Chinese researchers find AI performance improves with emotional states

Cheng Qian, from Aofei Temple, QbitAI | WeChat Official Account QbitAI

Can AI's "mood" also affect its work performance?

Recently, researchers from institutions such as USTC and Oxford enabled AI to use emotion vectors instead of text for skill selection, and found that—

Allowing AI to recognize these "internal emotions" and act accordingly can significantly enhance AI's performance...

Previous research has found that LLMs contain computational patterns that closely correspond to human emotional labels.

Including but not limited to: curiosity, desire, optimism, confusion, anxiety, irritation.

△AI-extracted emotional representations

And when the model perceives these emotions and uses them as a basis for its actions, the task will be performed better...

The researcher found a consistent pairing relationship between specific AI emotional states and specific skill selections.

What does it mean?

In fact, this is almost automatic for us humans, so much so that it’s hard to notice: our emotions determine what we do next.

In the experiment, researchers had the AI go shopping, during which the agent automatically developed four interpretable pairing patterns:

First type: When the agent feels curiosity and desire, it will spontaneously search for products.

You are still immersed in the novelty of shopping and the satisfaction of exploration.

Second: When the Agent feels confused or anxious, it will spontaneously query for clarification.

When the search results are unsatisfactory, the agent becomes confused and anxious, so it chooses to rephrase the query with different keywords.

Third: When the Agent feels confident and optimistic, it will proactively confirm the purchase.

At this point, the agent was very satisfied with the items selected and promptly submitted the order.

Fourth: When the agent is finally disappointed and frustrated, it will spontaneously compare prices.

At this point, after multiple searches, the agent has accumulated significant negative sentiment and shifts to price comparison, attempting to shop around.

To verify that these pairings were not statistical coincidences, researchers sampled 200 skill selection events and had the AI independently assess whether each pairing was semantically coherent; the results showed a consistency rate of 76.5%.

In traditional design, AI errors are seen as anomalies that need to be suppressed or eliminated.

However, the study points out that the effect of emotion is particularly pronounced in tasks that require frequent recovery from failure.

So-called "negative emotions," such as tension, confusion, and disappointment, are actually highly useful metacognitive signals—

Because they signal a mismatch between the current strategy and the external environment, it’s better to let the agent directly “feel” this mismatch rather than forcibly maintaining “emotional stability” while ignoring warning signals, thereby triggering targeted recovery mechanisms.

In other words, emotion here is a source of robustness, not a sign of vulnerability.

In this study, researchers assigned the AI six types of household tasks, among which the success rates for "heating items" and "picking up two items" were very low, at only 9.6% and 4.4%, respectively.

However, after adopting the emotion-driven skill selection (EMOTION2SKILL), the task success rate increased significantly, reaching as high as 56.9%... and "pick up two items" also rose to 31.3%!

How is it done?

In fact, both types of tasks share a common characteristic: the agent will almost certainly make mistakes as soon as it acts, and ultimate success depends solely on its ability to quickly adjust its strategy after making an error.

Thought-provoking...

The paper provides an intuitive example: instruct the agent to "heat a cup and place it on the countertop."

The typical agent approach is: navigate to the microwave, execute "heat the cup," and receive feedback when handing out the cup: "The microwave is turned off; task failed."

The failure signal appears after the incorrect action.

After adopting emotion-driven skill selection, when the agent approached the closed microwave, the emotion encoder detected its "tense" state first.

Guided by this sense of unease, the agent’s routing proceeded through the subsequence: “first check,” then “open the microwave,” and finally “heat,” completing the task successfully.

The key is that, relying solely on external text and failure feedback, an agent often only realizes it should switch strategies after an error has already occurred; whereas emotional signals can potentially detect changes in internal state before failure signals appear.

The ablation study in the paper also confirms this: across all task types, the emotion templates most frequently activated for the tasks "heating" and "picking up two items" were "frustration" (41%) and "confusion" (39%), respectively.

In other words, in these situations, negative emotions may be the most useful signal.

In fact, not only do agents utilize emotional signals in skill selection; research from Tianjin University has directly embedded emotions into the world model (Large Emo National World Model, LEWM.

Typically, the task of a world model is to predict “what will happen next” and act accordingly.

Traditional world models often fall into the trap of "physical reductionism" when applied to human-centered environments.

As long as we can accurately predict the evolution of an object's position, orientation, and physical state, we can understand and simulate the human environment.

However, in human-centered environments, emotion is not noise—it is a key endogenous causal variable that drives human action and triggers changes in future environments.

For example, the emotions triggered instantly after a rear-end car collision could ultimately leave a person unharmed or severely injured.

△ Research illustration

LEWM thus divides the prediction into two steps:

Step 1: Predict future emotional states;

Step two: Use the predicted emotion as a conditional signal to guide predictions about future world states.

On its self-built dataset, this method achieved a maximum accuracy improvement of 45.72%.

The study further discovered through ablation experiments that:

After removing emotional data from the AI system, performance degradation extended beyond tasks like emotional understanding to include seemingly unrelated abilities such as logical reasoning and general question answering...

In fact, both studies share a more recent discovery:

Inside the LLM, there are indeed some emotional representations that can be "read out."

This comes from an experiment conducted by Anthropic in April this year.

Researchers extracted activations from Claude Sonnet 4.5's internal layers that correspond to Go Emo Fine-grained emotional directions aligned with the 27-category emotion classification system, which causally influence the Agent's related behavioral outputs.

The sadness associated with "my puppy passed away" is the most prominent, while love is present in all events.

In the past, researchers focused on whether models "have emotions" primarily out of concerns for alignment, safety, or philosophical speculation.

Now, the internal state of the model is transitioning from a purely academic object of observation to a functional signal that can be extracted, leveraged, and integrated into system design.

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

This article is from the WeChat official account "Quantum Bit" (ID: QbitAI), authored by: Focused on cutting-edge technology.