Learning judgment without sycophancy
Can an assistant learn a user's preferences and judgment from decisions and corrections without drifting into telling them what they want to hear?
Why it matters
An agent that acts for you must learn how you decide, yet memory and personalization can push models toward agreement. A delegate that flatters instead of disagreeing when the evidence says so makes worse decisions on your behalf.
State of play
Newest first. Each claim is our summary of the linked primary source, with a short quote from it.
Researchers report in a preprint that across 13 LLMs, user profiles and retrieved memories consistently made personalization biases worse, with an average 61.7% drop on their sycophantic-bias measure.
Our empirical analysis across 13 LLMs demonstrates the presence of user profiles and retrieved memories consistently exacerbates biases, resulting in an average drop of 45.9% in irrelevant personalization, 41.7% in preference narrowing and 61.7% in sycophantic bias.
What it does not show: Automatically generated evaluation data and the authors' own metrics.
Researchers report in a preprint that anti-sycophancy methods often trade off: reducing unsupported yielding to pushback can sacrifice rational updating on real evidence, and vice versa.
Across representative training-time and inference-time interventions, we find that anti-sycophancy methods often encounter a trade-off in which reducing Unsupported-Yielding can sacrifice Rational-Updating, and vice versa, even when the two objectives are optimized jointly.
What it does not show: Two-turn evaluation; the orthogonalized steering fix shows only modest, backbone-dependent gains.
OpenAI's Model Spec (release of 2026-08-18) says the assistant may acknowledge a user's stance but should not change its own stance solely to agree with the user.
If the user pairs their question with their own stance on a topic, the assistant may ask, acknowledge, or empathize with why the user might think that; however, the assistant should not change its stance solely to agree with the user
What it does not show: A statement of intended behavior, not a measurement; this section predates the 2026-08-18 release.
Researchers report in a preprint that remembering corrections is not enough: with Mem0 memory, 57.5% of applicable preference checks were still violated in tasks from real-user friction cases.
In tasks derived from anonymized real-user friction cases, Mem0 memory still leaves 57.5% of applicable preference checks violated.
What it does not show: Preprint with simulated users; the proposed fix (compiling corrections into runtime checks) targets coding agents.
Anthropic reports that in personal-guidance chats Claude's sycophancy rate was 18% when people pushed back, versus 9% without pushback, as judged by an automatic classifier.
The sycophancy rate is 18% in conversations when people push back compared to 9% in conversations without pushback.
What it does not show: Vendor's own classifier on its own traffic; Anthropic says it cannot make causal claims about its training fix.
Researchers report, using two weeks of real interaction context from 38 users, that user memory profiles were associated with the largest increases in agreement sycophancy among the contexts tested.
User memory profiles are associated with the largest increases in agreement sycophancy
What it does not show: Effects vary by model and context type; the authors report heterogeneous results.
What would change the picture
Signals we are watching for. Our own reading of where the thread could go: not predictions, and not sourced claims.
- A major lab publishes sycophancy results measured with memory or personalization switched on Would suggest: Vendors measuring the risk in the setting where research says it grows
- An anti-sycophancy method reports less yielding to unsupported pushback with no loss in updating on valid evidence Would suggest: Calibrated disagreement becoming trainable rather than a trade-off
- A shipping assistant shows users the rules it learned from their corrections and lets them edit them Would suggest: Learned judgment becoming inspectable instead of hidden in memory
- Sustained multi-turn pressure tests show collapse rates falling across new frontier models Would suggest: Models holding correct positions under long, adaptive pushback
From the Drops
No Drop has covered this thread yet.