Society of Thought

When a computer program is supposed to answer a difficult question, it can do so in two ways. It can work through the problem once and output the first result. Or it can pursue several solution paths at the same time, compare them with one another, and form a joint answer from them. The second way is called Society of Thought, meaning roughly a “society of thoughts.” The name goes back to an old idea of the researcher Marvin Minsky: intelligence does not arise from a single clever component, but from many simple parts negotiating with each other. Applied to today’s systems, this means: instead of one voice, you let several compete against each other.

Why multiple voices produce better answers

A single line of reasoning can veer off in the wrong direction early on. Everything that follows then builds on that mistake, and yet the result still sounds convincing. This is exactly a well-known problem with today’s language models, meaning programs that generate text word by word. They occasionally invent facts and don’t notice it themselves. If, on the other hand, five attempts calculate independently of one another, an outlier stands out.

The effect is the same as with group work at school. Four students work on a problem separately, then compare their results. If three arrive at the same answer and one gets something different, it is usually the individual who is mistaken. Experts call this principle self-consistency. It costs nothing in additional training, only more computing time when answering.

It is important to distinguish this from a related term. In chain of thought, a model thinks through comprehensible intermediate steps, but only along a single path. Society of Thought operates one level above that: it coordinates many such paths. Chain of thought is the individual student, Society of Thought is the entire study group.

From proposal to joint answer

In the first step, the system generates several proposed solutions. The same model can do this multiple times, with slightly randomized word choice. But they can also be different models that were trained differently. Some systems even give the voices fixed roles: one proposes, one looks for counterarguments, one checks the calculation.

In the second step, the proposals are merged. The simplest variant is a vote: the most common answer wins. More elaborate variants let the voices discuss over several rounds. Each one sees the others' answers and may correct its own. At the end, another model summarizes the result.

The price for this is obvious. Five lines of reasoning cost roughly five times as much computing time as one. With millions of requests per day, this is a real cost factor. There is also a weakness that is easily overlooked: if all the voices come from the same model, they often share the same biases. Then everyone agrees and is collectively wrong. Diversity within the group is therefore not a luxury, but the precondition for the method to work at all.

Where the approach is found in today’s products

This principle becomes visible above all in systems advertised for difficult tasks. When a provider offers a “thinking mode” or a particularly thorough setting, some form of multiple computation is often behind it. The answer then noticeably takes longer, sometimes half a minute. In math problems and in programming, the hit rate rises measurably as a result.

In business news, the idea usually appears under the keyword agents. This refers to programs that independently handle subtasks. If you have several of these work together and check one another, that is essentially the same idea. Companies use such setups when mistakes would be costly, for example when reviewing contracts.

A typical misconception is worth mentioning in closing. Many consider Society of Thought to be a specific product or a model architecture. But it is a construction principle that is placed around existing models. You can switch it on when accuracy matters, and leave it out when speed is more important.