Researchers can read an AI's 'mood' well enough to pick the right tool

A new arXiv preprint shows that a 27-dimensional vector pulled from a model's internal activations at every step can be used to pick which tool an AI agent invokes next. The authors call the vector an "emotion": a residual-stream signal, a short fingerprint of the activations a transformer passes between layers. It is not a claim that the model feels anything.

The paper, Emotion2Skill, comes from Bohan Lin's team. For most of the past three years, "AI interpretability" has meant looking inside a model after the fact: probing what a representation encodes, charting where a concept lights up, and labelling the internal circuits that produce a behavior. The team argues the field is now moving from observation to orchestration. Prior work had used internal-state vectors mostly to describe what a model was doing. Emotion2Skill uses the same kind of vector to decide what the model should do next. The vector is collapsed into a short, confidence-gated summary and pasted back into the prompt the agent uses to choose its next skill. A "skill" here is one of a library of named routines, like search, compare, and place an order.

The numbers are author-reported and worth taking seriously with the asterisk attached. On Qwen3-8B, the paper reports a 26.9% success-rate gain on WebShop and a 25.5% average gain on ALFWorld over a zero-shot baseline, with consistent results on the larger Qwen3-14B. The two benchmarks are synthetic: WebShop is a text-only shopping simulator, and ALFWorld is a household-task simulator built on top of a text adventure. The repository, github.com/BoHan-LIN04/Emotion2Skill, is public, though its README is bare at the moment of writing. There is no peer review, no third-party reproduction, and the "emotion" label is the authors' own coinage for a linear representation in activation space.

The more interesting second move is skill evolution. The team also watches the trajectory of the "mood" vector across a run. When the vector makes an abrupt jump that lines up with a failed skill call, the system treats that as a targeted signal that this particular skill's standard operating procedure needs rewriting. (A "standard operating procedure" here is the running playbook an agent updates after each attempt.) The paper argues this replaces the binary "did the run succeed" signal of earlier skill-evolution pipelines with a finer-grained internal one. Co-activation analysis, the authors write, shows the chosen skill and the internal-state vector are semantically coherent rather than just statistically coupled.

LLM interpretability has spent years producing tools that describe what a model is doing. This preprint is the first to argue, with benchmark numbers attached, that a probed internal state can also tell the model what to do next. If the result holds up, the question shifts away from whether AI agents can be "emotional" in any literal sense and toward whether an agent's next move can come from its own wiring rather than from the text it has produced so far. The next signals to look for are a v2, an independent re-run, and a benchmark outside the synthetic-agent set.