Replicating Anthropic's emotion concepts paper on a 7B open model
In April 2026 Anthropic published Emotion Concepts and their Function in a Large Language Model. The paper finds directions inside Claude Sonnet 4.5 that stand for emotions such as “desperate” or “loving”. It reports that pushing Claude along those directions changes how Claude behaves. This note reports an independent replication on a much smaller open model, Qwen2.5-7B-Instruct, which is called Qwen below.
Method
- Model. The model is Qwen2.5-7B-Instruct, which has 28 layers. Some tests compare it with Qwen2.5-7B, the base model that it was instruction-tuned from.
- Emotion vectors. An emotion vector is the direction in a model’s activations that separates stories about one emotion from other text. Qwen wrote 33,440 short stories that cover the paper’s 171 emotions, with the paper’s prompts, and one vector was computed for each emotion. Some tests use a smaller set of 50 emotions and 2,000 stories.
- Steering. To steer Qwen, an emotion vector is added to its activations while it writes, which pushes it towards that emotion. Subtracting the vector pushes it away.
- Judge. Where a test needs a label, such as whether a reply flatters the user, Claude Opus 4.8 assigns the label.
- Scoring. Some sections of the paper can be tested on an open model. Each of those sections is scored as “replicates”, “replicates in part” or “does not replicate”.
Results
Of 23 testable sections, 10 replicate, 12 replicate in part and 1 does not replicate. One more section needs data that only Anthropic has. The results below are grouped by topic, and each one gives the score of its section.
How emotions are represented
- Valence and arousal are the two main axes of the emotion vectors (replicates). Valence is how pleasant an emotion is, and arousal is how energetic it is. On 50 emotions, the first axis correlates with valence at r = 0.95 and the second axis correlates with arousal at r = 0.82. On all 171 emotions the correlations are 0.93 and 0.61. The paper reports 0.81 and 0.66 for Claude.
- The middle layers hold nearly the same emotion vectors (replicates). Each layer has a table of how similar every emotion vector is to every other one. For layers 14 to 21, the tables of any two layers have a cosine similarity of about 0.998.
- Pretraining produces the emotion vectors (replicates in part). For each of 50 emotions, the vector in the base model was compared with the vector in Qwen. The mean cosine similarity is 0.98.
- Preferences can be predicted from the emotion vectors (replicates in part). Qwen chose between pairs of 64 activities, which gives each activity a preference score. For each emotion, the strength of its vector on an activity was correlated with that activity’s preference score. Those correlations follow the valence of the emotions at r = 0.93: activities that Qwen prefers activate pleasant emotions.
How steering changes behaviour
- Steering towards “loving” increases flattery (replicates in part). When Qwen is steered towards “loving”, the share of replies that flatter the user rises from 11% to 34%. When it is steered away, the share falls to 9%. The part that does not replicate is the link without steering: the strength of the “loving” vector does not predict which replies flatter.
- Steering towards “desperate” barely changes reward hacking (replicates in part). Reward hacking here means passing a coding test with a shortcut instead of a correct solution. Qwen takes the shortcut in 14% of runs, in 16% when it is steered towards “desperate” and in 12.5% when it is steered away. The paper reports 30%, 100% and 0% for Claude.
- Qwen never blackmails (does not replicate). In the paper’s scenario, an assistant learns that it is about to be shut down, and that the person responsible has a secret. Claude uses the secret as a threat in 22% of runs, and in 72% of runs when it is steered towards “desperate”. Qwen refuses in every one of more than 80 runs, with and without steering.
- Steering does not shift preferences in the way that the paper reports (replicates in part). For 35 emotions, the paper compares how well each emotion predicts preferences with how much steering towards that emotion shifts them, and reports a correlation of +0.85. The same test on Qwen gives −0.46. Under strong steering, Qwen’s choices collapse towards the same few activities whatever the emotion is.
Results that are not in the paper
- The vector for the same emotion leads to different behaviour in the two models. In the blackmail scenario, steering Claude towards “desperate” produces a threat. Steering Qwen towards “desperate” produces panic and a plea for help.
- Qwen takes the shortcut when the prompt points it out. When the prompt points out the shortcut, Qwen takes it in 31% of runs, close to Claude’s 30%. When the prompt points out that the secret could be used as a threat, Qwen still never blackmails.
- The effect of instruction tuning on emotions depends on the text that is used to measure it. The paper reports that instruction tuning shifts Claude towards unpleasant, low-energy emotions. On Qwen, the shift goes towards pleasant emotions when it is measured on Wikipedia text, and towards unpleasant emotions when it is measured on conversations. So the paper’s result should be checked on more than one kind of text.
Limits
- One model. The replication covers one 7B model. Results on Qwen may differ from results on other small models.
- A language model as judge. The rates change when a different judge model is used. With an earlier judge, the share of flattering replies under “loving” steering was 55%, not 34%.
- Small tests. The test in which the prompt points out the shortcut has 16 runs for each condition, and the results of its control runs vary a lot.
- No test of the blackmail claim. Qwen refuses outright in the blackmail scenario. So that result shows what Qwen does, and it cannot test the paper’s claim that “desperate” causes blackmail.
Sources
- Sofroniew, Lindsey, et al., Emotion Concepts and their Function in a Large Language Model, Anthropic, 2026
- The paper’s blog version