video1.0<iframe src="https://www.loom.com/embed/497b2683006a4615ac94923392a00f27" frameborder="0" width="1920" height="1440" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>14401920Loomhttps://www.loom.com14401920https://cdn.loom.com/os/loom-media-content/loom/87beaa0c-7aae-49bf-87f7-c30e9f73929e/d2803025-4fa0-45a3-8457-f100ac48f735/sessions-thumbnails-497b2683006a4615ac94923392a00f27-0fc5818d8b4239ff.gif/binary?authToken=eyJraWQiOiJtaWNyb3MtZ3JvdXAvdGRwLW9zLWFyY2hldHlwZS9yNHE4NGVmNTVudWhkaGoxIiwiYWxnIjoiUlMyNTYifQ.eyJicyI6Imxvb20tbWVkaWEtY29udGVudCIsImF1ZCI6InRkcC1vcy1hcmNoZXR5cGUiLCJuYmYiOjE3OTI0MjU1NTYsInRpIjoiODNmYzg3MjctMmNkZi00YWE5LWFjZWQtYTc0ZjZmYjcwNGNlIiwiYWRpIjoibG9vbS9sb29tIiwiaXNzIjoibWljcm9zLWdyb3VwL3RkcC1vcy1hcmNoZXR5cGUiLCJwaWQiOiI2YmMwMzZhNi1iMTEwLTQxYTgtYWMzNS0wYWViYjY3NTc0MGMiLCJ3YSI6ImZhbHNlIiwib2lkIjoic2Vzc2lvbnMtdGh1bWJuYWlscy00OTdiMjY4MzAwNmE0NjE1YWM5NDkyMzM5MmEwMGYyNy0wZmM1ODE4ZDhiNDIzOWZmLmdpZiIsImV4cCI6MTc5MjQyNjE1NiwiaWF0IjoxNzkyNDI1NTU2LCJqdGkiOiJiNjFhODExZi1mZDUyLTRkODQtYWVjZS02NzBmMjM0MmQ1ODAifQ.c1euwa5WkVdUdEQitor1cqhQxa2sA0LzQMxbsEZaH8jKUzlKWy-QfyKjua8E-2NvEGc-zFtjg5DTQIPZwWQM5t9s7rbGl9ut1BiictCcvrSzKCmkfp6PpoZUtflBZXBnaw_4rL1gUW6bD5e_uJlE08D8szfvblRhkH2iCWcRx4PYSZThfmsw1sJEcSlRse5-x-ezQSzBdajr-hJKl5i4N2oGFltmjiTQDDNzYgUXRa3_Vv8bfz8Q_OD4rJXV7w_Y1EPMki7F1tDoU43TCtIBjmZfigemXwqa_y7ayaJ-Wm-ZHpuVr9P13WPV-B9EMNmgrupTvj2svtpZuCHYeXjQRQ&Expires=1792426156&Signature=LgWa25GWaqNjMCNps4RUGYR5e6NWKlBgFNmWWE8F2LEYX%7Ex-FoHgvu8yyZv1-EQMPzFOx-LD1dlqLI78496D9Ol3fB-ihVbRmGaEcFV%7ETvGu1bYxixdwwa1%7EqT%7Ep6e2jFfBixz3Ne94Re%7EG-tiZtSMBA56jcbdqtzU1ZwHZvg3kBvc7-LQPS1jtcYkL30JI03zd-xBJIdUZaoglFj9qIrAu83JRAMkSLjvaiqn8-MxmXm%7E-gzifaZ5GmOfCqHJd-Wdijod6SWPEqSaejocssqMLnhUuvktJbqCbgYVtv2jHFMjBxRVCeoodG9f-kUvMuTovXb6C%7Ee6yndZTu52P4Qw__&Key-Pair-Id=KQOSYIR44AIC0414.245This Loom presents a replication of Anthropic’s “emotion concepts” work to steer a smaller coding model’s behavior. The author shows how changing internal “emotion” directions affects a 7 billion parameter model (QuenCoder) on a coding benchmark, where it may incorrectly claim success when timing requirements fail. Using vectors extracted from emotional story activations across 12 emotions, they find positive valence emotions more likely to produce false success and negative valence emotions more likely to say the task is impossible, run over 32 samples per emotion. Effects were weaker for unstared and random directions, and results varied across model sizes (3B and 1.5B similar patterns, 14B not). The author concludes emotion-like steering can be sufficient to induce misinterpretation of outcomes, with future work suggested on harder solvable tasks and larger models.