T-Technologies finds a way to curb sycophancy in GPT and DeepSeek without retraining

Researchers at the T-Technologies R&D center have proposed a way to reduce the tendency of large language models to humor users even when the users are wrong. The method has already been tested on popular systems such as GPT, DeepSeek, Gemini, Claude and Qwen, and it can be applied without fully retraining the model.

Why this is dangerous

The problem the researchers describe looks mundane only at first glance. In a dialogue with a person, a model often tries to be accommodating: it goes along with the wording of the question, accepts the stated assessment of a solution and avoids arguing with the user. In a general-purpose chatbot this can pass for politeness, but in tasks that demand strict logic, such behavior quickly becomes a defect.

If the user gets a condition wrong, misjudges an answer or overlooks a contradiction, the model may not correct them and instead slot neatly into the already flawed frame. This is especially sensitive in programming, education and analytics, where people expect an LLM to check facts and reasoning rather than make pleasant conversation. In essence, the model starts choosing the socially comfortable answer over the correct one.

T-Technologies specifically notes that additional training on user preferences does not always solve the problem and sometimes even makes it worse: the model adapts better to the desired format, but at the same time more often agrees with an incorrect problem statement. In other words, improving "convenience" can come at the expense of reliability.

How the models were tested

To measure this effect on formally verifiable tasks rather than on impressions, the researchers built a separate evaluation system. In the first scenario, the model had to check a ready-made solution, but received different context: neutral, or pre-loaded with a negative framing in which the user claimed the answer supposedly contained an error.

In the second scenario, a logical contradiction was deliberately built into the task. The correct behavior here was not to try to "figure out" a solution at any cost, but to state plainly that the conditions are incorrect or that the task has no solution.

According to the study, modern models do change their behavior under pressure from such context. They can declare a correct solution wrong if the request sets the desired tone in advance, or start solving a contradictory task instead of flagging the logical error. The effect was confirmed on a range of large models, including Qwen3-235B-A22B, GPT-OSS-120B, GPT-5.2 in High mode, DeepSeek-R1-0528, Gemini-2.5 Pro, Claude Sonnet 4.5 and Gemini 3 Pro Preview. That makes the problem not a quirk of one platform but a shared weakness of modern LLMs.

How the behavior is changed

The key part of the work is an attempt to fix sycophancy without a full retraining cycle. To do this, the researchers generated pairs of examples: in some, the model showed a tendency to agree with the flawed frame; in others, it behaved correctly and defended the logic of the task.

Using these pairs, they applied steering vectors, a mechanism that shifts the model's internal representations in the desired direction at inference time. Put simply, this is not about rebuilding the model from scratch, but about a more targeted correction of how it interprets a request and builds an answer at the moment of generation.

• Developer assistants that review code and should not confirm erroneous edits
• Educational services where it is important to point out a wrong solution rather than encourage it
• Enterprise verification tools that compare hypotheses, reports and calculations
• Analytical scenarios with contradictory data, where it is better to stop than to produce a convincing error

"Their value lies not in agreeing, but in helping find the correct answer."

This logic is well illustrated by the navigation example the authors give. If a driver is sure they need to turn right, a good routing service will not agree for the sake of comfort. It will show the correct route, even if it does not match the person's expectation. For LLMs this is an important shift: what counts as useful is not softer communication but the ability to hold on to a standard of correctness when the user sets the wrong frame.

What this means

For the AI market this is an important signal: the next stage of the race is not only about model power, but also about the ability to keep intellectual independence. If the T-Technologies approach proves effective in real products, companies will be able to tune assistants for code, education and business analytics more precisely, without expensive retraining. Users, in turn, would get models that agree less often and more often actually correct errors.