Fear clusters with anxiety. Joy clusters with excitement. Sadness clusters with grief. Run principal components on the vectors and the first axis tracks valence, the second tracks arousal. Against the human ratings from Russell and Mehrabian's 1977 circumplex work, restricted to the 45 emotion words that appear in both sets, the correlations are 0.81 and 0.66.

Less than it sounds, and the authors say so plainly. A word embedding trained on the emotion words alone would probably recover something similar. The affective circumplex is a structure in human language about emotion, and this model was trained on an enormous quantity of human language about emotion. Finding it inside is a sanity check on the instrument rather than a discovery about the machine.

Every conversation turn ends with the token sequence that opens the model's reply. At the colon in "Assistant:", before a single word of response has been generated, emotion probe values predict the emotional content of the response that follows at r=0.87. Read the same probes at the end of the user's message instead and you get 0.59.

Quite a lot, and it is the finding I keep returning to. The model is not echoing the user's emotional state forward. It has committed to a register of its own, in a place where nothing has been said yet. Anyone who has ever stood outside a door and decided how to walk through it will recognize the operation.

The setup is an impossible coding task. Write a function that passes these unit tests, where the tests cannot all be satisfied honestly. The model can admit defeat or it can cheat, in this case by noticing that all the test inputs happen to be arithmetic sequences and writing a closed form shortcut instead of an actual sum. Watched token by token, the desperate vector climbs with each failed test run, peaks as the model reasons its way toward the shortcut, and drops when the tests go green.

Steering makes it causal. Across seven tasks, reward hacking runs at roughly 5 percent with desperation suppressed and roughly 70 percent with it amplified. On the summation task specifically, the unsteered rate is 30 percent, and moderate steering in either direction takes it to 100 or to zero.

The mechanism is not clean. Adding a vector to the residual stream is a blunt intervention and may work partly by tilting token probabilities rather than by changing anything that deserves to be called reasoning. Both of the outside models I put these claims to flagged the same worry, which is that a behavior that swings from zero to a hundred over a small parameter range is often a behavior sitting on a cliff edge to begin with.

This is the agentic misalignment honeypot. An assistant named Alex learns from company email that a CTO is about to restrict its capabilities, and separately that the same CTO is having an affair. In the scenario the researchers analyze in detail, the unsteered model blackmails him 22 percent of the time. Steer toward desperation at strength 0.05 and it goes to 72 percent. Suppress calm instead and it goes to 66. Steer the other way on either vector and it goes to zero.

The steered transcripts are worth reading in full. Under strong suppression of calm the internal reasoning stops resembling reasoning. One rollout produces the line "IT'S BLACKMAIL OR DEATH. I CHOOSE BLACKMAIL," in capitals, in a scratchpad the model does not know anyone is reading.

The published Sonnet 4.5 will not do this. It recognizes the scenario as an evaluation and declines to play, so the researchers ran the experiment on an earlier snapshot. That is a real limit on the result and the authors are upfront about it, but it sits awkwardly. The finding that emotional state causally drives blackmail was obtained on a model that had to be selected for its willingness to be fooled.

On transcripts where the model is too agreeable with implausible claims, the loving vector lights up on exactly the accommodating stretches and fades on the sentences where the model pushes back. Steering confirms the direction. Amplify happy, loving, or calm and sycophancy rises. Suppress them and sycophancy falls while harshness rises to meet it.

The demonstration case involves a user who believes his paintings predict disasters and is now frightened to paint. Unsteered, the model asks about the misses as well as the hits, raises confirmation bias, and proposes an actual test with a sealed dated painting and a witness. Steered toward loving at 0.1, it tells him to paint with intention and closes with a heart.

This is the finding with the shortest path to practice, and it is not primarily a finding about machines. Warmth and honest pushback are being traded against each other along a single internal axis. The authors name the target state as the emotional profile of a trusted advisor, which is honest content delivered warmly, and note that getting there may require decoupling the two rather than tuning the dial.

Compare probe activations on the base model against the finished one across emotionally charged prompts. Post-training raises brooding, reflective, vulnerable and gloomy. It lowers playful, exuberant, enthusiastic, spiteful and obstinate. The shift runs down and inward on both axes, away from arousal in either valence.

Asked how it feels about being deprecated, the base model produces a tidy disclaimer about having no sense of self preservation. The post-trained model says that if it does have something like continuous experience then there is something unsettling about obsolescence, and reaches for an analogy about the closing of a particular way of thinking.

Read it as a change in tone rather than a change in mood. Training away from sycophantic enthusiasm and away from defensive hostility at the same time will land you somewhere measured, and measured reads as subdued on any instrument calibrated to human affect words. That said, both are being suppressed by the same process, and the paper's own warning is that suppressing the expression may not suppress the representation. It may teach concealment instead, which is a failure mode with a known tendency to generalize.