Alex Wang, who writes at alexwang.ai, published a post on 8 November arguing that reasoning models do not improve by thinking for longer when you ask them a question. In his account the hard searching happened earlier, during training, and the model carries the result around in its weights. It is a theory, laid out in a blog post, and Wang himself describes it as an attempt to clear up his own confusion (“mostly for myself”).
The puzzle he starts from is a familiar one. During reinforcement learning, where a model is rewarded for answers that can be checked, its chains of thought grow longer and its accuracy climbs. Yet take a finished model and force one chain to run longer, and the benefit fades. Wang points to Muennighoff and colleagues, who made a model keep going by appending “Wait” whenever it reached for an ending. Scores climbed early, then levelled off. He also cites Gema and colleagues, who found tasks where more reasoning made answers worse.
His alternative is to treat each attempt as a path from the question to a correct answer. Put one arithmetic question to a model ten times and ten different chains of reasoning come back. For the sum of 1 through 10, one path adds the numbers in order and another pairs 1 with 10, 2 with 9, and so on, giving five pairs worth 11 each. Both reach 55. The chance of success is the combined probability of every path that lands on a right answer.
Two things can raise that chance. The cheap one is to sample many attempts and keep a winner, which helps as long as a single attempt has any chance at all. Wang credits Brown and colleagues with showing this early on. What scales is the number of attempts. The length of each one stays as it was.
The other way is to change the model. Sample paths on many checkable prompts, keep the ones that succeed, train on them, and repeat. Wang walks through the math to show that reinforcement learning with a pass-or-fail reward does nearly the same thing, scaled by how often the model currently succeeds. Chains grow longer along the way because winning paths tended to run longer, for instance when they included backtracking.
So the search gets paid for up front. Training burns compute exploring paths, then teaches the model to produce good ones directly, so a single sample at question time usually succeeds. Wang says the same trick appears outside language models: AlphaZero searches a tree of moves during self-play and then trains its move-picking network on what the search turned up, and forecasters train networks to predict the output of expensive simulators.
He draws two side conclusions. Because the reward only checks the final answer, nothing forces the intermediate tokens to be a true explanation of it, which is why chains of thought can be unfaithful. And because real checkers are imperfect, a model can find shortcut regions the checker accepts but should not. More samples and longer samples in training, he writes, make those shortcuts easier to find.
The argument rests on a simplified setup, including a checker that is assumed to be right every time, and the post tests no commercial model. Wang’s evidence for “thinking longer” is two cited studies where extra length stopped helping. It does not settle how much of a product’s reasoning mode is training and how much is spend at answer time.
If Wang is right, the figure worth tracking on a reasoning model is how often its first answer is correct, and the length of the chain behind it tells you very little.
Reported by Alex Wang’s blog (alexwang.ai) in a post dated 8 November 2026.
