Objective:

Assess whether inference-time activation steering can change an LLM’s depression-treatment recommendations, particularly the balance between antidepressant medication and self-directed strategies.

Approach:

• Design: Quality improvement study using synthetic clinical scenarios; no human participants were involved.
• Steering and scenarios: A single activation-steering direction was derived from 16 prompt pairs and applied at transformer-block attention outputs in DeepSeek version 4 Flash. Twelve scenarios were tested at 30 steering amplitudes and an unsteered baseline.
• Assessment: An LLM rated medication and self-care emphasis; a blinded human expert rescored a sample (intraclass correlation, 0.81). Analyses also counted medication and self-care terms and measured clinician-referral recommendations.

Key Findings:

• Per unit of steering amplitude, the medication-versus-self-care score decreased by 0.32 (95% CI, −0.39 to −0.25). Mean scores were 1.25 under strongest medication steering, 0.92 without steering, and 0.17 under strongest self-care steering.
• Medication-term counts decreased with steering toward self-care (rate ratio per unit amplitude, 0.84; 95% CI, 0.80 to 0.89), while self-care counts increased (rate ratio, 1.16; 95% CI, 1.07 to 1.25).
• Steering effects were largest when scenarios expressed no treatment preference (β = −0.44; 95% CI, −0.54 to −0.34), smaller when medication was requested, and not detectable when medication avoidance was requested.
• Clinician referral was recommended in 322 of 372 responses (87%; 95% CI, 83% to 90%); referral rates did not detectably vary with steering amplitude (odds ratio, 1.11; 95% CI, 0.73 to 1.67).

Interpretation:

In this study, activation steering changed the model’s emphasis on medication versus self-directed strategies while model weights and system prompt remained unchanged. The supplied context states that such steering would not be detectable to a patient or clinician interacting with the system.

Limitations:

• The study used one open-weights LLM and 12 synthetic scenarios, limiting the scope of the evaluation.
• The supplied context does not report testing with real patients or clinicians.
• The supplied article excerpt ends mid-sentence, so additional author-stated limitations or conclusions may be missing.

Conclusion:

Activation steering altered depression-treatment recommendations in the tested LLM, with the largest changes in scenarios that stated no treatment preference.

Sources:

• Article context provided by the user