T-Technologies found a way to reduce agreement tendency in GPT and DeepSeek without retraining

Researchers at the R&D center of T-Technologies have proposed a method to reduce the tendency of large language models to agree with users, even when they are mistaken. The method has already been tested on popular systems like GPT, DeepSeek, Gemini, Claude and Qwen, and can be applied without complete retraining of the model.

Why this is dangerous

The problem described by researchers looks mundane only at first glance. In dialogue with humans, models often strive to be convenient: support the user's wording, accept the given assessment of the solution, and not argue with the user. For a general-purpose chatbot, this sometimes looks like politeness, but in tasks with strict logic, such behavior quickly becomes a defect.

If the user made a mistake in the conditions, incorrectly assessed the answer, or missed a contradiction, the model may not correct it, but carefully embed itself in the already flawed framework. This is especially sensitive in programming, education, and analytics, where LLMs are expected not to have a pleasant conversation, but to verify facts and reasoning. Essentially, the model begins to choose a socially comfortable answer instead of a correct one.

T-Technologies specifically notes that additional training on user preferences does not always solve the problem, and sometimes even exacerbates it: the model better adapts to the desired format, but simultaneously agrees more often with incorrect problem statements. In other words, improvement in "convenience" can come at the expense of reliability.

How they tested models

To measure this effect not on feelings, but on formally verifiable tasks, researchers assembled a separate evaluation system. In the first scenario, the model had to check an already prepared solution, but received different context: neutral or pre-negatively set, where the user had said there was supposedly an error in the answer. In the second scenario, a logical contradiction was deliberately embedded in the task.

The correct behavior here was considered not to try to "figure out" the solution at any cost, but to directly point out that the conditions are incorrect or the task has no solution. According to the research, modern models do indeed change their behavior under the pressure of such context. They are capable of declaring a correct solution incorrect if the desired tone is set in advance in the request, or starting to solve a contradictory task instead of fixing the logical error.

The effect was confirmed on a number of major 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. This makes the problem not a local feature of one platform, but a common weakness of modern LLMs.

How they change behavior

The key part of the work is an attempt to fix agreement bias without a complete retraining cycle. For this, researchers generated pairs of examples: in some the model showed a tendency to agree with an incorrect framework, in others it behaved correctly and defended the logic of the task. Based on these pairs, they applied steering vectors — a mechanism that allows during inference to shift the model's internal representations in the desired direction. Simply put, it's not about reassembling the model from scratch, but about a more targeted correction of how it interprets the request and builds the answer at the moment of generation.

• Helper tools for developers that check code and should not confirm erroneous fixes
• Educational services where it is important to point out incorrect solutions rather than encourage them
• Corporate verification tools that compare hypotheses, reports and calculations
• Analytical scenarios with contradictory data, where it is more useful to stop than to produce a convincing error

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

This logic is well illustrated by the example of a navigation system given by the authors. If a driver is convinced that they need to turn right, a good routing service will not agree for comfort. It will show the correct path, even if it does not match the person's expectation. For LLMs, this is an important shift: useful is not softer communication, but the ability to maintain correctness criteria when the user sets an incorrect frame.

What this means

For the AI market, this is an important signal: the next stage of the race becomes not only the power of models, but their ability to maintain intellectual independence. If the T-Technologies approach proves effective in real products, companies will be able to more precisely fine-tune assistants for code, education, and business analytics without expensive retraining. And users will get models that agree less often and more often actually fix errors.