article points out the irony that the removal code may have been written by Claude. The piece frames this as an ongoing battle between watermarking for traceability against misinformation and the desire for unmarked, owned content from paying users. It questions the practicality of mandatory technical markings when AI-generated text becomes indistinguishable from human writing, suggesting the open-source community's rapid development of countermeasures will continually outpace regulatory efforts.

marsbit7m ago

Claude's Watermark Has Been Cracked, Gaining 11k Stars, But Installation Is Refused

marsbit7m ago

SafePal Leaks Data of Nearly 40,000 Hardware Wallet Buyers: Private Keys Intact, Yet Danger Moves Closer to the Physical

Hardware wallet manufacturer SafePal has disclosed a data breach affecting approximately 39,798 customers who placed orders between March 2025 and April 2026. The leak exposed personal information including names, email addresses, phone numbers, physical delivery addresses, and purchase records. The company confirmed that private keys, recovery phrases, wallet passwords, and financial details were not compromised, as the cold storage systems operate in an isolated environment separate from the e-commerce servers. However, the breach poses significant risks beyond digital theft. Attackers now possess a high-value list of confirmed hardware wallet owners, effectively marking them as likely holders of substantial cryptocurrency. This enables highly targeted social engineering attacks, such as phishing emails referencing real order details, fake hardware deliveries, or phone scams impersonating SafePal support. The company has already identified and taken down over 30 related phishing sites. A critical aspect of the incident is the delayed disclosure timeline. SafePal acknowledged receiving initial user reports of phishing attempts in May but treated them as isolated. A full investigation began in July, with a public announcement not made until August, leaving users exposed for approximately three months. Furthermore, a configuration error prevented a data-purge routine from deleting old order information as intended, potentially increasing the scope of the leaked data. The incident highlights a structural paradox in the hardware wallet industry: while the devices are designed to secure private keys offline, the necessary e-commerce process collects sensitive personal data that, if leaked, makes the user a target. This mirrors a similar breach suffered by Ledger in 2020. Affected users are advised to be extremely vigilant. They should verify if they are impacted via SafePal's dedicated page, treat all unsolicited communications (emails, calls, physical mail) referencing SafePal as suspicious, and never share recovery phrases. Users who may have entered sensitive information on a phishing site must create a new wallet immediately. The breach underscores that in cryptocurrency security, the most vulnerable link is often the human user, not the cryptographic technology.

marsbit27m ago

SafePal Leaks Data of Nearly 40,000 Hardware Wallet Buyers: Private Keys Intact, Yet Danger Moves Closer to the Physical

marsbit27m ago

30 Years After Being Crushed by AI, People Have Fallen Back in Love with Chess

On May 11, 1997, IBM's "Deep Blue" defeated chess champion Garry Kasparov, marking the first time a machine triumphed in a top-level intellectual game. The narrative of human defeat by AI seemed cemented when AlphaGo beat Lee Sedol in Go in 2016, a game once considered AI's final frontier. Yet, nearly 30 years after AI's dominance began, chess is experiencing unprecedented popularity. Chess.com boasts over 250 million registered users and 10 million daily active players. Its CEO, Erik Allebest, attributes this resurgence to several waves: the pandemic, the Netflix series *The Queen's Gambit*, and viral AI chess bots like "Mittens" on social media. Crucially, each surge left a permanently higher user base. The key insight is that AI liberated the game. When machines unequivocally became the best, the pressure to "win" as the ultimate human was removed. Chess returned to its core: the intrinsic joy of play—the thrill of a tactical combo, the tension of a time scramble, the curiosity of post-game analysis. AI, now serving as an always-available coach and anti-cheat tool, became infrastructure that enhanced rather than killed the experience. In contrast, Go, deeply rooted in East Asian elite culture and often pursued for mastery and status, suffered a "collapse of meaning" at the professional level after AlphaGo. Players began mimicking AI moves, erasing distinctive styles and narrative. While some Go players gained fame as online personalities, it didn't translate to widespread engagement with the game itself. The divergence highlights a fundamental question in the age of AI: is the motivation for an activity about *winning* or *playing*? Activities where the process itself is the reward, like chess, can thrive when the pressure of being the best is gone. AI may rightly take over tasks done purely for outcome, but it cannot replace the human experience of simply enjoying the game.

marsbit42m ago

30 Years After Being Crushed by AI, People Have Fallen Back in Love with Chess

marsbit42m ago

Exiting Top Ten Shareholders of Kweichow Moutai, 'Long-term Capital' Portfolio Adjustments Revealed: National Social Security Fund Enters 12 New Stocks, Invests in Hard Tech Mthreads

With the ongoing release of semi-annual reports, the second-quarter investment moves of long-term institutional investors like the National Social Security Fund (NSSF) and insurance capital are becoming clear. Central Huijin Asset Management and China Securities Finance Corp., often referred to as the "national team," are no longer among the top ten shareholders of Kweichow Moutai. Data shows that as of August 14th, the NSSF held positions in 33 A-share companies, with a total portfolio value exceeding 11 billion yuan. It added 12 new stocks in Q2, spanning sectors like chemicals, food & beverage, and semiconductors. The NSSF maintains its stable investment style, favoring companies with solid performance and attractive valuations. Apart from exiting Moutai, Huijin and China Securities Finance also left the top ten shareholder lists of companies like Ping An Bank and Dahua Technology. Insurance capital heavily invested in 41 companies in Q2, with a total holding value over 26 billion yuan. They showed a continued preference for cyclical sectors like non-ferrous metals and chemicals, as well as high-dividend-yield stocks. Analysts note that these long-term funds act as market stabilizers and investment bellwethers. Their presence is reshaping the market ecology, steering focus towards fundamental corporate value and away from speculative trading. A notable move was the NSSF's new investment in Moore Threads, a loss-making but leading domestic GPU design company listed on the STAR Market, indicating interest in hard technology sectors. Looking ahead, analysts expect long-term capital to continue a dual-strategy: maintaining a foundation in high-dividend-value stocks while gradually increasing exposure to growth areas aligned with industrial upgrading, such as advanced manufacturing and tech self-sufficiency. The market's recovery is seen as gaining a firmer footing after recent adjustments.

marsbit52m ago

Exiting Top Ten Shareholders of Kweichow Moutai, 'Long-term Capital' Portfolio Adjustments Revealed: National Social Security Fund Enters 12 New Stocks, Invests in Hard Tech Mthreads

marsbit52m ago

Data of 54,000 wallet users leaked, Clarity odds just 10%: Hodler’s Digest, Aug. 16

Galaxy Digital has slashed the odds of the CLARITY Act passing in 2026 from 75% to just 10%, citing limited Senate session days. If it fails, the SEC and CFTC plan to issue their own crypto rules, though an SEC meeting was abruptly cancelled. High-profile meetings at the White House are planned to discuss the bill. Amid growing hack fears, crypto companies are urging AI labs to grant developers early access to advanced AI models for cybersecurity, following a $116M Coldcard wallet theft. Data breaches at Trezor and SafePal have exposed over 54,000 users' personal information. The CFTC is clashing with states over regulating prediction markets like Kalshi, ordering it to ignore a New York restraining order to maintain a national market, while a Washington state judge ruled against it. The Ethereum Foundation is revising its post-quantum plan, moving away from the Poseidon hash function, and scoping its next major upgrade, Hegotá, for next year. Tether received its first full clean audit opinion from KPMG, showing reserves exceeding liabilities by $6.814 billion. Marketwise, major cryptos saw weekly declines. Predictions include a possible Bitcoin bottom in October, while analysts dispute the feasibility of BTC reaching $1M by 2030. Glassnode notes Bitcoin is in its longest capitulation phase since FTX's collapse. Three men were charged for an alleged Bitcoin kidnapping plot in Missouri.

cointelegraph1h ago

Data of 54,000 wallet users leaked, Clarity odds just 10%: Hodler’s Digest, Aug. 16

cointelegraph1h ago

Trading

Spot

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片